A COMPARISON OF 2ND-ORDER CLASSIFIERS FOR SAR SEA-ICE DISCRIMINATION

In this paper we present results of an analysis of the relative utilit y of statistical, structural, and frequency based second-order texture methods for discrimination of sea ice types in synthetic aperture rad ar (SAR) data. Algorithms were trained using a calibration data set an d robustness of the methods were assessed by directly computing ice cl asses within a validation data set Classification using a first-order approach (average grey level) produced Kappa classification accuracies of 51.0 and 33. 0 percent for the calibration and validation data. Th e first-order approach is provided primarily as a reference from which to compare the second-order approaches because the test conditions we re selected to be specifically difficult (i.e., different incidence an gle ranges between calibration and validation images) for any approach using image tone or the relative scattering cross section. Results fr om the second-order approaches indicate that the two spatial domain st atistical approaches, Grey Level Co-Occurrence Matrix (GLCM) and the N eighboring Grey Level Dependence Matrix (NGLDM) provided high classifi cation accuracies under the difficult test conditions examined here. T he GLCM results achieved a Kappa Coefficient of 84. 0 and 81.0 percent for the calibration and validation sets. The NGLDM achieved a Kappa C oefficient of 83. 0 and 76. 0 percent for the calibration and validati on data sets. These results are statistically equivalent between the c alibration and validation data sets and between the GLCM and NGLDM sch emes. The Spatial/Spatial Frequency (S/Sf) approach appears to be sens itive to the training conditions generated from the calibration data s et and therefore do not provide statistically reproducible results bet ween the calibration (87 percent) and validation (18 percent) test con ditions. Results from the Primitive Texture Value (PTV) method suggest poor operational capabilities both due to low calibration (65 percent ) and validation (58 percent) accuracies.