Aggregated flood susceptibility mapping in Upper Chao Phraya River Basin using Shannon’s Entropy, Machine Learning, and Stacking ensemble methods

Abstract

Flood susceptibility mapping in large and heterogeneous basins requires methods capable of representing spatial variability that conventional basin-wide models often overlook. This study develops a sub-basin aggregation framework for the Upper Chao Phraya River Basin, Thailand, integrating localized susceptibility modelling into a unified basin-scale product. Thirteen flood conditioning factors were initially selected and objectively weighted using Shannon’s Entropy (SE), which reduced to 6–9 distinct hydrological drivers in the basin and each sub-basin. Nine machine learning algorithms (RF, KNN, SVM, DT, LR, ANN, NB, CART, and MLP) and a Stacking ensemble were applied to each basin, followed by three aggregation strategies: (1) SE-based sub-basin aggregation, (2) Stacking-based aggregation, and (3) best ML-based sub-basin aggregation. Results show that aggregated models outperform single basin-wide models, with the best ML-based aggregation achieving the highest accuracy (AUC = 0.973). Meanwhile, the SE-based aggregation produced the most balanced susceptibility map (44.3% Very Low; 15.7% Very High), highlighting a trade-off between predictive performance and spatial realism. The framework effectively captures sub-basin heterogeneity—for instance, Curvature was dominant in the Wang sub-basin, whereas Elevation prevailed elsewhere. Overall, the proposed aggregation strategy offers a scalable and transferable approach for large-scale modelling and supports data-driven flood management in complex river systems.