Optimizing Machine Learning Models for Event-conditioned Flood Susceptibility Mapping in Binh Dinh Region

Abstract

Abstract: Flooding is a major natural hazard in Vietnam, causing significant human and economic losses, particularly in Bình Định (now the eastern part of Gia Lai Province), where complex terrain and dense river networks increase vulnerability. This study integrates the Puma Optimization (PO) meta-heuristic algorithm with Random Forest (RF), Extreme Gradient Boosting (XGB), and Support Vector Machine (SVM) models to improve hyperparameter tuning, addressing limitations of conventional methods such as Random Search (RS) and Particle Swarm Optimization (PSO). The framework was implemented on Google Earth Engine using 16 conditioning factors and 1,978 flood/non-flood samples from the 06–10/2022 flood events. Results show that the PO–XGB model achieved the highest performance (R² = 0.84), outperforming PSO–XGB (0.81) and RS–XGB (0.55), with low errors (MAE = 0.07; RMSE = 0.11). The results demonstrate that the PO algorithm achieves stable convergence and high accuracy in capturing the complex nonlinear relationships inherent in flood processes. The integration of next-generation optimization algorithms shows strong potential to significantly enhance the reliability of flood susceptibility maps under the 2022 wet-season context, while effectively overcoming the limitations of standalone approaches. This study not only contributes methodologically to the application of artificial intelligence in hydrology but also provides a scientific basis and decision-support tools for sustainable spatial planning and disaster risk reduction at the local scale.