Quantitative assessment of the efficiency of supervised classification using coherent and incoherent polarimetric SAR observables

Abstract

Supervised classification of Earth surface, using polarimetric SAR data, was used as a test case for comparing the classification efficiency achieved on one hand by using coherent polarimetric observables and on the other incoherent ones. The coherent observables were derived from the diagonalization of the pixel [S] matrix and the incoherent observables were obtained from the coherency matrix using Cloude and Pottier's entropy/alpha. Investigations revealed that both categories of observables, i.e., coherent and incoherent, yield comparable efficiencies. A more interesting aspect of the results relates the averaging over pixels. It was found that classification efficiency improved generally when the number of pixels in the averaging window was increased. The improvement in classification performance was observed not only in the incoherent case but, surprisingly, also in the coherent case. Such observations seem to imply that neighboring pixels measured with SAR are not entirely uncorrelated.