An Optimization-based Ensemble Framework for Spatially Consistent Multisource SAR Classification

In this work, we propose a Binary Integer Linear Programming (BILP) framework for multisource Synthetic Aperture Radar (SAR) image classification using a stochastic-distance-based region classifier. The method integrates statistical outputs from individual sensor image classifications with their respective reliability measures, jointly exploiting multisensor evidence and spatial coherence. Experiments using Spaceborne Imaging Radar-C/X images (C and L bands) demonstrated the applicability of the proposed approach for multisource combination and spatially consistent classification. Moreover, sensitivity analysis revealed that moderate spatial regularization provided an appropriate balance between accuracy and spatial detail, avoiding oversmoothing effects. The results showed that integrating multisensor statistical information with spatial context through an ensemble-based optimization procedure is a viable strategy for robust SAR classification. Furthermore, classification performance varied across classes due to the distinct responses of C- and L-band sensors to different scattering mechanisms and land-cover patterns.