Assessing the transferability of LSTM-based streamflow models under varying source basin diversity and target data availability (Mangla Basin, Pakistan)

This study focuses on the Mangla Basin in Pakistan, a data-scarce catchment with limited streamflow observations, where reliable hydrological modeling is challenging. We evaluate the potential of transfer learning (TL) to improve streamflow prediction by leveraging pretrained long short-term memory (LSTM) models developed using large-sample datasets, including CAMELS-US (531 basins) and Caravan (∼5000 basins). Model performance was assessed using progressively increasing target-basin training data, corresponding to approximately 2–10 years (2000–2014), while validation and testing periods were fixed. Results show that under severely limited training data (20%), both local and TL models perform poorly (validation NSE ≈ 0.1–0.3), reflecting insufficient hydrological learning. As training data increases, TL models improve substantially, achieving validation NSE values of 0.89 for CAMELS-US and 0.87 for Caravan-based models at 80% training length, consistently outperforming the local model. At full training length, TL performance declines slightly but remains marginally superior, suggesting local models adapt more to basin-specific dynamics with additional data. Across all scenarios, negative FHV values indicate systematic underestimation of high flows under data scarcity. Overall, findings demonstrate that transfer learning enhances streamflow prediction in data-scarce basins, although its relative advantage diminishes as local data availability increases. • Transfer learning improved streamflow prediction in data-scarce basins. • Extreme data scarcity limits effective model transferability. • Using more basins in source-model training does not significantly enhance transfer learning benefits. • Study provides insights for model transferability across diverse basins.