Characterizing the influence of remotely sensed wetland and lake water storage on discharge using LSTM models

Globally, many wetlands and lakes are at risk for further loss, which can amplify downstream consequences of flood and drought events. We derived remotely sensed based time series of surface water storage (SW_storage) to determine when and where accounting for SW_storage dynamics improves predictions of river discharge. We trained four long short-term memory (LSTM) models, that differed in their inclusion of storage data and catchment characteristics, to simulate daily river discharge (2016–2023) for select watersheds across the conterminous United States. Adding SW_storage to a meteorology-only or meteorology-and-catchment characteristics model improved upon model Nash-Sutcliffe efficiency (NSE) in 80.6% of the watersheds. Residuals during low-flow (Q70) events decreased by 47.6% when adding storage to meteorological data. Improvements were most consistent in ecoregions with a greater abundance of non-floodplain lakes and wetlands. This effort represents the first exploration to train a multi-watershed LSTM on landscape-scale remotely sensed time series of SW_storage.