E-MSLDNet: A Dual-Enhanced Network with Edge-Aware and Multiscale Feature Fusion for Precision Landslide Detection in Complex Terrain

Precise boundary measurement and quantitative delineation of landslides from high-resolution remote sensing imagery faces significant challenges, including scale heterogeneity, boundary measurement uncertainty, and inadequate precision in complex terrain environments. To address these challenges, we propose E-MSLDNet, an Edge-aware and Multiscale Feature Fusion Dual-enhanced Landslide Detection Network that systematically tackles landslide segmentation difficulties through three novel contributions. Firstly, we develop a multiscale feature aggregation attention module (MSFAA) that innovatively combines parallel multi-receptive field processing with dynamic channel attention mechanisms, enabling robust representation of landslides across diverse scales. Secondly, we design an edge-aware attention module (EAAM) that employs originally conceived direction-sensitive attention mechanisms to achieve precise boundary capture in ambiguous regions. Additionally, we propose an edge enhancement fusion module (EEFM) that dynamically integrates edge priors with semantic features through our novel spatially-guided calibration strategy, significantly improving boundary segmentation accuracy. Furthermore, we pioneer the application of HD95 and FOM metrics in landslide detection to establish quantitative benchmarks for boundary assessment. Experimental results on four Chinese Academy of Sciences landslide datasets demonstrate the superiority of E-MSLDNet. It achieves IoU gains of 1.81-4.72% and Boundary F1 (BF1) score improvements of up to 10.27 percentage points over baseline methods. The HD95 metric is reduced by 11.08% and 15.17% compared to the second-best models, respectively. Furthermore, the FOM metric remains optimal on the challenging Jiuzhaigou dataset. Ablation studies validate the complementary contributions of our proposed multiscale learning and edge enhancement mechanisms, effectively advancing landslide detection performance in complex terrains.