Multi-Machine Learning Ensemble Regionalization of Hydrological Parameters for Enhances Flood Prediction in Ungauged Mountainous Catchments

Abstract. Machine learning-based parameter regionalization is an important method for flood prediction in ungauged mountainous catchments. However, single machine learning parameter regionalization often exhibits limitations in prediction accuracy and robustness. Therefore, this study proposes a multi-machine learning ensemble regionalization method that integrates Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), and Extremely Randomized Trees (ERT) methods (GBM-KNN-ERT) to regionalize the sensitive parameters of the Topography-Based Subsurface Storm Flow (Top-SSF) model. Validated across 80 mountainous catchments in southwestern China, the GBM-KNN-ERT method demonstrates superior performance with 90 % of ungauged catchments achieving the Nash-Sutcliffe Efficiency (NSE) above 0.9, representing a 67.44 % improvement over single machine learning parameter regionalization. Notably, the GBM-KNN-ERT method shows improved robustness to climate change and changes in the number of donor catchments compared to other regionalization methods. An optimal balance between accuracy and computational efficiency was achieved using 20–40 high quality donor catchments (NSE greater than 0.85). This study provides systematic evidence that multi-machine learning ensemble can effectively address regionalization challenges in ungauged mountainous regions, offering a reliable tool for water resource management and flood disaster mitigation.