Land Cover Classification Using Combinations of L- and C-band Sar and Optical Satellite Images

Whereas C-band SAR data mainly provides information on the upper layer of land covers, properties of the lower surfaces can be captured using L-band SAR because of its higher capability of penetration. Moreover, optical imagery illuminates different characteristics of ground surfaces compared to radar images since they are obtained at different frequencies of the electro-magnetic spectrum. Hence, integration of Land Cband SAR together with optical images can potentially enhance the quality of land cover mapping. This study investigates classification of various land cover features in South of Vietnam using a combination of multi-temporal ALOS/ PALSAR (L-band) radar, ENVISAT/ASAR (C-band) radar and SPOT multi-spectral optical satellite data. The incorporation of Grey Level Co-occurrence Matrix (GLCM) textured measures for land cover classification is also studied. Correlation matrix and discriminant functions such as Transformed Divergence and Jefferies-Matusita distances are used to analyze correlation and separability within a dataset. The classification processes are carried out for selected single, and various combined datasets in succession using a traditional parametric Maximum Likelihood classifier and a non-parametric classifier such as the Support Vector Machine (SVM). Results indicate that integration of multi-frequency SAR data gave significant improvement in classification performance. The combination of SAR and optical data provided the highest classification accuracy (94.99%). The application of textured information together with original data can improve the classification accuracy in certain cases. Finally, the SVM classification technique has demonstrated its advantages over the Maximum Likelihood classification, particularly for handling complex datasets.