Polarimetric synthetic aperture radar image processing for land cover classification
Lee, Ken Yoong
School of Computer Engineering
This thesis presents the processing design and development towards an improved land cover classification of multi-look POLSAR data. In this context, three related processing aspects, namely edge detection, speckle suppression and region-based seg¬mentation, were investigated. To detect edges in speckle-corrupted POLSAR data, two new edge detectors, which are based separately on the Roy’s largest eigenvalue and trace ratio, were proposed. Their capabilities were compared with three other edge detectors, which use the likelihood ratio, Dowson-Landau metric and Euclidean distance, respectively. Applied to nine-look NASA/JPL POLSAR C- and L-band data, both the proposed edge detectors exhibited their satisfactory performance in detecting edges while coping with speckle noise disturbance. For the speckle suppression, a novel spatially adaptive speckle filter was developed, which aims at preserving image features during filtering process. In the experiments, the proposed filter showed its good performance in speckle removal, image feature retention and radiometric preservation. Comparisons with the boxcar, Lee-refined and annealing filters were carried out. In the area of region-based segmentation, an existing hybrid segmen¬tation algorithm was extended for multi-look POLSAR data. The capabilities of the extended algorithm were examined and bench¬marked against the HSWO algorithm. From the results, it was noticed that the segmentation outputs from both the algorithms relied strongly on the user-defined region number, which is employed as the termination rule. Finally, the classification of NASA/JPL POLSAR data over a selected area of Kuala Muda in Peninsular Malaysia was carried out based separately on scattering mechanisms and statistical distribution. For the scattering property-based unsupervised classification, three different schemes were examined, namely van Zyl’s scattering classification, Freeman-Durden scattering model and Cloude-Pottier target decomposition. The derived scattering mechanisms of different land cover classes were studied. In the supervised complex Wishart classification, the obtained overall accuracies were 63% and 67% for both the C- and L-band, respectively. An improved accuracy of 75% was attained by using the dual-frequency input. Furthermore, comparison between per-pixel and per-segment classification approaches was conducted, where the per-segment approach was expectedly found to improve the classification accuracy.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision