Explainable Machine Learning-Based Multi-Hazard Risk Assessment and Upazila-Level Intervention Prioritization in the Chattogram Region, Bangladesh

The increasing occurrence of natural hazards, either individually or in combination, has intensified disaster risk worldwide, with the Chattogram Region of Bangladesh representing a prominent multi-hazard hotspot where earthquakes, landslides, and floods frequently coexist. An interpretable GIS–ML-based multi-hazard risk assessment (MHRA) framework was developed to generate earthquake, landslide, and flood susceptibility zones (EQSZ, LSSZ, and FSZ) using ensemble machine learning algorithms, with model predictions explained through SHapley Additive exPlanations (SHAP). The resulting susceptibility maps were integrated into a Multi-Hazard Index, which was then combined with a Vulnerability Index to produce the final Multi-Hazard Risk Zone (MHRZ). GBM achieved the highest predictive performance for earthquake (AUC = 0.799) and landslide (AUC = 0.786), whereas RF performed best for flood susceptibility (AUC = 0.917). SHAP identified shear-wave velocity (Vs30), Soil Moisture Index, and elevation as the most influential predictors of earthquake, landslide, and flood susceptibility, respectively. Although the eastern hill districts exhibited the highest multi-hazard susceptibility, the MHRZ identified the densely populated upazilas of Chattogram and Cox's Bazar as the highest-risk areas because socio-economic vulnerability outweighed comparatively moderate hazard susceptibility. Approximately 55.10% of the study area fell within the very low to low-risk classes, whereas only 21.49% fell within the high-to-very-high-risk classes, highlighting the critical role of socio-economic vulnerability in shaping regional disaster risk. The proposed MHRA framework offers a methodologically transferable basis for evidence-based disaster risk reduction (DRR) by enabling upazila-level policy screening, preliminary DRR intervention prioritization, and resilient spatial planning in multi-hazard, data-scarce regions.