Your SlideShare is downloading. ×
chanussot.pdf
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×

Saving this for later?

Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime - even offline.

Text the download link to your phone

Standard text messaging rates apply

chanussot.pdf

471
views

Published on

Published in: Technology, Education

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
471
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
13
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. Unsupervised classification and spectral unmixing for sub-pixel labelling A.Villa , ,† , J.Chanussot , J.A. Benediktsson , C.Jutten GIPSA-lab, Signal & Image Dept., Grenoble Institute of Technology, France. Faculty of Electrical and Computer Engineering, University of Iceland, Iceland. † Aresys, Politecnico di Milano, Italy. IEEE IGARSS 2011 Vancouver, Canada - 2011
  • 2. A new approach to classification Experiments Conclusions Hyperspectral Images Widely used in remote sensing: λ √ Wide spectral range and large number of wavelengths - Trees - Grass √ Very high spectral resolution VIS NIR 0.4 μm 2.4 μm × Tradeoff between spectral and spatial resolution Jocelyn Chanussot Gipsa-Lab 2 / 21
  • 3. A new approach to classification Experiments Conclusions Challenges Low spatial resolution → appearance of mixed pixels • Common in hyperspectral images Pure pixel: • Traditional classifiers inadequate, 100% grass partially addressed by mixed pixel techniques Mixed pixel: • Critical for land cover mapping 70% metal sheet 30% grass Joint use (full + mixed techniques) desirable, but little investigated [Wang and Jia, 2010]. Jocelyn Chanussot Gipsa-Lab 3 / 21
  • 4. A new approach to classification Experiments Conclusions Challenges Low spatial resolution → appearance of mixed pixels • Common in hyperspectral images Pure pixel: • Traditional classifiers inadequate, 100% grass partially addressed by mixed pixel techniques Mixed pixel: • Critical for land cover mapping 70% metal sheet 30% grass Incorporation of spectral unmixing in the classification process: • Does it provide accuracy improvement? • Is it possible to improve the classification map spatial resolution? Jocelyn Chanussot Gipsa-Lab 3 / 21
  • 5. A new approach to classification Experiments Conclusions 1 A new approach to classification 2 Experiments 3 Conclusions Jocelyn Chanussot Gipsa-Lab 4 / 21
  • 6. A new approach to classification Experiments Conclusions Context Traditional techniques neglect sub-pixel and spatial information Additional information provided by unmixing not fully exploited 0.6 0.9 Pure pixel: 100% grass 1 0.9 0.8 0.6 1 Mixed pixel: 1 1 0.8 70% metal sheet 30% grass 0.9 0.6 1 Original image Classification Unmixing Finer resolution? How to jointly use full and mixed pixel techniques? Jocelyn Chanussot Gipsa-Lab 5 / 21
  • 7. A new approach to classification Experiments Conclusions Proposed Approach Low resolution hyperpspectral data Unmixing Classes Abundances identification maps Classification "Upsampled" classification map Spatial regularization Final map Jocelyn Chanussot Gipsa-Lab 6 / 21
  • 8. A new approach to classification Experiments Conclusions Proposed Approach 1. Abundances fractions are computed from a HSI Step 1: Low resolution hyperpspectral data Pure pixel: 100% grass Step 1 Step 2 Mixed pixel: Classes Abundances 70% metal sheet identification maps 30% grass Step 2: "Upsampled" 0.6 classification map 0.9 1 0.9 0.8 Spatial regularization 0.6 1 1 1 0.8 Final map 0.9 0.6 1 Jocelyn Chanussot Gipsa-Lab 7 / 21
