Spatio-Temporal Machine Learning for Flood Risk Assessment Under SSP Scenarios: A Case Study of Maha Sarakham, Thailand
TL;DR
This study developed and evaluated machine learning models for flood severity prediction within a hexagonal grid system under rainy season conditions in Maha Sarakham, Thailand and recommends XGBoost for operational forecasting due to its temporal stability, computational efficiency, and interpretability.
Abstract
Flooding is a destructive natural hazard intensified by climate change, posing challenges to sustainable disaster risk management. This study developed and evaluated machine learning models for flood severity prediction within a hexagonal grid system (H3, resolution 8) under rainy season conditions in Maha Sarakham, Thailand. Four models Random Forest (RF), XGBoost, Gradient Boosting (GB), and Support Vector Machine (SVM) were trained using 11 environmental variables across historical years (2018, 2021, 2022) and tested on a projected year (2025) under SSP scenarios. XGBoost demonstrated the most stable performance (accuracy > 0.95 across all years), while SVM achieved high historical accuracy (0.970 average) but failed to detect positive flood cases in 2025 (recall = 0), highlighting the importance of temporal validation. Topographic variables were the most consistent predictors, but NSMI (soil moisture) emerged as the top SHAP predictor in 2025 (r = 0.52), suggesting a shift in flood-generating mechanisms under climate change. A polarization pattern was observed: flood-affected area declined to 6.8% in 2025 (79% reduction from 2022), yet maximum flood point counts remained high at 14.0, indicating more concentrated but intense flooding. Under SSP projections, using the historical baseline (27.5%), SSP1-2.6 (45.2%) and SSP2-4.5 (45.0%) indicate increased flood risk relative to the historical baseline through 2040. The SSP5-8.5 projection (3.4%) is identified as a model extrapolation artifact through formal out-of-distribution assessment (Mahalanobis distance = 8.72, p < 0.001) and is therefore excluded from policy recommendations. Although GRU and LSTM achieved marginally higher AUC values in retrospective validation, we recommend XGBoost for operational forecasting due to its temporal stability, computational efficiency, and interpretability. We further recommend integrating real-time soil moisture monitoring into early warning systems and shifting to hotspot-targeted adaptation strategies.