Danyang Zhao (University of Stuttgart, Institute of Geodesy, Germany); Omid Elmi (University of Stuttgart, Institute of Geodesy, Germany); Junyang Gou (Institute of Geodesy and Photogrammetry, ETH Zurich, Zurich, Switzerland, Switzerland)

The Surface Water and Ocean Topography (SWOT) mission, through its Ka-band Radar Interferometer (KaRIn), provides high-resolution observations of surface water extent and elevation, offering new opportunities for hydrological applications. However, the standard surface type classification in SWOT’s pixel cloud (PIXC) product can be unreliable over complex inland water bodies. Misclassification, caused by high noise levels, and the challenge of defining the extent of the water body based on the classes can impair the accurate estimation of key hydrological variables, including river water surface elevation, slope, and width, as well as lake water height and surface area, as provided in the vector products (RiverSP and LakeSP).
To improve classification accuracy, we developed and evaluated a series of deep learning models that vary in both architecture and input features. These models account for different pixel cloud sizes and incorporate combinations of SWOT observables, such as coherence, backscatter power (Sigma0), phase noise standard deviation, surface height, and auxiliary indicators derived from satellite imagery. We assessed model performance over rivers and lakes in Germany by generating modified RiverSP and LakeSP products based on the reclassified PIXC data and comparing them against results from the original classification.
For rivers in particular, we analyzed the water surface profile at each time epoch by extracting longitudinal river profiles and assessing variability after removing the static river shape. The model with the lowest root mean square error (RMSE) in profile fluctuation was considered the best-performing. Preliminary results over the Peene and Warnow rivers show an average RMSE improvement of approximately 1 m compared to the original classification.
All tested models so far outperformed vector products derived from the original SWOT classification in capturing river and lake features. Models that incorporate multiple observables consistently outperformed those relying on a single feature. In particular, approaches that include spatial context and surface elevation information yielded the most reliable classifications. We test and show more results at the Science Team meeting.
These findings highlight the importance of refined classification strategies, including the development of machine learning and deep learning methods, for enhancing the hydrological value of SWOT data products.