Hierarchical cross attention achieves pixel precise landslide segmentation in submeter optical imagery

Accurate landslide segmentation using remote sensing imagery is a critical component of geohazards response systems, particularly in time-sensitive tasks such as post-earthquake landslide damage assessment and emergency resource allocation. However, current methodologies struggle with two persistent challenges in sub-meter true-color imagery: fine-grained inter-class confusion between landslides and spectrally analogous terrain features, and within-landslide heterogeneity where localized damage signatures coexist with macro-scale deformation patterns within individual landslide bodies. To overcome these, we propose the Cross-Attention Landslide Detector (CALandDet), which improves the model's ability to distinguish between landslide and background features by sharply capturing global landslide feature information and integrating global landslide feature information with local information via a cross-attention feature enhancement mechanism. Ablation experiments show that CALandDet outperforms baselines, as evidenced by a 4.89% enhanced F1 score and an 8.73% greater Intersection over Union (IoU). In comparative experiments, it outperforms the other models by 8.05-10.78% in IoU and 1.05-8.9% in F1 score, achieving an IoU of 82.65% and an F1 score of 81.64%. Furthermore, the Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirm that the decision regions generated by the CALandDet model exhibit a higher spatial consistency with the actual landslide areas, effectively capturing indicative features including surface textures, sliding debris, accumulation bodies, and vegetation destruction. The proposed method may serve as a reference for future advancements in landslide segmentation and other remote sensing segmentation tasks.