Globally, many waterbodies and floodplains have been lost, degraded, or are at risk for further loss, which may have unintended consequences for rivers, including exacerbating flood and drought conditions. We explored how including surface water storage time series in deep learning models influences our ability to predict river discharge. We utilized Sentinel-1 and Sentinel-2 algorithms to generate time series of surface water extent. Surface water extent (m2) was converted to storage (m3) using topographic estimates of depression probability and depth. These surface water storage estimates were then tested with meteorological data and catchment characteristics in four Long Short-Term Memory (LSTM) models, each containing a different combination of variable groups, to simulate daily river discharge (2016-2023) for 72 watersheds across the conterminous United States.