Exploring the impact of introducing the TRIGRS physical model into machine learning model on the rainfall-induced shallow landslide-susceptibility assessment

Abstract

Yunyang County is located in the Three Gorges Reservoir area, with high annual rainfall, and fragile geological conditions, leading to frequent geological disasters. Previous studies on rainfall-induced shallow landslide susceptibility often relied on machine learning models that depend on available data, overlooking landslide mechanisms and key topographical features, making it hard to link them to underlying processes. To resolve this challenge, this study incorporates a physical model (TRIGRS: Transient Rainfall Infiltration and Grid-Based Regional Slope-Stability Model) based on the infinite slope theory into the machine learning framework. First, we introduce a BP model and three BP coupling models combined with the AO (Aquila Optimizer), AVOA (African Vultures Optimization Algorithm), and IHAOAVOA (Improved Hybrid Aquila Optimizer and African Vultures Optimization Algorithm) to assess landslide susceptibility in Yunyang County, Chongqing. Model performance is assessed using the ROC (Receiver Operating Characteristic) curve, and the results indicate that the IHAOAVOABP exhibited the best performance. Subsequently, the safety factor (Fs), a measure of slope stability, for the study area is calculated using the TRIGRS. Finally, the TRIGRS is integrated with the BP, AOBP, AVOABP, and IHAOAVOABP models in proportions, resulting in four integrated models. The results indicate that the integrated model combining the machine learning model with TRIGRS has significantly better predictive performance than the single machine learning model. Among these, IHAOAVOABP-TR demonstrated the best performance, with its %LRclass index of 90.98%, an improvement of 10.05% over IHAOAVOABP. This research combines machine learning models with physical models to more effectively assess rainfall-induced shallow landslide susceptibility.