Estimation of Crop Acreage From Satellite Imagery Daniel Kuntz EENG 510 – Image Processing Fall 2014 Overview of Project Goals ● ● 2 Use satellite imagery to of the San Luis Valley to: – Estimate where crops are and where crops are not in a given scene – Segment fields from each other so that the number of fields and the area of each field can be determined – Classify each field by crop Why is this information useful? – Market estimation, getting early estimation of crop yields – Land management, legacy Landsat images can provide a wealth of information about land use if they can be interpreted effectively Agriculture in the San Luis Valley The San Luis Valley (SLV) makes a good test case for a proof of concept ● Fields are irrigated by center pivot and are mostly uniform circles with a north / south road grid running between them ● Climate is arid and fields are not located next to dense vegetation ● There are three main crops: potatoes, barley and alfalfa. 3 Field Detection Approach The first step in the process is to determine where the fields are. The Normalized Difference Vegetation Index (NVDI) and temporal averaging were used for this. ● The NVDI is defined as: NVDI = ● r IR −r RED , where: r x =reflectance in band x r IR +r RED For good results we need a temporal average of NVDI, which we can then segment. n 1 NVDI i=NVDI in each scene NVDI AVG= ∑ NVDI i , where: n i=1 n=number of scenes 4 Field Detection Results 5 Field Segmentation Approach The best method for segmenting the fields while preserving their area this particular case was to find the roads running between the fields. This is done by: ● Determining a box which encompasses both roads and fields, but not empty space: – ● Finding candidate areas where the roads might be – ● Morphological closing with a square element slightly bigger than each field Subtract the field area from the “box” area described above Finding the roads – Hough transform on the candidate areas 6 Field Segmentation Results 7 Field Segmentation Results: Good 8 Field Segmentation Results: Bad 9 10 Field Segmentation: Analysis ● We can get a big picture of how well the segmentation works by calculating the percentage of unsegmented fields: C = Number of fields with area > 700 (Multi-field blobs) A = Total area of fields with area > 700 A K = (Number of fields represented by the blobs) 620 K P = (Percentage of under-segmented Fields) N fields +( K−C ) ● We can see that this technique could use a bit of refinement N_fields 1272 C 72 A 92167 K 149 P 11.05 Classification Approach 11 Classification by crop was attempted for each field using the Support Vector Machine (SVM) classification method. Ground truth data for each field was obtained from USGS CropScape Website. ● ● ● Classification was attempted by using mean reflectance and variance in reflectance in all 7 bands along with the month of the image. Because of time and computational restraints, a subset of 96 fields was chosen for classification Chosen classification groups were: Potatoes, Barley, Alfalfa, Potatoes/Barely, and Other. Classification Approach Method for obtaining result: 1) Set was divided randomly into 80% training and 20% test datasets 2) SVM was trained with the 80% of data and tested with the other 20% 3) Step 1-2 was repeated 100 times and the mean and variance of all outcomes was taken. 4) After all trials had been run, the mean and variance of the result for each trial was recorded 12 Classification Results ● ● 13 Results achieved for crop classification were about 48% of fields properly identified, (much better than randomly guessing). Given the small dataset of about 1260 data points, and 96 fields, it is indicated that this method may be feasible with more work. Computer Program Output: After 100 trials, correct outcome stats: Mean: 0.48190476190476184 Var: 0.0009165784832451503 Aside: USGS Classifications 2008 Classifications 14 2012 Classifications Conclusion / Possible Improvements ● Overall Outcome – ● 15 Results of this project are not very impressive, but given the large amount of room for improvement, it proves the feasibility of something similar on a larger scale Possible Improvements – Using a “watershed” segmentation algorithm may provide a better, more general way to segment the fields. – A much larger dataset would be very beneficial. – Some way to gauge the start of the growing season would reduce the variability in phrenology of crops Questions ● Please ask some! 16