Transfer learning using the global Caravan dataset for developing a local river streamflow prediction model

Effective water resource and flood risk management depends on reliable streamflow forecasting. However, the accuracy of such forecasts is often limited by sparse monitoring networks and insufficient historical data. To address this issue, we explore the potential of a multi-basin training approach using the global Caravan hydrological dataset to improve local streamflow forecasting. As a case study, we focus on the Uba River basin in East Kazakhstan. The developed models are evaluated against two baselines: GR4J hydrological model and an LSTM model trained exclusively on local data. Results indicate that our approach enhances forecasting accuracy and outperforms the baseline models, with the best model achieving Nash-Sutcliffe efficiency value of 0.8187 compared to 0.72 of GR4J and 0.7602 of LSTM trained exclusively on local data. These findings indicate that multi-basin training with global datasets can enhance local streamflow forecasting in data-scarce regions.