1. A Number-of-Classes-Adaptive Unsupervised Classification Framework for SAR Images Bin Liu, Hao Hu, Kaizhi Wang, Xingzhao Liu, and Wenxian Yu Remote Sensing Technology Institute Shanghai Jiao Tong University
2. Content Problem Description & Introduction Framework Feature Extraction & Image Representation Estimation of the Number of Classes Final Classification Incorporation of Spatial Relations between Patches Implementation Experiments & Results Conclusions & Future Work July, 2011 IEEE IGARSS 2011 2
3. Content Problem Description & Introduction Framework Feature Extraction & Image Representation Estimation of the Number of Classes Final Classification Incorporation of Spatial Relations between Patches Implementation Experiments & Results Conclusions & Future Work July, 2011 IEEE IGARSS 2011 3
4. Problem Description Radar as the third eye – Synthetic Aperture Radar (SAR) A day-or-night, all-weather means of remote sensing High resolution images and useful information about the earth Several spaceborne platforms continuously deliver enormous amounts of SAR data TerraSAR-X, Germany RADARSAT-2, Canada COSMO-SkyMed, Italy ALOS-PALSAR, Japan … Develop automatic/semi-automatic systems for SAR image interpretation and target recognition July, 2011 IEEE IGARSS 2011 4
5. Problem Description & Introduction SAR image classification Fundamental to exploiting the enormous amounts of SAR data Akey requirement in both military and civil sectors Ahighly desired goal for developing intelligent databases Anecessary process for target detection and recognition Develop an automatic system to divide the SAR images into basic land covers: Water, built-up areas, vegetated areas, … An important problem: The number of classes in the image is generally UNKNOWN A Number-of-Classes-Adaptive (NoCA) Unsupervised Classification Framework for SAR Images July, 2011 IEEE IGARSS 2011 5
6. Content Problem Description & Introduction Framework Feature Extraction & Image Representation Estimation of the Number of Classes Final Classification Incorporation of Spatial Relations between Patches Implementation Experiments & Results Conclusions & Future Work July, 2011 IEEE IGARSS 2011 6
7. Feature Extraction & Image Representation SAR image partitioned into NP patches, m * m Feature Extraction Grey Histogram: calculated with BGHbins Texture Histogram: Inspired by Reigber et al., using the filter coefficient of the Lee speckle filter to describe the texture inhomogeneity. calculated with BTHbins x*- local mean var(x) - local variance σn2is equal to 1 over the number of looks July, 2011 IEEE IGARSS 2011 7 A. Reigber, M. Jäger, W. He, L. Ferro-Famil, and O. Hellwich, “Detection and classification of urban structures based on high-resolution SAR imagery,” in Proc. Urban Remote Sens. Joint Event, Paris, France, 2007, pp. 1–6.
8. Feature Extraction & Image Representation Image Representation The N * M SAR image including NPpatches An NP * NP dissimilarity image fGH(∙) and fTH(∙) denote grey and texture histograms, respectively Dis(∙) is the city block distance αis the fusion factor July, 2011 IEEE IGARSS 2011 8
10. Estimation of the Number of Classes Reordering Cattell: reorder the objects suitably image better able to highlight the potential class structure in the data Different methods of implementing visual representation of pairwise dissimilarity information – the Reordered Dissimilarity Image (RDI) Using the Visual Assessment of cluster Tendency (VAT) algorithm to transform the dissimilarity image into the RDI July, 2011 IEEE IGARSS 2011 10 J. C. Bezdek and R. Hathaway, “VAT: a tool for visual assessment of (cluster) tendency,” in Proc. Int’l Joint Conf. Neural Networks (IJCNN ’02), Honolulu, HI, May 2002, pp. 2225–2230
11. Estimation of the Number of Classes Reordering July, 2011 IEEE IGARSS 2011 11
12. Estimation of the Number of Classes Extraction The RDI can highlight the potential classes as a set of dark blocks along the diagonal of the image The Dark Block Extraction (DBE) method to automatically extract dark blocks. Using several common image and signal processing techniques Perform image segmentation on the RDI to obtain a binary image, and then apply the directional morphological filters to the binary image Apply a distance transform to the filtered binary image, and then project the pixel values along the main diagonal axis of the image to form a projection signal Smooth the projection signal, and use the first-order derivative of the projection signal to detect the major peaks and valleys of the projection signal July, 2011 IEEE IGARSS 2011 12 L. Wang, C. Leckie, K. Ramamohanarao, and J. Bezdek, “Automatically determining the number of clusters in unlabeled data sets,” IEEE Trans. Knowledge and Data Eng., vol. 21, no. 3, pp. 335–350, Mar. 2009.
