Estimation of base and surface flow using deep neural networks and a hydrologic model in two watersheds of the Chesapeake Bay

River water arrives by two routes — rain running straight off the surface, which drives flood peaks, and water that has soaked into the ground and seeps out over weeks, which decides whether a river survives a dry summer — and conventional watershed models are poor at telling the two apart. Nineteen years of records from two watersheds feeding Chesapeake Bay were used here to set a standard process-based model against two learning-based ones: a single model covering each whole basin, and a finer version that treats every sub-basin separately. The basin-wide learner modestly beat the conventional model, by 14%, 49% and 38% for slow, fast and total flow in the Tuckahoe Creek watershed, but the sub-basin version was better still in both watersheds, exceeding the conventional model by 39%, 79% and 50% at Greensboro. The advantage came from letting the model see how a watershed is put together internally instead of treating it as a single unit. The authors are careful to add that no one model won everywhere across the nineteen years, and suggest using several in combination.

Download paper here