Filtering effects on polarimetric SAR image classification

Abstract

Feature extraction from SAR images is usually impeded by the presence of speckle noise. This becomes more serious in the case of polarimetric SAR system. A polarimetric filter recently proposed by Lee et al. [1997] emphasizes not introducing additional cross-talk and statistical correlation between channels, preserving polarimetric information and not degrading the image quality. This paper exams its effects on the image classification by a supervised fuzzy dynamic learning neural network trained by a Kalman filter technique. Based on the available ground truth, the classification performance were evaluated using the original and filtered SAR images. Two independent test sites are selected for this purpose. The first case is a P-band JPL polarimetric SAR data over Les Landes for tree age classification. A total of 12 classes between 5 to 44 years of age were to be classified, along with a bare soil type. The second test site is over Flevoland of the Netherlands. This agricultural site consists of 11 landcover types. Again, the polarimetric SAR data were acquired with JPL P, L, C bands airsar system. For the first case, it was found that the overall classification accuracy was able to improve from 69% to about 86% with kappa coefficient up from 0.46 to 0.76. Substantial improvement was also confirmed for the second case. In particular, when classification was performed using only single frequency. This shows that the polarimetric information are well preserved. By visual inspection from classified map, the land cover boundaries were also delineated more clearly. As for fuzzy neural network performance, among the tested cases, the fuzzy index equal to 2 gets the best results.