Abstract
Landslides are among the geological disasters that frequently occur worldwide and significantly restrict the sustainable development of society. Therefore, it is of great practical significance to perform landslide susceptibility assessment. In addressing issues such as limited training samples, inadequate utilization of spatially effective features, and high computational costs associated with existing methods, we propose a landslide susceptibility assessment method (DS-DRN), which uses a lightweight dense residual network with emphasis on deep spatial features. To minimize computational costs, we design a depthwise separable residual module that optimizes traditional convolution on residual branches into depthwise separable convolution. Additionally, to prevent vanishing gradient and improve the reuse rate of landslide feature information, dense connections are employed to construct a deep feature extraction module. Finally, the output of the model is fed into the Softmax classifier for landslide susceptibility prediction. Taking Ya'an City in Sichuan Province as the study area, we compare the proposed DS-DRN method with three widely used deep learning methods: CNN, CPCNN-RF, and U-net. Evaluating model accuracy and performance, the DS-DRN method exhibits the highest prediction accuracy while also saving computational costs. Therefore, the proposed model can better fit the complex nonlinear relationship in landslide susceptibility, effectively mine deep spatial features, and address the high computational costs associated with complex networks.