Challenges and Successes in Building a Global, Operational, Theory-Guided Machine Learning Model

Given that much of the world is ungauged and that even within the United States, we have entire regions with sparse data, we need more insight into future river flows on many different time horizons. At the AGU Fall Meeting 2022, we presented some challenges and successes of building an operational, deep-learning hydrological modeling system to scale commercially and globally to guide water management efficiently.

The HydroForecast model is a theory-guided, statistical prediction model that sources globally and publicly accessible data inputs from remote-sensing observations and meteorological datasets. In our presentation, we shared how we use an operational neural network distributed modal to make predictions in data-sparse regions and construct historical reanalysis streamflow records in ungauged, human-altered basins.

Our distributed model delineates sub-basins and incorporates upstream gauge observations to accurately predict runoff and routing through river networks. See our AGU Fall Meeting 2022 poster to learn more about the model and its capabilities.
In order to protect habitats and ecosystems for riverine and aquatic species, we need more information about historical streamflow in basins with alterations and ones without gauges. Through a project with The Nature Conservancy California, we developed 20-year daily records of streamflow in over 300 basins across the state.