Abstract: Carrying out regional landslide susceptibility assessment is the key to landslide meteorological early warning and risk assessment. Aiming at the fact that many current susceptibility studies do not consider the relationship between the occurrence of landslides and adjacent environments, a regional landslide susceptibility modeling framework based on convolutional neural network (CNN) is proposed. Taking Wanzhou district of the Three Gorges Reservoir area as an example, 12 factors such as slope and aspect are selected to construct an evaluation index system, and the influence of factors on landslide development is analyzed by information method. The local two-dimensional matrix is used to construct the dataset, CNN is used for susceptibility modeling. At the same time, the impact of the size of the local two-dimensional matrix to the accuracy when constructing samples is explored. The results show that landslides are more likely to occur the closer to the reservoir zone, and the water system and human engineering activities have a greater impact on the development of landslides. The accuracy of the CNN model is 0.925, which is significantly higher than that of the machine learning model, and the accuracy can be improved by increasing the local two-dimensional matrix size when constructing the sample. The CNN model has advantages in multidimensional spatial data processing, considering the influence of landslide location and its adjacent environment, and it is an accurate and reliable regional landslide susceptibility evaluation method.

Key words: reservoir landslide, susceptibility mapping, convolutional neural network, Three Gorges Reservoir area