A Watershed-Aware Earth Observation Foundation Model for Hydrologic Science

Kshitij Dahal; Saurav Kumar; Laxman Bokati

Earth observation (EO) models today can classify land cover or segment images, but they do not understand how watersheds function. They are not trained to learn hydrologic processes such as runoff generation, channel and river networks, or water storage components as spatial processes. We aim to develop a new foundation model trained on unlabeled satellite data with self-supervised learning so it can learn hydrologic processes directly from EO data.

Our model captures both spatial and temporal patterns, like existing EO foundation models (for example, Prithvi), but it also understands the structure and behavior of watersheds with masked autoencoders. We first test the model’s hydrologic understanding by asking it to estimate key watershed properties such as watershed area, concentration time, and other physical descriptors. The second test focuses on streamflow forecasting, where we expect the model to improve predictions, especially in arid and semi-arid watersheds.

The model uses MODIS daily snow cover 250 m resolution, combined with elevation data to inform spatial hydrologic processes. Preliminary results show a 4% increase in forecast skill for daily streamflow across 404 USGS stations in Arizona. This work aims to move EO foundation models from patches-level understanding toward process-level understanding of the Earth system, bridging satellite data and hydrologic science.

Remote Sensing of HABs and Watershed-Level Drivers: A Decadal Chlorophyll-a Assessment

Hamid Norouzi; Jillian Greene; Leonid Metlitsky; Reginald Blake

Harmful algal blooms (HABs) remain a significant water-quality concern across inland watersheds, and many regions still lack the consistent monitoring needed to understand how blooms develop and change over time. This work presents a satellite-driven framework for estimating Chlorophyll-a concentrations and evaluating long-term HAB-related trends using multi-sensor remote sensing, watershed-scale predictors, and physics-informed machine-learning methods. The approach relies on Sentinel-2 MSI and Landsat 8/9 OLI observations processed with inland-water atmospheric correction to generate Chlorophyll-a estimates for each satellite overpass. We applied this method across New York State, producing more than a decade of reconstructed Chlorophyll-a records for thousands of lakes and ponds. This long time series makes it possible to examine climate-related variability and identify watershed-level factors, such as land use, thermal conditions, hydrologic setting, and weather patterns, that help explain differences in bloom behavior across regions. Comparisons with available in-situ measurements show strong consistency across sensors and waterbody types. To support managers and decision-makers, the framework is paired with an operational web platform that provides near-real-time Chlorophyll-a maps, historical time series, and event notifications for each overpass. This reduces monitoring gaps and offers a practical early-warning resource for agencies working to assess bloom conditions. The overall approach is scalable to other regions and watershed systems and can be integrated into ongoing water-quality and climate-adaptation programs.

Earth Observation Fusion and Assimilation for Enhanced Hydrologic Representation of Water-Scarce Transboundary watersheds

Saman Ebrahimil Saurav Kumar

Growing water scarcity under a warming, drying climate in the U.S. Southwest is intensifying competition over limited water resources and threatening the long-term sustainability of irrigated agriculture practices. These challenges are acute in highly managed transboundary river systems, where water fluxes are shaped by human interventions and cross-border data asymmetries persist; the Middle Rio Grande (MRG) is one such system. In this basin, pressure on limited water resources and shifts in agricultural practices from seasonal to perennial strongly control evapotranspiration and flow conditions, while management actions (reservoir operations, canal diversions, groundwater pumping) are poorly documented, with sparse and inconsistent records across the border. This work evaluates how Earth observation (EO) fusion and assimilation can compensate for missing management information and improve hydrologic model performance in such basins. A cross-border crop and land-cover/land-use dataset is constructed from multisensory satellite imagery, and a basin-scale hydrologic model is implemented for a recent decade under three configurations: static land cover, annually updated EO-derived land cover, and dynamic land cover with ensemble assimilation of high-resolution EO-derived leaf Area Index (LAI). The modeling experiments are designed to test whether dynamic EO-derived forcing and LAI assimilation improve evapotranspiration timing and patterns, sharpen irrigation representation, and reduce uncertainty in key states and fluxes. The anticipated outcome is a transferable EO fusion and assimilation framework that indirectly reveals undocumented anthropogenic signals such as irrigation timing, crop portfolio shifts, and changes in conjunctive use, and supports modeling and management of arid transboundary basins under growing water scarcity.