Multimodal Remote Sensing Image Registration via Radiometric-Agnostic Geometric Discrepancy Estimation

Abstract

Multimodal remote sensing image registration is inherently challenged by sensor-induced nonlinear radiometric differences, which can confound true geometric misalignment and weaken the reliability of conventional similarity metrics. This article proposes RAGDE-Net, a coarse-to-fine registration framework centered on a radiometric-agnostic Geometric Discrepancy Evaluator (GDE). The GDE is trained offline to predict dense geometric discrepancy responses that are sensitive to spatial misalignment while being robust to cross-modal appearance changes. To obtain supervision without manual dense annotations, structure-preserving pseudocross-modal pairs are generated by a Modality Translation Generator constrained with multiscale modality-independent neighborhood descriptors, and controllable synthetic perturbations are used to derive pixel-wise discrepancy labels. During online registration, the pretrained GDE is shared by the affine and deformable stages, providing a unified geometry-oriented supervision signal and response-guided input modulation that reduce the objective inconsistency between global affine estimation and local deformable refinement. We further construct TRSR, a tri-modal benchmark of co-registered visible, SAR, and infrared images. Experiments on TRSR and the public OSdataset demonstrate that RAGDE-Net consistently outperforms representative state-of-the-art methods, reducing reprojection error by 31.2% on TRSR visible-SAR, 20.3% on OSdataset visible-SAR, and 48.2% on TRSR visible-infrared registration.