Hillshade-Based Gated Matching Network for Robust Optical–DEM Cross-Modal Registration

Abstract

Cross-modal registration of optical images and digital elevation models (DEMs) is fundamental for remote sensing data fusion and terrain analysis. However, optical products and DEM retrieved from remote sensing cloud computing platforms according to the same geographic range may still exhibit residual spatial misalignment, and the large modality gap between spectral textures and terrain elevation makes robust correspondence establishment challenging. To address these problems, we propose a hillshade-based gated matching network (HGM-Net) and construct an optical–DEM cross-modal registration dataset. HGM-Net generates hillshade maps from DEM to reduce the modality gap, employs independent adapters and a shared encoder to extract terrain structural features from DEM and textural features from optical images respectively, introduces a multiorder gated detector to achieve robust detection of high-confidence keypoints, and designs a shadow-weighted loss to enhance the stability of keypoint detection in local shadow regions. Experiments on both directly cropped initially geo-registered multisource data and affine-perturbed data demonstrate that the proposed method achieves robust optical–DEM matching and fine registration.