Machine Learning for Watershed Science: Capturing Bedrock-to-Canopy Co-Variability and Hydrologic Dynamics
Dr. Haruko Wainwright is an assistant professor of Nuclear Science and Engineering, and Civil and Environmental Engineering at the Massachusetts Institute of Technology.
Abstract: Watershed science requires characterization and monitoring across interconnected bedrock-to-canopy compartments. Particularly, in snow-dominated mountainous watersheds of the western United States, subsurface water storage following snowmelt plays a central role in streamflow generation and ecosystem productivity. Recent advances in remote sensing technologies—including airborne LiDAR, hyperspectral imaging, and electromagnetic surveys—along with the growing availability of low-cost distributed in situ sensors have greatly expanded observational capabilities. However, integrating these heterogeneous, multiscale datasets into a coherent understanding of watershed function remains a major challenge. This talk presents machine-learning approaches for capturing spatial heterogeneity and temporal dynamics of hydrological and ecosystem processes in snow-dominated mountainous watersheds. First, unsupervised learning—specif cally clustering—is an powerful approach to characterize bedrock-to- canopy co-variability in heterogeneous watersheds where ground-truth data are limited. We introduce a watershed zonation framework that integrates multiple spatial data layers to identify subsystems with distinct distributions of bedrock-through-canopy properties, providing a semi-quantitative basis for understanding the bedrock-to-canopy co-evolution as well as guiding monitoring and sampling strategies. Supervised regression is then employed to spatially distribute difficult-to-measure parameters, such as soil and bedrock properties, at the watershed scale. Second, we present a physics-informed machine- learning framework that integrates real-time in situ sensor data with vadose-zone flow simulations for near-surface soil-moisture monitoring and forecasting. This approach combines ensemble Kalman filtering for data assimilation with a Flow Map Learning–based emulator to accelerate model prediction during assimilation. Together, these applications demonstrate how ML and AI methods can enhance data integration, computational efficiency, and predictive capability in watershed science.
About: Dr. Haruko Wainwright is an assistant professor of Nuclear Science and Engineering, and Civil and Environmental Engineering at the Massachusetts Institute of Technology. Before joining MIT, she was a Staff Scientist in the Earth and Environmental Sciences Area at Lawrence Berkeley National Laboratory. Her research focuses on environmental informatics, aiming to improve understanding and predictions in Earth and environmental systems through mechanistic modeling and machine learning.