Dual-Branch Fusion of Convolutional Neural Network and Graph Convolutional Network for PolSAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) images contain useful information, which can lead to extensive land cover interpretation and a variety of output products. In contrast optical imagery, there are several challenges in extracting beneficial features from PolSAR data. Deep learning (DL) methods provide solutions address feature extraction challenges. The convolutional neural networks (CNNs) graph (GCNs) drive image characteristics by deploying kernel abilities considering neighborhood (local) information graphs long-range similarities. A novel dual-branch fusion CNN mini-GCN is proposed this study for classification. To fully utilize the capacity, different spatial-based polarimetric-based incorporated into branches model. performance method verified comparing classification results multiple state-of-the-art approaches on airborne (AIRSAR) dataset Flevoland San Francisco. approach showed 1.3% 2.7% improvements overall accuracy compared conventional with these AIRSAR datasets. Meanwhile, it enhanced its one-branch version 0.73% 1.82%. Analyses over data further indicated effectiveness model using varied training sampling ratios, leading promising 99.9% 10% ratio.