ExaSheds: Advancing Watershed System Science using Machine Learning-Assisted Simulation

This project develops and applies a process-guided deep learning (PGDL) model to generate reliable projections of future river water temperatures in the Delaware River Basin and Upper Neversink Watershed

watershed

An area of land that channels rainfall and snowmelt to creeks, streams, and rivers, and eventually flows to the lowest point in a basin.

Source: USGS Source

under scenarios of altered climate and land use. The PGDL approach injects a deep neural network with process knowledge in the forms of process-based model inputs, architectural constraints, training data, and/or feedback during model training. PGDL has shown exceptional accuracy in preliminary applications to predict lake and river water temperatures and river discharges, with better generalizability than other deep learning methods for predictions in new locations and time periods. However, PGDL has been implemented in diverse and sometimes incompatible ways, and deep learning methods continue to evolve rapidly, such that there is much potential to introduce new PGDL techniques, and assess and optimize existing ones, to achieve even better performance for environmental prediction.

PGDL will be applied in this project to predict water temperature, an ecological master variable governing rates of biological activity, chemical reactions, and habitat suitability in stream and river networks. USGS project members will collaborate on two of the six themes of the larger, multi-institution ExaSheds project also funded by DOE: Theme 3, hybrid ML-physics models for hydrology,

hydrology

The science that deals with water, its properties, circulation, or distribution on Earth's surface, and in the atmosphere through evapotranspiration.

Source: NASA earth observatory Source

and Theme 6, demonstration and integration. Collaboration on Theme 3 will explore new PGDL techniques in the context of recent but separate successes by several team members in the areas of hybrid Bayesian neural networks, high-resolution process-based temperature modeling, multifidelity modeling, and process-guided model architectures and training procedures. Collaboration on Theme 6 will yield multi-decadal PGDL projections for scenarios of climate change,

climate change

A change in the average conditions — such as temperature and rainfall — in a region over a long period of time.

Source: NASA Climate Kids Source

land use change, and/or extreme events such as droughts and heat waves. Our overarching goal is to push past the limits of current PGDL methods, achieving deeper integration of process knowledge and greater reliability of projections, to support evidence-based decision making in a changing world.