Abstract: In response to the problem that the traditional landslide susceptibility analysis methods only consider the impact factor information of the landslide point itself and ignore the surrounding spatial information, a method that combines recursive feature elimination and a 2D convolutional neural network is proposed. Firstly, the recursive feature elimination method is used to rank and filter the landslide impact factors. Subsequently, the 2D feature factor set is cropped and fed into a 2D CNN with L2 regularization, Dropout, and other optimization methods, and the spatial information around the landslide is taken into account to predict the landslide susceptibility while ensuring the prediction accuracy and generalization ability of the model. In this paper, the Jiuzhaigou area is taken as the experimental area, and 14 relevant factors such as elevation and lithology are selected as landslide-influencing factors to predict the probability of landslide occurrence and draw a landslide susceptibility map. Finally, a logistic model and three SVM models with different kernel functions (linear kernel function, radial basis kernel function, and sigmoid kernel function) are used for comparison and validation. The experimental results show that the proposed method has the highest accuracy and AUC, which proves the validity and reliability of the proposed method.

Key words: landslide susceptibility, recursive feature elimination, 2D convolutional neural network, L2 regularization, support vector machine