An approach for predicting landslide susceptibility and evaluating predisposing factors

Abstract

Effectively leveraging landslide spatial location information is crucial for improving the accuracy of deep learning in predicting landslide susceptibility and exploring the impacts of predisposing factors. Current single deep learning models for landslide susceptibility assessment require enhancements in both prediction accuracy and robustness. Inclusion of non-interrelated positional information among samples leads to reduced prediction accuracy and challenges in quantifying landslide risk covariates. This study proposes a landslide susceptibility assessment method that integrates ensemble learning with geographically weighted concepts. Using a stacking method, a 1D convolutional neural network (1D-CNN), a recurrent neural network (RNN), and a long short-term memory (LSTM) network were combined to form the CRNN-LSTM ensemble model. Additionally, we constructed a deep learning geographically weighted regression (GW-DNN) model based on the deep learning principles and geographically weighted regression to quantify the impacts of landslide-predisposing factors.The experimental results show that the CRNN-LSTM model achieved AUC values of 0.977 and 0.961 on the training and validation sets, significantly outperforming the individual classifiers (AUC of 0.944 and 0.940 for the 1D-CNN model, 0.950 and 0.948 for the RNN model, and 0.956 and 0.952 for the LSTM model). Additionally, the GW-DNN model achieved R2 coefficients of 0.876 and 0.860 during the training and validation phases. These findings indicate that our proposed method not only highly accurately predicts landslide susceptibility but also provides a precise quantitative assessment of the impact of landslide-predisposing factors at specific spatial points (landslide units) in high-risk areas. These findings offer valuable technical support for landslide disaster prevention and mitigation.

Keywords

• Landslide susceptibility prediction
• Ensemble learning
• Deep learning geographically weighted regression
• Risk assessment
• Quantitative analysis