Abstract:
Convolutional neural network (CNN) models are widely used in landslide susceptibility assessment due to their powerful feature extraction capabilities, and traditional CNN is no longer able to meet the requirements. Therefore, this paper proposes a multi-scale convolutional neural networks (MSCNN) model that can take into account deep and shallow features. By increasing the depth of the model and expanding the receptive field of samples, the MSCNN can tap deeper and more stable features to improve the reliability of landslide susceptibility assessment in complex scenarios. In this study, Shenzhen City is selected as the research area, and 12 landslide conditioning factors of landslides in Shenzhen City were selected based on systematic and representative principles. A multi-scale convolutional neural network landslide susceptibility assessment model is constructed and compared with methods such as multilayer perceptron (MLP), support vector machine (SVM), and random forest (RF). The results show that the
AUC value (0.99) of the MSCNN model constructed in this paper is higher than that of MLP (0.97), SVM (0.91), and RF (0.85), which proves that the proposed MSCNN model has excellent prediction ability. The area of extremely high susceptibility in Shenzhen City is approximately 105.3 km², accounting for 4.98% of the total area of the study area, mainly distributed in Longgang District with steep slopes, sparse vegetation cover, and frequent human engineering activities. Slope, surface roughness, and surface relief are identified as the main conditioning factors affecting landslides in Shenzhen City. The landslide susceptibility mapping implemented in this paper reflects the current distribution of landslide disasters in Shenzhen City, providing data support and key technical support for future landslide disaster prevention and control in Shenzhen City.