Integrating multi-criteria models into geospatial regression to explore spatially varying flood risk

Along with hazard, socio-economic vulnerability is an important consideration for flood risk assessment to ensure sustainable development. Risk is commonly quantified utilizing Multi-criteria Decision Making (MCDM) or Machine Learning (ML) methods. Incorporating the spatially biased nature of risk as a non-stationary criterion can significantly enhance realism. This research develops an integrated approach that incorporates MCDM-based risk scores into Geographically Weighted Regression (GWR) framework to consider the influences of neighborhoods on the homogeneity of prediction. The study also evaluates the proposed approaches with popular ML models. The methodological development focuses on identifying suitable criteria following statistical analysis, evaluating both knowledge-based and data-driven criteria weights, and validating the models against a known scenario from a recent flood event in the Ganges tidal plain of Bangladesh. Random Forest (RF) exhibited the strongest statistical association with the observed flood impact as indicated by a high adjusted R 2 (0.6484) in out-of-fold cross-validation. It also displayed higher classification accuracy (91.39%), though the ranking outcome had higher sensitivity to criteria perturbation. Compensatory and non-compensatory MCDMs with Analytic Hierarchy Process (AHP) based weighting schemes show classification accuracies of 65.00-66.39%, close to those of SVM (0.6694) and ANN (0.6806), with the least effect from input perturbation. Various MCDM-GWR and ML models predict 26–50% and 25–39% of the area as ‘high’ and ‘very high’ risk zones. Lower uncertainty in prediction for 57% area indicates moderate agreement among the models. A comparative analysis of these models, prioritizing socio-economic challenges, will encourage policymakers to adopt best practices in response, mitigation, and adaptation.