remote sensing Article Change Detection in Synthetic Aperture Radar Images Using a Multiscale-Driven Approach Olaniyi A. Ajadi *, Franz J. Meyer and Peter W. Webley Geophysical Institute, University of Alaska Fairbanks, 903 Koyukuk Drive, P.O. Box 757320, Fairbanks, AK 99775, USA; fjmeyer@alaska.edu (F.J.M.); pwwebley@alaska.edu (P.W.W.) * Correspondence: oaajadi@alaska.edu; Tel.: +1-270-293-9069 Academic Editors: Richard Gloaguen and Prasad S. Thenkabail Received: 24 March 2016; Accepted: 2 June 2016; Published: 8 June 2016 Abstract: Despite the significant progress that was achieved throughout the recent years, to this day, automatic change detection and classification from synthetic aperture radar (SAR) images remains a difficult task. This is, in large part, due to (a) the high level of speckle noise that is inherent to SAR data; (b) the complex scattering response of SAR even for rather homogeneous targets; (c) the low temporal sampling that is often achieved with SAR systems, since sequential images do not always have the same radar geometry (incident angle, orbit path, etc.); and (d) the typically limited performance of SAR in delineating the exact boundary of changed regions. With this paper we present a promising change detection method that utilizes SAR images and provides solutions for these previously mentioned difficulties. We will show that the presented approach enables automatic and high-performance change detection across a wide range of spatial scales (resolution levels). The developed method follows a three-step approach of (i) initial pre-processing; (ii) data enhancement/filtering; and (iii) wavelet-based, multi-scale change detection. The stand-alone property of our approach is the high flexibility in applying the change detection approach to a wide range of change detection problems. The performance of the developed approach is demonstrated using synthetic data as well as a real-data application to wildfire progression near Fairbanks, Alaska. Keywords: change detection; SAR; decision support; image decomposition; image analysis; Bayesian inferencing 1. Introduction and Background Multi-temporal images acquired by optical [1] or radar remote sensing sensors [2] are routinely applied to the detection of changes on the Earth’s surface. As both of these two sensor types have their own imaging and sensitivity characteristics, their performance in change detection varies with the properties of the changing surface features [3]. In recent years, synthetic aperture radar (SAR) data have gained increasing importance in change detection applications, because SAR is an active sensor, operating without regard to weather, smoke, cloud cover, or daylight [4]. So SAR has shown to be a valuable data source for the detection of changes related to river ice breakup [5], earthquake damage [6], oil spill detection [7], flood [8], and forest growth assessment [9]. In this paper we are interested in developing an unsupervised change detection method from series of SAR images. Despite extensive research that was dedicated to this topic throughout the last decade [7], automation and robust change detection from a series of SAR images remains difficult for the following reasons: (A) Due to the complicated nature of the surface backscatter response in SAR data, most SAR-based change detection methods require repeated acquisitions from near-identical vantage points. These repeated acquisitions provide a static background against which surface change can be identified with reasonable performance [10]. This requirement severely limits the temporal sampling that can be achieved with modern SAR sensors with revisit times in the order of tens of days, Remote Sens. 2016, 8, 482; doi:10.3390/rs8060482 www.mdpi.com/journal/remotesensing Remote Sens. 2016, 8, 482 2 of 27 and strongly limits the relevance of the method for many dynamic phenomena; (B) The multiplicative speckle statistic associated with SAR images limits the change detection performance because it renders the identification of a suitable detection threshold difficult and it adds noise to the detection result; (C) The lack of automatic and adaptive techniques for the definition of change detection thresholds has hindered the development of a fully automatic change detection approach. Finally; (D) the lack of accuracy in delineating the boundary of changed regions has limited the performance of SAR data for applications that require high location accuracy of change information. Most of the recently proposed SAR-based change detection techniques utilize the concept of ratio or difference images for suppressing background information and enhancing change information. Published methods differ in their approach to extracting a final binary change detection map from the image ratio or difference data. In [1], two automatic unsupervised approaches based on Bayesian inferencing were proposed for the analysis of the difference image data. The first approach aims to select an adaptive decision threshold to minimize the overall change detection error, using the assumption that the pixels of the change map are spatially independent and that the gray value probability density function (PDF) of the difference image map is composed of two Gaussian distributions representing a change and a no-change class. This approach was further extended in [11,12] by adding an expectation-maximization (EM) algorithm to estimate the statistical parameters of the two Gaussian components. The second approach in [1], which utilizes a Markov random field (MRF) model, is taking into account contextual information when analyzing the change map. In [13], an MRF is used to model noiseless images for an optimal change image using the maximum a posteriori probability computation and the simulated annealing (SA) algorithm. The SA algorithm generates a random sequence of change images, such that a new configuration is established, which depends only on the previous change image and observed images by using the Gibbs sampling procedure [13]. A computationally efficient approach for unsupervised change detection is proposed in [14]. The approach initially start by generating an h ˆ h non-overlapping image block from the difference image. Eigenvector space is created using Principal Component Analysis (PCA) on the h ˆ h non-overlapping image blocks. In addition, a feature vector space over the entire difference image is created by projecting overlapping h ˆ h data blocks around each pixel onto eigenvector space. Furthermore, a k-means algorithm is employed to cluster the feature vector space into two clusters and assigns each pixel in the final change detection map to the cluster that minimizes the Euclidean distance between the pixel’s feature vector and the mean feature vector of clusters [14]. Several recent publications have utilized wavelet techniques for change detection from SAR. Analyzing an image in the wavelet domain helps to reduce the problems caused by the speckle noise. Wavelet domain analysis has been applied to unsupervised change detection for SAR images [15–18]. In [15], a two-dimensional discrete stationary wavelet transform (2D-SWT) was applied to decompose SAR ratio images into different scale-dependent images, each of which is characterized by a tradeoff between speckle noise suppression and preservation of image details. The undecimated discrete wavelet transform (UDWT) was proposed by [16] to decompose the difference image. For each pixel in the difference image, a feature vector is extracted by locally sampling the data from the multiresolution representation of the difference image [16]. The final change detection map is obtained using a binary k-means algorithm to cluster the multi-scale feature vectors, while obtaining two disjoint classes: change and no-change. Individual decompositions of each input image using the dual-tree complex wavelet transform (DT-CWT) are used in [17]. Each input image is decomposed into a single low-pass band and six directionally-oriented high-pass bands at each level of decomposition. The DT-CWT coefficient difference resulting from the comparison of the six high-pass bands of each input image determines classification of either change or no-change classes, creating a binary change detection map for each band. These detection maps are then merged into a final change detection map using inter-scale and intra-scale fusion. The number of decomposition scales (levels) for this method must be determined in advance. This method boasts high performance and robust results, but has a high Remote Sens. 2016, 8, 482 3 of 27 computational cost. In [18], a region-based active contour model with UDWT was applied to a SAR difference image for segmenting the difference image into change and no-change classes. More recently, [19] used a curvelet-based change detection algorithm to automatically extract changes from SAR difference images. Even though these papers have provided a solution to a subset of the limitations highlighted above, not all the limitations are solved in a single paper. The work done by [1,11,12] does not provide an effective way to select the number of scales needed for the EM algorithm in order to avoid over- or under-estimation of the classification. Also, the noise filtering methods employed do not preserve the detailed outline of a change feature. The disadvantages of the method in [13] are its high computational complexity and reduced performance in the presence of speckle noise. Although the method in [14] achieves a good result with low computational cost, the performance of the employed PCA algorithm decreases when the data is highly nonlinear. While [15–19] provided promising results in a heterogeneous image, the various methods have several disadvantages. They do not consider image acquisitions from multiple geometry, are not fully preserving the outline of the changed regions, and require manual selection of detection thresholds. To improve upon previous work, we developed a change detection approach that is automatic, more robust in detecting surface change across a range of spatial scales, and efficiently preserves the boundaries of change regions. In response to the previously identified limitation (A), we utilize modern methods for radiometric terrain correction (RTC) [10,20,21] to mitigate radiometric differences between SAR images acquired from different geometries (e.g., from neighboring tracks). We show in this paper that the application of RTC technology allows combining multi-geometry SAR data into joint change detection procedures with little reduction in performance. Thus, we show that the addition of RTC results in improved temporal sampling with change detection information and in an increased relevance of SAR for monitoring dynamic phenomena. To reduce the effects of speckle on image classification (limitation (B)), we integrate several recent image processing developments in our approach: We use modern non-local filtering methods [22] to effectively suppress noise while preserving most relevant image details. Similarly, we perform a multi-scale decomposition of the input images to generate image instances with varying resolutions and signal-to-noise ratios. We use a 2D-SWT in our approach to conduct this decomposition. To fully automate the classification operations required in the multi-scale change detection approach (limitation (C)), we model the probability density function of the change map at each resolution level as a sum of two or more Gaussian distributions (similar to [11]). We developed an automatic method to identify the number of Gaussian processes that make up our data and then use probabilistic Bayesian inferencing with EM algorithm and mathematical morphology to optimally separate these processes. Finally, to accurately delineate the boundary of the changed region (limitation (D)), we utilize measurement level fusion techniques. These techniques used the posterior probability of each class at each multi-scale image to compose a final change detection map. Here we tested five different techniques including (i) product rule fusion; (ii) sum rule fusion; (iii) max rule fusion; (iv) min rule fusion; and (v) majority voting rule fusion. These individual methods are combined in a multi-step change detection approach consisting of a pre-processing step, a data enhancement and filtering step, and the application of the multi-scale change detection algorithm. The details of the proposed approach are described in Section 2. A performance assessment using a synthetic dataset and an application to wildfire mapping is shown in Sections 3 and 4, respectively. A summary of the presented work is shown in Section 5. 2. Change Detection Methodology The work presented here is motivated by the challenges associated with change detection and the desire to develop an improved algorithm for unsupervised change detection in SAR images that can provide change information at high temporal sampling rates. We aim at using information at Remote Sens. 2016, 8, 482 4 of 27 Remote Sens.resolution 2016, 8, 482 levels to obtain high accuracy change detection maps in both heterogeneous4 and of 27 different homogeneous regions, contributing to a change detection method that is applicable to a wide range presented in Figure 1. The overall contribution of our methodology has two components. First, we of change situations. To achieve this goal, a multi-step processing workflow was developed that is developed a set of new techniques that are utilized within our change detection workflow to presented in Figure 1. The overall contribution of our methodology has two components. First, we streamline processing and improve detection performance. As part of these techniques we (1) developed a set of new techniques that are utilized within our change detection workflow to streamline developed an efficient way to improve classification performance by combining EM algorithms with processing and improve detection performance. As part of these techniques we (1) developed an mathematical morphology—this is a novel contribution to this field of research; (2) we integrated a efficient way to improve classification performance by combining EM algorithms with mathematical low computational complexity way to achieve high accuracy in preserving the boundary of changed morphology—this is a novel contribution to this field of research; (2) we integrated a low computational regions using measurement level fusion techniques; and (3) we combine modern non-local filtering complexity way to achieve high accuracy in preserving the boundary of changed regions using and 2D-SWT to provide robustness against noise. The second contribution comes from a novel way measurement level fusion techniques; and (3) we combine modern non-local filtering and 2D-SWT to of combining our own technology with other published processing procedures to arrive at a new, provide robustness against noise. The second contribution comes from a novel way of combining our efficient and highly flexible change detection workflow that can be applied to a wide range of change own technology with other published processing procedures to arrive at a new, efficient and highly detection problems. flexible change detection workflow that can be applied to a wide range of change detection problems. 0 Figure thethe original log-ratio image, XkLRX is the kth Figure 1. 1. Workflow Workflowof ofthe theproposed proposedapproach. approach.Here, Here,XLR X is is original log-ratio image, is the K´1 decomposed image, andand XLRX is the highest decomposed level. kth decomposed image, is the highest decomposed level. requires some parameters to It is worth worth mentioning mentioningthat thatthe theautomatic automaticprocess processofofour ourapproach approach requires some parameters be be set set beforehand. The The parameters that need to be set include (i) the neighborhood size of the filtering to beforehand. parameters that need to be set include (i) the neighborhood size of the step; (ii)step; the number of multi-scale decomposition levels;levels; (iii) the element of the filtering (ii) the number of multi-scale decomposition (iii)structuring the structuring element of morphological filter; and,and, finally, (iv) (iv) the maximum number of allowed change classes. Please note the morphological filter; finally, the maximum number of allowed change classes. Please that while we identified optimal settings for these parameters, we found that the performance of our note that while we identified optimal settings for these parameters, we found that the performance algorithm does not critically depend on the exact choice for these variables. This is true for the following reasons: (i) as non-local means filtering is conducted very early in the workflow, the impact of changes in the neighborhood size is mitigated by subsequent processing steps such as multi-scale Remote Sens. 2016, 8, 482 5 of 27 of our algorithm does not critically depend on the exact choice for these variables. This is true for the following reasons: (i) as non-local means filtering is conducted very early in the workflow, the impact of changes in the neighborhood size is mitigated by subsequent processing steps such as multi-scale decomposition and the application of mathematical morphology. Hence, we found that varying the neighborhood size from its optimal value changed system performance only slowly; (ii) in empirical tests it was found that using six decomposition levels was a good compromise between processing speed and classification accuracy. Adding additional levels (beyond six) did not result in significant performance improvement but added computational cost. Reducing the number of layers leads to a slow decrease of change detection performance, yet this reduction of performance does not become significant unless the number of bands drops below four; (iii) from an analysis of a broad range of data from different change detection projects we found (1) that a 20 ˆ 20 pixel-sized structuring element of the morphological filter led to the most consistent results; and (2) that change detection performance changed slowly with deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice of the window size is not critical for change detection success; finally, (iv) the maximum number of allowable change classes is a very uncritical variable as it merely sets an upper bound for a subsequent algorithm that automatically determines the number of distinguishable classes in a data set (see Section 2.3.3). By presetting this variable to 20 classes we ensure that almost all real life change detection scenarios are captured. There is no need to change this variable unless unusually complex change detection situations with more than 20 radiometrically distinguishable change features are expected 2.1. SAR Data Pre-Processing The ultimate goal of the pre-processing step is to perform image normalization, i.e., to suppress all image signals other than surface change that may introduce radiometric differences between the acquisitions used in the change detection analysis. Such signals are largely related to (i) seasonal growth or (ii) topographic effects such as terrain undulation that arise if images were not acquired from near-identical vantage points. In order to enable a joint change detection analysis of SAR amplitude images acquired from different observation geometries, we attempt to mitigate relative geometric and radiometric distortions. In a calibrated SAR image, the radar cross-section (RCS) of a pixel can be modeled as [4]: σ “ σ0 pθi q ˆ Aσ pθi q (1) where Aσ is the (incidence angle–dependent) surface area covered by a pixel, θi is the local incidence angle, and σ0 is the normalized RCS. According to (1), images acquired from different geometries will differ due to the look angle dependence of both σ0 and Aσ . In areas that are dominated by rough surface scattering and for moderate differences ∆θi of observation geometries, we can often assume that ∆σ0 p∆θi q ! ∆Aσ p∆θi q [4]. Under these conditions, the geometric dependence of σ can largely be removed by correcting for ∆Aσ p∆θi q. This correction is called radiometric terrain correction [20], which is completed by the following steps: ‚ ‚ In the first step, geometric image distortions related to the non-nadir image geometry are removed by applying a “geometric terrain correction” step [10] using a digital elevation model (DEM). Secondly, to remove radiometric differences between images, we use radiometric terrain normalization [10]. This normalization also utilizes a DEM to estimate, pixel by pixel, and compensate Aσ pθi q, for the radiometric distortions. In areas dominated by rough surface scattering, the application of RTC allows for combining SAR data acquired from moderately different incidence angles into joint change detection procedures with little reduction in performance. Hence, it can lead to significant improvements in the temporal sampling with change detection data. An example of the effect of geometric and radiometric terrain correction is shown in Figure 2. Figure 2a shows an original ALOS PALSAR image over the Tanana Flats region in Alaska while the Remote Sens. 2016, 8, 482 6 of 27 image after geometric and radiometric terrain correction is presented in Figure 2b. The normalized data is now largely free of geometric influences, reducing differences between images acquired from different Remote Sens.geometries 2016, 8, 482 and enhancing the performance of change detection from multi-geometry data. 6 of 27 Figure 2. Figure 2. Example Example of of (a) (a) image image affected affected by by radiometric radiometric and and geometric geometric distortions; distortions; (b) (b) radiometric radiometric and and geometric normalized normalized image image enabling enabling change change detection detection analysis analysis from from multiple multiple geometries. geometries. geometric It is worth mentioning that the RTC utilized in our approach is most effective when dealing with It is worth mentioning that the RTC utilized in our approach is most effective when dealing natural environments that are dominated by rough surface scattering. For these target types, the with natural environments that are dominated by rough surface scattering. For these target types, surface scattering properties change slowly with incidence angle and differences in measured radar the surface scattering properties change slowly with incidence angle and differences in measured brightness are dominated by geometric effects. However, RTC will be less useful for areas dominated radar brightness are dominated by geometric effects. However, RTC will be less useful for areas by targets with very oriented scattering characteristics (e.g., urban environments). For these regions, dominated by targets with very oriented scattering characteristics (e.g., urban environments). For RTC correction may not lead to significant reduction of radiometric differences between images from these regions, RTC correction may not lead to significant reduction of radiometric differences between different incidence angles. Furthermore, limitations exist for regions with complex small-scale images from different incidence angles. Furthermore, limitations exist for regions with complex topography, if this topography is not sufficiently captured in the available DEM. small-scale topography, if this topography is not sufficiently captured in the available DEM. 2.2. Data 2.2. Data Enhancement Enhancement 2.2.1. Logarithmic Scaling and Ratio Image Formation To suppress suppress image image background background structure and improve the detectability detectability of potential surface changes from SAR data, a ratio image is formed between a newly acquired image Xii and a reference R.. Using image XR Using ratio ratio images images in in change change detection detection was was first first suggested suggested by by [3] and has since been the basis of many change detection detection methods methods [12,23,24]. [12,23,24]. The The reference reference image image X XRR and image Xii are selected such that the effects of seasonal variations as well as spurious changes of surface reflectivity on the change detection product are minimized. Before ratio image formation, all data are geometrically and radiometrically calibrated calibratedfollowing followingthe theapproach approachininSection Section2.1. 2.1.The The resulting ratio image then radiometrically resulting ratio image cancan then be be modeled as [3]: modeled as [3]: Oir “=xTir (2) (2) where and Tir is where Oir isis the the observed observed intensity intensity ratio, ratio, x isis aamultiplicative multiplicative speckle speckle contribution, contribution, and is the the underlying true intensity ratio. The observed intensity ratio image has the disadvantage that the underlying true intensity ratio. The observed intensity ratio image has the disadvantage that the multiplicative noise is is difficult difficultto toremove. remove.Therefore, Therefore,aalogarithmic logarithmicscaling scalingisisapplied appliedtoto Oir, ,resulting resultingin: in: multiplicative noise (` )+ 10 pxq 10log( ′ “=10log X 1LR 10log pTir q ) (3) The application of logarithmic scaling and ratio image formation helps to transform our data The application of logarithmic scaling and ratio image formation helps to transform our data into a near normal distribution which closely resembles a Gaussian distribution. To suppress the into a near normal distribution which closely resembles a Gaussian distribution. To suppress now-additive noise in the log-scaled ratio image ( ` ) , ˘we applied a fast non-local means the now-additive noise in the log-scaled ratio image X 1 LR , we applied a fast non-local means filtering procedure. filtering procedure. 