Soil moisture dynamics play a critical role in slope stability, especially for rainfall-induced groupoccurring landslides. With the growing availability of remote sensing–derived soil moisture products, there is increasing potential to improve landslide hazard assessment. However, few studies have explicitly incorporated both the spatial and temporal dynamics of soil moisture into hazard modeling. This study introduces a novel framework that integrates a Residual-Sparse Autoencoder (ResSAE) with Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) algorithms to enhance landslide susceptibility prediction using remotely sensed soil moisture data. Spatio-temporal soil moisture information for the study area in Nanping, China, is obtained from three open-access datasets: SMCI1.0, ERA5-Land, and SMAPL4. Results show that antecedent soil moisture features extracted by ResSAE substantially improve prediction accuracy. The influence of rainfall, antecedent period length, and dataset source is further evaluated. Further analysis reveals that antecedent soil moisture over the prior seven days captures most of the hydrological memory relevant for slope failure, while additional rainfall data contribute only marginal gains. Optimal performance is achieved with ERA5-Land for RF, SMAP-L4 for SVM, and SMCI1.0 for ANN.Overall, the study highlights the importance of incorporating spatio-temporal soil moisture into hazard assessment. The proposed approach enables efficient and cost-effective predictions, supporting near-real-time applications and offering potential to strengthen regional to global rainfall-induced landslide prevention and mitigation strategies.
 

Spatio-temporal soil moisture dynamics; Rainfall-triggered landslide clusters; Machine learning algorithms; Landslide hazard assessment