Detecting Landslides Beyond Hydrology: Spectral Indices and Geochemical Controls

Abstract Extreme rainfall represents one of the main triggers of shallow landslides in mountainous and hilly environments; however, the rapid identification of rainfall-induced slope failures remains challenging due to the lack of systematic post-event inventories and the limited availability of field data. This gap constrains the development of reliable diagnostic and predictive frameworks, particularly in regions increasingly affected by extreme precipitation events. This study aims to evaluate the capability of satellite-derived spectral indices to detect rainfall-triggered landslides and to interpret their spectral responses within a geological and environmental geochemical context. The analysis focuses on the extreme precipitation event of 1–3 March 2011 in the Marche Region (Central Italy), an area characterized by widespread marly–clayey and Plio–Pleistocene fine-grained formations. Multitemporal Landsat 5 imagery at 30 m spatial resolution acquired between 2007 and 2011 was used to derive seven spectral indices related to vegetation cover, soil exposure, moisture conditions, and surface mineralogy (NDVI, NDWI, NBR, NDSI, BSI, SCI, and CMR). Owing to multicollinearity among predictors, each index was analyzed independently using two complementary statistical approaches: Binary Logistic Regression (BLR), representing a parametric linear framework, and the QUEST decision tree algorithm, designed to capture non-linear relationships and threshold effects. Vegetation-related indices (NDVI and NBR) also show strong predictive capability, with AUC values ranging from 0.94 to 0.96 across both models. Moisture-sensitive indices exhibit a predominantly non-linear response: NDWI performs markedly better under QUEST than BLR (AUC 0.88 vs. 0.58), while NDSI shows the opposite pattern (AUC 0.81 with BLR vs. 0.54 with QUEST), highlighting the complementary strengths of the two modeling frameworks. Importantly, both BLR and QUEST converge on consistent and robust threshold values, providing objective criteria for post-event landslide detection. The identified spectral thresholds are physically interpreted in terms of short-term geochemical weakening processes affecting clay-rich and marly materials. Intense rainfall promotes hydration of expandable clay minerals, partial dissolution of carbonate cement, and rapid soil–water–mineral interactions, leading to reduced soil cohesion and enhanced slope instability. By integrating remote sensing, statistical modeling, and environmental geochemistry, this study offers a practical and transferable approach for rapid post-event landslide mapping. The framework is intended to support hazard assessment in sedimentary terrains prone to rainfall-induced failures, with careful attention to its potential applicability in similar settings. Graphical Abstract Graphical abstract descriptions: The graphical abstract schematically illustrates the conceptual framework and methodological workflow of the study, integrating geological, geochemical, remote sensing, and statistical components into a unified representation. The upper section depicts the geomorphological and lithological setting of the Marche Region in Central Italy, highlighting marly–clayey formations and Plio–Pleistocene deposits that are highly susceptible to rainfall-triggered slope instability. Intense precipitation is shown to initiate landslide processes through infiltration into fine-grained materials, activating physico-chemical mechanisms such as clay hydration, carbonate dissolution, and cation exchange, which lead to modifications in pore-water chemistry and progressive reduction of soil shear strength. The central portion of the diagram connects these subsurface processes to surface responses detectable by satellite remote sensing, introducing key spectral indices including the Soil Composition Index (SCI), Bare Soil Index (BSI), and Normalized Difference Vegetation Index (NDVI), which capture variations in mineralogical composition, bare soil exposure, and vegetation cover associated with landslide occurrence. The lower section presents the analytical phase, where spectral information is processed through binary logistic regression and decision tree models to quantify landslide probability and identify threshold values distinguishing stable and unstable terrain. Finally, the integration of geochemical interpretation, satellite data, and predictive modeling is linked to landslide hazard assessment, emphasizing the interdisciplinary approach adopted in the study. The graphical abstract thus condenses the entire research structure into a single coherent visual scheme, enabling an immediate perception of the scientific logic, methodological sequence, and interpretative outcomes of the investigation.