Authors: Hongxing Liu, Yitao Pu, Dan Tian, and Matthew LaFevor – University of Alabama; Song Shu – Appalachian State University; Shujie Wang – The Pennsylvania University; Lei Wang – Louisiana State University; Habin Su – Texas A&M University – Kingsville
Title: Near-Real-Time Monitoring of Reservoir across the CONUS Using Multi-Sensor Satellite Observations and Deep Learning Techniques
Presentation Type: Poster Presentation
Abstract: Effective reservoir monitoring is critical for water resource management, yet the majority of reservoirs across the contiguous United States remain ungauged or poorly monitored, limiting the applicability of national reservoir operation models. This study presents a scalable, satellite-driven framework for near-real-time monitoring of key reservoir variables by integrating multi-sensor remote sensing observations and deep learning techniques. We first develop a high-resolution, temporally consistent georeferenced reservoir inventory using September 2020 Sentinel-1 SAR imagery, addressing inconsistencies in existing datasets and providing a uniform baseline for over 4,000 reservoirs. The resulting reservoir dataset is spatially joined with the National Inventory of Dams (NID), enabling comprehensive analysis of reservoir morphological characteristics and supporting regional assessments of water resource distribution. Building on this foundation, reservoir-specific hypsometric rating curves are constructed by integrating water surface elevation from SWOT altimetry with water surface area derived from SAR and optical imagery. The hypsometric rating curves enable reservoir water surface elevation to be estimated from surface area observations from SAR and optical satellites, which are far more frequent than SWOT and other altimetry missions. This approach enables near-real-time estimation of reservoir water surface areas, water levels, and storage volume change by integrating high-temporal-resolution SAR and optical planimetric observations with critical but less frequent SWOT and altimetry vertical measurements. We have generated high-frequency water surface area time series by integrating Sentinel-1 SAR and Landsat-8/9 and Senetinal-2 optical imagery using the Segment Anything Model 2 (SAM2), a deep learning foundation model that leverages both spatial and temporal information for efficient large-scale segmentation and quantification of reservoirs. The derived time series enable the analysis of seasonal variability, long-term trends, and extreme hydrologic events across reservoirs. The resulting data products provide critical inputs for NOAA’s national reservoir operation models, improving their calibration, spatial coverage, and applicability to ungauged systems. This work demonstrates how emerging satellite missions and deep learning methodologies can transform reservoir monitoring into an operational, data-rich system supporting water security, hazard mitigation, and climate resilience.