Analysis of effective window size in texture-based classification of 2007-2010 ALOS PALSAR 25m mosaic images

This study aims to develop a land cover texture-based classification scheme applicable for ALOS PALSAR imageries of the upper Marikina watershed acquired 2007 – 2010. From the raw dual polarization bands of HH+HV that has a ground resolution of 25m, additional bands HH/HV and NL was computed for surface texture normalization. The classification scheme was based on texture analysis using grey level cooccurrence matrix with parameters of mean, variance and angular second moment to extract imageries feature statistics. Varied window sizes from 3x3 to 29x29 in odd series was produced to generate texture-window size bands (TWS bands). Using Support Vector Machine for land cover classification, each classified TSW bands’ accuracy was computed and yielded an initial result of stability on the NL band at ~78%-81% on window sizes 15-29. Additional 2,744 TWS bands with permuted window sizes of the 3 texture variable of NL band was produced and classified for accuracy assessment. In general, high dependence on variance texture variable was observed for classified TWS bands with high accuracy. These TWS bands has large window size that caused generalization of classification. For land cover change detection, given an illogical transition of land cover due to misclassification from the SVM classification, a drastic land cover was observed especially on the forest cover of the watershed from 2007-2010. Index Terms — Polarimetric Synthetic aperture radar; image texture analysis