2.2.2. Fast Non-Local Means Filtering Approach As we are interested in developing a flexible change detection method that can be applied to a wide range of change situations, we are interested in preserving the original image resolution when filtering the data. Non-local means filters are ideal for this task as they identify similar image patches Remote Sens. 2016, 8, 482 7 of 27 2.2.2. Fast Non-Local Means Filtering Approach As we are interested in developing a flexible change detection method that can be applied to a wide range of change situations, we are interested in preserving the original image resolution when filtering the data. Non-local means filters are ideal for this task as they identify similar image patches in a dataset and use those patches to optimally suppress noise without sacrificing image resolution. The non-local means concept was first published in [25]. The algorithm uses redundant information to reduce noise, and restores the original noise-free image by performing a weighted average of pixel values, considering the spatial and intensity similarities between pixels [22]. Given the log-ratio image 1 (see Section 2.2), we interpret its noisy content over a discrete regular grid, Ω. XLR X 1LR “ X 1LR px, yq | px, yq P Ω ( (4) The restored image content X LR pxi , yi q at pixel pxi , yi q is then computed as a weighted average of all of the pixels in the image, according to: ` ˘ X LR xi , x j “ ÿ ` ˘ w pi, jq X 1LR x j , y j (5) px j y j qPΩ ` ˘ The weight w pi, jq measures the similarities between two pixels pxi , yi q and x j , y j and is given by: 1 ´ spi,jq w pi, jq “ e h2 (6) D piq where h controls the amount of filtering, D piq is the normalization constant, and s pi, jq is the weighted Euclidean distance (gray-value distance) of two neighborhood pixels (Pi and Pj ) that are of equal size. According to [25], the similarity between two neighborhood pixels (Pi and Pj ) is based on the similarity of their intensity gray level. Neighborhoods having similar gray-level pixels will have larger weights in the average. To compute the similarity of the intensity gray level, we estimated s pi, jq as follow: s pi, jq “ k Pi ´ Pj k2a (7) The standard deviation of the Gaussian kernel is denoted as a. To ensure that the averaging in Equations (5)–(7) is more robust, we set the neighborhood size for weight computation to 5 ˆ 5 pixels and the size of the searching region to 13 ˆ 13 pixels. We referred to Ω as the searching window. The optimal neighborhood sizes were determined in empirical tests. These tests also showed that the choice of neighborhood size does not critically affect the filtering performance. We implemented the modern non-local filtering methods to effectively suppress the noise while preserving most relevant details in the image. 2.3. Change Detection Approach The workflow of our change detection approach is shown in the lower frame of the sketch in Figure 1 and includes three key elements. In an initial step, a multi-scale decomposition of the input ratio images is conducted using a 2D-SWT and resulting in K image instances. Secondly, a multi-scale classification is performed at each of the K levels, resulting in K classification results per pixel. In our approach, the classifications are performed automatically and adaptively using an EM algorithm with mathematical morphology. Finally, in a third step, we conduct a measurement level fusion of the classification results to enhance the performance of our change detection. 2.3.1. Multi-Scale Decomposition As previously mentioned, often our ability to detect change in SAR images is limited by substantial noise in full resolution SAR data. Here, we utilize multi-scale decomposition of our input images Remote Sens. 2016, 8, 482 8 of 27 to generate image instances with varying resolutions and signal-to-noise ratios, and also to further reduce residual noise that remains after an initial non-local means filtering algorithm (Section 2.2) was applied. Multi-scale decomposition is an elegant way to optimize noise reduction while preserving the desired geometric detail [2]. In our approach, 2D-SWTs are used to decompose the log-ratio images [26]. Figure 3 shows Remote Sens. 2016, 8,that 482 in our implementation of the wavelet decomposition we apply a set of high-pass 8 of 27 and low-pass filters first to the rows and then to the columns of an input image at resolution level klevel ´ 1, resulting in fourindecomposition imagesimages at resolution level k. level The four decomposition images − 1, resulting four decomposition at resolution . The four decomposition LL and (2) three high-frequency detail images (X LH , X HL , and include (1) a lower image X LR images include (1) resolution a lower resolution image and (2) three high-frequency detail LR images , LR ( HH ), where the superscripts LL, LH, HL, HH indicate in which order low-pass (L) and high-pass (H) X LR , and ), where the superscripts , , , indicate in which order low-pass ( ) and filters were( applied. high-pass ) filters were applied. Figure 3. 3. Multi-scale Multi-scale decomposition decomposition of of input input log-ratio log-ratio image image X into into aa lower lower resolution resolution and and detail detail Figure LR images; X LL isisthe thelower lowerresolution resolutionimage imageand and(X ( LH , ,X HL,, X HH) )are the high frequency detail images, images; LR LR LR LR are the high frequency detail images, which captures horizontal, vertical and diagonal directions, respectively. which captures horizontal, vertical and diagonal directions, respectively. We chose the discrete stationary wavelet transform ( ) over the discrete wavelet transform We chose the discrete stationary wavelet transform (SWT) over the discrete wavelet transform ( ), as the is computationally more efficient, shift invariant, and un-decimated ( (DWT), as the SWT is computationally more efficient, shift invariant, and un-decimated (SWT adjusts adjusts the filters (up-sampling) at each level by padding them with zero in order to preserve the the filters (up-sampling) at each level by padding them with zero in order to preserve the image size). image size). To gain a greater flexibility in the construction of wavelet bases, we select a wavelet To gain a greater flexibility in the construction of wavelet bases, we select a wavelet decomposition decomposition filter from the biorthogonal wavelet family. The biorthogonal wavelet was selected filter from the biorthogonal wavelet family. The biorthogonal wavelet was selected because of its because of its symmetric capabilities, which is often desirable since it exhibits the property of linear symmetric capabilities, which is often desirable since it exhibits the property of linear phase, which phase, which is needed for image reconstruction. Another reason for using the biorthogonal wavelet is needed for image reconstruction. Another reason for using the biorthogonal wavelet was that, was that, rather than having one wavelet and one scaling function, biorthogonal wavelets have two rather than having one wavelet and one scaling function, biorthogonal wavelets have two different different wavelet functions and two scaling functions that may produce different multiresolution wavelet functions and two scaling functions that may produce different multiresolution analyses. In analyses. In addition, the biorthogonal wavelet has good compact support, smoothness and good addition, the biorthogonal wavelet has good compact support, smoothness and good localization. localization. We choose the fifth-order filter, which has a filter length of 12. The fifth-order filter was We choose the fifth-order filter, which has a filter length of 12. The fifth-order filter was selected selected to avoid distortions along the image boundary. Using , the log-ratio image is to avoid distortions along the image boundary. Using SWT, the log-ratio image X LR is recursively recursively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum decomposed into six resolution levels. Empirical analysis suggested six to be the optimum level for level for multi-scale decomposition. multi-scale decomposition. To reduce computation time, we focused only on the lower resolution images LL per To reduce computation time, we focused only on the lower resolution images X LR per decomposition level. Discarding the detail images is allowed as the information contents at a certain decomposition level. Discarding the detail images is allowed as the information contents at a certain resolution level are recovered at a higher level. Hence, the exclusion of the detail images does not resolution level are recovered at a higher level. Hence, the exclusion of the detail images does not affect affect the change detection approach. The final multi-scale decomposition image stack then the change detection approach. The final multi-scale decomposition X MS image stack then contains contains the lower resolution images at each level, as below: the lower resolution images at each level, as below: (8) .., ……….., != ) K´1 k X MS “ X 0LR .., X LR . . . . . . . . . .., X LR (8) 2.3.2. Classification by Expectation-Maximization (EM) Algorithm with Mathematical Morphology 2.3.2. Classification by Expectation-Maximization (EM) Algorithm with Mathematical Morphology After the multi-scale decomposition, each decomposed image is inserted into a mathematical After theframework. multi-scaleMathematical decomposition, each decomposed into a mathematical morphology morphology defines aimage familyisofinserted morphological filters, which morphology framework. Mathematical morphology defines a family of morphological filters, are nonlinear operators that aim at emphasizing spatial structures in a gray-level image. Forwhich more details on mathematical morphology the reader is referred to [27]. Morphological filters are defined by a structuring element ( ), which is based on a moving window of a given size and shape centered ( , ). In image processing, erosion ( , ) and dilation ( , ) are the basic on a pixel operators used, and are defined as follows [27]: analyses. In wavelet addition, the biorthogonal wavelet has good compact support, smoothness and good . In addition, the contains the lower resolution images at each level, as below: biorthogonal has good compact support, smoothness and good Figure 3. Multi‐scale decomposition of input log‐ratio image into a lower resolution and detail ime, we focused only on wavelet the scaling lower resolution images per tion level are recovered at a higher level. Hence, the exclusion of the detail images does not different functions and that two scaling functions that may produce different multiresolution ent wavelet functions and two functions may produce different multiresolution localization. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was ion. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was We chose the discrete stationary wavelet transform ( ) over the discrete wavelet transform .., ……….., (8) images; is the lower resolution image and ( , decomposition , ) are the high frequency detail images, g the detail images is allowed as the information contents at a certain the change detection approach. The final multi‐scale compact image stack analyses. In the addition, the has biorthogonal wavelet has good support, smoothness and good is ses. In addition, the biorthogonal wavelet good compact support, smoothness and good selected to avoid distortions along the image boundary. Using log‐ratio image to avoid distortions along image boundary. Using , the log‐ratio image then is , the ( ), as the is computationally more efficient, shift invariant, and un‐decimated ( which captures horizontal, vertical and diagonal directions, respectively. t a higher level. Hence, the exclusion of the detail images does not ns the lower resolution images at each level, as below: localization. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was zation. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was recursively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum ely decomposed into six resolution levels. Empirical analysis suggested six to be the optimum adjusts the filters (up‐sampling) at each level by padding them with zero in order to preserve th then proach. The final multi‐scale decomposition image 2.3.2. Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology selected to avoid distortions along the stack image boundary. Using is ed to avoid distortions along the image boundary. Using , the log‐ratio image , the is log‐ratio image level for multi‐scale decomposition. multi‐scale decomposition. (8) ) over the discrete wavelet transform . . , size). … …To … .gain ., image a greater flexibility in the construction of wavelet bases, we select a wavel We chose the discrete stationary wavelet transform ( Remote Sens. 2016, 8, 482 9 of 27 mages at each level, as below: recursively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum sively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum decomposition of input log‐ratio image into a lower resolution and detail To reduce computation time, we focused only on the lower resolution images per reduce computation time, we focused only on the lower resolution images per After the multi‐scale decomposition, each decomposed image is inserted into a mathematical decomposition filter from the biorthogonal wavelet family. The biorthogonal wavelet was selecte ( ), as the is computationally more efficient, shift invariant, and un‐decimated ( level for multi‐scale decomposition. or multi‐scale decomposition. wer resolution image and ( decomposition level. Discarding the detail images is allowed as the information contents at a certain osition level. Discarding the detail images is allowed as the information contents at a certain (8) ) are the high frequency detail images, . . , morphology framework. Mathematical morphology defines a family of morphological filters, which … … … . ,. , , because of its symmetric capabilities, which is often desirable since it exhibits the property of linea adjusts the filters (up‐sampling) at each level by padding them with zero in order to preserve the Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology To time, reduce time, we focused only on images the lower resolution images per ontal, vertical and diagonal directions, respectively. To reduce computation we computation focused only on the lower resolution per resolution level are recovered at a higher level. Hence, the exclusion of the detail images does not on level are recovered at a higher level. Hence, the exclusion of the detail images does not are nonlinear operators that aimconstruction at emphasizing spatial structures a gray-level image. are nonlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more phase, which is needed for image reconstruction. Another reason for using the biorthogonal wavel image Remote Sens. 2016, 8, 482 size). To gain a greater flexibility in the of wavelet bases, we in select a wavelet 9 of 27 For more Remote Sens. 2016, 8, 482 decomposition level. Discarding the detail images is allowed as the information contents at a certain mposition level. Discarding the detail images is allowed as the information contents at a certain affect the change detection approach. The final multi‐scale decomposition image then 9 of 27 e decomposition filter from the biorthogonal wavelet family. The biorthogonal wavelet was selected change detection approach. The final multi‐scale decomposition image stack then After the multi‐scale decomposition, each decomposed image is inserted into a mathematical details on mathematical morphology the reader is referred to [27]. Morphological filters are defined details on mathematical morphology the reader is referred to [27]. Morphological filters arestack defined was that, rather than having one wavelet and one scaling function, biorthogonal wavelets have tw on‐Maximization (EM) Algorithm with Mathematical Morphology Remote Sens. 2016, 8, 482 9 of 27 ete stationary wavelet transform ( ) over the discrete wavelet transform resolution level are recovered at a higher level. Hence, the exclusion of the detail images does not ution level are recovered at a higher level. Hence, the exclusion of the detail images does not Remote Sens. 2016, 8, 482 9 of 27 contains the lower resolution images at each level, as below: the lower resolution images at each level, as below: hology framework. Mathematical morphology defines a family of morphological filters, which , which is based on a moving window of a given size and shape centered by a structuring element pSq, which isopening based on aand moving window of a given size and shape different wavelet functions two scaling functions that may produce different multiresolutio because of its symmetric capabilities, which is often desirable since it exhibits the property of linear ,opening and closing , closing , and centered The by a structuring element two morphological filters used are , and , , and The two morphological filters used are Remote Sens. 2016, 8, 482 9 of 27 computationally more efficient, shift invariant, and un‐decimated ( Remote Sens. 2016, 8, 482 affect approach. the change detection approach. decomposition The final multi‐scale decomposition image stack then the change detection The final multi‐scale image stack then mposition, each decomposed image is inserted into a mathematical nlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more on a pixel , jq. . In image processing, erosion ,…are .and dilation , jqq are the basic jqq pXcompact on a pixelthey are the concatenation of erosion and dilation. They are defined as follows: X LR pi, In processing, are the basic analyses. In the wavelet has good support, smoothness goo phase, which is needed for image reconstruction. Another reason for using the biorthogonal wavelet LR LR they are the concatenation of erosion and dilation. They are defined as follows: (8) (8) . . , pXused …pi, … .and , , dilation are biorthogonal . The . , image …… …addition, . . , used erosion , pi, and closing , and , and two morphological filters opening 8, 482 2016, 8, 482 9 of 27 9 of 27 ampling) at each level by padding them with zero in order to preserve the and closing , , and The two morphological filters opening contains the lower resolution images at each level, as below: ns the lower resolution images at each level, as below: matical morphology defines a family of morphological filters, which Remote Sens. 2016, 8, 482 s on mathematical morphology the reader is referred to [27]. Morphological filters are defined operators used, and are defined as follows [27]: operators used, and are defined as follows [27]: localization. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter wa was that, rather than having one wavelet and one scaling function, biorthogonal wavelets have two , and closing , and The two morphological are opening they are the concatenation of erosion and dilation. They are defined as follows: , and two morphological filters used are opening , 9 of 27 greater flexibility in the construction of wavelet bases, we filters select used a The wavelet Remote Sens. 2016, 8, 482 they are the concatenation of erosion and dilation. They are defined as follows: Remote Sens. 2016, 8, 482 m at emphasizing spatial structures in a gray‐level image. For more Remote Sens. 2016, 8, 482 ,. . , distortions ,…, …the . different (11), the log‐ratio ructuring element , which is based on a moving window of a given size and shape centered , (8) Using 9 of 27 (11) (8) . . , … . , . . , … … … selected to avoid along image boundary. image different wavelet functions and two scaling functions that may produce multiresolution , , and and closing closing , , , and , and morphological wo morphological filters filters used used are opening are opening they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as fo , , , ∀ ∈ , (9) m the biorthogonal wavelet family. The biorthogonal wavelet was selected 2.3.2. Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology , and (9) closing , The pi, two morphological filters used are opening assification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology ology the reader is referred to [27]. Morphological filters are defined Remote Sens. 2016, 8, 482 pixel , . In image processing, erosion , and dilation , are the basic pX jqq “ min tX pi, jq , x u @ x P S pi, jq , , (11) s s Remote Sens. 2016, 8, 482 9 of 27 LR LR recursively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum analyses. In addition, the biorthogonal wavelet has opening good compact smoothness and good , , closing (11) , and 9 of 27 9 of he concatenation of erosion and dilation. They are defined as follows: ncatenation of erosion and dilation. They are defined as follows: , are and The two morphological filters are support, Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 ,opening and , closing , and 9 of 27 The two morphological filters used are c capabilities, which is often desirable since it exhibits the property of linear , , 9 of 27 and closing The two used morphological filters used they are the concatenation of erosion and dilation. They are defined as follows: hich is based on a moving window of a given size and shape centered Remote Sens. 2016, 8, 482 9 of 27 ors used, and are defined as follows [27]: , , opening , 9 of 27 , (11) , , (12) After the multi‐scale decomposition, each decomposed image is inserted into a mathematical er the multi‐scale decomposition, each decomposed image is inserted into a mathematical level for multi‐scale decomposition. localization. We choose the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was , , (12) Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 9 of 27 , , , ∀ ∈ , (10) 2.3.2. Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology they are the concatenation of erosion and dilation. They are defined as follows: Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology for image reconstruction. Another reason for using the biorthogonal wavelet they are the concatenation of erosion and dilation. They are defined as follows: , and closing , The two morphological filters used are opening they are the concatenation of erosion and dilation. They are defined as follows: ocessing, erosion , and dilation , are the basic pX pi, jqq “ max tX pi, jq , x u @ x P S pi, jq (10) Remote Sens. 2016, 8, 482 s s , and closing , , and The two morphological filters used are opening LR LR , along , ,, boundary. , computation (11) ,(11) logy framework. Mathematical morphology defines a family of morphological filters, which To reduce time, we only on the lower images p selected to avoid morphology framework. Mathematical morphology defines a family of morphological filters, which distortions the image Using , are the log‐ratio image is ,focused (12) , closing , opening ∀ used used ∈ filters ,are (9) , opening ,and closing and ,,, closing ,and and , resolution closing , ,and , and , , , and , an The two The ,morphological two The morphological two The morphological two filters morphological filters used are opening filters opening used opening are , , (12) ving one wavelet and one scaling function, biorthogonal wavelets have two , and closing , and The two morphological filters used are they are the concatenation of erosion and dilation. They are defined as follows: The effect of opening on a gray‐level image tends to suppress regions that are brighter than the as follows [27]: The effect of opening on a gray‐level image tends to suppress regions that are brighter than the they are the concatenation of erosion and dilation. They are defined as follows: After the multi‐scale decomposition, each decomposed image is inserted into a mathematical After the multi‐scale decomposition, each decomposed image is inserted into a mathematical Remote Sens. 2016, 8, 482 9 of 27 , closing , p pX LR are nonlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more and , pi, , , and The two morphological filters used are opening (11) inear operators that aim at emphasizing spatial structures in a gray‐level image. For more , used , , p , pX LR, pi, (11) decomposition level. Discarding the detail images is allowed as the information contents at a certa recursively decomposed into six resolution levels. Empirical analysis suggested six to be the optimum ,closing , jqqq, , (12) , jqqq and , and The two morphological filters are opening The two morphological filters used are opening and closing and Remote Sens. 2016, 8, 482 9 of 27 they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as follows: ons and two scaling functions that may produce different multiresolution , to and closing , The two morphological filters used are opening they are the concatenation of erosion and dilation. They are defined as follows: surroundings while closing suppresses regions that are darker than the surroundings. In order surroundings while closing suppresses regions that are darker than the surroundings. In order to The effect of opening on a gray‐level image tends to suppress regions that are brighter than the morphology framework. Mathematical morphology defines a family of morphological filters, which hology framework. Mathematical morphology defines a family of morphological filters, which details on mathematical morphology the reader is referred to [27]. Morphological filters are defined n mathematical morphology the reader is referred to [27]. Morphological filters are defined The effect of opening on a gray‐level image tends to suppress regions that are brighter than the they are the concatenation of erosion and dilation. They are defined as follows: , , , ∀ ∈ , resolution level are recovered at a higher level. Hence, the exclusion of the detail images does n level for multi‐scale decomposition. (9) , , , , (12) (12) , , , ∀ ∈ , ote Sens. 2016, 8, 482 9 of 27 (10) they are the concatenation of erosion and dilation. They are defined as follows: , , they are the concatenation of erosion and dilation. They are defined as follows: , , , , (11) he biorthogonal wavelet has good compact support, smoothness and good they are the concatenation of erosion and dilation. They are defined as follows: preserve the spatial structures of our original image, opening by reconstruction followed with closing closing , , and The two closing morphological filters used are opening preserve the spatial structures of our original image, opening by reconstruction followed with closing surroundings while closing suppresses regions In order The effect of opening on a gray‐level image tends to suppress regions that are brighter than the onlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more , the ,The effect of opening on a gray‐level image tends to suppress region ,The , final , that are darker , images , , , and the surroundings. (11),to (11) (11) to (1 by a structuring element , which is based on a moving window of a given size and shape centered cturing element , which is based on a moving window of a given size and shape centered surroundings while suppresses regions that are darker than In order than and closing and The two morphological filters affect the change detection approach. multi‐scale image the To reduce are nonlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more computation time, we focused on lower resolution ,decomposition per , ,only ,used (11) , the surroundings. (12), stack , , are opening (12) , , the fifth‐order filter, which has a filter length of 12. The fifth‐order filter was by reconstruction was applied. The sequence of first doing opening by reconstruction followed by , , (11) they are the concatenation of erosion and dilation. They are defined as follows: by reconstruction was applied. The sequence of first doing opening by reconstruction followed by preserve the spatial structures of our original image, opening by reconstruction followed with closing surroundings while closing suppresses regions that are darker than the surroundings. In order to effect of opening on a gray‐level image tends to suppress regions that are brighter than the of opening on a gray‐level image tends to suppress regions that are brighter than the , , surroundings while closing suppresses regions that are darker than the (11) details on mathematical morphology the reader is referred to [27]. Morphological filters are defined s on mathematical morphology the reader is referred to [27]. Morphological filters are defined on a pixel , . In image processing, erosion , and dilation , are the basic , and closing , , and The two morphological filters used are opening el , . In image processing, erosion , and dilation , are the basic , , , ∀ ∈ , pX pi, jqq “ p pX pi, jqqq (11) preserve the spatial structures of our original image, opening by reconstruction followed with closing The effect of opening on a gray‐level image tends to suppress regions that are brighter th (10) Remote Sens. 2016, 8, 482 they are the concatenation of erosion and dilation. They are defined as follows: contains the lower resolution images at each level, as below: decomposition level. Discarding the detail images is allowed as the information contents at a certain LR , , , , rtions along the image boundary. Using , the log‐ratio image LR is closing by reconstruction is reconstruction particularly designed to reduce noise in detection masks ,the surroundings. (12) without closing by is particularly designed to order reduce noise in without detection masks by reconstruction was applied. The sequence of first doing opening by reconstruction followed by preserve the spatial structures of our original image, opening by reconstruction followed with closing while ngs while closing closing suppresses regions suppresses regions that are darker that are darker than than the surroundings. In , order In to to preserve the spatial structures of our original image, opening by reconstru by a structuring element , which is based on a moving window of a given size and shape centered The effect of opening on a gray‐level image tends to suppress regions that are brighter than the tructuring element , which is based on a moving window of a given size and shape centered The effect of opening on a gray‐level image tends to suppress regions that are brighter than the operators used, and are defined as follows [27]: by reconstruction was applied. The sequence of first doing opening by reconstruction followed by surroundings while closing suppresses regions that are darker or y are the concatenation of erosion and dilation. They are defined as follows: The effect of opening on a gray‐level image tends to suppress regions that are bri s used, and are defined as follows [27]: resolution level are recovered at a higher level. Hence, the exclusion of the detail images does not , , , , , , , , than the surroundings. (12) (12)(11) (12)In (1 ,achieving … . . , boundary , in , it , is [27]. (12) d into six resolution levels. Empirical analysis suggested six to be the optimum . . , … … incurring loss of details in mask outlines [27]. Hence, relevant for , , (11) , ( incurring loss of details mask outlines Hence, it is relevant for achieving boundary closing by reconstruction is particularly designed to reduce noise in detection masks without by reconstruction was applied. The sequence of first doing opening by reconstruction followed by patial structures of our original image, opening by reconstruction followed with closing he spatial structures of our original image, opening by reconstruction followed with closing by reconstruction was applied. The sequence of first doing opening by on a pixel , . In image processing, erosion , and dilation , are the basic surroundings while closing suppresses regions that are darker than the surroundings. In order to , , pixel , . In image processing, erosion , and dilation , are the basic pX pi, jqq “ p pX pi, jqqq (12) (12) , and closing The two morphological filters used are opening surroundings while closing suppresses regions that are darker than the surroundings. In order to LR LR , , (12) closing by reconstruction is particularly designed to reduce noise in detection masks without preserve the spatial structures of our original image, opening by reconstruction followed with c surroundings while closing suppresses regions that are darker than the surrounding The effect of opening on a gray‐level image tends to suppress regions that are brighter affect the change The effect of opening on a gray‐level image tends to suppress regions that are brighter than the detection approach. The , final decomposition image then , ,∈ for achieving ,∈, particularly , reconstruction , ∀ noise , stack (9)reduce ,closing , of multi‐scale ∀details (9) omposition. preservation. The method requires an original image and a marker image. If the marker image ,reconstruction (11) preservation. The method requires an original image and a marker image. If the marker image incurring loss in , closing mask outlines [27]. Hence, it is relevant boundary by is designed to relevant reduce in detection masks without truction was applied. The sequence of first doing opening by reconstruction followed by ion was applied. The sequence of first doing opening by reconstruction followed by The effect of opening on a gray‐level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter than th by is particularly designed to noise In in preserve the spatial structures of our original image, opening by reconstruction followed with closing operators used, and are defined as follows [27]: tors used, and are defined as follows [27]: they are the concatenation of erosion and dilation. They are defined as follows: preserve the spatial structures of our original image, opening by reconstruction followed with closing incurring loss of details in mask outlines [27]. Hence, it is for achieving boundary by reconstruction was applied. The sequence of first doing opening by reconstruction follow preserve the spatial structures of our original image, opening by reconstruction followe surroundings while closing suppresses regions that are darker than the surroundings. The effect of opening on a gray‐level image tends to suppress regions that are brighter than the contains the lower resolution images at each level, as below: surroundings while closing suppresses regions that are darker than the surroundings. In order to The effect of opening on a gray-level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter than the tation time, we focused only on the lower resolution images per , , (12) The effect of opening on a gray‐level image tends to suppress regions that are brighter than the , , and the is obtained after erosion has been initially applied to the original image preservation. The method requires an original image and a marker image. If the marker image incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving boundary y construction reconstruction is particularly is particularly designed designed to reduce to reduce noise noise in detection in detection masks masks without without surroundings surroundings surroundings while surroundings while closing closing suppresses regions while while suppresses regions closing closing suppresses regions suppresses regions that are darker that are darker that are darker than that are darker the surroundings. than the surroundings. than the surroundings. than In the surroundings. order In order to In to order In order to incurring loss of details in mask outlines [27]. Hence, it is relevan by reconstruction was applied. The sequence of first doing opening by reconstruction followed by 2.3.2. Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology , detection , and the masks is obtained after erosion has been initially applied to the original image ,the surroundings. , In reduce (12) w by reconstruction was applied. The sequence of first doing opening by reconstruction followed by preservation. The method requires an original image and a marker image. If the marker image closing by particularly in preserve the spatial structures of our original image, opening by reconstruction followed wit surroundings while closing suppresses regions that are darker than to noise preserve the spatial structures of our original image, opening by reconstruction followed with closing ,… ,, darker ∈ designed , ,surroundings. (9) , while ,,. by reconstruction was applied. The sequence of first doing opening by reconstructio ∀ ,…∈regions is , that are darker (9) , ∀ than ∀than to order (10) , while ,The effect of opening on a gray‐level image tends to suppress regions that are brighter th ∀reconstruction , that are darker surroundings closing suppresses that than the In order to (10) closing suppresses regions the surroundings. order to scarding the detail images is allowed as the information contents at a certain ,∈the surroundings. , In order (8) ,are . it ,, relevant …∈for . to . ,preservation. The method requires an original image and a marker surroundings while closing suppresses regions In to preservation. The method requires an original image and a marker image. If the marker image loss of details of closing details in surroundings mask in mask outlines outlines [27]. [27]. Hence, Hence, it is is relevant achieving for achieving boundary boundary preserve the spatial structures of our original image, opening by reconstruction followed with closing preserve the spatial structures of our original image, opening by reconstruction followed with closing preserve the spatial structures of our original image, opening by reconstruction followed with closing preserve the spatial structures of our original image, opening by reconstruction followed with closin closing by reconstruction is particularly designed to reduce noise in detection masks without , , (12) , , and the is obtained after erosion has been initially applied to the original image original image is reconstructed by a series of iterative dilations of , then the resulting filter is by reconstruction is particularly designed reduce noise in detection masks without incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving closing by reconstruction is particularly designed to reduce noise in detection m by reconstruction was applied. The sequence of first doing opening by reconstruction foll preserve the spatial structures of our original image, opening by reconstruction followed with closing surroundings while closing suppresses regions that are darker than the surroundings. or original image is reconstructed by a series opening of iterative dilations of followed , then ,the , and the is obtained after erosion has been initially applied to the original image The effect of opening on a gray‐level image tends to suppress regions that are brighter than the by reconstruction was applied. The sequence of first doing opening by reconstruction followed by After the multi‐scale decomposition, each decomposed image is inserted into a mathematic preserve the spatial structures of our original image, by reconstruction withresulting closing filter is In bou preserve the spatial structures of our original image, opening by reconstruction followed with closing overed at a higher level. Hence, the exclusion of the detail images does not The effect of opening on a gray‐level image tends to suppress regions that are brighter than the preserve the spatial structures of our original image, opening by reconstruction followed with closing The method requires an original image and a marker on. The method requires an original image and a marker image. If the marker image image. If the marker image by reconstruction was applied. The sequence of first doing opening by reconstruction followed by by reconstruction was applied. The sequence of first doing opening by reconstruction followed by by reconstruction was applied. The sequence of first doing opening by reconstruction followed by by reconstruction was applied. The sequence of first doing opening by reconstruction followed b incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving boundary , , and the is obtained after erosion has been initially applied to the original image opening by reconstruction : incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving boundary is obtained after erosion has been initially applied to the original image preservation. The method requires an original image and a marker image. If the marker loss of in opening mask [27]. Hence, it filter is the relevant for filter achiev by is designed to masks reduce noise in is detection masks by reconstruction was applied. The sequence of first doing opening by reconstruction followed by preserve the spatial structures of our original image, opening by reconstruction followed with c opening by reconstruction : of image is reconstructed by a ,doing series iterative dilations of , without then resulting is ,sequence , particularly ∀of ∈of outlines , detection surroundings while closing suppresses regions that are darker than the surroundings. In order to (10) , original , incurring ,by ∀reconstruction details ∈ of ,stack (10) by multi‐scale reconstruction is closing particularly to reduce noise in morphology framework. Mathematical morphology defines a family of morphological filters, whic by reconstruction applied. The first by reconstruction followed by original image is was reconstructed a series iterative dilations , then the resulting The effect of opening on a gray‐level image tends to suppress regions that are brighter than the by reconstruction was applied. The sequence of first doing opening by reconstruction followed by tion approach. closing The final decomposition designed image then surroundings while suppresses regions that are darker than the surroundings. In masks order to by reconstruction was applied. The sequence of first doing opening by reconstruction followed by , noise ,in masks closing closing by reconstruction by closing reconstruction closing by reconstruction is by particularly reconstruction is closing particularly is designed particularly is designed particularly to reduce designed to reduce designed noise to reduce in masks to detection reduce in noise detection noise masks detection in without detection without masks without withou preservation. The method requires an original image and a marker image. If the marker image 2.3.2. Classification by Expectation‐Maximization (EM) Algorithm with Mathematical Morphology , , , and the , and the d after erosion has been initially applied to the original image er erosion has been initially applied to the original image preservation. The method requires an original image and a marker image. If the marker image preservation. The method requires an original image and a marker image. If the incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving closing by reconstruction is particularly designed to reduce noise in detection without by reconstruction was applied. The sequence of first doing opening by reconstruction follow opening by reconstruction : original image is reconstructed by a series of iterative dilations of , then the resulting filter is original image is reconstructed by a series of iterative dilations of preserve the spatial structures of our original image, opening by reconstruction followed with closing , , ab is obtained after erosion has been initially applied to the original image incurring loss of details in mask outlines [27]. to Hence, is relevant for achieving boundary are nonlinear operators that aim at emphasizing spatial structures in a gray‐level image. For mo opening by reconstruction : designed closing by reconstruction is particularly reduce noise in detection masks without incurring roundings while closing suppresses regions that are darker than the surroundings. In order to closing by reconstruction is in particularly designed to reduce noise in detection masks without ution images at each level, as below: preserve the spatial structures of our original image, opening by reconstruction followed with closing ,mask , it is ,[27]. (13) closing by reconstruction is particularly designed to reduce noise in detection masks without , , , (13) incurring incurring loss incurring of loss details incurring of details loss of mask loss in details of outlines details in outlines mask [27]. in mask outlines Hence, outlines Hence, it [27]. it relevant [27]. Hence, is relevant Hence, it for is achieving it relevant for is achieving relevant for boundary achieving for boundary achieving boundary boundar , , and the is obtained after erosion has been initially applied to the original image preservation. The method requires an original image and a marker 9 of 27 image. If the mark incurring loss details in mask outlines [27]. Hence, it is relevant for filter achieving closing by reconstruction is particularly designed to reduce noise detection masks wf , boundary is obtained after erosion has been initially applied to the original image opening by reconstruction : is e mage is reconstructed is After the multi‐scale decomposition, each decomposed image is inserted into a mathematical reconstructed by of a by series a series of iterative of iterative dilations dilations of of , then , opening by reconstruction then the resulting the resulting filter is : , and the by reconstruction was applied. The sequence of first doing opening by reconstruction followed by is obtained after erosion has been initially applied to the original image Remote Sens. 2016, 8, 482 preservation. The method requires an original image and a marker image. If the marker image details on mathematical morphology the reader is referred to [27]. Morphological filters are define loss of details in mask outlines [27]. Hence, it relevant achieving boundary preservation. original image is reconstructed by a series of is iterative dilations of in , then the resulting The effect of opening on a gray‐level image tends to suppress regions that are brighter th serve the spatial structures of our original image, opening by reconstruction followed with closing incurring loss of details in in mask outlines [27]. Hence, it it is boundary by reconstruction was applied. The sequence of first doing opening by reconstruction followed by incurring loss of details mask outlines [27]. Hence, is relevant relevant for achieving boundary ,,for , it is ,achieving (13) preservation. The method requires an original image and a marker preservation. The method requires an original image and a marker preservation. The method requires an original image and a marker preservation. The method requires an original image and a marker for image. If the marker image image. If the marker image image. If the marker image image. If the marker imag (8) . . , … … … . . , , , (13) preservation. The method requires an original image and a marker image. If the marker image incurring loss of details in mask outlines [27]. Hence, relevant for achieving construction y reconstruction : : initially applies dilation to the original image, Moreover, closing by reconstruction original image is reconstructed by a series of iterative dilations of , then the resulting filter is closing by reconstruction is particularly designed to reduce noise in detection masks without ,bou is obtained after erosion has been initially applied to the original image original is obtained after erosion has been initially applied to the original image image is reconstructed by a series of iterative dilations of q image. , then resulting filter is : initially applies dilation the original image, Moreover, closing by reconstruction by a structuring element , which is based on a moving window of a given size and shape centere morphology framework. Mathematical morphology defines a family of morphological filters, which opening by reconstruction p ,designed The method requires an original image and a marker Ifthe marker image is obtained original image is reconstructed by to a series of noise iterative dilations of , without then the res surroundings while closing suppresses regions that are darker than the surroundings. In , to , and the reconstruction was applied. The sequence of first doing opening by reconstruction followed by preservation. The method requires an original image and a marker the image. If the marker image closing by reconstruction is particularly reduce in detection masks preservation. The method requires an original image and a marker image. If the marker image , , (13) , , or , , , and the , , and the , , and the , , and th is obtained after erosion has been initially applied to the original image is obtained after erosion has been initially applied to the original image is obtained after erosion has been initially applied to the original image is obtained after erosion has been initially applied to the original image , and closing , , and The two morphological filters used are opening preservation. The method requires an original image and a marker image. If the marker opening by reconstruction : then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving boundary , , and the is obtained after erosion has been initially applied to the original image opening by reconstruction : initially applies dilation to the original image, Moreover, closing by reconstruction on a pixel , . In image processing, erosion , and dilation , are the bas then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: are nonlinear operators that aim at emphasizing spatial structures in a gray‐level image. For more opening by reconstruction : after erosion has been initially applied to the original image p “ pX pi, jqqq, and the original original image is reconstructed by a series of iterative dilations of , then the resulting preserve the spatial structures of our original image, opening by reconstruction followed with c ing by reconstruction is particularly designed to noise in without LRis 9 of 27 loss details outlines [27]. masks Hence, it relevant for achieving boundary detection initially applies to resulting the , original image, Moreover, closing by reconstruction original image is reconstructed series iterative dilations of , dilation then the is , and the is obtained after erosion has been initially applied to the original image ,incurring , , in of mask ,of by , a , reduce (13) ,, then , and the is obtained after erosion has been initially applied to the original image ,(13) ,filter original image, , of Remote Sens. 2016, 8, 482 9 of 27 they are the concatenation of erosion and dilation. They are defined as follows: original original image image original is reconstructed original is image reconstructed image is reconstructed by of is a iterative reconstructed series by a series of by iterative of a series by iterative a dilations series of , iterative dilations of of iterative dilations of dilations of , reconstruction then , the then resulting the resulting , the filter the then resulting filter is the resulting is filter is dila preservation. The method requires an original image and a marker dilation image. If the marker image , filter , an is obtained after erosion has been initially applied to the original image xpectation‐Maximization (EM) Algorithm with Mathematical Morphology initially applies to Moreover, closing by reconstruction details on mathematical morphology the reader is referred to [27]. Morphological filters are defined then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: operators used, and are defined as follows [27]: initially applies Moreover, closing by opening by reconstruction : image is reconstructed by a series dilations of , then the resulting filter is opening by original is reconstructed by a series of iterative dilations of then the resulting filter is by reconstruction was applied. The sequence of first doing opening by reconstruction follow urring loss of image details in mask outlines [27]. Hence, it is relevant for achieving boundary ens. 2016, 8, 482 9 of 27 preservation. The method requires an original image and a marker image. If the marker image opening by reconstruction then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: , , , the the (14)filter (13) original image is reconstructed by a : ,a series of of iterative is is ,, of of , , , then ,, ,, dilations (13) resulting (14) original image is reconstructed by series iterative dilations , then resulting filter , , opening by reconstruction opening by reconstruction opening by reconstruction opening by reconstruction : : : : : initially applies dilation dilation to by original image, the original image, r, closing over, closing by reconstruction by reconstruction , , then , and the is obtained after erosion has been initially applied to the original image by a structuring element , which is based on a moving window of a given size and shape centered then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: ,original Moreover, closing and closing , to , and ogical filters used are opening opening by reconstruction qq: reconstruction p p initially applies image is reconstructed by a series of iterative of dilation the resulting fi closing by reconstruction particularly to dilations reduce noise in detection masks w then reconstructs the original image by applying a series of iterative eros servation. The method requires an original image and a marker is image. If the marker image to the original i reconstruction ,the , designed , and the is obtained after erosion has been initially applied to the original image le decomposition, each decomposed image is inserted into a mathematical , t, , and , , initially applies , , (11) opening by reconstruction Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 9 of 27 closing The two morphological filters used , ∀ , , ∈ , , ,and (14) ( opening by reconstruction : opening ,are the basic t : are , , , , , , (14) (13) on a pixel , . In image processing, erosion , and dilation nstructs the original image by applying a series of iterative erosion. The resulting filter is: cts the original image by applying a series of iterative erosion. The resulting filter is: , and closing , , and e two morphological filters used are opening opening by reconstruction ploss pi, jqqq “, mint ,reconstruction (13) incurring of by in Hence, is relevant for dilation achieving bou tion of erosion and dilation. They are defined as follows: pX of . It is worth mentioning that in this Both filtering processes stop when : outlines initially applies dilation to the original image, Moreover, closing by reconstruction details initially applies the Moreover, closing by image reconstruction LR LR original is reconstructed series iterative the filter is to mask Xdilation pi, dilations Both filtering processes stop when it initially applies the o Moreover, closing by then reconstructs the original image by applying a series of iterative erosion. The resulting filt ,,jquto , and the btained after erosion has been initially applied to the original image . Mathematical morphology defines a family of morphological filters, which ,a by , dilations [27]. ,, . It is worth mentioning that in this original image, (13) , resulting ,of , ,, then (13) (13) original image is reconstructed , , then the (13) resulting is they are the concatenation of erosion and dilation. They are defined as follows: , of (14) , (1 ,filter , , , of , ,iterative , a series operators used, and are defined as follows [27]: preservation. The method requires an original image and a marker image. If the marker e the concatenation of erosion and dilation. They are defined as follows: paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad , , , then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: (13) opening by reconstruction : then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: paper, we used a fixed square shape with a size of 20 × 20 pixels. From an analysis of a broad dilation Both filtering processes stop when ,. It is worth mentioning that in this filter , resulting initially applies dilation by reconstruction then reconstructs the original image by applying a series of iterative erosion. The resul , the , toto ,the ∀,. It is worth mentioning that in this (13) ∈ , closing (1 ,and and closing , origin , , and The two The morphological two morphological filters used used are opening are that aim at emphasizing spatial structures in a gray‐level image. For more ginal image is reconstructed by a closing of reconstruction iterative , then is the filters (11) Both filtering processes stop when initially applies original image, Moreover, closing ,dilation qq,, : initially Moreover, by applies original image, then (12) to the , opening by reconstruction ,series , p p, of , Moreover, closing , , dilations (14) (14) ,Moreover, closing , by reconstruction ,,opening ,original image, ,original image, , to the , the initially applies dilation initially applies initially applies dilation initially applies dilation to the to dilation dilation to original image, the original imag Moreover, closing Moreover, closing by Moreover, closing reconstruction by reconstruction by reconstruction by reconstruction range of data from different change detection projects, we found (1) that most consistent results were , , (11) , , an is obtained after erosion has been initially applied to the original image range of data from different change detection projects, we found (1) that most consistent results were paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad . It is worth mentioning that in this Both filtering processes stop when initially applies to the original image, Moreover, closing by reconstruction . It is w Both filtering processes stop when then reconstructs the original image by applying a series of iterative erosion. The resulting morphology the reader is referred to [27]. Morphological filters are defined they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as follows: ∀series , , erosion. ning by reconstruction (9) paper, we used a fixed square shape then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: reconstructs the: original image by applying The, to resulting filter , ,by ,, iterative the , (11) ∈initially applies dilation original image, Moreover, closing reconstruction , , , ,, awith a size of 20 × 20 pixels. From an analysis of a broad , of (14) is:, (13) (13) dilation to the Moreover, closing by reconstruction , (14) then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: achieved with a 20 pixel window; and (2) that change detection performance changed slowly with , a series of , dilations original image, , ,initially applies The effect of opening on a gray‐level image tends to suppress regions that are brighter than the achieved with a 20 pixel window; and (2) that change detection performance changed slowly with range of data from different change detection projects, we found (1) that most consistent results were paper, we used a fixed square shape ,initially applies with a size of 20 × 20 pixels. From an analysis of a broad then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: original Moreover, closing image . It is worth mentioning that in this . It is worth mentioning that in this ring processes stop when paper, we used a fixed square shape , . It is worth mentioning that with a size of 20 × 20 pixels. F dilation to the original ifi by reconstruction then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: t filtering processes stop when , which is based on a moving window of a given size and shape centered is reconstructed by iterative of , then the resulting range of data from different change detection projects, we found (1) that most consistent results were Both filtering processes stop when then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: , , (12) , , , then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: , , , , (11) deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice t t , , , (13) , , , (14) surroundings while closing suppresses regions that are darker than the surroundings. In order to achieved with a 20 pixel window; and (2) that change detection performance changed slowly with d a fixed square shape used a fixed square shape ,deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice range of data from different change detection projects, we found (1) that most consistent results were with a size of 20 × 20 pixels. From an analysis of a broad ,Both filtering processes stop when ,with a size of 20 × 20 pixels. From an analysis of a broad ,pi, , . It is worth mentioning that in this , then reconstructs the original image by applying a series of iterative erosion. The resulting filte Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 9 of 27 ,LR,pi, , ,, X ,: LR mage processing, erosion and dilation , jqqq are the basic opening by reconstruction p reconstruction pX “, maxt (14) ∀ ∈ . It is worth mentioning that in this Both filtering processes stop when initially applies Moreover, closing by achieved with a 20 pixel window; and (2) that change detection performance changed slowly with paper, we used a fixed square shape , ,initially applies with a size of 20 × 20 pixels. From an analysis of a dilation original image, . It is worth mentioni Both filtering processes stop when , range of data from different change detection projects, we found (1) that m ,jqu 9 of 27 ,to the (12) , dilation , ,(10) (14) (14) (14) (1 to the original image, Moreover, closing by reconstruction , , (12) , , , (14) of the window size is not critical for change detection success. The new multi‐scale decomposition preserve the spatial structures of our original image, opening by reconstruction followed with closing of the window size is not critical for change detection success. The new multi‐scale decomposition deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice achieved with a 20 pixel window; and (2) that change detection performance changed slowly with rom different change detection projects, we found (1) that most consistent results were ata from different change detection projects, we found (1) that most consistent results were paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad , , achieved with a 20 pixel window; and (2) that change detection perform , , , with a size of 20 × 20 pixels. From an analy , , (14) defined as follows [27]: paper, we used a fixed square shape with a size of 20 × 20 pixels. From an analysis of a broad then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: ing on a gray‐level image tends to suppress regions that are brighter than the (14) deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice range of data from different change detection projects, we found (1) that most consistent result paper, we used a fixed square shape . It is worth mentioning th , Both filtering processes stop when initially applies original image, Moreover, closing by reconstruction dilation t to the . It is worth mentioning that in this Both filtering processes stop when then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: ,is , in this ,,, and t “ t´ 1 and t´ 1closing , The effect of opening on a gray‐level image tends to suppress regions that are brighter than the stack now contains morphological filtered images at each level in , as below: by reconstruction was applied. The sequence of first doing opening by reconstruction followed by , , , , , and closing and , , , and The two The morphological two morphological filters filters used used are opening are opening Both filtering processes stop when “ . It worth mentioning that (12) stack now contains morphological filtered images at each level in , as below: of the window size is not critical for change detection success. The new multi‐scale decomposition deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice with a 20 pixel window; and (2) that change detection performance changed slowly with a 20 pixel window; and (2) that change detection performance changed slowly with . It is worth mentioning that in this . It is worth mentioning that in this . It is worth mentioning that in this . It is worth mentioning that in th Both filtering processes stop when Both filtering processes stop when Both filtering processes stop when Both filtering processes stop when deviation from the 20 pixel setting. Hence, while 20 pixels was found to range of data from different change detection projects, we found (1) that most consistent results were range of data from different change detection projects, we found (1) that most consistent results were osing suppresses regions that are darker than the surroundings. In order to e effect of opening on a gray‐level image tends to suppress regions that are brighter than the of the window size is not critical for change detection success. The new multi‐scale decomposition achieved with a 20 pixel window; and (2) that change detection performance changed slowl range of data from different change detection projects, we found (1) that most consisten paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis o . It is worth mentioning that in this Both filtering processes stop when n reconstructs the original image by applying a series of iterative erosion. The resulting filter is: , with a size of 20 × 20 pixels. From an analysis of a broad , paper, we used a fixed square shape , , ∀ ∈ , (9) . It is worth mentioning that in this Both filtering processes stop when surroundings while closing suppresses regions that are darker than the surroundings. In order to , , , (14) . It is worth mentioning that in this Both filtering processes stop when closing by is particularly designed to ,ˆreduce noise detection masks without paper, we used areconstruction fixed square shape S. .pi, size 20 pixels. From an analysis. It is worth mentioning that of, as below: a broad they are the concatenation of erosion and dilation. They are defined as follows: they are the concatenation of erosion and dilation. They are defined as follows: stack now contains morphological filtered images at each level in of the window size is not critical for change detection success. The new multi‐scale decomposition m the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice paper, we used a fixed square shape paper, we used a fixed square shape paper, we used a fixed square shape , jq with with a size of 20 × 20 pixels. From an analysis of a broad , …by a…with a size of 20 × 20 pixels. From an analysis of a broad with a size of 20 × 20 pixels. From an analysis of a broad with a size of 20 × 20 pixels. From an analysis of a broa achieved with a 20 pixel window; and (2) that change detection performance changed slowly with , in (14) achieved with a 20 pixel window; and (2) that change detection performance changed slowly with uctures of our original image, opening by reconstruction followed with closing ndings while closing suppresses regions that are darker the surroundings. order paper, we used a fixed square shape stack now contains morphological filtered images at each level in deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact achieved with a 20 pixel window; and (2) that change detection performance change range of data from different change detection projects, we found (1) that most consistent res paper, we used a fixed square shape ,Both filtering processes stop when with a size of 20 × 20 pixels. From an analysis of a broad , than …of the window size is not critical for change detection success. The new .reconstruction . , ,of. . 20 , as below: initially applies dilation to the Moreover, closing (15)original i , (15) ,, …In …… . . , to range of data from different change detection projects, we found (1) that most consistent results were The effect of opening on a gray‐level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter tha paper, we used a fixed square shape , , outlines with a size of 20 × 20 pixels. From an analysis of a broad preserve the spatial structures of our original image, opening by reconstruction followed with closing paper, we used a fixed square shape with a size of 20 × 20 pixels. From an analysis of a broad incurring loss of details in mask [27]. Hence, it is relevant for achieving boundary range of data from different change detection projects, we found (1) that most consistent results were stack now contains morphological filtered images at each level in , as below: w size is not critical for change detection success. The new multi‐scale decomposition ndow size is not critical for change detection success. The new multi‐scale decomposition range of data from different change detection projects, we found (1) that most consistent results were range of data from different change detection projects, we found (1) that most consistent results were range of data from different change detection projects, we found (1) that most consistent results were range of data from different change detection projects, we found (1) that most consistent results we stack now contains morphological filtered images at each level in deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice , , , (14) deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice applied. The sequence of first doing opening by reconstruction followed by e the spatial structures of our original image, opening by reconstruction followed with closing of the window size is not critical for change detection success. The new multi‐scale decompo range of data from different change detection projects, we found (1) that most consistent results were paper, we used a fixed square shape with a size of 20 × 20 pixels. From an analysis of a then reconstructs the original image by applying a series of iterative erosion. The resulting filte … (15) (11) …. It is worth mentioning that in this … . . . It is worth mentioning that in this , (11) ,The importance of opening and closing by reconstruction is that it filters out darker and brighter , Both filtering processes stop when , surroundings ∀ surroundings , while ∈ achieved with a 20 pixel window; and (2) that change detection performance changed slo (10) achieved with a 20 pixel window; and (2) that change detection performance changed slowly with , , deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, th , suppresses regions …… . . , . ,. , that are darker while closing closing than the surroundings. than the surroundings. In order In ord to range of data from different change detection projects, we found (1) that most consistent results were (15) . . , , suppresses regions by reconstruction was applied. The sequence of first doing opening by reconstruction followed by …that are darker Both filtering processes stop when range of data from different change detection projects, we found (1) that most consistent results were The importance of opening and closing by reconstruction is that it filters out darker and brighter preservation. The method requires an original image and a marker image. If the marker image achieved with a reduce 20 pixelrange of data from different change detection projects, we found (1) that most consistent results window; and (2) that change detection performance changed slowly with w contains morphological filtered images at each level in now contains morphological filtered images at each level in , as below: , as below: achieved with a 20 pixel window; and (2) that change detection performance changed slowly with achieved with a 20 pixel window; and (2) that change detection performance changed slowly with achieved with a 20 pixel window; and (2) that change detection performance changed slowly with achieved with a 20 pixel window; and (2) that change detection performance changed slowly wi of the window size is not critical for change detection success. The new multi‐scale decomposition of the window size is not critical for change detection success. The new multi‐scale decomposition tion is particularly designed to noise in detection masks without nstruction was applied. The sequence of first doing opening by reconstruction followed by stack now contains morphological filtered images at each level in , as below: of the window size is not critical for change detection success. The new multi‐scale d deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exa achieved with a 20 pixel window; and (2) that change detection performance changed slowly with paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad (15) . . , … … … . . , . . , … … … . . , deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice preserve the spatial structures of our original image, opening by reconstruction followed with closing particularly Hence, while . It is worth mentioning that in this Both filtering processes stop when achieved with a 20 pixel window; and (2) that change detection performance changed slowly with elements smaller than , while preserving the original boundaries of structures in our image. closing by is obtained after erosion has been initially applied to the original image reconstruction is preserve the spatial structures of our original image, opening by reconstruction followed with cl to reduce noise in ,to detection masks without , with a size of 20 × 20 pixels. From an analysis of a broad achieved with a 20 pixel window; and (2) that change detection performance changed slowly with elements smaller than , designed , while preserving the original boundaries of structures in our image. The importance of opening and closing by reconstruction is that it filters out darker and brighter deviation from the, designed 20 pixel setting. pixels was found be optimal, the of , ,choice , and the , exact deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choic paper, we used a fixed square shape stack now contains morphological filtered images at each level in , as below: The importance of opening and closing by reconstruction is that it filters out darker and brighter stack now contains morphological filtered images at each level in , as below: ails mask outlines [27]. Hence, it is relevant for achieving boundary by in reconstruction is deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice particularly to reduce noise in 20detection masks without stack now contains morphological filtered images at each level in , as below: of the window size is not critical for change detection success. The new multi‐scale decom deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice achieved with a 20 pixel window; and (2) that change detection performance changed slowly range of data from different change detection projects, we found (1) that most consistent results were (15) (15) . . , . . , … … … … . … . , … . . , of the window size is not critical for change detection success. The new multi‐scale decomposition by reconstruction was applied. The sequence of first doing opening by reconstruction followed by by reconstruction was applied. The sequence of first doing opening by reconstruction followe deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice er, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad Note that morphological filtering leads to a quantization of the gray‐value space such that each image X . . , … … … . . , incurring loss of details in mask outlines [27]. Hence, it is relevant for achieving boundary range of data from different change detection projects, we found (1) that most consistent results were , , , , (12) (12) deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice Note that morphological filtering leads to a quantization of the gray‐value space such that each image elements smaller than , , while preserving the original boundaries of structures in our image. The importance of opening and closing by reconstruction is that it filters out darker and brighter The importance of opening and closing by reconstruction is that it fil the window size is not critical for change detection success. The new multi-scale decomposition of the window size is not critical for change detection success. The new multi‐scale decomposition of the window size is not critical for change detection success. The new multi‐scale decomposition of the window size is not critical for change detection success. The new multi‐scale decomposition of the window size is not critical for change detection success. The new multi‐scale decompositio MD , , while preserving the original boundaries of structures in our image. original image is reconstructed by a series of iterative dilations of , then the resulting filter is hod requires an original image and a marker image. If the marker image ng loss of details elements smaller than in mask outlines [27]. Hence, it is relevant for achieving boundary stack now contains morphological filtered images at each level in , as below: of the window size is not critical for change detection success. The new multi‐scale decomposition deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact c achieved with a 20 pixel window; and (2) that change detection performance changed slowly with stack now contains morphological filtered images at each level in (15)masks