From Diagnosis to Protection: A Data-Driven Framework for Assessing Flash Flood Vulnerability of Huizhou Covered Bridge Heritage

This study investigates the structural vulnerability of Huizhou covered bridges under mountain-flood events to enhance the resilience of traditional architectural heritage. A dataset of 116 flood-induced damage cases involving 89 bridges over a five-year period was analyzed to identify damage patterns and dominant risk factors. A machine-learning framework based on the Random Forest (RF) algorithm was developed to quantify factor importance and to elucidate nonlinear impact mechanisms. Results show that bridges with stone-arch foundations and brick-timber truss systems exhibit the highest flood resistance. Bridges located near transportation nodes or drainage outlets experience lower damage probabilities due to shielding effects from surrounding buildings. Six factors — openness ratio, protection level, building height, structural type, state of repair, and ground-floor area — collectively account for over 75% of the total relative importance, indicating their dominant role in flood vulnerability. Factors with importance values below 0.05 were removed through Area Under the Curve (AUC) and Accuracy (ACC) optimization, resulting in optimal model performance. Based on the identified damage mechanisms and risk drivers, the study proposes targeted mitigation strategies integrating engineering and non-engineering measures. These findings provide quantitative evidence to support the protection, management, and preventive conservation of Huizhou covered-bridge heritage.