Spatio-temporal forecasting of landslide hazard in Chongqing National Transmission Protection Regions, China

Previous studies on landslide susceptibility (LS) and hazard (LH) assessments have overlooked a crucial measure for evaluating validation effectiveness, which involves the use of post-temporal landslide inventories instead of solely relying on historical data. Therefore, this study constructed a regional LH prediction and validation framework at different timescales using the landslide data from 2002 to 2021. Firstly, based on updated landslide data and conditioning factors, the datasets for susceptibility modelling were built for various periods. Then, the performances of traditional ma chine learning, ensemble and deep learning models were compared. Finally, the multi-period rainfall data and susceptibility results were used in innovative LH forecasting. Our study results show that the ensemble learning model (random forest) has stronger generalisation ability in different time periods, with all Area Under Curve (AUC) values exceeding 0.9. The hazard modelling framework based on rainfall erosion intensity and susceptibility can effectively forecast LH, and the shortened timescale for rainfall erosion intensity can improve the accuracy of LH forecasting. At the very high hazard level, the percentage of landslides forecast increased from 25.23% to 42.86%. This study comprehensively explores spatio-temporal dynamic changes in LH, accurately identifying areas posing threats to the National Transmission Line Protection Regions (NTLPR).