th .detection . .reduce . . , designed ……… . . , to , as below: closing closing by Both filtering processes stop when reconstruction by reconstruction particularly designed reduce to noise noise in detection wi , while preserving the original boundaries …(15) . It is worth mentioning that . ,. , , while preserving the original boundaries of structures in our image. …is …particularly …at .is . ,each th without ge of data from different change detection projects, we found (1) that most consistent results were in preservation. The method requires an original image and a marker of the window size is not critical for change detection success. The new multi‐scale decomposition can be normalized into the gray‐value range of [0, 255] without loss of information. At the k masks ,, as below: … … . , in image. If the marker image achieved with a 20 pixel window; and (2) that change detection performance changed slowly with of the window size is not critical for change detection success. The new multi‐scale decomposition in Note that morphological filtering leads to a quantization of the gray‐value space such that each image can be normalized into the gray‐value range of [0, 255] without loss of information. At the k elements smaller than mportance of opening and closing by reconstruction is that it filters out darker and brighter rtance of opening and closing by reconstruction is that it filters out darker and brighter elements smaller than , below: stack now contains morphological filtered images level , as below: in X MS , as opening by reconstruction stack now contains morphological filtered images at each level in stack now contains morphological filtered images at each level in stack now contains morphological filtered images at each level in of the window size is not critical for change detection success. The new multi‐scale decompo stack now contains morphological filtered images at each level in , as below: , as below: , as below: Note that morphological filtering leads to a quantization of the gray‐value space such that each image The importance of opening and closing by reconstruction is that it filters out darker and br : vation. The method requires an original image and a marker image. If the marker image stack now contains morphological filtered images at each level in , , and the n has been initially applied to the original image deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice The effect of opening on a gray‐level image tends to suppress regions that are brighter than the The effect of opening on a gray‐level image tends to suppress regions that are brighter than the incurring incurring loss of loss details of details in mask in mask outlines outlines [27]. [27]. Hence, Hence, it relevant is for for achieving boundary boun , with a size of 20 × 20 pixels. From an analysis of a th stack now contains morphological filtered images at each level in , as below: ieved with a 20 pixel window; and (2) that change detection performance changed slowly with , a lossless normalization is applied, leading to a float value between 0 and 255, such that: achieving . ., as below: , it is … …, and the . . ,relevant deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice stack now contains morphological filtered images at each level in level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: in paper, we used a fixed square shape can be normalized into the gray‐value range of [0, 255] without loss of information. At the k Note that morphological filtering leads to a quantization of the gray‐value space such that each image smaller than ler than level in , The importance of opening and closing by reconstruction is that it filters out darker and brighter , while preserving the original boundaries of structures in our image. , in , while preserving the original boundaries of structures in our image. , … is obtained after erosion has been initially applied to the original image (15) th The importance of opening and closing by reconstruction is that it filters out darker and brighter . Note that morphological filtering leads to a quantization of the gray‐value ., …….., !that are darker ) of the window size is not critical for change detection success. The new multi‐scale decomposition can be normalized into the gray‐value range of [0, 255] without loss of information. At the k elements smaller than ,… , while preserving the original boundaries of structures in our i The importance of opening and closing by reconstruction is that it filters out darke stack now contains morphological filtered images at each level in , as below: , , and the ned after erosion has been initially applied to the original image surroundings surroundings while while closing closing suppresses regions suppresses regions that are darker than the surroundings. than the surroundings. In order In order to to (15) (15) (15) (1 . . , . . … , … … … . . . … , . , … . . , . … , … … … . . … , … . . , nstructed by a series of iterative dilations of , then the resulting filter is preservation. The method requires an original image and a marker preservation. The method requires an original image and a marker image. If the marker image image. If the marker i range of data from different change detection projects, we found (1) that most consistent results K ´ 1 0 k th iation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice (15) . . , … … … . . , of the window size is not critical for change detection success. The new multi‐scale decomposition , dilations , the resulting filter is (15) level in phological filtering leads to a quantization of the gray‐value space such that each image morphological filtering leads to a quantization of the gray‐value space such that each image Xa MD X MD .., . . can be normalized into the gray‐value range of [0, 255] without lo . . .…. ..,.. ,.of .., X ,MD (15) (13) ,“ , while preserving the original boundaries of structures in our image. elements smaller than , can be normalized into the gray‐value range of [0, 255] without loss of information. At the k , while preserving the original boundaries of structures in our image. original image elements smaller than is in reconstructed by , a lossless normalization is applied, leading to a float value between 0 and 255, such that: series of iterative , then level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: Note that morphological filtering leads to a quantization of the gray‐value space such that each elements smaller than The importance of opening and closing by reconstruction is that it filters out darker and . in . .X , . ,MD … (15) ……the … …55 . . , while preserving the original boundaries of structures , The importance of opening and closing by reconstruction is that it filters out darker and brighter stack now contains morphological filtered images at each level in , as below: preserve the spatial structures of our original image, opening by reconstruction followed with closing preserve the spatial structures of our original image, opening by reconstruction followed with closing tion : l image is reconstructed by a series of iterative dilations of , then resulting filter is achieved with a 20 pixel window; and (2) that change detection performance changed slowly th . .th, he window size is not critical for change detection success. The new multi‐scale decomposition … … … . . , (16) , (16) , and the , , an is obtained after erosion has been initially applied to the original image is obtained after erosion has been initially applied to the original image stack now contains morphological filtered images at each level in , as below: 55 level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: normalized into the gray‐value range of [0, 255] without loss of information. At the k n be normalized into the gray‐value range of [0, 255] without loss of information. At the k The importance of opening and closing by reconstruction is that it filters out darker and brighter The importance of opening and closing by reconstruction is that it filters out darker and brighter The importance of opening and closing by reconstruction is that it filters out darker and brighter The importance of opening and closing by reconstruction is that it filters out darker and bright level in , , while preserving the original boundaries of structures in ou , a lossless normalization is applied, leading to a float value b Note that morphological filtering leads to a quantization of the gray‐value space such that each image Note that morphological filtering leads to a quantization of the gray‐value space such that each image opening by reconstruction : Note that morphological filtering leads to a quantization of the gray‐value space such th in can be normalized into the gray‐value range of [0, 255] without loss of information. At The importance of opening and closing by reconstruction is that it filters out darker and brighter elements smaller than , elements smaller than , while preserving the original boundaries of structures in our image. by reconstruction was applied. The sequence of first doing opening by reconstruction followed by by reconstruction was applied. The sequence of first doing opening by reconstruction followed by initially applies dilation to the original image, Moreover, closing by reconstruction g by reconstruction : The importance of opening and closing by reconstruction is that it filters out darker and brighter deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact lossless normalization is applied, leading to a float value between 0 and 255, such that: stack now contains morphological filtered images at each level in , as below: The importance of opening and closing by reconstruction is that it filters out darker and brighter th 55 of th The importance opening and closing by, reconstruction is that filters out darker brighter dilations .(13) , a . .series …… … .it . ,iterative , a lossless normalization is applied, leading to a float value between 0 and 255, such that: elements smaller than elements smaller than elements smaller than elements smaller than ,,image , while preserving the original boundaries of structures in our image. , while preserving the original boundaries of structures in our image. , reconstructed , while preserving the original boundaries of structures in our image. , while preserving the original boundaries of structures in our imag original is reconstructed is by a .series by iterative , then , the then resulting the (15) resulting filter is in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k (15) (16)fil in Note that morphological filtering leads to a quantization of the gray‐value space such that each image can be normalized into the gray‐value range of [0, 255] without loss of information. At the k 55 level in , a lossless normalization is applied, leading to a float value between 0 and 255, suc in can be normalized into the gray‐value range of [0, 255] without loss of informa Note that morphological filtering leads to a quantization of the gray‐value space such that ea elements smaller than , while preserving the original boundaries of structures in our image. The importance of opening and closing by reconstruction is that it filters out darker and br ,then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: ,image , oforiginal of and , of …of … … . .dilations , without (16) closing closing by reconstruction by reconstruction is ,particularly is particularly designed designed to reduce to reduce noise noise in detection in detection masks masks without elements smaller than , ,jq, , while preserving the original boundaries of structures in our image. of the window size is not critical for change detection success. The new multi‐scale decompo , , , (13) elements smaller than , while preserving the original boundaries of structures in our image. 55 elements smaller than S pi, while preserving the original boundaries of structures in our image. Note 55 Note that morphological filtering leads to a quantization of the gray‐value space such that each image Note that morphological filtering leads to a quantization of the gray‐value space such that each image Note that morphological filtering leads to a quantization of the gray‐value space such that each image Note that morphological filtering leads to a quantization of the gray‐value space such that each imag opening by reconstruction opening by reconstruction : : level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: After the mathematical morphology the posterior probability one “no‐ level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: (15) of level in , a lossless normalization is applied, leading to a float value between 0 and in … can be normalized into the gray‐value range of [0, 255] without loss of information. Note that morphological filtering leads to a quantization of the gray‐value space such that each image elements smaller than , step, , while preserving the original boundaries of structures in our i it . .outlines , [27]. …step, …morphology . we . , Hence, ,the , calculate ,[27]. After mathematical we calculate the posterior probability of one “no‐(16) th (13)boundary in of can be normalized into the gray‐value range of [0, 255] without loss of information. At the k th, as below: incurring incurring loss of loss details details in The importance of opening and closing by reconstruction is that it filters out darker and brighter mask in mask outlines Hence, it is relevant is relevant for achieving for achieving boundary Note that morphological filtering leads to a quantization of the gray‐value space such that each image leads stack now contains morphological filtered images at each level in The importance of opening and closing by reconstruction is that it filters out darker and brighter th th Note that morphological filtering leads to a quantization of the gray‐value space such that each image initially applies dilation to the original image, g by reconstruction 55 55 that morphological filtering to a quantization of the gray-value space such that each image in in in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k can be normalized into the gray‐value range of [0, 255] without loss of information. At the k in in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k can be normalized into the gray‐value range of [0, 255] without loss of information. At the k th change” and potentially several “change” classes at every one of the resolution levels, resulting in (16) (16) , , , (14) of 55 level in , a lossless normalization is applied, leading to a float value between 0 and 255, s in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k Note that morphological filtering leads to a quantization of the gray‐value space such that each change” and potentially several “change” classes at every one of the resolution levels, resulting in After the morphology mathematical step, we calculate the posterior probability one “no‐ elements smaller than , , while preserving the original boundaries of structures in our image. morphology initially applies dilation to the original image, Moreover, closing by reconstruction level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: th th preservation. The method requires an original image and a marker preservation. The method requires an original image and a marker image. If the marker image , image. If the marker image After the mathematical step, we calculate the posterior probability of one “no‐ , , , , , (13) in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k The importance of opening and closing by reconstruction is that it filters out darker and brighter th elements smaller than , , while preserving the original boundaries of structures in our image. in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k The initially applies dilation to the original image, oreover, closing by reconstruction riginal image by applying a series of iterative erosion. The resulting filter is: X MD can be normalized into the gray-value range ofassumed [0, 255] without of information. At the kcalculate level in level in level in level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: , a lossless normalization is applied, leading to a float value between 0 and 255, such that: , a lossless normalization is applied, leading to a float value between 0 and 255, such that: , a lossless normalization is applied, leading to a float value between 0 and 255, such tha 55 55 posterior probabilities per pixel. various classes are to calculate be a loss mixture of Gaussian . . , the … … . . , a (16) 55 (16) level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: in can be normalized into the gray‐value range of [0, 255] without loss of information. At posterior probabilities per pixel. The various classes are assumed to … be mixture Gaussian change” and potentially several “change” classes at every one of the resolution levels, resulting in After the mathematical morphology step, we posterior probability of one “no‐ After the mathematical morphology step, we of the poster Note that morphological filtering leads to a quantization of the gray‐value space such that each image then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: change” and potentially several “change” classes at every one of the resolution levels, resulting in , , and the , , and the is obtained after erosion has been initially applied to the original image is obtained after erosion has been initially applied to the original image level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: ments smaller than , , while preserving the original boundaries of structures in our image. to a float value . It is worth mentioning that in this Both filtering processes stop when level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: constructs the original image by applying a series of iterative erosion. The resulting filter is: level in Xdensity distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio ,Note that morphological filtering leads to a quantization of the gray‐value space such that each image a step, lossless normalization isposterior applied, leading between 0 and 255, such that: MD density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio 55 initially applies be initially applies dilation dilation to the to original image, the original im Moreover, closing Moreover, closing by reconstruction by reconstruction level in , a lossless normalization is applied, leading to a float value between 0 and 255, such posterior probabilities per pixel. The classes are assumed to be a mixture of th Gaussian change” and potentially several “change” classes at every one of the resolution levels, resulting in mathematical the mathematical morphology morphology step, we calculate we calculate posterior the probability probability of various one of assumed one “no‐ “no‐ change” and potentially several “change” classes at every one of the re in can be normalized into the gray‐value range of [0, 255] without loss of information. At the k thof 55 ,paper, we used a fixed square shape , per the (14) posterior probabilities pixel. The various classes are to a mixture of Gaussian After the mathematical morphology step, we calculate the posterior probability The importance of opening and closing by reconstruction is that it filters out darker and br (16) e that morphological filtering leads to a quantization of the gray‐value space such that each image in , can be normalized into the gray‐value range of [0, 255] without loss of information. At the k on , with a size of 20 × 20 pixels. From an analysis of a broad original original image image is reconstructed is reconstructed by a series by a series of iterative of iterative dilations dilations of of , then , the then resulting the resulting filter filter is is 55 55 55 55 , , , (14) ´ ¯ data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio (16) (16) (16) (1 55 then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: then reconstructs the original image by applying a series of iterative erosion. The resulting filte data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio After posterior probabilities per The various classes are assumed to (16) be a mixture of Gaussian nd potentially several “change” classes at every one of the otentially several “change” classes at every one of the we resolution levels, resulting in the resolution levels, resulting in posterior probabilities per pixel. The various are assumed to the mathematical morphology we calculate the posterior probability of one “no‐ level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: After density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio the range of data from different change detection projects, we found (1) that most consistent results were mathematical step, calculate the posterior probability one “no‐ thof , morphology ,pixel. , After (14) change” and potentially several “change” classes at every one of the classes resolution levels, resul morphology step, calculate the posterior probabilit elements smaller than , step, , while preserving the original boundaries of structures in our i 55 k mathematical k can be normalized into the gray‐value range of [0, 255] without loss of information. At the k we level in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: 55 (16) opening by reconstruction opening by reconstruction : : X ´ min X (16) data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of . It is worth mentioning that in this cesses stop when MD MD 55 data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio robabilities or probabilities per pixel. per pixel. The The various various classes classes are assumed are assumed to be to a be mixture a mixture of Gaussian of Gaussian density distributions. While an approximation, assuming Gaussian char change” and potentially several “change” classes at every one of the resolution levels, resulting in k change” and potentially several “change” classes at every one of the resolution levels, resulting in data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio posterior probabilities The various are to be a probability mixture of (14) Ga change” and potentially several “change” classes at every one of the resolution leve the step, mathematical morphology we , calculate of Note that morphological filtering leads to a quantization of the gray‐value space such that each per pixel. Both filtering processes stop when el in , a lossless normalization is applied, leading to a float value between 0 and 255, such that: ` kwe ` . It is worth mentioning that in this XAfter ˆ (16) the mathematical morphology calculate probability of the one achieved with a 20 pixel window; and (2) that change detection performance changed slowly with ,˘˘posterior ,55 step, classes , posterior “no‐ , , assumed MD “ . It is worth mentioning that in this k the Gaussian characteristics is only weakly affecting the performance of our change detection approach. square shape , After probabilities with a size of 20 × 20 pixels. From an analysis of a broad mathematical The normal th filtering processes stop when max X ´ min X Gaussian characteristics is only weakly affecting the performance of our change detection approach. data is near and closely resembles a Gaussian distribution. Hence, our assumption of data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio stributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio utions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio After After the mathematical the After After the mathematical morphology the mathematical morphology step, morphology step, we morphology calculate we step, calculate the step, we posterior calculate the we posterior calculate probability the posterior probability the posterior of probability one of “no‐ probability one “no‐ of one of “no‐ one “n data is not uncommon. Previous research [3] has found that the statisti 55 posterior probabilities per pixel. The various classes are assumed to be a mixture of Gaussian , , , , , , After posterior per pixel. various classes are assumed to be a mixture of Gaussian (13) (13) MD MD (16) data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of density distributions. While an approximation, assuming Gaussian characteristics for SAR log posterior probabilities per pixel. The various classes are assumed to be a mixtur change” and potentially several “change” classes at every one of the resolution levels, res the mathematical morphology step, we calculate the posterior probability of one “no‐ in can be normalized into the gray‐value range of [0, 255] without loss of information. At paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad 55 of one “no‐ change” and potentially several “change” classes at every one of the resolution levels, resulting in deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice (16) After the mathematical morphology step, we calculate the posterior probability Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing erent change detection projects, we found (1) that most consistent results were we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad After the mathematical morphology step, we calculate the posterior probability of one “no‐ Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing Gaussian characteristics is only weakly affecting the performance of our change detection approach. data is change” and potentially several “change” classes at every one of the near normal and closely resembles a Gaussian assumption of of of common. Previous research [3] has found that the statistical distribution of log‐ratio t uncommon. Previous research [3] has found that the statistical distribution of log‐ratio change” and potentially several “change” classes at every one of the change” and potentially several “change” classes at every one of the change” and potentially several “change” classes at every one of the distribution. resolution levels, resulting in calculate resolution levels, resulting in resolution levels, resulting in posterior resolution levels, resulting data is per near normal and a our Gaussian distribution. density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio The various to closely a resembles . It is worth mentioning that in this Hence, Both filtering processes stop when Both filtering processes stop when Gaussian characteristics is only weakly affecting the performance of our change detection approach. data is not uncommon. Previous research [3] has found that the statistical distribution of log density distributions. While an approximation, assuming Gaussian characteristics for After posterior probabilities pixel. classes are assumed to probability be a mixture change” and potentially several “change” classes at every one of the resolution levels, resulting in the mathematical morphology step, we the one level in , a lossless normalization is applied, leading to a float value between 0 and 255, suc 55 range of data from different change detection projects, we found (1) that most consistent results were posterior probabilities per pixel. The various classes are assumed be mixture of . It is worth mentioning that i Gaussian of the window size is not critical for change detection success. The new multi‐scale decomposition (16) change” and potentially several “change” classes at every one of the original image, resolution levels, resulting in The initially applies The initially applies dilation dilation to the to the original image, Moreover, closing Moreover, closing by reconstruction by reconstruction steps and, depending on the results of this study, may modify our approach in the future. To el window; and (2) that change detection performance changed slowly with f data from different change detection projects, we found (1) that most consistent results were change” and potentially several “change” classes at every one of the resolution levels, resulting in steps and, depending on the results of this study, may modify our approach in the future. To Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing Gaussian characteristics is only weakly affecting the performance of our change detection approach. normal ear normal and and closely closely resembles resembles a Gaussian a Gaussian distribution. distribution. Hence, Hence, our our assumption assumption of of posterior posterior probabilities posterior probabilities posterior per probabilities pixel. per probabilities pixel. per various pixel. per various classes pixel. The classes various The are various assumed are classes assumed classes to are be assumed to are a mixture be assumed a mixture to of be Gaussian to a of mixture be Gaussian a mixture of Gaussian of Gaussia Gaussian characteristics is only weakly affecting the performance of our data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio paper, we used a fixed square shape paper, we used a fixed square shape , with a size of 20 × 20 pixels. From an analysis of a broad , with a size of 20 × 20 pixels. From an analysis of a b Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing data is near normal and closely resembles a Gaussian distribution. Hence, our assumpt data is not uncommon. Previous research [3] has found that the statistical distributio density distributions. While an approximation, assuming Gaussian characteristics for SAR posterior probabilities per pixel. The various classes are assumed to be a mixture of Gaussian change” and potentially several “change” classes at every one of the resolution levels, result achieved with a 20 pixel window; and (2) that change detection performance changed slowly with After the mathematical morphology step, we calculate the posterior probability of one “no‐ density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio stack now contains morphological filtered images at each level in , as below: density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio posterior probabilities per pixel. The various classes are assumed to be a mixture of of Gaussian After the mathematical morphology step, we calculate the posterior probability of one then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: then reconstructs the original image by applying a series of iterative erosion. The resulting filter is: automate the calculation of the posterior probabilities, we employ an EM approach. The importance pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice ed with a 20 pixel window; and (2) that change detection performance changed slowly with posterior probabilities per pixel. The various classes are assumed to be a mixture Gaussian automate the calculation of the posterior probabilities, we employ an EM approach. The importance steps and, depending on the results of this study, may modify our approach in a the future. To Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing acteristics is only weakly affecting the performance of our change detection approach. characteristics is only weakly affecting the performance of our change detection approach. density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐rat Still, we are currently assessing the benefits of using non‐Gaussian desc data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of “no‐ data is data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio near normal and closely a distribution. Hence, our assumption of range of data from different change detection projects, we found (1) that most consistent results were range of data from different change detection projects, we found (1) that most consistent results 55 steps and, depending on the results of this study, may modify our approach in the future. To Gaussian characteristics is only weakly affecting the performance of our change detection app data is Gaussian near normal and closely resembles a resolution levels, resulting in Gaussian distribution. Hence, our a data is not uncommon. Previous research [3] has found that the statistical distribution of density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio resembles posterior probabilities per pixel. The various classes are assumed to be mixture of Gau deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice change” and potentially several “change” classes at every one of the density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio After the mathematical morphology step, we calculate the posterior probability of one “no‐ change” and potentially several “change” classes at every one of the resolution levels, resulting in of integrating mathematical morphology into our EM algorithm framework is to suppress the effect not critical for change detection success. The new multi‐scale decomposition on from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio (15) . . , … … … . . , of integrating mathematical morphology into our EM algorithm framework is to suppress the effect automate the calculation of the posterior probabilities, we employ an EM approach. The importance steps and, depending on the results of this study, may modify our approach in the future. To urrently assessing the benefits of using non‐Gaussian descriptions in these processing re currently assessing the benefits of using non‐Gaussian descriptions in these processing data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio data is not uncommon. Previous research [3] has found that the statistical distribution of log‐rat steps and, depending on the results of this study, may modify our Gaussian characteristics is only weakly affecting the performance of our change detection approach. Gaussian characteristics is only weakly affecting the performance of our change detection approach. achieved with a 20 pixel window; and (2) that change detection performance changed slowly with achieved with a 20 pixel window; and (2) that change detection performance changed slowly automate the calculation of the posterior probabilities, we employ an EM approach. The importance Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these proc data is near normal and a Gaussian distribution. our assuma data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio density distributions. While an approximation, assuming Gaussian characteristics for SAR log of the window size is not critical for change detection success. The new multi‐scale decomposition , Gaussian characteristics is only weakly affecting the performance of our change detect , per various , distribution. resembles ,The ,, closely probabilities pixel. various classes are assumed to be a mixture of Hence, Gaussian (14) (14) data is near posterior normal and closely resembles a Gaussian Hence, our assumption of nge” and potentially several “change” classes at every one of the resembles resolution levels, resulting in data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio posterior probabilities per pixel. The classes are assumed to be a mixture of Gaussian of background clutter that may constitute false positives after applying the EM algorithm. ns morphological filtered images at each level in , as below: window size is not critical for change detection success. The new multi‐scale decomposition data is not uncommon. Previous research [3] has found that the statistical distribution of log‐ratio of background clutter that may constitute false positives after applying the EM algorithm. of integrating mathematical morphology into our EM algorithm framework is to suppress the effect automate the calculation of the posterior probabilities, we employ an EM approach. The importance pending , depending on the on results the results of this of this study, study, may may modify modify our our approach approach in the in future. the future. To To data is data near is normal near data normal data is and near is closely and near normal closely normal resembles and resembles closely and a closely Gaussian a Gaussian resembles distribution. a Gaussian distribution. a Gaussian Hence, distribution. Hence, distribution. our assumption our Hence, assumption Hence, our of assumption our of assumption of automate the calculation of the posterior probabilities, we employ an EM Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact choice deviation from the 20 pixel setting. Hence, while 20 pixels was found to be optimal, the exact c of integrating mathematical morphology into our EM algorithm framework is to suppress the effect steps and, depending on the results of this study, may modify our approach in the Still, we are currently assessing the benefits of using non‐Gaussian descriptions in the Gaussian characteristics is only weakly affecting the performance of our change detection a data is near normal Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these processing and closely resembles a the Gaussian distribution. Hence, our assumption of data is not uncommon. Previous research [3] has found that the statistical distribution of log stack now contains morphological filtered images at each level in , as below: The importance of opening and closing by reconstruction is that it filters out darker and brighter density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio Gaussian characteristics is only weakly affecting the performance of our change detection approach. After mathematical morphology step, we calculate the posterior probability of futu one posterior probabilities per pixel. The various classes are assumed to be a mixture of Gaussian data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of density distributions. While an approximation, assuming Gaussian characteristics for SAR log‐ratio ack now contains morphological filtered images at each level in , as below: data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of of integrating mathematical morphology into our EM algorithm framework is to suppress the effect calculation of the posterior probabilities, we employ an EM approach. The importance the calculation of the posterior probabilities, we employ an EM approach. The importance Gaussian characteristics is only weakly affecting the performance of our change detection approach. Gaussian characteristics is only weakly affecting the performance of our change detection approach. Gaussian characteristics is only weakly affecting the performance of our change detection approach. Gaussian characteristics is only weakly affecting the performance of our change detection approac of steps the results depending may modify . It is worth mentioning that in this Both filtering processes stop when Both filtering processes stop when of integrating mathematical morphology into our EM algorithm framew and, on of . It is worth mentioning that in this this study, may modify our approach the future. To assumpti steps and, depending the this study, our approach in the future. To in of the window size is not critical for change detection success. The new multi‐scale decomposition of the window size is not critical for change detection success. The new multi‐scale decompo of background clutter that may constitute false positives after applying the EM algorithm. automate the calculation of the posterior probabilities, we employ an EM approach. The impo and, on the results this study, may modify our approach in t Still, we are currently assessing the benefits of using non‐Gaussian descriptions in these pr Gaussian characteristics is only weakly affecting the performance of our change detection approach. data near normal and closely resembles a of Gaussian distribution. Hence, our (15) is . . ,steps …on …of background clutter that may constitute false positives after applying the EM algorithm. … .depending . ,results Remote Sens. 2016, 8, 482 10 of 27 After the mathematical morphology step, we calculate the posterior probability of one “no-change” and potentially several “change” classes at every one of the K resolution levels, resulting in K posterior probabilities per pixel. The various classes are assumed to be a mixture of Gaussian density distributions. While an approximation, assuming Gaussian characteristics for SAR log-ratio data is not uncommon. Previous research [3] has found that the statistical distribution of log-ratio data is near normal and closely resembles a Gaussian distribution. Hence, our assumption of Gaussian characteristics is only weakly affecting the performance of our change detection approach. Still, we are currently assessing the benefits of using non-Gaussian descriptions in these processing steps and, depending on the results of this study, may modify our approach in the future. To automate the calculation of the posterior probabilities, we employ an EM approach. The importance of integrating mathematical morphology into our EM algorithm framework is to suppress the effect of background clutter that may constitute false positives after applying the EM algorithm. The EM algorithm focuses on discrimination between the posterior probability of one no-change pωu q and potentially several change classes pωc q. For each level in X MD , we model the probability density function ppX MD q of the normalized image series X MD as a mixture of N density distributions. This mixture contains the probability density functions, denoted ppX MD |ωc q and ppX MD |ωu q, and the prior probability, Ppωc q and Ppωu q. At the kth level in X MD , the probability density function (PDF) is modeled as: ´ 1´ ´ ¯ Nÿ ¯ (17) p X kMD “ ppX kMD |wcn qP pwcn q ` ppX kMD |ωu qP pωu q n“1 ř The first summand in Equation (17) ( nN “´ 11 ppX kMD |wcn qP pwcn qq represents the mixture of N ´ 1 change PDFs described by their respective likelihood ppX kMD |wcn q and prior probabilities P pwcn q, while the second summand describes the PDF of a single no-change class. It is worth noting that all PDFs in Equation (17) are assumed to be of Gaussian nature, such that the mean µωc and variance σω2 c are sufficient to define the density function associated with the change classes ωc , and mean µωu and variance σω2 u can be used to describe the density function related with the no-change classes ωu . The parameters in Equation (17) are estimated using an EM algorithm. Given X kMD is the kth level image in X MD , we inferred which class pωi “ pωc , ωu qq each pixel in X kMD belongs to, using ` ˘ θs “ µωc , σω2 c , µωu , σω2 u which is our current (best) estimate for the full distribution, and θ as our improved estimate. The expectation step at the sth iteration is calculated by the conditional expectation: ” ´ ¯ı ÿ ´ ¯ ´ ¯ Q pθ|θs q “ E In P ωi , X kMD |θ|X kMD , θs “ P ωi |X kMD , θs In P ωi , X kMD |θ (18) The maximization step maximizes Q pθ|θs q to acquire the next estimate: θps ` 1q “ argmaxQ pθ|θs q (19) θ The iterations cease when the absolute differences between the previous and current variables are below a tolerance value (ε). In our paper, we empirically set the tolerance value ε to 10´6 . Once the iterations cease, the final optimal θps ` 1q is used to calculate the posterior probability using Bayes’ formula. The EM algorithm (see Algorithm 1) is applied separately to all K levels of the multi-scale decomposition series, X MD , resulting in a stack of posterior probability maps ! ) PPMD “ PP0 , . . . .PPk , . . . ., PPK ´ 1 of depth K where each map contains the posterior probability of change and no-change classes, respectively. Our EM algorithm is illustrated as follows: iterations cease, the final optimal Ѳ is used to calculate the posterior probability using Bayes’ formula. The algorithm (see algorithm 1) is applied separately to all levels of the multi‐scale decomposition series, , resulting in a stack of posterior probability maps of depth where each map contains the posterior probability of change ,…. ,…., and no‐change classes, respectively. Our EM algorithm is illustrated as follows: Remote Sens. 2016, 8, 482 11 of 27 Algorithm 1. (Expectation‐Maximization) Algorithm 1. (Expectation-Maximization) Begin initialize Ɵ°, ε, s = 0 Begin initialize θ˝ , ε, s = 0 do s ← s + 1 do s Ð s + 1 E step: compute Q(Ɵ|Ɵs) E step: compute Q(θ|θs ) M step: Ɵs+1 ← argmax Q(Ɵ|Ɵs) M step: θs+1 Ð argmax Q(θ|θs ) s+1|Ɵs) − Q(Ɵ s|Ɵ s−1) ≤ ε s |θ s´1 ) ď ε until Q(Ɵ until Q(θs+1 |θs ) ´ Q(θ s+1 return Ӫ ← Ɵ return Ð θ s+1 end end 2.3.3. Selection of Number of Change Classes 2.3.3. Selection of Number of Change Classes In order order toto execute execute the the expectation expectation maximization maximization algorithm algorithm from from Section Section 2.3, 2.3, the the number number ofof In change classes that are present in an image has to be known. Selection of classes is a difficult task, change classes that are present in an image has to be known. Selection of classes is a difficult task, and care should be taken to avoid over‐ or under‐classifying the data. Various methods have been and care should be taken to avoid over- or under-classifying the data. Various methods have been proposed for selecting number of classes to fitbest the and data, and examples include the pNq to best proposed for selecting thethe number of classes the fit data, examples include the penalty penalty method [28], the cross‐validation method [29] and the minimum description length approach method [28], the cross-validation method [29] and the minimum description length approach [30]. In Remote Sens. 2016, 8, 482 of 27 [30]. In our paper, we developed a selection approach that identifies the number of required classes our paper, we developed a selection approach that identifies the number of required classes pNq11using a sum of square error (SSE) approach. The SSE approach utilized the measured data PDF, which is ( ) using a sum of square error (SSE) approach. The SSE approach utilized the measured data PDF, the statistical distribution of the highest decomposition level image, and the estimated PDF, which is which is the statistical distribution of the highest decomposition level image, and the estimated PDF, the statistical distribution estimated after applying the EM algorithm to the highest decomposition which is the statistical distribution estimated after applying the EM algorithm to the highest level image. The highest decomposition level image was used because at this level, most of the noise decomposition level image. The highest decomposition level image was used because at this level, in the image was filtered out. This approach seeks to minimize the sum of the square of the differences most of the noise in the image was filtered out. This approach seeks to minimize the sum of the square between the measured data PDF and the estimated PDF, as follows: of the differences between the measured data PDF and the estimated PDF, as follows: NN ÿ SSE “ pmii ´ f ii q2 (20) (20) = ( − ) i“2 where NN isisthe number of classes, mii is theisoriginal measured PDF and f ii isand the estimated where theoverall overall number of classes, the original measured PDF is the PDF. An example of the dependence of SSE on the number of classes used in f is shown in Figure estimated PDF. An example of the dependence of on the number of classesii used in is shown4. Initially, start with classes change (one and one no-change and calculate thecalculate SSE. For in Figure we 4. Initially, wetwo start with (one two classes change and oneclass) no-change class) and each extra class that is added into the procedure, its corresponding SSE is estimated. Plotting each the SSE. For each extra class that is added into the procedure, its corresponding SSE is estimated. class sequentially against its corresponding SSE leads to a continuous decrease of approximation error Plotting each class sequentially against its corresponding SSE leads to a continuous decrease of as NN gets larger. approximation error as NN gets larger. Figure Figure 4. 4. Selection Selectionof ofclasses classesused usedfor for final final classification. classification. The data from Section 4 is used in this example and SSE was carried out using a maximum of 20 classes, with each classes being added sequentially. The knee point on the curve in Figure 4 suggests that 3 is a good candidate for . 2.3.4. Measurement Level Fusion Remote Sens. 2016, 8, 482 12 of 27 The data from Section 4 is used in this example and SSE was carried out using a maximum of 20 classes, with each classes being added sequentially. The knee point on the curve in Figure 4 suggests that 3 is a good candidate for N. 2.3.4. Measurement Level Fusion We developed a measurement level fusion technique to accurately delineate the boundary of the changed region by using the posterior probability of each class at each multi-scale image to compose a final change detection map (M). We developed and tested five measurement level fusion methods that are briefly explained below: 1. Product rule fusion: This fusion method assumes conditionally statistical independence, and each pixel is assigned to the class that maximizes the posterior conditional probability, creating a change detection map. Each pixel is assigned to the class such that ωi “ argmax K ! ź K ! ) ) ź Ppωi |X kMD pi, jqq “ argmax PPi ωi εtωc ,ωu u i “ 0 2. (21) ωi εtωc ,ωu u i “ 0 Sum rule fusion: This method is very useful when there is a high level of noise leading to uncertainty in the classification process. This fusion method assumes each posterior probability map in X MD does not deviate much from its corresponding prior probabilities. Each pixel is assigned to the class such that K ! ) ÿ ωi “ argmax PPi (22) ωi εtωc ,ωu u i “ 0 3. Max rule fusion: Approximating the sum in Equation (22) by the maximum of the posterior probability, we obtain K ! ) ωi “ argmax max PPi (23) ωi εtωc ,ωu u i “ 0 4. Min rule fusion: This method is derived by bounding the product of posterior probability. Each pixel is assigned to the class such that K ! ) PPi min ωi εtωc ,ωu u i “ 0 ωi “ argmax 5. (24) Majority voting rule fusion: This method assigns a class to the pixel that carries the highest number of votes. Each pixel in each posterior probability map pPPMD q is converted to binary, i.e., $ ’ K ’ ( ’ & 1, i f max PPi ∆ibn “ ’ i“0 ’ ’ % 0, otherwise K ř ωi “ argmax ∆ibn (25) ωi εtωc ,ωu u i “ 0 The measurement level fusion with the lowest overall error, highest accuracy, and highest kappa coefficient is selected as the best fusion method and is used as our final change detection map. The next section shows how the best fusion method was selected. Remote Sens. 2016, 8, 482 13 of 27 3. Performance Assessment Using Synthetic Data To assess the performance of the developed change detection approach under both controlled and uncontrolled conditions, we have conducted two types of validation studies, both of which are presented in this paper. In this section, we summarize change detection results on a synthetic dataset to evaluate the performance and limitations of the technique under controlled conditions. Subsequently, in the next section, we show an application of the developed change detection technique to wildfire mapping. An area in Alaska is chosen and the change detection results are compared to ground truth measurements for validation. 