Daily River Discharge Prediction Using LSTM with Satellite Data: A Case Study of the Upper Kampar Watershed, Indonesia

Abstract

River discharge forecasting is crucial for flood mitigation and water resources management in tropical watersheds with scarce observational data. In this study, the multivariate LSTM model is developed to predict the daily river discharge in the Upper Kampar Watershed, Indonesia by considering the CHIRPS precipitation, ERA5 meteorological variables, and GloFAS daily discharge estimates. Data for the period 2015-2024 were aggregated for thirteen administrative sub-district units within the watershed. These administrative units were used as spatial reporting units, not physical sub-basins. Data were normalized in min-max and transformed into seven-day sliding windows and divided chronologically into training (80%) and testing (20%) datasets. The model performance was evaluated by RMSE, MAPE and NSE. The RMSE values of the LSTM model ranged between 0.69 and 18.44 m 3 /s, the MAPE values ranged between 7.61% and 11.44%, and the NSE values ranged between 0.89 and 0.97, which confirmed a good agreement between the predicted and reference discharge. The results indicate that the use of CHIRPS, ERA5 and GloFAS data with an LSTM model constitutes a promising solution for river discharge forecasting in data-scarce tropical watersheds.