Noise-Robust Multimodal Remote Sensing Image Matching Network with Edge-Aware Attention Fusion

Abstract

Multimodal remote sensing image matching is challenged by nonlinear radiometric differences, geometric deformation, sensor noise, and weak cross-modal feature repeatability. We present an edge-aware coarse-to-fine Transformer network. A Feature Enhancement Transformer (FET) performs linear self- and cross-attention, while a Multiscale Edge Enhancement Module (MEEM) injects stable boundary cues and suppresses modality-specific interference. Bidirectional confidence and mutual-nearest-neighbor filtering select reliable coarse matches, followed by local correlation refinement. On optical-depth, infrared-optical, optical-map, and synthetic aperture radar (SAR)-optical test groups, the method achieves an average root-mean-square error (RMSE) of 1.38 pixels, 776.3 correct matches, and a runtime of 0.28 s. It nearly doubles the second-best correct-match count while retaining Local Feature Transformer (LoFTR)-level accuracy and subsecond efficiency.