Item

Abstract

Aspect-dependent landslide initiation, in which sunny slopes are more susceptible to shallow failures than shady slopes, has received increasing attention, yet the role of aspect-dependent soil thickness remains underexplored. We develop a hybrid framework that integrates aspect-dependent factors (ADF) into machine learning (ML) algorithms to predict soil thickness and couples the best-performing map with the physically-based RegionGrow3D (R3D) model to assess rainfall-induced shallow landslide susceptibility. The framework is applied to the Niangniangba loess hilly region of northwestern China, where insolation-driven slope-aspect contrasts strongly influence landslide activity. Two modeling strategies are employed: (1) empirical models, including the Geomorphological Index Soil Thickness (GIST) model and its improved version (GIST(ADF)), which incorporates ADF such as sunny-shady slope gradient index (SSSI) and sunny-shady slope position index (SSP); and (2) ML models, including Extreme Gradient Boosting (XGB), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN), trained on both empirical and environmental factors. Results show that, among the tested models, the XGB model incorporating ADF (XGB(ADF)) achieved the highest performance (R2 = 0.83), effectively capturing soil-thickness variations between sunny and shady slopes. When incorporated into R3D, the ADF-based thickness map increased classification skill (AUC from 0.83 to 0.92) and accuracy (from 71.7 % to 89.3 %) compared with models lacking ADF. These findings demonstrate that incorporating aspect-dependent soil thickness substantially improves the physical realism and predictive accuracy of landslide-susceptibility modeling, providing a stronger process-based foundation for hazard assessment and risk reduction in complex terrains.