Deep learning estimation of river discharge based on satellite observations in Mediterranean catchments

Satellite observations have frequently been used for river discharge estimation, particularly in ungauged catchments. The largest challenge for producing continuous time series of river discharge, e.g. with daily time steps, is typically the sporadic nature of satellite observations. Various methods, including spatio-temporal densification of satellite-derived water levels along river networks, have been proposed to address this issue. However, these estimates often suffer from high uncertainties.

Here, we present a novel approach, using both satellite-derived water levels (SWL) and reflectance indices (SRI) to estimate river discharge across 46 river stations in the Mediterranean region. We utilize Long Short-Term Memory (LSTM), known for their efficiency in modeling complex temporal relationships. While LSTM models have been widely applied in rainfall-runoff modeling within the hydrology community, few studies have explored satellite-derived river states as inputs due to their uncertainties and temporal discontinuities.

Gap filling was necessary for SWL and SRI datasets, originally available at intervals ranging from roughly 5 to 30 days. This was accomplished based on freely available discharge from the European Flood Awareness System (EFAS). For each catchment, we compiled daily dynamic variables. Besides the gap-filled SWL and SRI data, this included observed river discharge, as well as precipitation, temperature and potential evapotranspiration from global datasets.

For benchmarking purposes, we set up and calibrated lumped hydrological models for the same 46 catchments, using the same climate data as forcing. Results show that LSTM models outperformed lumped hydrological models in many catchments when using only climate variables as inputs, i.e. when being informed by the same dynamic data as the lumped rainfall-runoff models. The performance of LSTM models can be further improved with the inclusion of SRI and SWL. Shapley Additive Explanations (SHAP) analysis indicated that while climate variables are the most informative for discharge estimation, SRI and SWL also contribute significantly, but varying across individual stations.

The method integrates satellite-derived river states for improved river discharge estimation, while still allowing ingestion of climate input data. This goes beyond conventional hydrological models being forced by climate data only, or also existing densification algorithms for SWL, only using satellite observations.