Intelligent identification and susceptibility analysis of rainfall-induced shallow landslides in mountainous areas based on high-resolution remote sensing and deep learning

In recent years, extreme rainfall events have become more frequent, leading to a significant increase in shallow landslide disasters, which trigger debris flows and increase the risk of subsequent chain disasters. Therefore, conducting research on the identification and spatial distribution patterns of shallow landslides is crucial for mitigating and preventing geological disasters induced by rainfall. In this study, we utilized high-resolution remote sensing imagery and introduced deep learning methods to develop a rapid identification approach for shallow landslides. We constructed a shallow landslide dataset using high-resolution optical satellite data from the Beijing-3(BJ-3) satellite. A Mask Region-based Convolutional Neural Network (Mask R-CNN) deep learning model was employed to train an automatic landslide identification model, generating a landslide inventory. For the identified shallow landslides, we conducted a correlation analysis between rainfall and landslide occurrence. Spatial statistical analysis was used to examine the spatial distribution patterns, while and an Information value method coupling support vector machine (I-SVM) model was applied to assess the susceptibility of shallow landslides to extreme rainfall. The study area was selected in Puwa Township, Fangshan District, Beijing. Using the automatic landslide identification model, we identified 1,237 landslides with a precision of 88.04%, a recall rate of 69.40%, and a processing time of 303 s. After manual screening, 13 landslide susceptibility evaluation factors were selected for further spatial statistical analysis and susceptibility assessment. The results indicate that landslides primarily occur on concave steep slopes at elevations ranging from 540 to 1,080 m. Under given rainfall conditions, the probability of shallow landslide occurrence is higher in Dong Village, Anzigang Village, and Dongniwa villages. The Area Under Curve(AUC) of the landslide susceptibility model is 97%, indicating a high evaluation precision. The research conducted in this paper provides valuable references and recommendations for the rapid identification and disaster prevention of shallow landslides.