  • 9. A new approach to classification Experiments Conclusions The proposed approach M = Mixed pixel M Proposed method M M The abundances computation is in two steps, to take the spatial information into account: 1. Pixels with an abundance over a certain threshold are considered ’pure’ M M 2. Abundances of ’mixed’ pixels are M M computed by selecting as endmembers pixels spatially close M Jocelyn Chanussot Gipsa-Lab 8 / 21
  • 10. A new approach to classification Experiments Conclusions The proposed approach M = Mixed pixel Proposed method The abundances computation is in two steps, to take the spatial information into account: 1. Pixels with an abundance over a certain threshold are considered ’pure’ 2. Abundances of ’mixed’ pixels are computed by selecting as endmembers pixels spatially close Jocelyn Chanussot Gipsa-Lab 8 / 21
  • 11. A new approach to classification Experiments Conclusions Proposed Approach 2. Creation of a finer classification map Step 2: 0.6 Low resolution hyperpspectral data 0.9 1 0.9 0.8 0.6 1 Step 2 1 1 0.8 Classes Abundances identification maps 0.9 0.6 1 Step 3 "Upsampled" Step 3: classification map Spatial regularization Final map Jocelyn Chanussot Gipsa-Lab 9 / 21
  • 12. A new approach to classification Experiments Conclusions Proposed Approach 3. Final spatial regularization Step 3: Low resolution hyperpspectral data Classes Abundances identification maps Step 3 "Upsampled" Step 4: classification map Step 4 Spatial regularization Final map Jocelyn Chanussot Gipsa-Lab 10 / 21
  • 13. A new approach to classification Experiments Conclusions Spatial regularization Criterion: minimization of the total perimeter of the connected areas (e.g., belonging to the same class) M 0,5 0,3 0,2 M 0,5 0,3 0,2 0,7 M 0,3 M 0,6 0,4 0,7 M 0,3 M 0,6 0,4 0,9 0,8 0,9 0,6 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,1 0,2 M 0,4 M 0,1 0,1 0,2 0,9 0,9 0,9 0,9 M 0,1 M 0,7 0,1 0,3 M 0,1 M 0,7 0,1 0,3 M 0,9 0,1 M 0,9 0,1 Criterion not satisfied Criterion satisfied Jocelyn Chanussot Gipsa-Lab 11 / 21
  • 14. A new approach to classification Experiments Conclusions Spectral unmixing based approach [Villa Novelties introduced: et al., 2010] 1. Retrieve classes with unsupervised 1. VCA for class retrieval clustering (→ more robust to outliers) 2. FCLS for abundance determination 2. Include spatial information (→ use more accurate 3. Simulated Annealing for spatial endmembers) regularization Jocelyn Chanussot Gipsa-Lab 12 / 21
  • 15. A new approach to classification Experiments Conclusions VCA vs. K-MEANS 6000 6000 5000 5000 4000 4000 3000 3000 2000 2000 1000 1000 0 0 4000 4000 2000 2000 6000 4000 5000 6000 0 2000 3000 4000 5000 0 1000 2000 3000 0 1000 0 VCA K-MEANS Jocelyn Chanussot Gipsa-Lab 13 / 21
  • 16. A new approach to classification Experiments Conclusions 1 A new approach to classification 2 Experiments 3 Conclusions Jocelyn Chanussot Gipsa-Lab 14 / 21
  • 17. A new approach to classification Experiments Conclusions How to verify the results? Decrease original resolution Final map Proposed (sub-pixel precision) approach Jocelyn Chanussot Gipsa-Lab 15 / 21
  • 18. A new approach to classification Experiments Conclusions Experiments on real data ROSIS University data set AISA data set • Classification of a metal sheet roof • 400×500 pixels area, six classes of (120×90 pixels) interest • 1.3 m spatial resolution, 103 • 6 m spatial resolution, 252 spectral spectral bands. bands • Spatial resolution of the original • Spatial resolution of the original data data degraded of a factor 3 degraded of a factor 5 Jocelyn Chanussot Gipsa-Lab 16 / 21
  • 19. A new approach to classification Experiments Conclusions Real data sets ROSIS data set: Original Image K-means (93.75%) VCA+SU (96.95%) KM+SU (95.89%) Jocelyn Chanussot Gipsa-Lab 17 / 21
  • 20. A new approach to classification Experiments Conclusions Real data set AISA data set: K-means (51.61%) VCA+SU (59.69%) KM+SU (75.72%) Jocelyn Chanussot Gipsa-Lab 18 / 21
