Flood susceptibility mapping in the Yom River Basin, Thailand: stacking ensemble learning using multi-year flood inventory data

Accurate assessment models of flood susceptibility are crucial for informing risk management strategies related to severe threats posed by floods. This study assessed flood susceptibility in the Yom River Basin, Thailand, using conventional and ML methods. SE was compared with KNN, SVM, DT, RF, and a Stacking ensemble model (SVM, DT, RF). A point-based flood inventory was sampled from multi-year flood polygons using a method considering flood frequency and inundation size. Results showed all ML models, except KNN, outperformed SE. RF achieved AUCs of 96.0% (test) and 96.1% (verification), while Stacking achieved 99.9% (test) and 96.1% (verification). Stacking also outperformed in accuracy (0.982, 0.893), precision (0.974, 0.915), F1 (0.990, 0.866), sensitivity (0.982, 0.890), specificity (0.974, 0.920), and kappa (0.964, 0.786). These findings highlight the potential of using ensemble ML techniques to significantly improve flood susceptibility mapping and risk management in data-limited regions such as the Yom River Basin.