Colin Gleason (University of Massachusetts Amherst, United States); Konstantinos Andreadis (University of Massachusetts Amherst, United States)

SWOT provides the opportunity to improve hydrologic models of river discharge by incorporating real-time data on surface water fluxes and storages. Traditionally, we would need to either first invert discharge from SWOT observations (i.e, from the DAWG) and then assimilate these estimates into a model or have a model capable of representing water surface heights and widths as the global state variable. The latter approach is rare and difficult globally, so the former is expected. However, the DAWG’s discharge inversions necessarily add error to SWOT’s high quality information. If our goal is discharge estimation from a model, it is best to use SWOT’s primary measurements together with meteorological data. Now, more than two years into the mission, we have enough data to try a nontraditional approach: train deep learning models that can use the primary SWOT observations directly. We combine a Long Short-Term Memory model (LSTM; which has been shown to be very effective in global, large-scale hydrologic modeling) with a cross-attention transformer (which is commonly used in multi-modal and language translation models) to take full advantage of SWOT data. Preliminary results on 130 basins unseen during training shows improvement in median Kling Gupta Efficiency from 0.05 to 0.34 as a result of including SWOT data in the model. We hypothesize that these improvements will not be uniform across locations, but correlate with the amount of human modification to river systems (i.e. where climate forcings alone are not sufficient to predict discharge). In addition to improving prediction accuracy of models, we will investigate the model’s learned use of measurement quality flags to evaluate how much information is lost by these different sources of error.