Multi-scale convolutional neural networks (CNNs) for landslide inventory mapping from remote sensing imagery and landslide susceptibility mapping (LSM)

Abstract

Accurate landslide susceptibility mapping (LSM) relies on a detailed landslide inventory and relevant influencing factors. In this study, Sentinel 2 remote sensing imagery is employed to establish a comprehensive landslide inventory leveraging attention U-Net backbone networks in Zhangjiajie City of Hunan Province, China. Subsequently, the refined landslide inventory, with more precise boundaries, is integrated into LSM process. A multi-scale sampling three-dimensional convolutional neural network (3D-CNN) is introduced into LSM, facilitating the extraction of multi-scale neighbourhood characteristics and deep information of relevant topographical, hydrological, meteorological, geological, and human activity factors. Experimental results demonstrate that this method achieves the highest accuracy and area under the receiver operating characteristic (ROC) curve. Moreover, its recall and F1-score significantly higher surpass those of other small-, medium-, and large-scale models, with the F1-score being more than 10% higher. This superior performance is attributed to its proficiency in discerning the nonlinear spatial correlation between landslide occurrences and influencing factors. The comprehensive consideration of scale characteristics through the multi-scale sampling strategy outperforms the single-scale CNN models across all evaluation metrics. This study furnishes a suite of high-precision methodologies for landslide hazard assessment in Zhangjiajie City, thereby offering invaluable support to decision-makers engaged in large-scale land use planning and geologic disaster prevention.

Keywords

• Landslide susceptibility mapping
• deep learning
• convolutional neural networks
• multi-scale sampling
• landslide inventory identification