  • 21. A new approach to classification Experiments Conclusions 1 A new approach to classification 2 Experiments 3 Conclusions Jocelyn Chanussot Gipsa-Lab 19 / 21
  • 22. A new approach to classification Experiments Conclusions Conclusions and Perspectives New method to improve spatial resolution of thematic maps: • Unsupervised clustering to define classes • Integration of spatial information to locally model abundances • Simulated Annealing proposed for spatial regularization Clustering less sensitive to extreme pixels, VCA better in highly mixed scenarios Next step: Incorporate spectral variability of the classes Jocelyn Chanussot Gipsa-Lab 20 / 21
  • 23. A new approach to classification Experiments Conclusions Unsupervised classification and spectral unmixing for sub-pixel labelling A.Villa , ,† , J.Chanussot , J.A. Benediktsson , C.Jutten GIPSA-lab, Signal & Image Dept., Grenoble Institute of Technology, France. Faculty of Electrical and Computer Engineering, University of Iceland, Iceland. † Aresys, Politecnico di Milano, Italy. IEEE IGARSS 2011 Vancouver, Canada - 2011 Jocelyn Chanussot Gipsa-Lab 21 / 21
  • 24. A new approach to classification Experiments Conclusions Challenges Hyperspectral images issues: 1 Curse of dimensionality 2 Exploitation of contextual information 3 Presence of mixed pixels • Common in hyperspectral images Pure pixel: 100% grass • Traditional classifiers inadequate Mixed pixel: • Usually not considered for 70% metal sheet classification! 30% grass Jocelyn Chanussot Gipsa-Lab 22 / 21
  • 25. A new approach to classification Experiments Conclusions Context Traditional approaches to image analysis are full pixel and mixed pixel techniques • Full pixel techniques are traditional classification algorithms • Mixed pixel techniques are spectral unmixing, soft classification, . . . Joint use is desirable, but little investigated [Wang and Jia, 2010]. Incorporation of spectral unmixing in the classification process: • Does it provide accuracy improvement? • Is it possible to improve the classification map spatial resolution? Jocelyn Chanussot Gipsa-Lab 23 / 21
  • 26. A new approach to classification Experiments Conclusions Linear Spectral Unmixing Abundances estimation through spectral unmixing: • Goal: find extreme pixels (endmembers) that can be used to "unmix" other pixels. • Each "mixed" pixel is a combination of endmember fractional abundances. Jocelyn Chanussot Gipsa-Lab 24 / 21
  • 27. A new approach to classification Experiments Conclusions Linear Spectral Unmixing Abundances estimation through spectral unmixing: • Goal: find extreme pixels (endmembers) that can be used to "unmix" other pixels. • Each "mixed" pixel is a combination of endmember fractional abundances. Jocelyn Chanussot Gipsa-Lab 24 / 21
  • 28. A new approach to classification Experiments Conclusions Context Traditional techniques neglect information Additional information provided by unmixing not fully exploited 0.6 0.9 Pure pixel: 100% grass 1 0.9 0.8 0.6 1 Mixed pixel: 1 1 0.8 70% metal sheet 30% grass 0.9 0.6 1 Original image Classification Unmixing Finer resolution? How to jointly use full and mixed pixel techniques? Jocelyn Chanussot Gipsa-Lab 25 / 21
  • 29. A new approach to classification Experiments Conclusions The proposed approach M = Mixed pixel Proposed method M We propose a technique in four steps: M M 1. Preliminary classification with probabilistic classifier (SVM) 2. Choose suitable endmember candidates and perform unmixing M M 3. Split every pixel into n sub-pixels, and assign them to a class M M 4. Perform spatial regularization in order M to correctly locate sub-pixels A. Villa, J. Chanussot, J.A. Benediktsson, C. Jutten, "Spectral Unmixing for the Classification of Hyperspectral Images at a Finer Spatial Resolution", IEEE J. Sel. Topics Sign. Proc., vol. 5, n. 3, 2011 Jocelyn Chanussot Gipsa-Lab 26 / 21
