Spatiotemporal changes of landslide susceptibility in mountainous areas in response to rainfall and its future prediction — A case study of Sichuan Province, China

Abstract

In recent decades, global warming has led to significant changes in the spatial and temporal distribution of rainfall. In the mountainous areas of southwestern China, rainfall-induced landslides are one of the most common natural disasters, which are highly destructive and unpredictable. In order to adapt to and mitigate the effects of rainfall variability on landslides in mountainous areas, there is an urgent need to strengthen the future prediction of rainfall scenarios as well as research on the response of landslide susceptibility to rainfall variability. Hence in this study, annual cumulative rainfall (ACR) is utilized as a dynamic factor to generate annual-scale landslide susceptibility maps (LSM) for Sichuan Province, China. The study's main objectives are to reveal the dynamic response relationship between landslide susceptibility and rainfall and to consider the future prediction of rainfall for the future prediction of LSM. To achieve this objective, firstly, the historical landslide geospatial database was prepared for five time categories: 2000-2005, 2006-2010, 2011-2015 and 2016-2020. Second, the Extreme Learning Machine (ELM) model was utilized to generate LSMs for the years 2000-2020. Third, the rainfall for the year 2030 was predicted using the Long Short-Term Memory Network (LSTM) model and the month-by-month rainfall dataset for the years 1901-2020. Fourth, the rainfall prediction map was used as a driving parameter for the future prediction of LSM. Finally, the sensitivity of landslide susceptibility to changes in rainfall was discriminated using the elasticity model. The results showed that the very high landslide susceptibility in the study area increased by 5.95%, 7.09%, and 1.29% in 2005, 2015, and 2020, respectively, as compared to 2000, while the index decreased by 3.97% in 2010. The sensitivity coefficient of landslide susceptibility to changes in rainfall from 2000 to 2020 was 1.35, indicating that a 1% change in rainfall would result in an average change in landslide susceptibility of 1.35%. In the future landslide susceptibility projections, the percentage of landslide susceptibility for each category in 2030 is 9.75%, 25.15%, 30.42%, 24.51%, and 10.16%, respectively, with a 5.11% increase in the very high landslide susceptibility category. The reliability of the results was verified using the ROC curve and area under the curve.