Fig 1.
Architecture for transformer network based on dual-stage DEM guidance and fusion optimization.
Fig 2.
Dual-Stage DEM-Guided Fusion module.
Fig 3.
Terrain-aware Pixel-wise Adaptive Context Enhancement module.
Fig 4.
(Points of interest are shown over an open basemap. Basemap: Natural Earth II, public domain. Made with Natural Earth. Figure © Authors, CC BY 4.0.).
Table 1.
Parameters of the backbone structures of D2FLS-Net -L, D2FLS-Net -B and D2FLS-Net -S.
Table 2.
Experimental results on the Bijie dataset.
Table 3.
Experimental results on the Landslide4Sense2022 dataset.
Table 4.
Experimental results for different stage combinations.
Fig 5.
Visual comparison on the Bijie dataset.
Blue = landslide, Yellow = background. (Contains modified Copernicus Sentinel-2 data, processed by the authors. No TripleSat pixels are reproduced; masks were trained/tested using the Bijie dataset. This replacement image is similar but not identical to the original and is for illustrative purposes only. Figure © Authors, CC BY 4.0).
Table 5.
Quantitative comparison on the Bijie dataset.
Fig 6.
Visual comparison on the Landslide4Sense2022 dataset. blue = landslide, yellow = backgroun (CC BY 4.0, https://doi.org/10.5281/zenodo.10463239).
Table 6.
Quantitative comparison on the Landslide4Sense2022 dataset.
Table 7.
The time complexity.
Fig 7.
Receiver Operating Characteristic curve with five-fold cross-validation.
(a) Bijie dataset; (b) Landslide4Sense 2022. Shaded bands denote 95% confidence intervals across folds; the dashed diagonal indicates chance. Legends report mean AUC ± s.d. over the five folds.