Zimeena Rasheed, Akshay Aravamudan, Ali Gorji Sefidmazgi, Georgios C. Anagnostopoulos, Efthymios Nikolopoulos

Flood prediction across scales and more specifically in ungauged areas remains still a great challenge that limits the efficiency of flood risk mitigation strategies and disaster preparedness. Building upon the recent success of Machine Learning (ML) models on streamflow prediction, this work presents a prototype ML-based framework for flood warning and flood peak prediction. The fundamental elements of the proposed system consist of a) a LSTM model for classifying storm events to threat/no-threat given a threshold based on the 90th flow percentile and b) the flood peak prediction models. The selected ML-models for flood peak prediction are the Histogram based Gradient Boosting Regressor and the Random Forest. One of the strengths, and reason for selection, of these decision-tree models is their degree of interpretability. This is exploited in the study to help us spatially disentangle the role of both the static and dynamic drivers of flood peak response. Our analysis is presented for 18 distinct hydroclimatic regions across the contiguous US. Results reveal a significant regional dependence on both predictive performance and dominant flood predictors, which emphasize the variability in the complexity of a catchment’s hydrologic behavior as well as its impact on modeling flood response. Evaluation of the drivers of flood peaks noted distinct dependencies among the dynamic and static predictors considered in our models for flood peaks of different severity. Specifically, low-moderate flood events showed a clear preponderance for the static catchment attributes over dynamic predictors like precipitation whereas precipitation was the dominant driver for the high severity flood peaks. The proposed flood peak prediction models were compared against a state-of-the-art LSTM model and were shown to consistently outperform in ungauged basins. Overall, the proposed system classified storms correctly for >85% in all cases and exhibited a percent relative difference in flood peak estimates of <30% in most cases.