  • 30. A new approach to classification Experiments Conclusions The proposed approach M = Mixed pixel Proposed method We propose a technique in four steps: 1. Preliminary classification with probabilistic classifier (SVM) 2. Choose suitable endmember candidates and perform unmixing 3. Split every pixel into n sub-pixels, and assign them to a class 4. Perform spatial regularization in order to correctly locate sub-pixels A. Villa, J. Chanussot, J.A. Benediktsson, C. Jutten, "Spectral Unmixing for the Classification of Hyperspectral Images at a Finer Spatial Resolution", IEEE J. Sel. Topics Sign. Proc., vol. 5, n. 3, 2011 Jocelyn Chanussot Gipsa-Lab 26 / 21
  • 31. A new approach to classification Experiments Conclusions The proposed approach Proposed method 0,5 0,3 0,2 M We propose a technique in four steps: 0,7 M 0,6 1. Preliminary classification with M 0,3 0,4 probabilistic classifier (SVM) 2. Choose suitable endmember candidates and perform unmixing 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,1 0,2 3. Split every pixel into n sub-pixels, and assign them to a class 0,9 0,9 M 0,1 M 0,7 0,1 0,3 4. Perform spatial regularization in order M 0,9 0,1 to correctly locate sub-pixels A. Villa, J. Chanussot, J.A. Benediktsson, C. Jutten, "Spectral Unmixing for the Classification of Hyperspectral Images at a Finer Spatial Resolution", IEEE J. Sel. Topics Sign. Proc., vol. 5, n. 3, 2011 Jocelyn Chanussot Gipsa-Lab 26 / 21
  • 32. A new approach to classification Experiments Conclusions The proposed approach Proposed method 0,5 0,3 0,2 M We propose a technique in four steps: 0,7 M 0,6 1. Preliminary classification with M 0,3 0,4 probabilistic classifier (SVM) 2. Choose suitable endmember candidates and perform unmixing 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,1 0,2 3. Split every pixel into n sub-pixels, and assign them to a class 0,9 0,9 M 0,1 M 0,7 0,1 0,3 4. Perform spatial regularization in order M 0,9 0,1 to correctly locate sub-pixels A. Villa, J. Chanussot, J.A. Benediktsson, C. Jutten, "Spectral Unmixing for the Classification of Hyperspectral Images at a Finer Spatial Resolution", IEEE J. Sel. Topics Sign. Proc., vol. 5, n. 3, 2011 Jocelyn Chanussot Gipsa-Lab 26 / 21
  • 33. A new approach to classification Experiments Conclusions Simulated Annealing Minimize a given Cost Function introducing random perturbations: • decreases of the CF are always accepted • increases of the CF accepted with a probability inversely proportional to the degradation • probability of ’bad solutions’ decreases as the search continues Simulated Annealing optimization avoids local minima leading to global optimal solution Jocelyn Chanussot Gipsa-Lab 27 / 21
  • 34. A new approach to classification Experiments Conclusions Simulated Annealing Cost function to be minimized: total perimeter of the connected areas (e.g., belonging to the same class) M 0,5 0,3 0,2 M 0,5 0,3 0,2 0,7 M 0,3 M 0,6 0,4 0,7 M 0,3 M 0,6 0,4 0,9 0,8 0,9 0,6 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,2 0,1 M 0,4 M 0,1 0,2 0,1 0,9 0,9 0,9 0,9 M 0,1 M 0,7 0,1 0,3 M 0,1 M 0,7 0,1 0,3 M 0,9 0,1 M 0,9 0,1 Cost function not optimized Cost function optimized Jocelyn Chanussot Gipsa-Lab 28 / 21
  • 35. A new approach to classification Experiments Conclusions Simulated Annealing M 0,5 0,3 0,2 M 0,5 0,3 0,2 0,7 M 0,3 M 0,6 0,4 0,7 M 0,3 M 0,6 0,4 0,9 0,8 0,9 0,6 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,2 0,1 M 0,4 M 0,1 0,2 0,1 0,9 0,9 0,9 0,9 M 0,1 M 0,7 0,1 0,3 M 0,1 M 0,7 0,1 0,3 M 0,9 0,1 M 0,9 0,1 Initial condition Iteration 1 M 0,5 0,3 0,2 M 0,5 0,3 0,2 0,7 M 0,3 M 0,6 0,4 0,7 M 0,3 M 0,6 0,4 0,9 0,8 0,9 0,6 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,2 0,1 M 0,4 M 0,1 0,2 0,1 0,9 0,9 0,9 0,9 M 0,1 M 0,7 0,1 0,3 M 0,1 M 0,7 0,1 0,3 M 0,9 0,1 M 0,9 0,1 Iteration n Final result Jocelyn Chanussot Gipsa-Lab 28 / 21
