A hybrid DeepGR4J and regionalization framework for lake water balance assessment in data-scarce regions: Application to Lake Tana, Ethiopia

Study region Lake Tana, the largest freshwater lake in Ethiopia and the headwater of the Blue Nile River, is located in the northwestern Ethiopian Highlands and is vital for water supply, hydropower, irrigation, and ecosystem services. The basin has strong rainfall seasonality, complex topography, and many ungauged catchments, making hydrological assessment difficult due to limited observations. Study focus This study develops a hybrid hydrological framework integrating the DeepGR4J rainfall-runoff model with regression-based regionalization to improve lake water balance estimation under data-scarce conditions. DeepGR4J combines the conceptual GR4J model with LSTM-based error correction to enhance runoff simulation, while regionalization transfers parameters from gauged to ungauged watersheds using physiographic and remote sensing descriptors. The framework is applied for 2000–2020 using CHIRPS precipitation, ERA5 temperature, and Hargreaves evapotranspiration estimates. New hydrological insights for the region Results show a mean annual inflow of 12.61 km³ yr⁻¹ and a combined water-balance outflow term of 12.67 km³ yr⁻¹, resulting in a net storage decline of −0.20 km³ yr⁻¹ and a residual imbalance of 0.06 km³ yr⁻¹. The model reproduces observed lake levels with high accuracy (R² = 0.94, NSE = 0.94, RMSE = 0.18 m). Findings highlight the strong contribution of ungauged catchments and demonstrate that hybrid conceptual-deep learning models with regionalization improve water balance estimation in data-scarce lake basins.