Prediction of rainfall-induced debris flow using Random Forests and Bayesian Optimization in Yingxiu Town, Wenchuan County, China

• Published in: Bulletin of engineering geology and the environment Vol. 83; no. 5; p. 156
• Main Authors: Liu, Renwen, Zhou, Wei, Yin, Huaqiang, Ren, Wenming, Feng, Jianguo, Deng, Jiulin
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.05.2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Automation; Bayesian analysis; Bayesian theory; Casualties; Datasets; Debris flow; Decision trees; Detritus; Earth and Environmental Science; Earth Sciences; Earthquakes; Foundations; Geoecology/Natural Processes; Geoengineering; Geology; Geotechnical Engineering & Applied Earth Sciences; Hydraulics; Hydrologic data; Hydrology; Landslides; Landslides & mudslides; Machine learning; Mathematical models; Mechanics; Mitigation; Mountain regions; Mountainous areas; Nature Conservation; Optimization; Original Paper; Performance prediction; Plant cover; Precipitation; Precipitation data; Probability theory; Rainfall; Rainfall forecasting; Risk reduction; Topography; Vegetation; Vegetation cover; China; Sichuan China; Bayesian Optimization; Random Forests; Debris flow; Prediction
• ISSN: 1435-9529; 1435-9537
• DOI: 10.1007/s10064-024-03649-2

Mountainous areas are susceptible to rainfall-induced debris flows. An accurate debris flow forecast helps reduce disaster damage and casualties. Conventional forecasting methods, which primarily concentrate on precipitation data, exhibit limitations by neglecting other crucial variables. This study aims to address this gap by constructing an advanced Random Forest model that integrates a diverse range of factors linked to rainfall-induced debris flows. To do this, we compiled data on Yingxiu Town, including historical debris flow events and 19 potential influencing factors. Through rigorous feature selection, we focused on the key factors from the dataset. Our refined Random Forest model integrates critical variables such as fractional vegetation cover, landslide ratio, and a suite of rainfall metrics encompassing average intensity, duration, and cumulative precipitation, accounting for both antecedent and current rainfall events. To enhance the model efficiency, we opted for an automated approach to optimize hyperparameters. This involved utilizing Bayesian optimization to determine the optimal configurations for our model. This customized approach yielded a model that excelled in predictive performance when compared to a more generalized model incorporating all potential factors. On the test dataset, the model achieved an F 1 score of 0.57, an accuracy of 0.9, and a significant area under the ROC curve (AUC) of 0.87. Beyond its predictive success, our study offered insights into the mechanics of hyperparameter tuning and provided a nuanced understanding of decision tree construction within the Random Forests algorithm. The contributions of our research offer valuable guidance for enhancing debris flow risk mitigation strategies.