3.1. Description of Synthetic Dataset A synthetic dataset of size 1152 ˆ 1152 pixels was generated from two SAR images acquired over the same area but at different times, and “change patches” were artificially introduced into the second of these images. A post-event image was generated by adding 2 dB to the radar cross-section of certain locations to form the change patches. Our approach was tested on synthetic data, because with synthetic data, we can control every aspect of the data, and also evaluate the accuracy with absoluteRemote reliability. the Sens. 2016,The 8, 482purpose of analyzing the synthetic dataset was to assess, more accurately, 13 of 27 robustness of the proposed approach and also to select the best measurement level fusion. with synthetic data, we can control every aspect of the data, and also evaluate the accuracy with It is worth mentioning that, with the exception of the border pixels, moderate changes in speckle absolute reliability. The purpose of analyzing the synthetic dataset was to assess, more accurately, noise dothe not significantly impact our classification This ismeasurement due to the mitigating robustness of the proposed approach and also toresult. select the best level fusion. effects of the wavelet decomposition and due to the way decisions at different decomposition levels are merged It is worth mentioning that, with the exception of the border pixels, moderate changes in speckle noise do not significantly impact our classification result. This is due to the mitigating effects of the in our approach. In homogeneous areas (e.g., within the change patches), the higher decomposition wavelet decomposition and due to the way decisions at different decomposition levels are merged levels carry the majority of the weight in the classification of these areas. As most speckleinnoise is our approach. In homogeneous areas (e.g., within the change patches), the higher decomposition removed at these higher decomposition levels, the classification result within homogeneous areas is levels carry the majority of the weight in the classification of these areas. As most speckle noise is very robust to speckle noise. Only in heterogeneous areas (the boundary pixels of our simulated data), removed at these higher decomposition levels, the classification result within homogeneous areas is where lower decomposition levelsOnly contribute more to areas the final classification very robust to speckle noise. in heterogeneous (the boundary pixelsresult, of our residual simulatedspeckle noise may have a measurable effect on performance. As there are only very few boundary data), where lower decomposition levels contribute more to the final classification result, residual pixels speckle may ahave a measurable effect on of performance. As there arewe only very few boundary and as we onlynoise applied very moderate change radar cross-section, believe that neglecting the pixels of and as we is only applied a very moderateFigure change5aofshows radar the cross-section, wegenerated believe thatbetween recalculation speckle a well-justified decision. ratio image neglecting the recalculation of speckle is a well-justified decision. Figure 5a shows the ratio image the synthetic pre-event image and the post-event images, while Figure 5b depicts the ground truth generated between the synthetic pre-event image and the post-event images, while Figure 5b depicts change the map. The faint change signatures in Figure 5a indicate that a challenging change detection ground truth change map. The faint change signatures in Figure 5a indicate that a challenging situation was change generated. detection situation was generated. Figure 5. (a) Ratio image calculated from pre- and post-event images; (b) ground truth change map. Figure 5. (a) Ratio image calculated from pre- and post-event images; (b) ground truth change map. 3.2. Performance Evaluation for Selecting Best Fusion Method 3.2. Performance Evaluation for Selecting Best Fusion Method For selecting the best performing fusion method, quantitative measurements are derived from binary change detection map obtained the different measurement level fusion Forthe selecting the best performing fusion using method, quantitative measurements aretechniques. derived from the In generating the confusion matrix which the accuracy of our classification result,techniques. the binary change detection map obtained usingshows the different measurement level fusion following defined quantities are computed: (i) the false alarm ( ), which signifies the no-change In generating the confusion matrix which shows the accuracy of our classification result, the pixels incorrectly classified as change. The false alarm rate ( ) is computed in percentage as following defined quantities are computed: (i) the false alarm pFAq, which signifies the no-change = / × 100%, where is the total number of unchanged pixels in the ground truth change detection map; (ii) the missed alarm ( ), which signifies the changed pixels incorrectly classified as no-change. The missed alarm rate ( ) is computed in percentage as = / × 100%, where is the total number of changed pixels in the ground truth change detection map; (iii) the overall error ( ), which is the percentage ratio of incorrect classification made (addition of both the false alarm rate and missed alarm rate). Hence, =( + )/( + ) × 100%; (iv) the overall Remote Sens. 2016, 8, 482 14 of 27 pixels incorrectly classified as change. The false alarm rate pFARq is computed in percentage as FAR “ FA{N1 ˆ 100%, where N1 is the total number of unchanged pixels in the ground truth change detection map; (ii) the missed alarm pMAq, which signifies the changed pixels incorrectly classified as no-change. The missed alarm rate pMARq is computed in percentage as MAR “ MA{N0 ˆ 100%, where N0 is the total number of changed pixels in the ground truth change detection map; (iii) the overall error pOEq, which is the percentage ratio of incorrect classification made (addition of both the false alarm rate and missed alarm rate). Hence, OE “ pFA ` MAq { pN1 ` N0 q ˆ 100%; (iv) the overall accuracy pOAq, which is calculated by adding the number of pixels classified correctly and dividing it by the total number of pixels. Therefore OA “ 100% ´ OE; and (v) the kappa coefficient, which measures the agreement between classification pixels and ground truth pixels [31]. A kappa value of 1 represents perfect agreement while a value of 0 represents no agreement. Table 1 shows the quantitative performance of each of the measurement level fusion techniques. It can be seen that product rule fusion achieved the best performance with an overall accuracy of 98.97%. Table 1. Overall accuracy, overall error, kappa coefficient, false positive and false negative of the measurement level fusion using Figure 5a. Product rule fusion Sum rule fusion Max rule fusion Min rule fusion Majority voting rule fusion Overall Accuracy (%) Overall Error (%) Kappa Coefficient False Positive (%) False Negative (%) 98.973 98.941 98.952 67.256 98.826 1.035 1.058 1.042 32.743 1.173 0.906 0.904 0.905 0.157 0.891 0.361 0.380 0.365 34.076 0.340 11.532 11.626 11.582 11.972 14.157 3.3. Comparison to Alternative Change Detection Methods To evaluate the performance of our proposed algorithm relative to the state-of-the-art, we conducted an extensive qualitative (Figure 6) and quantitative (Table 2) comparison of the results of our change detection approach to results of other recently published change detection methods. The following alternative methods were used in the comparison: (a) Method A: our proposed approach is used without fast non-local means filtering; (b) Method B: our proposed approach is used without 2D-SWT and measurement level fusion—the approach utilized mathematical morphology and the EM algorithm for classification; (c) Method C: our proposed approach is used without morphological filtering; (d) Method D: our proposed approach is used without the EM algorithm—the approach employed the thresholding algorithm proposed by [32] and majority voting rule fusion for classification; (e) Method E: semi-supervised change detection, based on using a kernel-based abnormal detection into the wavelet decomposition of the SAR image [33]; (f) Method F: image denoising using fast discrete curvelet transform via wrapping with the EM algorithm to produce the change detection map [34]; (g) Method G: using UDWT to obtain a multiresolution representation of the log-ratio image, then identifying the number of reliable scales, and producing the final change detection map using fusion at feature level (FFL_ARS) on all reliable scales [15]; (h) Method H: implementing probabilistic Bayesian inferencing with the EM algorithm to perform unsupervised thresholding over the images generated by the dual-tree complex wavelet transform (DT-CWT) at various scales, and moreover, using intra- and inter-scale data fusion to produce the final change detection map [12]; (i) Method I: obtaining a multiresolution representation of the log-ratio image using UDWT, then applying the Chan–Vese (region-based) active contour model to the multiresolution representation to give the final change detection map [18]. Based on Table 2 and Figure 6, one can observe that the change detection result from Method G showed the lowest performance of all tested methods with an overall accuracy of 68.412% and a kappa coefficient of 0.162. The low performance of this approach is due to the effectiveness of the method employed to select the optimal resolution level (reliable scale). The result of Method G was computed by averaging the reliable scales for every pixel which led to the low performance. Thus, most reliable scales contain a large amount of both geometrical details and Remote Sens. 2016, 8, 482 15 of 27 speckle components (lowest decomposition level) and a low amount of geometrical details and speckle components (highest decomposition level). The proposed method yields the best performance with an overall accuracy of 98.973% and a kappa coefficient of 0.906. It is worth mentioning that the aim of Methods A–D is to show the importance of each element in our proposed approach (see gray shade area in Table 2). Remote Sens. 2016, 8, 482 15 of 27 Figure 6. Change detection map showing: (a) proposed approach without fast non-local means filter; Figure 6. Change detection map showing: (a) proposed approach without fast non-local means filter; (b) proposed approach without 2D-SWT; (c) proposed approach without morphological filtering; (d) (b) proposed approach without 2D-SWT; (c)(e)proposed approach morphological proposed approach without EM algorithm; SVM approach; (f) imagewithout denoising using fast discrete filtering; curvelet transform via wrapping EM algorithm; (g) FFL_ARS approach; DT-CWT with intra-using fast (d) proposed approach without EM with algorithm; (e) SVM approach; (f) (h) image denoising and inter-scale data fusion approach; (i) UDWT with Chan–Vese active contour model approach. discrete curvelet transform via wrapping with EM algorithm; (g) FFL_ARS approach; (h) DT-CWT with intra- and inter-scale data fusion approach; (i) UDWT with Chan–Vese active contour model approach. Table 2. Overall accuracy, overall error, kappa coefficient, false positive and false negative of our proposed approach with alternative methods using Figure 5a. Table 2. Overall accuracy, overall error, kappa coefficient, false positive and falseFalse negative of our Overall Overall Kappa False proposed approach with alternative methods using Figure 5a. Accuracy (%) Error (%) Coefficient Positive (%) Negative (%) Proposed approach 98.973 Method A Overall 98.038 Method Accuracy B (%)98.024 Method C 98.352 Proposed approach Method D 98.973 98.647 Method A Method E 98.038 93.583 Method B Method F 98.024 94.763 Method C Method G 98.352 68.412 Method D Method H 98.647 94.527 Method E Method I 93.583 95.549 Method F Method G Method H Method I 94.763 68.412 94.527 95.549 1.035 1.961 Overall 1.975 Error (%) 1.648 1.0351.353 1.9616.416 1.9755.236 1.648 31.587 1.3535.473 6.4164.450 5.236 31.587 5.473 4.450 0.906 0.844 Kappa 0.827 Coefficient 0.836 0.906 0.878 0.844 0.575 0.827 0.594 0.836 0.162 0.878 0.519 0.575 0.659 0.594 0.162 0.519 0.659 0.361 1.843 False 1.127 Positive (%) 0.109 0.694 0.361 5.708 1.843 3.741 1.127 32.7730.109 2.953 0.694 3.409 5.708 3.741 32.773 2.953 3.409 11.532 3.806 False 15.201 Negative (%) 25.622 11.532 9.942 3.806 17.453 15.201 28.547 25.622 13.101 9.942 44.750 17.453 20.678 28.547 13.101 44.750 20.678 The study area for this real-data experiment is situated around the Tanana Flats region of Alaska, just east of the communities of Healy, Clear, Anderson and Nenana, and approximately 26 miles southwest of Fairbanks, Alaska (Figure 7). The topography of the area is relatively flat and is comprised of low-vegetation tundra and taiga regions with an interspersed network of small streams. Remote Sens. 2016, 8, 482 16 of 27 The forest vegetation of the area is diverse, containing white spruce, black spruce, aspen and birch. During the hot summer climate, this area is very prone to wildfires. 2 each year 4. Performance Analysis through Application to Wildfire Mapping In Alaska, wildfires affect thousands of km [35]. The dry conditions, increased lightning strikes, and higher-than-normal temperatures cause atmospheric effects that impact the 4.1. Description of Area strength of wildfires [35]. The strongest occurrences of wildfires in Alaska happen in the mid-summer months, occurrences annually. According to the around Alaska the interagency coordination center, The and study area for thisvary real-data experiment is situated Tanana Flats region of Alaska, the mapped firecommunities in our study area is called theAnderson Survey Line fire, which started in the year 2001. This just east of the of Healy, Clear, and Nenana, and approximately 26 miles fire, causedofby lightning, consumed 796 km2, and lasted four monthsflat from June to 14 southwest Fairbanks, Alaska (Figurean7).area Theof topography of the area is relatively and20 is comprised October 2001. During the and winter of regions the yearwith 2002,anthe fire scar disappears and thisstreams. is likely due the of low-vegetation tundra taiga interspersed network of small The to forest freezing of of thethe soils and the addition of a seasonal snow cover Section 4.5). Bybirch. the beginning of vegetation area is diverse, containing white spruce, black(see spruce, aspen and During the the year 2004, the trees in the fire scar area have regenerated. hot summer climate, this area is very prone to wildfires. Figure 7. Study area for wildfire mapping. 4.2. Description of SAR and Reference Data Used in this Study In Alaska, wildfires affect thousands of km2 each year [35]. The dry conditions, increased lightning strikes, higher-than-normal temperatures cause transmit atmospheric that impact strength of Weand obtained eight, HH-polarized (horizontal andeffects horizontal receivethe polarization), wildfires [35]. The strongest of wildfires Alaska the mid-summer months, Radarsat-1 images acquiredoccurrences between 2000 and 2004inover thehappen area ofininterest. Only acquisitions and occurrences vary annually. According to theimages Alaska were interagency center, theinmapped between April and October of preand post-fire used tocoordination avoid seasonal effects change fire in our For study area called the Survey the Lineacquisition fire, whichisstarted in the year 2001. This fire, caused detection. ease of is visual comparison, not shown successively according to the by lightning, consumed area of post-fire 796 km2 ,images. and lasted four images months(images from 20showing June to 14 2001. year but according to thean preand Pre-fire noOctober fire scar) are During thefirst, winter ofare theshown year 2002, the fire8a–d, scar while disappears and thisimages is likely(images due to the freezing the displayed and in Figure the post-fire showing fireofscar) soils and the addition a seasonal snow cover (see Section 4.5). By the beginning of the year 2004, the can be found in Figureof8e–h. trees in the fire scar area have regenerated. 4.2. Description of SAR and Reference Data Used in this Study We obtained eight, HH-polarized (horizontal transmit and horizontal receive polarization), Radarsat-1 images acquired between 2000 and 2004 over the area of interest. Only acquisitions between April and October of pre- and post-fire images were used to avoid seasonal effects in change detection. For ease of visual comparison, the acquisition is not shown successively according to the year but according to the pre- and post-fire images. Pre-fire images (images showing no fire scar) are Remote Sens. 2016, 8, 482 17 of 27 displayed first, and are shown in Figure 8a–d, while the post-fire images (images showing fire scar) can be found in Figure 8e–h. Remote Sens. 2016, 8, 482 17 of 27 Figure 8. Pre-fire image acquired on: (a) 21 June 2000; (b) 6 June 2001; (c) 23 October 2002; (d) 13 April Figure 8. Pre-fire image acquired on: (a) 21 June 2000; (b) 6 June 2001; (c) 23 October 2002; 2004. Post-fire image acquired on: (e) 3 August 2001; (f) 7 October 2001; (g) 22 August 2002; (h) 6 June (d) 13 April 2004. Post-fire image acquired on: (e) 3 August 2001; (f) 7 October 2001; (g) 22 August 2002; 2003. The triangular gray area in F denotes no data available. (h) 6 June 2003. The triangular gray area in F denotes no data available. Remote Sens. 2016, 8, 482 18 of 27 Each image has a size of 3584 ˆ 5056 pixels, corresponding to an area of 44.8 ˆ 63.2 square kilometers. The pixel spacing in each image is 12.5 ˆ 12.5 m with a range resolution of 26.8 m and an azimuth resolution of 24.7 m. The eight Radarsat-1 images were used to generate seven different ratio images, and were used for the analysis of seven different detection scenarios (scenarios 1–7 in Table 3). Table 3 shows the seven independent scenarios that we investigated to analyze the performance of our approach. The image acquisitions highlighted in gray in Table 3 are the post-fire images, while the rest of the acquisitions are pre-fire images. Three of the analyzed scenarios are “negative tests” as they do not include fire activity and test our ability to correctly identify no-change scenarios. The four remaining scenarios include change signatures related to wildfires and allow us to quantify our ability to correctly detect fire perimeters. Table 4 shows the incidence angle of all the individual image acquisitions. The image acquisitions highlighted in gray in Table 4 are images with a different incidence angle. All the images were radiometrically corrected using the Alaska IFSAR DEM which has a resolution of 5 m. The images were calibrated and co-registered as well, with ASF MapReady software, then geocoded to latitude and longitude [21]. Notice that the pre- and post-fire comparisons in scenarios 2 and 5 collated images with different incidence angles. These comparisons were included to test our approach’s ability to increase temporal resolution through the combination of images of varying geometry (see Section 2.1). Table 3. Combination of image acquisitions used to generate the ratio images. Image 1 (X1) Image 2 (X2) Scenarios Change Figure 21 June 2000 6 June 2001 6 June 2001 6 June 2001 6 June 2001 6 June 2001 6 June 2001 6 June 2001 23 October 2002 13 April 2004 3 August 2001 7 October 2001 22 August 2002 6 June 2003 scenario 1 scenario 2 scenario 3 scenario 4 scenario 5 scenario 6 scenario 7 Negative Negative Negative Positive Positive Positive Positive 11a 11b 11c 10 11d 11e 11f Table 4. Incidence angle of all the individual image acquisitions. Image Acquisitions Incidence Angle 21 June 2000 6 June 2001 23 October 2002 13 April 2004 3 August 2001 7 October 2001 22 August 2002 6 June 2003 27.207˝ 27.238˝ 33.736˝ 27.375˝ 27.236˝ 22.796˝ 27.296˝ 27.336˝ To validate achieved detection results, we used a fire area history map generated by the Alaska Fire Service as ground truth. The fire area history map was prepared using ground-based GPS, which, although effective, suffers from the limitation that everything inside the fire area history map is considered to be wildfire. We modified the fire area history map by using optical imagery to visually digitizing areas that are not actually affected by wildfire. Other limitations the Alaska Fire Service faced include time and safety. 4.3. Description of Classification Processes Based on several experiments, the performance of our approach was measured qualitatively and quantitatively by comparing detections to a fire area history map. To execute the classification process, ratio images (scenarios 1–7) previously shown in Table 3 were filtered and decomposed into Remote Sens. 2016, 8, 482 19 of 27 six resolution levels. As previously described, 2D-SWT using the biorthogonal wavelets procedure was employed. Remote Sens. 2016, 8, 482 19 of 27 As an example of the decomposition process, Figure 9 shows the decomposed images for scenario the decomposition 9 shows theeach decomposed images for how 4. LookingAsatan theexample highest of resolution images in process, Figure 9 Figure (top left panel of sub-Figure), we see scenariobetween 4. Looking at the highest images is ininitially Figure 9 low (topdue left panel each sub-Figure), the contrast background andresolution change regions to theofsignificant noise in the seecontrast how the is contrast between background and change regions is decomposition initially low due to the significant data. we This continuously improving with the increasing level. This is due to noise in the data. This contrast is continuously improving with the increasing decomposition level. a continuous improvement of noise suppression through the K levels. At the same time, the loss of This is due to a continuous improvement of noise suppression through the levels. At the same contrast and detail level can be observed for higher decomposition levels, which is particularly strong time, the loss of contrast and detail level can be observed for higher decomposition levels, which is at theparticularly border between background and the changed area. and changed area. strong the at the border between background Figure 9. SWT decomposition of scenario 4 from level (level (level1–6) 1–6)ofofdecomposition decomposition is Figure 9. SWT decomposition of scenario 4 fromlevel level11to to6.6. Each Each level is showing the resolution lower resolution (LL) image three detail (high frequency)images images (LH, showing the lower (LL) image (top (top left)left) andand the the three detail (high frequency) (LH, HL, HH). 0 is the original SAR image. HL, HH). Level 0 isLevel the original SAR image. per pixel and to produce the final classification map . Table 5 shows the quantitative comparison of each fusion method to confirm our choice of product rule fusion as the best-performing method. It is evident that product rule fusion gave a better result with an overall accuracy of 99.932% and a kappa coefficient of 0.997. Figure 10 shows a qualitative comparison of the burned area detected by our algorithm to the available fire area information (red outlines in Figure 10) retrieved from the Remote Sens. 2016, 8, 482 20 of 27 archives of the Alaska interagency coordination center. Table 5. Overall accuracy, overall K error, kappa coefficient, false positive the andposterior false negative of the of For every of the decomposed resolution levels, we calculate probability measurement level fusion using scenario 4. one “no-change”, one “negative change” and one “positive change” class, resulting in K posterior probabilities per pixel. Product rule fusion is then used to Kappa fuse the K classification results achieved Overall Overall False False per pixel and to produce the final classification map M. Table 5 showsPositive the quantitative comparison Accuracy (%) Error (%) Coefficient (%) Negative (%) rule fusion 99.932 0.269method. of each Product fusion method to confirm our choice of 0.067 product rule0.997 fusion as the0.021 best-performing Sumthat ruleproduct fusion rule fusion 98.991 1.008 0.967 2.543 and a It is evident gave a better result with an overall 0.653 accuracy of 99.932% Max rule fusion 97.609 2.390 0.924 2.476 2.020 by our kappa coefficient of 0.997. Figure 10 shows a qualitative comparison of the burned area detected Min rule fusion 98.273 1.726 0.944 1.815 1.342 algorithm to the available fire area information (red outlines in Figure 10) retrieved from the archives Majority voting rule fusion 98.842 1.158 0.961 0.417 4.359 of the Alaska interagency coordination center. Figure Figure 10. 10. Overlay Overlay of of the the change change detection detection map map resulting resulting from from product product rule fusion applied to scenario 4 with fire area history. history. A good automatic detection and ground truth information can be of observed. Table 5. match Overallbetween accuracy,the overall error, kappa coefficient, false positive and false negative the Particularly note (1) the “cleanness” of the4.detection result outside of the burn scar and (2) the close measurement level fusion using scenario preservation of the burn scar boundary that was achieved. Both properties indicate our method’s Overall Overall Kappasuppression False False ability to simultaneously achieve high performance noise and outline preservation. Accuracy (%) Error (%) Coefficient Positive (%) Negative (%) Classification results for the remaining six scenarios are shown in Figure 11. These results show good Product rule fusion 99.932 0.067 0.997 0.021 0.269 performance for both the “negative” and the1.008 “positive” tests. As expected, no extended 2.543 change was Sum rule fusion 98.991 0.967 0.653 identifiedMax forrule thefusion scenarios that did not span a forest fire. Some change was identified along 97.609 2.390 0.924 2.476 2.020a stream Min rule fusion 98.273 1.726 0.944 1.815 1.342 that crosses the area. This change is likely real and due to changes of river flow patterns or water Majority voting rule fusion 98.842 1.158 0.961 0.417 4.359 level variations. Even though three classes (see Section 2.3.3) were used for the classification in areas with negative change, none of the classes have a mean similar to that of a fire scar class. All “positive” A good match between the automatic detection and ground truth information can be observed. Particularly note (1) the “cleanness” of the detection result outside of the burn scar and (2) the close preservation of the burn scar boundary that was achieved. Both properties indicate our method’s ability to simultaneously achieve high performance noise suppression and outline preservation. Classification results for the remaining six scenarios are shown in Figure 11. These results show good performance for both the “negative” and the “positive” tests. As expected, no extended change was identified for the scenarios that did not span a forest fire. Some change was identified along a stream that crosses the area. This change is likely real and due to changes of river flow patterns or water level variations. Even though three classes (see Section 2.3.3) were used for the classification in areas with Remote Sens. 2016, 8, 482 21 of 27 negative change, none Remote Sens. 2016, 8, 482of the classes have a mean similar to that of a fire scar class. All “positive” 21 of 27 tests reliably identify the burn scar and retrace its boundary. Figure 11b,d showed that the classification reliably identify the burn and retrace its boundary. 