Robust Pixel-by-Pixel Multimodal Remote Sensing Image Registration Using Geometry Preserving Dense Registration Network

The integration of multimodal remote sensing (RS) data is pivotal for advancing Earth observation capabilities, where high precision cross-modal image registration serves as the cornerstone for realizing its full potential. While sparse feature matching methods have achieved high registration accuracy by extracting corresponding points and solving the parameters of the assumed geometric transformation model, they still struggle to effectively address local nonrigid deformations between multimodal images. Additionally, image dense registration methods are prone to introducing significant mismatching noise, which frequently results in structural distortions of typical ground features (e.g., roads, buildings, etc.) within global deformation regions. To address that, we propose a novel geometry preserving dense registration network (GPDRNet) for multimodal RS image registration tasks. Our approach incorporates three key innovations: 1) a feature enhancement module that generates more discriminative cross-modal feature representations; 2) a geometry consistency loss that preserves critical structural and morphological characteristics of typical ground features within global deformation regions; and 3) a local deformation-aware post-processing module that first estimates global affine field, and then identifies local deformation regions through residual flow analysis. Experimental results validate the superior effectiveness and robustness of the proposed GPDRNet over existing state-of-the-art multimodal RS image registration methods. The code of the proposed method will be available at https://github.com/yeyuanxin110.