Scale-aware triple semantic deep network for polarimetric SAR image classification

Deep learning has demonstrated outstanding performance in polarimetric synthetic aperture radar (PolSAR) image classification. However, complex PolSAR scenes typically contain terrain objects of varying scales and orientations, and existing methods that rely on a single network architecture struggle to effectively capture both multi-scale terrain objects and precise edges. To address this challenge, we propose a scale-aware triple semantic deep network for PolSAR image classification, which is a novel “divide and conquer” framework. This network divides the scene into homogeneous, heterogeneous, and boundary regions, and then customizes specific subnetwork to learn scale-aware features for each region. For the boundary regions, a simple convolutional neural network (CNN) is designed to capture local details effectively. For homogeneous regions, a superpixel-based graph convolutional network (SGCN) is utilized to extract contextual features. For heterogeneous objects, a larger-scale $c$-hop SGCN is developed to capture global semantic information. These subnetworks are integrated with an attention mechanism to select relevant features and minimize redundancy. Experiments on four real PolSAR datasets demonstrate that our method outperforms state-of-the-art approaches, particularly in classifying heterogeneous objects and achieving precise edges.