Jun Li 1,2 , José M. Bioucas-Dias 2 and Antonio Plaza 1
1 Hyperspectral Computing Laboratory
University of Extremadura, Cáceres, Spain
2 Instituto de Telecomunicaçoes, Instituto Superior Técnico, TULisbon, Portugal
Contact e-mails: {junli, aplaza}@unex.es, bioucas@lx.it.pt
A New Subspace Approach for Hyperspectral Classification
Talk Outline:
1. Challenges in hyperspectral image classification
2. Subspace projection
2.1. Subspace projection-based framework
2.2. Considered subspace projection techniques: PCA versus HySime
2.3. Integration with different classifiers (LDA, SVM, MLR)
3. Experimental results
3.1. Experiments with AVIRIS Indian Pines hyperspectral data
3.2. Experiments with ROSIS Pavia University hyperspectral
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Challenges in Hyperspectral Image Classification
Concept of hyperspectral imaging using NASA Jet Propulsion Laboratory’s Airborne Visible Infra-Red Imaging Spectrometer
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 1
Challenges in Hyperspectral Image Classification
• Imbalance between dimensionality and training samples, presence of mixed pixels
Ultraspectral
(1000’s of bands)
Hyperspectral
(100’s of bands)
Multispectral
(10’s of bands)
Panchromatic
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 2
Challenges in Hyperspectral Image Classification
• The special characteristics of hyperspectral data pose several processing problems:
1. The high-dimensional nature of hyperspectral data introduces important limitations in supervised classifiers, such as the limited availability of
training samples or the inherently complex structure of the data
2. There is a need to address the presence of mixed pixels resulting from insufficient spatial resolution and other phenomena in order to properly model the hyperspectral data
3. There is a need to develop computationally efficient algorithms, able to provide a response in a reasonable time and thus address the computational requirements of time-critical remote sensing applications
• In this work, we evaluate the impact of using subspace projection techniques prior to
supervised classification of hyperspectral image data while analyzing each of the aforementioned items
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 3
A New Subspace Approach for Hyperspectral Classification
Talk Outline:
1. Challenges in hyperspectral image classification
2. Subspace projection
2.1. Subspace projection-based framework
2.2. Considered subspace projection techniques: PCA versus HySime
2.3. Integration with different classifiers (LDA, SVM, MLR)
3. Experimental results
3.1. Experiments with AVIRIS Indian Pines hyperspectral data
3.2. Experiments with ROSIS Pavia University hyperspectral
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Subspace Projection-Based Framework
• Hyperspectral image data generally lives in a lower-dimensional subspace compared with the input feature dimensionality
• This can be exploited to address ill-posed problems given by limited training samples
• The projection into such subspaces allows us to specifically avoid spectral confusion due to mixed pixels, thus reducing their impact in the subsequent classification process
J. Li, J. M. Bioucas-Dias and A. Plaza, “Spectral-spatial hyperspectral image segmentation using subspace multinomial logistic regression and Markov random fields,” IEEE Transactions on Geoscience and
Remote Sensing, in press, 2011.
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 4
Considered Subspace Projection Techniques: PCA versus HySime
• High-dimensional data can be transformed effectively according to its distribution in feature space (e.g. by finding the most important directions or axes, establishing those axes as the references of a new coordinate system which takes into account data distribution)
• Orders the resulting components in decreasing order of variance
Component 1
Component 2
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 5
Considered Subspace Projection Techniques: PCA versus HySime
• High-dimensional data can be transformed effectively according to its distribution in feature space (e.g. by finding the most important directions or axes, establishing those axes as the references of a new coordinate system which takes into account data distribution)
• Orders the resulting components in decreasing order of variance
Band PCA 1 Band PCA 2 Band PCA 3 Band PCA 4 Band PCA 5
Band PCA 6 Band PCA 7 Band PCA 8 Band PCA 9 Band PCA 10
Band PCA 11 Band PCA 12 Band PCA 13 Band PCA 14 Band PCA 15
Band PCA 16 Band PCA 17 Band PCA 18 Band PCA 19 Band PCA 20
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 6
Considered Subspace Projection Techniques: PCA versus HySime
• A recently developed method for subspace identification in remotely sensed hyperspectral data, which offers several additional features with regards to principal component analysis and other subspace projection techniques
Principal Component Analysis
• Seeks for the projection that best represents the original hyperspectral data in least square sense
• Reduces the original signal into subset of eigenvectors without computing any noise statistics
• The difficulty in getting reliable noise estimates from the resulting eigenvalues is that these eigenvalues still represent mixtures of signal sources and noise
HySime
• HySime finds the subset of eigenvectors and the correspondent eigenvalues by minimizing the mean square error between the original signal and its projection onto the eigenvector subspace
• Uses multiple regressions for the estimation of the noise and signal covariance matrices
• Optimally represents the original signal with minimum error
J. M. Bioucas-Dias and J. M. P Nascimento, “Hyperspectral subspace identification,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 46, no. 8, pp. 2435-2445, 2008.