11b,d showed that the this of firetests scars was not inhibited by thescar varying geometry of the pre-Figure and post-fire images. Thus, classification of fire scars was not inhibited by the varying geometry of the preand post-fire images. approach is able to increase the temporal sampling of our image acquisitions. The benefit of radiometric Thus, this approach is able to increase the temporal sampling of our image acquisitions. The benefit normalization for combining data with different observation geometry is quantified by comparing of radiometric normalization for combining data with different observation geometry is quantified scenario 5 with and without radiometric and geometric normalization. Employing radiometric and by comparing scenario 5 with and without radiometric and geometric normalization. Employing geometric normalization to scenario 5 gave an overall accuracy 97.78% and a kappa coefficient radiometric and geometric normalization to scenario 5 gave an of overall accuracy of 97.78% and a of 0.924.kappa Whencoefficient radiometric and geometric normalization is not normalization employed weishave an overallwe accuracy of 0.924. When radiometric and geometric not employed have of 87.08% a kappa coefficient 0.585. an and overall accuracy of 87.08%ofand a kappa coefficient of 0.585. Figure 11. Change detection map resulting from product rule fusion applied to (a) scenario 1; (b) Figure 11. Change detection map resulting from product rule fusion applied to (a) scenario 1; scenario 2; (c) scenario 3; (d) scenario 5; (e) scenario 6; (f) scenario 7. (b) scenario 2; (c) scenario 3; (d) scenario 5; (e) scenario 6; (f) scenario 7. 4.4. Comparison to Reference Change Detection Techniques 4.4. Comparison to Reference Change Detection Techniques Change detection results obtained from different approaches for scenario 4 are analyzed quantitatively in Table 6 and obtained qualitatively in Figure 12. Our proposed for approach yields Change detection results from different approaches scenario 4 an areoverall analyzed accuracy ofin99.932%, overall error of 0.067%, and a 12. kappa coefficient of 0.997. The performance of quantitatively Table 6anand qualitatively in Figure Our proposed approach yields an overall accuracy of 99.932%, an overall error of 0.067%, and a kappa coefficient of 0.997. The performance of Remote Sens. 2016, 8, 482 22 of 27 our approach without a fast non-local means filter gave a good result as well. However, the accuracy was reduced as a result of a false positive, causing the overall accuracy and kappa coefficient to drop to 98.402% and 0.946%, respectively. The performance decreased when wavelet transform was not utilized in our approach, giving an overall accuracy of 97.823% and a kappa coefficient of 0.927. Furthermore, without the application of the morphological filter the performance dropped, giving an overall accuracy of 97.963% and a kappa coefficient of 0.932. Also, in the absence of the EM algorithm in our approach, the accuracy dropped to 97.550% and the kappa coefficient dropped to 0.916. Given the results of Methods A–D (see gray shade area in Table 6), it is evident that each individual processing step is important for the effectiveness of our approach. With regard to other published techniques, Method E gave a lower accuracy when compared with our approach, with an overall accuracy of 93.816% and a kappa coefficient of 0.808. The lower accuracy is likely due to the region of interest definition which leads to spectral confusion and resulted in several numbers of false positives and false negatives in the final change detection result. The curvelet method (Method F) effectively filtered the ratio image; however, it only relies on the EM algorithm to classify the fire scar. The method has an overall accuracy of 93.054% and a kappa coefficient of 0.745. According to Method G, no filtering was done, along with the absence of the EM algorithm. The thresholding of fire scars was done manually, and this reduced the effectiveness of the method. Method G has an overall accuracy of 94.986% and a kappa coefficient of 0.830. The degradation in the accuracy of the change detection result from Method H, with respect to the proposed approach, is mainly because of the DT-CWT used. The DT-CWT is fairly robust to speckle noise, and the down-sampling process used is not able to detect the changes whose spatial supports are lost through the multi-scale decomposition. Method H gave an overall accuracy of 89.078% and a kappa coefficient of 0.605. Method I showed a decent detection with an overall accuracy of 96.587% and a kappa coefficient of 0.891. This is mainly due to the effective contour definition that partitions the ratio image into change and no-change. It is worth mentioning that all the methods used for comparative analysis were sent through the identical pre-processing chain including SAR data pre-processing, ratio image formation and logarithmic scaling, fast non-local means filtering and mathematical morphology filtering. While the fire scars are correctly delineated in Methods E–I, minute changes (false positive) in the background landscape were detected and amplified. Attempts to reduce the number of false positives for these techniques led to a significant distortion of the fire-scar boundaries and reduced the reliability of the final change detection map. Thus, our proposed approach produced a more consistent and more robust result as indicated by the information in Figure 12 and Table 6. Table 6. Overall accuracy, overall error, kappa coefficient, false positive and false negative of our proposed approach with alternative methods using scenario 4. Proposed approach Method A Method B Method C Method D Method E Method F Method G Method H Method I Overall Accuracy (%) Overall Error (%) Kappa Coefficient False Positive (%) False Negative (%) 99.932 0.067 0.997 0.021 0.269 98.402 97.823 97.963 97.550 93.816 93.054 94.986 89.078 96.587 1.597 2.176 2.036 2.450 6.183 6.945 5.013 10.921 3.412 0.946 0.927 0.932 0.916 0.808 0.745 0.830 0.605 0.891 0.172 0.797 0.713 0.248 5.526 1.105 3.964 3.959 2.887 7.761 8.144 7.754 11.972 9.021 32.206 4.659 41.035 5.681 Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 23 of 27 23 of 27 Proposed Approach a b c d e f g Figure 12. Cont. Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 h 24 of 27 24 of 27 i Figure Change detection detection map (a) (a) proposed approach without fast non-local means filter; Figure 12.12.Change mapshowing: showing: proposed approach without fast non-local means (b) proposed approach without 2D-SWT; (c) proposed approach without morphological filtering; (d) filter; (b) proposed approach without 2D-SWT; (c) proposed approach without morphological filtering; proposed approach without EM algorithm; (e) SVM approach; (f) image denoising using fast discrete (d) proposed approach without EM algorithm; (e) SVM approach; (f) image denoising using fast curvelet transform via wrapping with EM algorithm; (g) FFL_ARS approach; (h) DT-CWT with intradiscrete curvelet transform via wrapping with EM algorithm; (g) FFL_ARS approach; (h) DT-CWT with and inter-scale data fusion approach; (i) UDWT with Chan–Vese active contour model approach. intra- and inter-scale data fusion approach; (i) UDWT with Chan–Vese active contour model approach. 4.5. Electromagnetic Interpretation of Change Signatures 4.5. Electromagnetic Interpretation of Change Signatures After change features are detected using the aforementioned approach, the history of the After change are detected usinga the aforementioned approach, the history of the amplitude changefeatures can be analyzed to attempt geophysical interpretation of the observed change. amplitude change can be analyzed to attempt a geophysical interpretation of the observed change. In this section we show the kinds of analyses that can be conducted after a robust change detection In this section weapplied show the of analyses can be conducted after a robust change detection approach were to akinds time series of SARthat images. approach were applied to a time series of SAR images. In the presented wildfire scenarios, we observe (from the time-series of data that was analyzed) In the the burn presented scenarios, we observe (from the time-series of data thataffected was analyzed) that event wildfire led to a general increase of the average radar cross-section in the areas that the burn led to of a general increase oftothe cross-section in the This affected areas (Figure 13). event An increase about 4 dB relative theaverage pre-fire radar situation can be observed. is likely (Figure 13). increase of about 4 dB relative to the can burned be observed. This is likely due to anAn increase of double bounce scattering afterpre-fire the treesituation foliage was off, exposing the tree stumps and branches interaction with microwave signals. Figure 13exposing also showsthe dueremaining to an increase of double bounce for scattering after thethe tree foliage was burned off, that the tree radar cross-section in the affected area slowly decreases starting aboutFigure 1.5 years aftershows the remaining stumps and branches for interaction with the microwave signals. 13 also burn slow return toward pre-fire brightness is likely related regrowth in thethe that the event. radar This cross-section in the affected areaimage slowly decreases starting aboutto1.5 years after affected burn event.area. This slow return toward pre-fire image brightness is likely related to regrowth in the is worth mentioning that seasonal effects can be observed in the detection of the fire scar in affectedItarea. our While during the summer seasons the average radar brightness It isstudy wortharea. mentioning that seasonal effects can be observed in the detection of the fireremained scar in our significantly above the pre-burn level throughout the time series, it can be observed that the average study area. While during the summer seasons the average radar brightness remained significantly brightness value decrease during the winter periods, likely due to the freezing of the soils and the above the pre-burn level throughout the time series, it can be observed that the average brightness addition of a seasonal snow cover. After the winter periods, the average brightness value increases value decrease during the winter periods, likely due to the freezing of the soils and the addition of again near the radar cross-section of the previous summer. This is an interesting and significant a seasonal snow cover. After the winter periods, the average brightness value increases again near observation as it indicates that, for this test site and dataset, fire scar detection will be less successful the radar cross-section of the previous summer. This is an interesting and significant observation as it during winter. indicates that, for this test site and dataset, fire scar detection will be less successful during winter. To support our interpretation of the data, we plot a second time series in Figure 13 (red line) that To support our interpretation of theof data, we plot a second series in Figure 13 (redimages. line) that shows the average radar cross-section an area unaffected bytime fire in all preand post-fire shows average radar cross-section of an area by fire all preand post-fire images. of The The the seasonal brightness variations can also beunaffected observed for thisin area, supporting the assertion seasonal brightness variations can also be observed for this area, supporting the assertion of weather weather effects as the cause of the seasonal signal. effects as the cause of the seasonal signal. Remote Sens. 2016, 8, 482 Remote Sens. 2016, 8, 482 25 of 27 25 of 27 Figure 13. Time series showing fire scar area plotted over time (black line) and an area not affected by fire plotted over time (red line). The gray line represents a fire scar that cannot be seen as a result of of snow snow cover. cover. The The circle circle symbol symbol on on each each line line indicates indicates an an area area with with no no fire, fire, while while the the square square symbol symbol indicates a fire scar area. The gray shaded box area indicates winter months, and the indicates a fire scar area. The gray shaded box area indicates winter months, and the dashed dashed line line denotes denotes when when the the fire fire started. started. 5. Conclusions Conclusions A flexible and automatic change detection method was presented that is effective at identifying change signatures pairs of SAR The proposed approachapproach removes radiometric differences signaturesfrom from pairs of images. SAR images. The proposed removes radiometric between SAR imagesSAR acquired from different geometries by utilizingby radiometric terrain correction differences between images acquired from different geometries utilizing radiometric terrain (RTC) which enables change detection from different image geometries hence, improves the correction (RTC) which enables change detection from different imageand, geometries and, hence, temporal of surface change that can be achieved a given from database. Suppressing improves sampling the temporal sampling of surface change that canfrom be achieved a given database. background enhancing information by performing log-ratio operations, Suppressing information background and information andchange enhancing change information by performing log-ratio our approach detection performance while preserving change signature details. operations, ourdisplayed approach high displayed high detection performance while preserving change signature The integration of modern non-local filteringfiltering and 2D-SWT techniques provided robustness against details. The integration of modern non-local and 2D-SWT techniques provided robustness noise. classification performance, increased by integrating EM algorithm with mathematical againstThe noise. The classification performance, increased by an integrating an EM algorithm with morphology preservation of the geometric details the border regions, product mathematicaland morphology and preservation of the in geometric details in was the shown border when regions, was rule fusion was employed. Moreover, our approach gave a very high overall accuracy. In addition shown when product rule fusion was employed. Moreover, our approach gave a very high overall to analyzing the performance of our approach on synthetic data, we on used our algorithm toused conduct accuracy. In addition to analyzing the performance of our approach synthetic data, we our change detection in anchange area affected by wildfires. this by change detection analysis, we found that algorithm to conduct detection in an areaFrom affected wildfires. From this change detection aanalysis, fire scarwe could be that detected from the high available data.from In addition to accurately found a firewith scar high couldaccuracy be detected with accuracy the available data. In detecting the location and extentthe of the burn and scar,extent an analysis the scar, imageaninformation the addition to accurately detecting location of theof burn analysis of within the image information the detected scar revealed slow changes in image amplitudes time, most detected scar within revealed slow changes in image amplitudes over time, most likely relatedover to the regrowth likely related to the the burned regrowth of forest within the burned area. of forest within area. Comparison ofofour our approach to selected methods thatapproach (1) our performed approach Comparison approach to selected recentrecent methods showedshowed that (1) our performed a high overall accuracy and highpreservation; geometric preservation; (2)any neglecting any in of our the with a highwith overall accuracy and high geometric (2) neglecting of the steps steps in our approach result inchange an inferior change detection capability. approach will result inwill an inferior detection capability. maindrawbacks drawbacksofof proposed approach (1)assumption the assumption that the isimage is a The main thethe proposed approach are: are: (1) the that the image a mixture mixture of Gaussian distribution; and that the does approach does take fullof advantage of all the of Gaussian distribution; and (2) that the(2)approach not take fullnot advantage all the information information the speckle. Future workusing will explore Gammafor distribution for fitting the present in thepresent speckle.inFuture work will explore Gammausing distribution fitting the EM algorithm EM algorithm rather than the assumed Gaussian distribution. Non-uniform geometric and rather than the assumed Gaussian distribution. Non-uniform geometric and radiometric properties radiometric properties for all the areas of change in the synthetic image will be pursued as well. Remote Sens. 2016, 8, 482 26 of 27 for all the areas of change in the synthetic image will be pursued as well. Images with high varying geometry will be considered and analyzed. In addition, the development of the advanced approach for selecting the number of changed classes will be pursued. Acknowledgments: The authors want to thank the Alaska Satellite Facility (ASF) for providing access to the Radarsat-1 data used in this study. We furthermore thank the Geographic Information Network of Alaska (GINA) for the provision of high resolution Digital Elevation Models (DEMs) and for providing computing support. Thanks to Maria Tello Alonso for many fruitful discussions. The work presented here was funded through the NASA EPSCoR program under grant #NNX11AQ27A. Author Contributions: This research was written and revised by Olaniyi Ajadi with contributions from all coauthors. All authors have read and approved the final manuscript. Conflicts of Interest: The authors declare no conflict of interest. References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. Bruzzone, L.; Prieto, D.F. Automatic analysis of the difference image for unsupervised change detection. IEEE Trans. Geosci. Remote Sens. 2000, 38, 1171–1182. [CrossRef] Tello Alonso, M.; López-Martínez, C.; Mallorquí, J.J.; Salembier, P. Edge enhancement algorithm based on the wavelet transform for automatic edge detection in SAR images. IEEE Trans. Geosci. Remote Sens. 2011, 49, 222–235. [CrossRef] Dekker, R.J. Speckle filtering in satellite SAR change detection imagery. Int. J. Remote Sens. 1998, 19, 1133–1146. [CrossRef] Meyer, F.J.; McAlpin, D.B.; Gong, W.; Ajadi, O.; Arko, S.; Webley, P.W.; Dehn, J. Integrating SAR and derived products into operational volcano monitoring and decision support systems. ISPRS J. Photogramm Remote Sens. 2015, 100, 106–117. [CrossRef] Floyd, A.L.; Prakash, A.; Meyer, F.J.; Gens, R.; Liljedahl, A. Using synthetic aperture radar to define spring breakup on the kuparuk river, northern alaska. Arctic 2014, 67, 462–471. [CrossRef] Yun, S.-H.; Fielding, E.J.; Webb, F.H.; Simons, M. Damage Proxy Map from Interferometric Synthetic Aperture Radar Coherence. U.S. Patents 20120319893 A1, 20 December 2012. Tello, M.; Lopez-Martinez, C.; Mallorqui, J.J.; Danisi, A.; Di Martino, G.; Iodice, A.; Ruello, G.; Riccio, D. Characterization of local regularity in SAR imagery by means of multiscale techniques: Application to oil spill detection. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2007), Barcolona, Spain, 23–28 July 2007; pp. 5228–5231. D’Addabbo, A.; Refice, A.; Pasquariello, G.; Lovergine, F.P.; Capolongo, D.; Manfreda, S. A bayesian network for flood detection combining SAR imagery and ancillary data. IEEE Trans. Geosci. Remote Sens. 2016, 54, 3612–3625. [CrossRef] Siegert, F.; Hoffmann, A.A. The 1998 forest fires in east kalimantan (indonesia): A quantitative evaluation using high resolution, multitemporal ERS-2 SAR images and noaa-avhrr hotspot data. Remote Sens. Environ. 2000, 72, 64–77. [CrossRef] Loew, A.; Mauser, W. Generation of geometrically and radiometrically terrain corrected SAR image products. Remote Sens. Environ. 2007, 106, 337–349. [CrossRef] Bazi, Y.; Bruzzone, L.; Melgani, F. Image thresholding based on the EM algorithm and the generalized gaussian distribution. Pattern Recognit. 2007, 40, 619–634. [CrossRef] Celik, T. A bayesian approach to unsupervised multiscale change detection in synthetic aperture radar images. Signal Process. 2010, 90, 1471–1485. [CrossRef] Kasetkasem, T.; Varshney, P.K. An image change detection algorithm based on markov random field models. IEEE Trans. Geosci. Remote Sens. 2002, 40, 1815–1823. [CrossRef] Celik, T. Unsupervised change detection in satellite images using principal component analysis and-means clustering. IEEE Geosci. Remote Sens. Lett. 2009, 6, 772–776. [CrossRef] Bovolo, F.; Bruzzone, L. A detail-preserving scale-driven approach to change detection in multitemporal SAR images. IEEE Trans. Geosci. Remote Sens. 2005, 43, 2963–2972. [CrossRef] Celik, T. Multiscale change detection in multitemporal satellite images. IEEE Trans. Geosci. Remote Sens. 2009, 6, 820–824. [CrossRef] Remote Sens. 2016, 8, 482 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 27 of 27 Celik, T.; Ma, K.-K. Unsupervised change detection for satellite images using dual-tree complex wavelet transform. IEEE Trans. Geosci. Remote Sens. 2010, 48, 1199–1210. [CrossRef] Celik, T.; Ma, K.-K. Multitemporal image change detection using undecimated discrete wavelet transform and active contours. IEEE Trans. Geosci. Remote Sens. 2011, 49, 706–716. [CrossRef] Schmitt, A.; Wessel, B.; Roth, A. Curvelet-based change detection on SAR images for natural disaster mapping. Photogramm. Fernerkund. Geoinform. 2010, 2010, 463–474. [CrossRef] [PubMed] Small, D. Flattening gamma: Radiometric terrain correction for SAR imagery. IEEE Trans. Geosci. Remote Sens. 2011, 49, 3081–3093. [CrossRef] Gens, R.; Logan, T. Alaska Satellite Facility Software Tools: Manual; Geophysical Institute and University of Alaska Fairbanks: Fairbanks, AK, USA, 2003. Darbon, J.; Cunha, A.; Chan, T.F.; Osher, S.; Jensen, G.J. Fast nonlocal filtering applied to electron cryomicroscopy. In Proceedings of the 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro (ISBI 2008), Paris, France, 14–17 May 2008; pp. 1331–1334. Bazi, Y.; Bruzzone, L.; Melgani, F. An unsupervised approach based on the generalized gaussian model to automatic change detection in multitemporal SAR images. IEEE Trans. Geosci. Remote Sens. 2005, 43, 874–887. [CrossRef] Coppin, P.; Jonckheere, I.; Nackaerts, K.; Muys, B.; Lambin, E. Review articledigital change detection methods in ecosystem monitoring: A review. Int. J. Remote Sens. 2004, 25, 1565–1596. [CrossRef] Buades, A.; Coll, B.; Morel, J.-M. A review of image denoising algorithms, with a new one. Multiscale Model. Simul. 2005, 4, 490–530. [CrossRef] Wang, X.; Istepanian, R.S.; Song, Y.H. Microarray image enhancement by denoising using stationary wavelet transform. IEEE Trans. NanoBiosci. 2003, 2, 184–189. [CrossRef] Soille, P. Morphological Image Analysis: Principles and Applications; Springer Science & Business Media: New York, NY, USA, 2013. Zhu, X.X.; Bamler, R. Very high resolution spaceborne SAR tomography in urban environment. IEEE Trans. Geosci. Remote Sens. 2010, 48, 4296–4308. [CrossRef] Pham, D.L. Spatial models for fuzzy clustering. Comput. Vis. Image Underst. 2001, 84, 285–297. [CrossRef] Bischof, H.; Leonardis, A.; Selb, A. Mdl principle for robust vector quantisation. Pattern Anal. Appl. 1999, 2, 59–72. [CrossRef] ENVI User’s Guide. ENVI On-Line Software User’s Manual; ITT Visual Information Solutions: Boulder, CO, USA, 2008. Otsu, N. A threshold selection method from gray-level histograms. Automatica 1975, 11, 23–27. Mercier, G.; Girard-Ardhuin, F. Partially supervised oil-slick detection by SAR imagery using kernel expansion. IEEE Trans. Geosci. Remote Sens. 2006, 44, 2839–2846. [CrossRef] AlZubi, S.; Islam, N.; Abbod, M. Multiresolution analysis using wavelet, ridgelet, and curvelet transforms for medical image segmentation. J. Biomed. Imaging 2011, 2011, 4. [CrossRef] [PubMed] Wendler, G.; Conner, J.; Moore, B.; Shulski, M.; Stuefer, M. Climatology of alaskan wildfires with special emphasis on the extreme year of 2004. Theor. Appl. Climatol. 2011, 104, 459–472. [CrossRef] © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).