SAR and Optical Satellite Image Matching Incorporating Category-Supervised Semantic Features

The matching of SAR and optical satellite images is a crucial method for addressing the challenge of geolocalization in scenarios where real-time access to optical images is constrained (e.g., darkness, rain, fog, glare, etc.). To enhance the generalization capability of SAR-optical image matching, particularly under varying viewpoints (rotation), this paper presents an augmented Superpoint network architecture. This architecture incorporates semantic segmentation feature representation to facilitate a priori semantic description and to integrate it with the original network's feature descriptors. This integration is achieved by introducing self-attention and cross-attention layers, which enhance the feature vectors. The experiments demonstrate that, compared to the original architecture, the model reinforces the relationship for commonality matching and significantly increases the number of correctly matched points. Compared with other deep learning methods, including CMMNet and LoFTR, the matching scheme presented in this paper is less time-consuming and more robust with anti-rotation capability.