13. Estimation of the Number of Classes Extraction July, 2011 IEEE IGARSS 2011 13
14. Estimation of the Number of Classes Inversion In the projection signal, a peak between two neighboring valleys realistically represents a class in the data Suppose that there are Nx patches between the (x–1)th and the xth valleys Due to noise, the inversion step cannot simply determine that all the Nx patches belong to the xth class. In our method, the inversion step labels the β∙Nx patches nearest to the xth peak as elements of the xth class, where βis from 0 to 1 The other (1–β)∙Nx patches between the (x–1)th and the xth valleys are categorized as “undecided”, and their labels are determined in the final classification July, 2011 IEEE IGARSS 2011 14
15. Estimation of the Number of Classes Inversion July, 2011 IEEE IGARSS 2011 15
16. Estimation of the Number of Classes Estimate the number of classes and get initial classes July, 2011 IEEE IGARSS 2011 16
17. Final Classification After the estimation operation, several important initial class parameters The number of classes X β∙NP patches with class labels, belong to initial classes C1, C2, …, CX Final classification The commonly used technique Support Vector Machine (SVM) classifier is used LIBSVM Training data set – β∙NP patches with labels Features – Grey and texture histograms July, 2011 IEEE IGARSS 2011 17
18. Incorporation of Spatial Relations between Patches Simple yet effective The SAR image is partitioned into patches with overlaps Some parts of a patch may belong to many adjacent patches, and after the final classification, they might be assigned to different classes Majority vote July, 2011 IEEE IGARSS 2011 18 M. Liénou, H. Maître, and M. Datcu, “Semantic annotation of satellite images using latent Dirichlet allocation,” IEEE Geosci. Remote Sens. Lett., vol. 7, no. 1, pp. 78–82, Jan. 2010.
19. Implementation Procedures Preprocessing The single look SAR data are multilook-processed Patch Generation The SAR image is partitioned into image patches of m* m pixels with NOL pixels overlapping Preclustering Estimation The SAR image is represented by a dissimilarity image, which is reordered into a RDI image. Then, the DBE process extracts the number of classes, each class center, and initial class labels of some patches Final Classification Class labels of the “undecided” patches from the preclustering estimation operation are estimated and refined Final Decision A majority vote is used to ascribe common parts of overlapping patches to the likeliest class. Then, the class label of every pixel in the image is finally decided July, 2011 IEEE IGARSS 2011 19
20. Content Problem Description & Introduction Framework Feature Extraction & Image Representation Estimation of the Number of Classes Final Classification Incorporation of Spatial Relations between Patches Implementation Experiments & Results Conclusions & Future Work July, 2011 IEEE IGARSS 2011 20
22. Experiments and Results Performance of the proposed method Determine the number of classes = 3, √ The total accuracy is 92.97% July, 2011 IEEE IGARSS 2011 22 The SAR image The ground truth map The final classification map
23. Experiments and Results Performance of the proposed method Built-up areas misclassified as vegetated areas – 9.17% Vegetated areas misclassified as built-up areas – 6.57% It seems that built-up and vegetated areas are likely to be confused with each other. The classification result may be further refined by introducing more features and prior knowledge of the scene July, 2011 IEEE IGARSS 2011 23
25. Experiments and Results July, 2011 IEEE IGARSS 2011 25 Performance of the proposed method Determine the number of classes = 3, √ The total accuracy is 95.14% The SAR image The ground truth map The final classification map
27. Experiments and Results July, 2011 IEEE IGARSS 2011 27 Performance of the proposed method Determine the number of classes = 4, √ The total accuracy is 89.90% The SAR image The ground truth map The final classification map
28. Content Problem Description & Introduction Framework Feature Extraction & Image Representation Estimation of the Number of Classes Final Classification Incorporation of Spatial Relations between Patches Implementation Experiments & Results Conclusions & Future Work July, 2011 IEEE IGARSS 2011 28
29. Conclusions and Future Work An NoCA Unsupervised Classification framework for SAR images Extract the numbers of classes estimate each class center more accurately provide robust classification under various numbers of classes Based on image patches & incorporate relations between patches effective and efficient Pairwise-dissimilarity-based estimation operation, flexible multiple features can be selected and fused into the estimation operation Future Work Integration of statistical information Extract the number of classes and estimate each class center in randomly selected sub-scenes, and then apply the parameters to the whole image In the future, the framework needs to be applied on an enormous SAR image database to develop a fully operational procedure July, 2011 IEEE IGARSS 2011 29
30. Thank you for your attention ! July, 2011 IEEE IGARSS 2011 30