Climate change-adapted spatiotemporal prediction and monthly dynamic risk assessment of rainfall-induced landslides using 3ED-ConvLSTM

Climate change has increased the frequency and intensity of heavy rainfall, exacerbating rainfall-induced landslides (RILs). Research on dynamic risk assessment methods of RILs under climate change has become a major topic in natural hazard prevention and control. Most methods are limited to predicting static risk outcomes or evaluating uncertain future periods based on rainfall return periods. This study aims to improve RILs prediction and generate monthly risk assessments through an integrated spatiotemporal approach using a multi-modal dataset. We propose a 3ED- Convolutional Long Short-Term Memory model that synthesizes geological, topographical, and meteorological factors for dynamic risk assessment in Wanzhou, China. The model, trained using meteorological data from 1980 to 2015, predicts RILs from 2016 to 2023, and establishes a risk assessment by comparing its performance with Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models as benchmarks. The 3ED-ConvLSTM model shows superior spatial accuracy, predicting 125 landslides vs LSTM's 101 and CNN's 83 in high-risk areas, and annual temporal errors reduced by 0.5216 and 0.5577, respectively. The key findings reveal that rainfall, runoff, and temperature are the dominant triggers of RILs. This study generates 96 monthly risk scenarios, providing actionable insights for natural hazard risk prevention.