A Data-Driven Approach to Hurricane Debris Modeling

Abstract

The large amount of debris generated in the aftermath of hurricane and storm events can cause severe financial and logistical burdens to coastal communities. Existing debris estimation models mainly focus on wind-induced debris and produce estimates with errors of nearly 50%, highlighting the importance of developing more comprehensive models that can account for other types of debris while improving accuracy. Therefore, the objective of this study is to develop a probabilistic framework to estimate the presence and amount of waterborne debris following a severe storm using machine learning (ML) techniques as a function of relevant storm and landcover parameters. Machine learning techniques are leveraged to generate debris presence and volume models, employing pre- and post-event aerial and satellite imagery and a debris removal database for Hurricane Ike, respectively. The results show that the ensemble learning algorithms perform the best for both tasks, with a misclassification error of 5.56% for the debris presence predictive model, and a normalized root mean squared error (RMSE) value of 11.98 for the debris volume model, the lowest RMSE of the tested algorithms. Dual-layer ML models are also investigated, incorporating the debris presence as a predictor in the debris volume model. The results show a percent error of 11.29% for the dual-layer model and an approximately 5.4% increase in performance with respect to the model that does not incorporate debris presence. The generated debris volume and presence models will provide useful tools to inform decision-making, evaluate mitigation strategies, facilitate recovery efforts, and improve resource allocation following a storm event.