Rangel Daroya (UMass Amherst, United States); Jonathan Flores (UMass Amherst, United States); Colin Gleason (UMass Amherst, United States); Subhransu Maji (UMass Amherst, United States)

The SWOT satellite mission provides unprecedented hydrological measurements of water surface elevation and slope, and inundation extent of rivers globally. While water surface elevation and slope have been validated extensively, inundation extent and width have relied heavily on manually digitized water surfaces from aerial orthophotos, where labeling takes several hours per image, and thus, there are fewer samples to validate. To address this, our team developed a deep learning algorithm (RiverScope) that masks surface water in PlanetScope images. The algorithm was trained on 1,145 manually labeled PlanetScope images of rivers across the globe and the model was applied to masking a global sample of ~5000 SWORD reaches. Each reach has coincident SWOT and PlanetScope data within +12 hours of each other from April 1, 2023 – April 1, 2025, covering a total area of ~ 30 million square kilometers. Using the RiverScope masked images, we calculated effective width for each reach through time, using the SWORD reach length and area of water pixels. Preliminary results, using ~300 reaches within the Yukon River Basin, show that SWOT-derived widths from the RiverSP product have a mean average error (MAE) of 48% and 16% for RiverScope. This presentation will build on the preliminary results, highlighting the range of performance observed across the global sample.