MACHINE LEARNING-DRIVEN INSIGHTS INTO SOIL–VEGETATION–ATMOSPHERE INTERACTIONS ON CLIMATE-IMPACTED SLOPES

The selection of ecological indicators for assessing landslide susceptibility is highly context-dependent, varying significantly based on regional characteristics, data availability, and the specific objectives of the study. Despite this variability, certain indicators such as the Normalized Difference Vegetation Index (NDVI) and Land Use Land Cover (LULC) are consistently employed due to their relevance in reflecting vegetation health, land surface changes, and human-induced disturbances. NDVI helps capture variations in vegetation cover, which can influence slope stability, while LULC provides critical insights into land modification and land management practices that may exacerbate or mitigate landslide risks. Factors like Net Primary Productivity (NPP), Remote Sensing Ecological Index (RSEI) are less common . These indicators are often used in conjunction with other environmental, topographic, and climatic factors to develop a comprehensive understanding of landslide susceptibility across different landscapes. In May 2023, Emilia-Romagna region in Italy experienced an exceptionally intense rainfall event which triggered more than 65 thousand landslides (Geoportale: https://regione.emilia-romagna.it ). This offers a unique opportunity with its precisely located failure points (landslide) which gives fine temporal resolution for “pre‐ vs. post‐failure” ecological indices, statistical robustness from a large, heterogeneous sample, and rich spatial variability (from dense forests to agricultural terraces), immediate validation opportunities, and all these factors make it especially beneficial for evaluating how NPP, RSEI, and NDVI relate to both the initiation of landslides and the subsequent ecological recovery, further help to check which ecological feature is most suited for landslide susceptibility analysis under climate change scenarios . This study employs a data-driven framework on the Google Earth Engine (GEE) platform, integrating 20 geospatial and climatic parameters to model landslide susceptibility using Random Forest (RF) and LightGBM (LGBM) on GoogleColab used to introduce Explainable AI (XAI). Explainable AI (XAI) techniques, such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Diverse Counterfactual Explanations (DiCE), enable quantifying and visualizing how each ecological feature contributes to a model's predictions. This ensures that the selected variables are scientifically meaningful and operationally robust, while also revealing complex associations between features and landslide predictability by understanding the influence of each feature on the model's outputs. This XAI-driven insight not only helps in selecting the best-performing ecological factors but also supports detailed local interpretation of decisions and predictability. RF and LGBM are trained and tested on a subset of this specific event. In this study, RSEI, NPP and NDVI are integrated as variables along with other variables for generating a landslide susceptibility map of Region Emilia Romagna using RF. RF and LGBM offer different possibilities to understand the results of classification at the global level and local (pixel) level along with XAI, if only global explanation (feature importance’s averaged over the entire dataset) is required, RF’s built‐in feature importance can be adequate ; however, LGBM’s smoother continuous probabilities make local‐linear surrogates (LIME) more faithful and counterfactual searches (DiCE) more plausible. Faster predictions allow LIME and DiCE to sample more points, yielding richer, more robust explanations. LGBM offers finer control over calibration and local decision‐surface complexity, directly benefiting explainability tools. For these reasons, when primary goal is high‐quality, stable local explanations and counterfactuals, LightGBM is the preferred choice.