Spatio-Temporal Landslide Susceptibility Modeling Using High-Resolution Rainfall and Soil Moisture Data with Machine Learning

Shallow landslides threaten communities and infrastructure in mountainous regions, where climate change is intensifying and triggering rainfall events. There is a fundamental limitation for the current landslide detection approaches: methods based solely on rainfall duration and intensity achieve high detection rates but deliver excessive false alarm ratios, while methods using only terrain susceptibility fail to capture the temporal dynamics of triggering events. This persistent problem results in difficulties for the operational deployment of early warning systems. In this direction, the current research aims to develop an integrated machine learning framework that achieves notable false alarm reduction by explicitly coupling static terrain vulnerability with dynamic rainfall and soil moisture conditions. The methodology is based on a combination of the Random Forest algorithms with high- resolution satellite-derived precipitation and soil moisture data across 7,200 slope units in the northern Apennines of Italy. Using a comprehensive inventory of 412 landslide events documented over five years (2016–2021), the methodology follows three steps: (1) characterization of static environmental controls, including topography, geology, and land cover; (2) integration of dynamic meteorological variables to enable temporal prediction; and (3) development of a novel multi-criteria threshold framework designed to reduce false alarm rates. The core of the technique is a multi-criteria framework that fundamentally differs from existing approaches by requiring independent exceedance of two separate thresholds before announcing a warning. Unlike conventional methods that apply a single decision criterion—where terrain susceptibility or rainfall duration and intensity exceeds a threshold—the proposed framework implements an AND-logic operator: warnings are issued only when terrain susceptibility exceeds its calibrated threshold and soil moisture and rainfall simultaneously exceed their independent thresholds. In the current study, the proposed framework was evaluated against two single-criterion approaches that represent current practice: terrain-based thresholds alone and hydrological thresholds alone. The Combined Layers approach, which required simultaneous satisfaction of both criteria, achieved a 73.4% probability of detection with only a 10.8% false alarm ratio, representing a 74% reduction in false alarms compared to hydrological threshold methods and a 49% reduction compared to terrain-based methods. This performance improvement originated from filtering two categories of false alerts: vulnerable slopes experiencing insufficient rainfall and stable slopes withstanding elevated rainfall that their favourable characteristics accommodate. Only in the case of vulnerable ter- rain encountering critical rainfall and soil moisture, which represents the convergence of independent predisposing and triggering factors, does the system activate warnings. Overall, the proposed methodology provides a practical pathway toward reliable operational landslide early warning, achieving false alarm rates manageable for civil protection agencies (10.8%) while maintaining detection rates sufficient for public safety (73.4%). Since climate change intensifies extreme precipitation events across vulnerable mountainous regions, the current multi-criteria integration strategy offers a rigorous yet operationally feasible framework for warning systems that balance detection sensitivity with the reliability essential for sustained stakeholder confidence and effective risk reduction.