Flood Defense Standard Estimation Using Machine Learning and Its Representation in Large-Scale Flood Hazard Modeling

Abstract

We propose a machine learning-based approach to estimate the flood defense standard (FDS) for unlabeled sites. We adopted random forest regression (RFR) to characterize the relationship between the declared FDS and 10 explanatory factors contained in publicly available data sets. We compared RFR with multiple linear regression (MLR) and demonstrated the proposed approach in the conterminous United States (CONUS) and England, respectively. The results showed the following: (a) RFR performed better than MLR, with a Nash–Sutcliffe efficiency of 0.85 in the CONUS and 0.76 in England. Unsatisfactory performances of MLR indicated that the relationship between the FDS and explanatory factors did not obey an explicit linear function. (b) RFR revealed river flood factors had higher importance than physical and socio-economic factors in the FDS estimation. The proposed RFR achieved the highest performance using all factors for prediction and could not provide good predictions (NSE < 0.65) using physical or socio-economic factors individually. (c) We estimated the FDS for all unlabeled sites in the CONUS and England. Approximately 80% and 29% of sites were identified as high or highest standard (>100-year return period) in the CONUS and England, respectively. (d) We incorporated the estimated FDS in large-scale flood modeling and compared the model results with official flood hazard maps in three case studies. We identified obvious overestimations in protected areas when flood defenses were not taken into account; flood defenses were successfully represented using the proposed approach.