Geomorphology-informed multimodal learning with implicit–explicit constraints for non-salient old landslide mapping

Mapping non-salient old landslides remains challenging, particularly under severe optical–topographic resolution mismatch and regular-grid inference. Under these conditions, conventional multimodal methods often fail to effectively fuse topographic information with weak visual cues, leading to spatially fragmented and geomorphologically inconsistent predictions even when digital elevation models (DEMs) are combined with high-resolution optical imagery. To address this problem, we propose TUAFormer, a geomorphology-informed multimodal framework for spatially coherent mapping of non-salient old landslides. The framework introduces a dual-level constraint strategy that integrates implicit and explicit geomorphological information to address both multimodal inconsistency and patch-induced structural discontinuity. At the feature level, a Terrain-Unit-Aware Frequency Decoupling Fusion Module (TFDFM) incorporates macro-topographic structure into high-resolution visual features through frequency decomposition and surface-normal guidance. At the inference level, a slope-unit-based strategy is adopted to promote spatial coherence and topological plausibility. To further enhance degraded landslide representation, we design a SAM-based Adaptive Prior Injection Module (SAPIM) and a Boundary-aware Memory-augmented Pixel-wise Contrastive Learning strategy (BM-PCL). Experiments on three independent test regions in the Sichuan–Tibet corridor (F1-scores of 71.84%, 66.98%, and 71.64%) show that TUAFormer consistently outperforms representative baselines in both segmentation accuracy and geomorphological consistency, producing more spatially coherent landslide maps in complex mountainous environments.