Land cover classification using multi-temporal SAR data and optical data fusion with adaptive training sample selection

This paper proposes the classification framework based on the Bayesian theory with the single polarization multi-temporal synthetic aperture radar (SAR) and an optical data, and incorporates the proposed training sample selection (SS) methods. Within this framework, the combination with gray level co-occurrence matrix (GLCM)-based mean textural measure is investigated. The two procedures of the classification and proposed SS are united, where SS generates accurate and dispersed training samples. Extracted features from multi-temporal SAR data—namely, the average backscattering coefficient, backscatter temporal variability, and long-term coherence and reflectance values from optical data, are integrated with GLCM mean textural data in the framework. Classification results were generated by taking Osaka City, Japan, as the study area. The selected major classes were built-up areas, fields, woodlands, and water bodies. The most suitable data used for classification is the multi-temporal SAR and an optical data fusion with the mean textural, because of the supplement of different data used and the smoothing effect on the images of the texture. Moreover, the higher quality training samples obtained by using the combined SVM and NN-based SS method for training the Bayesian classifier generated the highest classification accuracies in all of tested cases.