Connecting Atmospheric and Land Surface Patterns with the Occurrence of New England Peak Flow Events Using Machine Learning Techniques

In the heavily populated Northeastern United States, floods damage infrastructure, disrupt trade and commerce, and mobilize pollutants. As a result, improving the near- and long-term forecasts of hydroclimate extremes in New England is of critical importance for regional water resource planning and emergency response. At present, however, forecasting of these events remains rudimentary, and has not harnessed recent advances in climate informatics and machine learning that track water as it cycles through the ocean, atmosphere, and land surface, and identify precursors that signal elevated flood hazard. Here we investigate the relationships between atmospheric and land surface patterns and the occurrence of peak flow events in New England by leveraging machine learning techniques. We first utilize the GAGESII dataset to identify moderate (return period 5–20 years) and major peak (return period 21–70 years) flow events. We then apply a random forest algorithm on high-resolution reanalysis (ERA5) to establish the key atmospheric and land surface conditions on the day of and up to five days prior to the peak flow events, and show that geopotential height anomalies (500 mb) and volumetric soil moisture represent master controls on the occurrence of floods. Finally, we use the MATLAB Self-Organizing Maps (SOM) toolbox on the ERA5 dataset to identify 6–8 distinct flood-generating mechanisms in New England, including midlatitude cyclones of different tracks, frontal systems, and convective systems — some of which are associated with high antecedent soil moisture, and others that are not. Ongoing work is linking these distinct flood-generating mechanisms to broader patterns in the atmosphere, ocean, and land surface. Our work enhances our understanding of the complex interactions between atmospheric dynamics and land surface conditions that contribute to riverine flood hazard in the New England region.