  • 36. A new approach to classification Experiments Conclusions Simulated Annealing Cost function to be minimized: total perimeter of the connected areas (e.g., belonging to the same class) M 0,5 0,3 0,2 M 0,5 0,3 0,2 0,7 M 0,3 M 0,6 0,4 0,7 M 0,3 M 0,6 0,4 0,9 0,8 0,9 0,6 0,9 0,8 0,9 0,6 M 0,4 M 0,1 0,1 0,2 M 0,4 M 0,1 0,1 0,2 0,9 0,9 0,9 0,9 M 0,1 M 0,7 0,1 0,3 M 0,1 M 0,7 0,1 0,3 M 0,9 0,1 M 0,9 0,1 Cost function not optimized Minimum cost function Jocelyn Chanussot Gipsa-Lab 28 / 21
  • 37. A new approach to classification Experiments Conclusions Experiment on real data AVIRIS Indian Pine data set • (145×145 pixels, 220 bands), 16 classes of interest • Spatial resolution of the original data degraded of a factor 2 • 10% of the labelled samples used as training set AVIRIS Hekla data set • (180×180 pixels, 157 bands), 9 classes of interest • Spatial resolution of the original data degraded of a factor 2 • 15% of the labelled samples used as training set Comparison with SVM 1vs1, RBF kernel 20 20 40 40 60 60 80 80 100 120 100 140 120 160 140 180 20 40 60 80 100 120 140 20 40 60 80 100 120 140 160 180 Indian Pine GT Hekla GT Jocelyn Chanussot Gipsa-Lab 29 / 21
  • 38. A new approach to classification Experiments Conclusions Evaluation of the results Jocelyn Chanussot Gipsa-Lab 30 / 21
  • 39. A new approach to classification Experiments Conclusions Evaluation of the results Jocelyn Chanussot Gipsa-Lab 30 / 21
  • 40. A new approach to classification Experiments Conclusions Evaluation of the results Jocelyn Chanussot Gipsa-Lab 30 / 21
  • 41. A new approach to classification Experiments Conclusions AVIRIS Indian Pine 20 10 40 20 60 30 80 40 100 50 120 60 140 70 20 40 60 80 100 120 140 10 20 30 40 50 60 70 Ground truth SVM map (OA = 72.31%) 20 20 40 40 60 60 80 80 100 100 120 120 140 140 20 40 60 80 100 120 140 20 40 60 80 100 120 140 Proposed, before SA (OA = 89.82%) Proposed, final map (OA = 91.10%) Jocelyn Chanussot Gipsa-Lab 31 / 21
  • 42. A new approach to classification Experiments Conclusions AVIRIS Hekla 10 10 20 20 30 30 40 40 50 50 60 60 70 70 80 80 90 90 10 20 30 40 50 60 70 80 90 10 20 30 40 50 60 70 80 90 Low res. GT SVM map (OA = 69.19%) 20 20 40 40 60 60 80 80 100 100 120 120 140 140 160 160 180 180 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 Proposed, before SA (OA = 78.90%) Proposed, final map (OA = 81.71%) Jocelyn Chanussot Gipsa-Lab 32 / 21
  • 43. A new approach to classification Experiments Conclusions A robust method AVIRIS Indian Pine (Complete) AVIRIS Indian Pine (full data set) 90 90 85 85 Proposed method Overall Accuracy (%) Overall Accuracy (%) Traditional SVM Traditional SVM Proposed Method 80 80 75 75 70 70 0.6 0.65 0.7 0.75 0.8 5 10 15 20 Treshold Pure Pixels Number of ’candidates endmember’ AVIRIS Hekla AVIRIS Hekla 82 82 80 80 78 78 Overall Accuracy (%) Overall Accuracy (%) Proposed Method 76 76 Traditional SVM 74 74 72 72 70 70 68 68 0.6 0.65 0.7 0.75 0.8 5 10 15 20 Treshold Pure Pixels Number of ’candidates endmember’ Jocelyn Chanussot Gipsa-Lab 33 / 21
  • 44. A new approach to classification Experiments Conclusions Conclusions and Perspectives New method to improve spatial resolution of thematic maps: • Spectral Unmixing considered to handle mixed pixels and abundances determination • Simulated Annealing proposed for spatial regularization Better definition of spatial structures with respect to full pixel classifiers when the image contains mixed pixels Large quantitative improvement Jocelyn Chanussot Gipsa-Lab 34 / 21