Integrating Explainable Machine Learning and Remote Sensing for Landslide Susceptibility Mapping and Carbon Stock Assessment under Climate Change

Landslides are among the most destructive geohazards, and their frequency and spatial footprint have expanded markedly in recent decades due to intensifying rainfall extremes, seismic disturbances, and accelerating human pressure on fragile terrain. Beyond their immediate impacts on settlements and infrastructure, these hazards strongly disrupt vegetation productivity, soil structure, and terrestrial carbon reservoirs particularly soil organic carbon (SOC) and above-ground biomass (AGB) which play a central role in the global carbon cycle. Despite increasing evidence of carbon mobilization during major disaster events, large-area, rainfall-triggered landslide clusters and flood-induced SOC redistribution remain poorly quantified in greenhouse-gas budgets and landscape-scale sustainability planning. These challenges highlight the need for scalable and interpretable hazard frameworks that can bridge data scarcity, cross-scale susceptibility transfer, and the quantification of landslide-driven ecosystem carbon disruption under extreme hydro-climatic conditions. To address these gaps, this research advances an integrated, multi-scale framework for landslide susceptibility mapping (LSM) and flood-affected ecological carbon diagnostics in the Emilia–Romagna region of Italy as a pilot case. Ecosystem carbon diagnostics reveal that the 2023 landslide cluster caused an estimated 35% reduction in AGB and decrease in SOC accumulation in 72% of the affected areas, indicating that extreme landslide clusters can temporarily reverse stable carbon-sink behavior, justifying inclusion in regional or national carbon accounting. Complementarily, flood-induced inundation, erosion, and oxygen-limited microbial shifts also alter SOC turnover and biomass productivity, degrading net carbon-retention capacity even in non-failure terrain units. This integrated perspective reframes hydro-slope disasters as coupled landscape disturbances that jointly govern spatial carbon depletion, redistribution, and short-term CO₂ re-emission potential To address training volume and transparency issue present study adopted approach that combines Random Forest (RF) classifier trained on both national and local landslide inventories, and enhanced using transfer learning (TL) to overcome severe data scarcity in rapid rainfall-triggered landslide training. The framework is further strengthened with attention-driven feature adaptation and multiple explainable AI (XAI) auditing layers like SHAP, and Integrated Gradients for feature-level attribution, and reasoning enabling both global transparency and pixel-scale interpretability. Model validation is performed using the catastrophic May 2023 extreme rainfall event, which triggered more than 65,000 landslides over 72 km² in Emilia–Romagna, providing a rare real-world benchmark for testing transferability and explanation stability. TL-enhanced models achieved satisfactory validation reliability, confirming that cross-domain learning can preserve spatial priors even when training data is limited, especially when source domains are large and heterogeneous. Attention analysis confirmed dynamic domain-shift adaptation, revealing that the model internally reweights local terrain-environmental drivers to support context-aware hazard reasoning. While transfer learning strengthens susceptibility generalization under limited training data, layered XAI and attention-weight audits restore physical reasoning, enabling spatial hazard predictions that are both accurate and process-traceable for carbon-aware planning .By jointly evaluating susceptibility transferability, internal explanation stability, and carbon-stock disruption from both landslides and flood, this study establishes a new direction for carbon-aware, interpretable, and transferable hazard mapping frameworks, supporting climate-resilient planning, risk communication, and ecosystem-focused disaster management.