D2FLS-Net: Dual-Stage DEM-guided Fusion Transformer for landslide segmentation.

Area of Science:

• Geosciences
• Remote Sensing
• Computer Vision

Background:

• Landslide segmentation from remote sensing imagery is critical for disaster management.
• Challenges include scale heterogeneity and subtle visual contrasts, hindering accurate segmentation.
• Existing methods often underutilize digital elevation model (DEM) data, limiting adaptability to terrain variability.

Purpose of the Study:

• To develop an advanced framework for accurate landslide segmentation by effectively fusing remote sensing imagery and DEM data.
• To improve fine-grained adaptability to terrain variability in landslide detection.
• To enhance the modeling of long-range contextual dependencies and terrain cues.

Main Methods:

• Proposed D2FLS-Net (Dual-Stage DEM-guided Fusion Transformer for landslide segmentation), a Swin-Transformer-based framework.
• Introduced a Dual-Stage DEM-Guided Fusion (DSDF) module for multi-stage DEM cue injection.
• Developed a Terrain-aware Pixel-wise Adaptive Context Enhancement (T-PACE) module for optimized feature representation.

Main Results:

• D2FLS-Net achieved state-of-the-art performance on the Bijie and Landslide4Sense 2022 datasets.
• On Bijie, mIoU reached 88.77%, surpassing SegFormer by 7.96%.
• On Landslide4Sense2022, mIoU reached 72.86%, surpassing SegFormer by 7.06%.

Conclusions:

• D2FLS-Net effectively integrates DEM and remote sensing imagery for superior landslide segmentation.
• The DSDF module reduces missed detections, while T-PACE refines pixel-level context selection.
• The proposed framework demonstrates significant improvements in landslide recognition and risk assessment.