Assessing regional flood risks under climate change: a machine learning and spatial clustering approach

Abstract

This paper introduces a machine-learning-based flood risk prediction model that integrates data of different granularities, with a focus on predicting flood probabilities under climate change scenarios. Initially trained on historical data, the model utilises CMIP6 projections to estimate monthly flood probabilities for 312 cities in China from 2025 to 2100. Additionally, the SKATER method was used to incorporate spatial adjacency, enabling effective risk zoning across cities. Among the four models tested, LightGBM consistently outperformed GLM, random forest, and neural network in terms of accuracy and adaptability. The results suggest that as climate conditions worsen, moving from SSP1-2.6 to SSP5-8.5, flood probabilities are expected to increase in Ningxia, Shaanxi, Henan, and Hebei, as well as in parts of the southwestern border regions. This study contributes to regional disaster risk management by effectively addressing low-frequency, low-resolution data, and demonstrates strong potential for application across diverse countries and regions worldwide.