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 7
Supervised Classification Framework Tested in this Work
• Includes subspace projection and supervised classification based on training samples:
PCA, HySime
Subspace projection
Randomly selected
Test
Samples
Training
Samples
Supervised classifier
Test classification accuracy
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 8
Integration with different classifiers (LDA, SVM, MLR)
Integration of subspace-based framework with different classifiers.-
• Three different supervised classifiers tested in this work:
1. Linear discriminant analysis (LDA): find a linear combination of features which separate two or more classes; the resulting combination may be used as a linear classifier (only linearly separable classes will remain separable after applying LDA)
2. Support vector machine (SVM): constructs a set of hyperplanes in high-dimensional space; a good separation is achieved by the hyperplane that has the largest distance to the nearest training data points of any class
3. Multinomial logistic regression (MLR): models the posterior class distributions in a Bayesian framework, thus supplying (in addition to the boundaries between the classes) a degree of plausibility for such classes
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 9
A New Subspace Approach for Hyperspectral Classification
Talk Outline:
1. Challenges in hyperspectral image classification
2. Subspace projection
2.1. Classic techniques for subspace projection: PCA versus HySime
2.2. Subspace projection-based framework
2.3. Integration with different classifiers (LDA, SVM, MLR)
3. Experimental results
3.1. Experiments with AVIRIS Indian Pines hyperspectral data
3.2. Experiments with ROSIS Pavia University hyperspectral
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Experimental Results Using Real Hyperspectral Data Sets
• Challenging classification scenario due to spectrally similar classes
• Early growth stage of the agricultural features (only around 5% coverage of soil)
• 145x145 pixels, 202 spectral bands, 16 ground-truth classes
• 10366 labeled pixels (random training subsets evenly distributed among classes)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
10
Experimental Results Using Real Hyperspectral Data Sets
• Classification results using 160 training samples (10 training samples per class)
• For the SVM classifier we used the Gaussian RBF kernel after testing other kernels
• The mean accuracies (after 10 Monte Carlo runs) and processing times are reported
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 11
Experimental Results Using Real Hyperspectral Data Sets
• Classification results using 240 training samples (15 training samples per class)
• For the SVM classifier we used the Gaussian RBF kernel after testing other kernels
• The mean accuracies (after 10 Monte Carlo runs) and processing times are reported
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 12
Experimental Results Using Real Hyperspectral Data Sets
• Classification results using 320 training samples (20 training samples per class)
• For the SVM classifier we used the Gaussian RBF kernel after testing other kernels
• The mean accuracies (after 10 Monte Carlo runs) and processing times are reported
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 13
Experimental Results Using Real Hyperspectral Data Sets
• Classification results using 320 training samples (20 training samples per class)
SVM (OA=65.36%) Subspace SVM (OA=70.33%) LDA (OA=50.74%) Subspace LDA (OA=54.90%)
Linear MLR (OA=60.38%) Subspace MLR (OA=67.53%) Ground-truth
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 14
Experimental Results Using Real Hyperspectral Data Sets
False color composition Ground-truth Training data
Overall classification accuracies and kappa coefficient (in the parentheses) using different training sets for the ROSIS Pavia
University
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 15
Conclusions and Hints at Plausible Future Research
• We have evaluated the impact of subspace projection on supervised classification of remotely sensed hyperspectral image data sets
• Two dimensionality reduction methods have been used: PCA and HySime, although many others are available and could be used: MNF, OSP, VD
• Three different supervised classifiers considered: LDA, SVM, MLR
• Experimental results indicate that different approaches for hyperspectral image classification approaches can benefit from subspace projection, particularly when very limited training samples are available
• Subspace projection can be naturally integrated with multinomial logistic regression (MLR) classifiers, which greatly benefit from dimensionality reduction
• Future work will focus on the evaluation of other subspace projection approaches and hyperspectral data sets
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 16
• IEEE J-STARS Special Issue on Hyperspectral Image and Signal Processing
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011 17
Jun Li 1,2 , José M. Bioucas-Dias 2 and Antonio Plaza 1
1 Hyperspectral Computing Laboratory
University of Extremadura, Cáceres, Spain
2 Instituto de Telecomunicaçoes, Instituto Superior Técnico, TULisbon, Portugal
Contact e-mails: {junli, aplaza}@unex.es, bioucas@lx.it.pt