Machine learning-based hydrograph modeling with LSTM: A case study in the Jatigede Reservoir Catchment, Indonesia

Predicting and comprehending river discharge is essential to Earth system science in order to manage water resources and preserve ecological sustainability. In data-scarce tropical regions, traditional hydrological models often face challenges due to limited observations and high variability in rainfall-runoff responses. This study explores the application of Long Short-Term Memory (LSTM) networks for hydrograph modeling in the Jatigede Reservoir Catchment, Indonesia, where satellite data is becoming more and more accessible, but gauging infrastructure is scarce. Utilizing 12 years of historical data, an LSTM model was developed to capture the complex non-linear dynamics between rainfall and runoff. The model successfully reproduced overall hydrograph patterns, obtaining a Nash-Sutcliffe Efficiency (NSE) of 0.60 and a Root Mean Squared Error (RMSE) of 12.16, while limitations were observed in simulating extreme hydrological events, primarily due to a lack of historical extreme event data. Sensitivity analysis revealed a significant decline in model performance when applied to a partial dataset, highlighting the importance of data representativeness in model calibration. While the LSTM model shows potential for surrogate hydrograph modeling in data-scarce regions, its ability to predict extreme events remains constrained. Future research should focus on incorporating additional extreme event data and enhancing model generalizability across diverse hydrological conditions.