Optimizing flood resilience in China’s mountainous areas: Design flood estimation using advanced machine learning techniques

China We developed machine learning (ML) models for design flood estimation in mountainous catchments (≤ 500 km²) across China. This process considered different ML algorithms (random forest, extreme gradient boosting, and support vector regression), model scopes (nation and hydrological zones), and feature input sets (1–14 features) to optimize model development strategies. Based on estimation performance and hyperparameter tuning efficiency, random forest was found to be the optimal algorithm. The optimal model scope resulted in five distinct models: a single lumped model encompassing six eastern zones and four separate zonal models for the western zones. Considering both accuracy and efficiency, the optimal number of input features ranged from 5 to 14 for different models. High estimation accuracy was observed in the Qinba-Dabie North, Southeast, Southwest, and Yunnan-Tibet Zone, with average RMSE , R ², MQE , and QR ranging from 55.90 to 103.97, 0.83–0.93, 45.62–65.77 %, and 55.90–60.98 %, respectively, for the test set across different return periods. The remaining zones exhibited moderate accuracy, with the Northwest Basin Zone demonstrating particularly low accuracy due to fewer catchments. Notably, catchments with areas > 100 km² demonstrated higher estimation accuracy, with an average 60 % reduction in MQE and a 30 % increase in QR compared to catchments of all sizes. This study provides crucial reference and data support for national flash flood prevention efforts. • Nationwide design flood estimation for mountainous catchments in China. • Optimized development strategies for machine learning models considering algorithms, model scopes, and feature input sets. • Model applicability evaluation across hydrological zones and catchments with distinct characteristics. • Comprehensive design flood results for nearly 300,000 small- and medium-sized mountainous catchments across China.