20 Years of Daily River Discharge Estimation by Using Long Short-Term Memory

Continuous monitoring of river discharge time series is essential for climate applications; however, it remains limited by sparse ground-truth measurements. As a result, there is an increasing demand for river discharge estimation based on satellite-derived observations. Nevertheless, generating daily discharge products remains challenging due to the irregular temporal resolution, which are not complementary for producing temporally continuous time series. Within the framework of the ESA River Discharge Climate Change Initiative (RD-CCI), this study addresses this limitation by producing daily river discharge estimates using Long Short-Term Memory (LSTM) networks that integrate multi-mission optical reflectance data and multi-mission altimetry-derived water level observations.

A major challenge in combining heterogeneous satellite missions is the irregular temporal sampling, which conflicts with the requirement of LSTM models for synchronized and regularly spaced input sequences. To address this issue, Akima interpolation was applied over short consecutive periods to harmonize temporal gaps across input features while preserving natural transitions in the time series. This approach significantly improved data continuity without introducing excessive artificial smoothing.

The LSTM model was implemented using a sliding window scheme of past time inputs to predict the one day ahead discharge value, and compared against other combined river discharge products available from the CEDA catalog (https://catalogue.ceda.ac.uk/uuid/dbba9cfe8d104648b19e39f4c2da1a27/). Input variables include reflectance data from multiple optical missions (Landsat 5, Landsat 7, Landsat 8, Landsat 9, Sentinel 2 Level-1C, Sentinel 3 OLCI, and MODIS on TERRA and AQUA) with orthometric heights obtained from multi-mission altimetry dataset from multiple missions (ERS-1, ERS-2, ENVISAT, Topex/Poseidon, Jason-1, Jason-2, Jason-3, Saral, Sentinel-3A and B, Sentinel-6A).

The LSTM approach was implemented across some diverse river basins, including the Amazon, Colville, Congo, Garonne, Lena, Limpopo, Mackenzie, Maroni, Mississippi, Niger, Ob, and Po rivers to produce daily-based river discharge estimation. Results across representative basins show Nash - Sutcliffe Efficiency values ranging from 0.11 in the Lena (Kyusur station, polar region) to 0.92 in the Amazon (Obidos station, tropical region). Kling–Gupta Efficiency varies between 0.22 for the Limpopo (Beithbrug station, arid region) and 0.95 for the Amazon, while relative Root Mean Square Error ranges from 288 % in arid basins to as low as 9 %in tropical regions. Overall, the results demonstrate that the LSTM model effectively captures the temporal dynamics of river discharge across diverse hydroclimatic regimes.