Flood Probability Prediction Model Based on Dual-Layer Blending Fusion
TL;DR
This study enriches the theoretical foundation of flood probability prediction models, offers a new approach for multivariable nonlinear flood forecasting, and provides a scientific basis for disaster mitigation and risk management.
Abstract
Floods are among the most devastating natural disasters, posing a major threat to human life and socioeconomic systems. Existing flood probability prediction models have limitations and lack flexibility. This study aims to develop an efficient flood probability prediction model for disaster prevention and policy making. In this paper, a two-layer fusion model based on CatBoost, support vector regression (SVR), and gradient boosting regression (GBR)—denoted as (CatBoost-SVR)-(CatBoost-GBR)—is proposed for flood probability prediction. First, the light gradient boosting machine (LightGBM), coupled with Shapley additive explanations (SHAP), is employed to evaluate the importance of 20 indicators affecting flood probability, and the 10 most influential features are selected as model inputs. Then, the base models are constructed using CatBoost, SVR, and GBR, and a blending approach is employed to perform two-layer fusion, synthesizing the strengths of the base models to improve prediction accuracy and stability. The results show that the proposed two-layer fusion model achieves high accuracy in predicting flood probability, with a root mean square error of 0.141 and an R 2 of 0.832. Compared with single-model and single-layer fusion approaches, the prediction accuracy is significantly improved, and the model effectively captures the changing trend of flood probability. The model is applied to predict flood probabilities across 745,305 samples; combined with a flood prediction index, a comprehensive risk score is calculated and classified into four risk levels. Based on the identified risk levels and the most influential features, flood management guidelines were developed, and corresponding prevention and mitigation recommendations were proposed. This study enriches the theoretical foundation of flood probability prediction models, offers a new approach for multivariable nonlinear flood forecasting, and provides a scientific basis for disaster mitigation and risk management.