Frederik Kratzert, Daniel Klotz, Sepp Hochreiter, and Grey Nearing
The study analyses the accuracy of rainfall-runoff predictions of a deep learning model based on the the Long Short-Term Memory (LSTM) network. The model was trained on data for basins in the continental US from the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset. The dataset contains basin-averaged daily meteorological products derived from three gridded data sources: DayMet, Maurer, and NLDAS. The authors trained an ensemble of LSTM models with randomly initialized weights and squared-error loss function. The inputs contained meteorological products from either one, or a combination of two, or all three data sources (DayMet, Maurer, NLDAS). The analysis was performed for a subset of 531 basins from the CAMELS dataset; comparison with a family of traditional hydrological models, including SAC-SMA, was performed for 447 basins of the subset. The analysis employed a range of performance metrics, comparisons between the observed and simulated hydrological signatures, and error characteristics. The results demonstrate that combining multiple meteorological products as inputs for the LSTM model significantly improves accuracy compared to using individual sources, and generally provides superior performance compared to traditional hydrological models.
The code is available on GitHub.
Hydrology and Earth System Sciences, 25, 5, 2685-2703, 2021-05-20.