Team
- Faculty: Jana Doppa (WSU); Kirti Rajagopalan (WSU), Yan Yan (WSU)
- Graduate students: Mohammed Amine (WSU, CS); Aryan Deshwal (WSU, CS); Bhupinderjeet Singh (Ag Biosys Eng, WSU)
- Undergraduate students: Wyatt Croucher (WSU, CS)
Objective
- Streamflow estimates are typically obtained via hydrological models which take meteorological inputs (temperature, precipitation, radiation, wind speed, humidity, etc.), land surface characteristics such as types of vegetative cover/land use (forests, grassland, ag land, urban areas, bare soil), and soil characteristics. These physics-based models solve differential equations related to the water and energy balance to calculate water and energy fluxes in the stream to quantify streamflow. These models also generalize to different regions but can be suboptimal in terms of streamflow prediction accuracy in many regions. Our goal in this project is to synergistically combine the benefits of domain knowledge in the form of physics-based models and spatiotemporal graph neural networks (GNNs) to improve the overall prediction accuracy and uncertainty quantification, especially in small-data settings.
Hypotheses
- Principles from representation learning and science-guided machine learning can aid in improving the precision and skill of streamflow forecasting.
- In small-data settings, importance-weighted training can improve the prediction accuracy and generalizability of models.
Related Publications
Sorry, no publications matched your criteria.
