CICRL-FLM: counterfactual inference causal representation learning network for fine-grained landslide mapping from satellite remote sensing images

Landslide mapping from satellite remote sensing images is of paramount importance for both risk assessment and disaster response. When performing landslide mapping in diverse and complex remote sensing scenarios, current deep-learning-based methods have a limited global semantic information modeling ability and may present spurious correlations during contextual modeling, thereby causing a high incidence of false positives and a decline in model robustness. In view of these challenges, we propose the counterfactual inference causal representation learning network for fine-grained landslide mapping from satellite remote sensing images (CICRL-FLM). It, for the very first time, investigates the problems of robustness degradation and spurious correlations in landslide segmentation models from the perspective of correlation and causation. In detail, the channel randomization transfer Swin based on the U-shaped network structure (CRT-SwinUNS) is designed, which consists of the attention blocks with spatial-channel cross and multi-scale convolution (SCMC-AB) and the channel randomization transfer Swin Transformer (CRT-SwinT). Besides, saliency map calculation (SMC) is introduced to guide the spatial and channel semantic cross-modeling at all levels and to promote the local spatial saliency and global background continuity of landslide areas for the fine-grained modeling of landslide areas and background boundaries at deeper layers, thus reducing errors in landslide boundaries. In addition, the counterfactual causal inference based on attention masks (AM-CCR) is constructed, which is the first counterfactual causal inference model in the field of remote sensing landslide mapping, so as to mitigate the spurious correlations present in contextual modeling. This strategy enhances model robustness while maintaining great interpretability. A satellite remote sensing landslide mapping (SRSLM) dataset comprising 25 study areas is provided to validate the effectiveness and robustness of CICRL-FLM. The experimental results demonstrate that our method achieves improvements of 4.56%, 3.21%, 5.49%, and 3.87% in four metrics, precision, recall, Intersection over Union, and F1-score, respectively, attaining the state-of-the-art advantages and exhibiting a strong generalizability. This study provides a robust and interpretable method for fine-grained landslide mapping from satellite remote sensing images.