Cross-regional extrapolation of landslide susceptibility mapping via transfer learning

• A transfer learning model for cross-regional landslide susceptibility mapping. • The comparison of cross-regional extrapolation performance of different models. • The SHAP algorithm is used to provide a global interpretation of the transfer learning model. Landslide susceptibility mapping (LSM) is an essential tool for the prevention and management of landslide-related disasters. Conventional machine learning-based LSM method faces significant limitations in cross-regional extrapolation. To address this challenge, this study develops a transfer learning (TL) model based on the Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) framework, specifically designed for cross-regional LSM. A total of 11 modelling scenarios is established to compare the cross-regional extrapolation performance of Random Forest (RF), CNN-BiLSTM, and TL models, with Wanzhou District and Wushan County in Chongqing used as case studies. The results indicate that, compared to the strategy of directly expanding training dataset used by RF and CNN-BiLSTM models, the pre-training and fine-tuning strategy employed by the TL model is more suitable for county-scale LSM and its cross-regional extrapolation. Additionally, the cross-regional extrapolation performance of the TL model improves as the volume of source domain data increases. Finally, the SHAP algorithm is used to provide a global interpretation of the TL #3 model, which demonstrates the best performance in cross-regional model extrapolation.