Knowledge-Guided Modulation for Terrain-Aware Landslide Detection Using Deformable Transformers

Abstract

Landslide detection using medium-resolution optical remote sensing imagery remains challenging in complex mountainous environments because of spectral ambiguity, vegetation cover, shadows, and background interference. Although recent deep learning methods have improved detection performance, most existing approaches remain primarily appearance-driven and do not explicitly exploit terrain-related priors that are closely associated with slope instability. To address this limitation, we propose a terrain-aware deformable transformer framework for landslide detection using multimodal remote sensing data, in which RGB imagery, DEM, and slope are jointly incorporated through a unified five-channel representation, and a knowledge-guided modulation module is introduced to enhance feature learning using terrain priors derived from DEM and slope. Here, “knowledge-guided” refers specifically to explicit topographic priors rather than complete geological or hydrological knowledge. Experimental results on the Bijie landslide dataset show that the proposed method outperforms several competitive baselines and achieves 72.9% AP@[0.5:0.95] and 77.2% AP75, while improving localization robustness in visually confusing mountainous scenes. These results indicate that terrain-aware feature modulation can improve geomorphological plausibility and detection accuracy for landslide inventory mapping, although further cross-region validation is still needed to assess broader generalization.