Accurate streamflow modeling in data-scarce catchments remains a significant challenge due to the limited availability of historical records. Transfer Learning (TL), increasingly applied in hydrology, leverages knowledge from data-rich catchments (sources) to enhance predictions in data-scarce catchments (targets), providing new possibilities of hydrological predictions. Most existing TL approaches pre-train models on large-scale meteoro-hydrological datasets and show good generalizability across multiple target catchments. However, for a specific target catchment, it remains unclear which source catchments contribute most effectively to the accurate prediction. Including many irrelevant sources may even degrade model performance. In this study, we investigated how source catchment selection affects TL performance by employing similarity-guided strategies based on three key factors, i.e., spatial distance, physical attributes, and flow regime characteristics. Using the CAMELS-GB dataset, we conducted comparative experiments by pre-training the networks with different ranked groups of the source catchments and fine-tuning them on three target catchments representing distinct hydrological environments. The results showed that carefully selected small subsets (fewer than 40, or even as few as 10) of highly similar catchments can achieve comparable or better TL performance than using all 668 available source catchments. All three target catchments yielded better NSE results from source catchments with closer spatial proximity and more consistent flow regimes. The TL performance of physical attribute similarity-based selection varied depending on the attribute combinations, with those related to land cover, climate, and soil properties leading to superior performance. These findings highlight the importance of similarity-guided source selection in hydrological TL. In addition, they demonstrate ways to reduce computational costs while improving modeling accuracy in data-scarce regions.

Streamflow gauging records remain sparse in many regions, limiting hydrological predictions. Transfer Learning (TL), which uses knowledge from data‐rich sources to improve predictions in data‐scarce targets, has emerged as a promising solution. Most TL applications in hydrology pre‐train deep learning (DL) models on large meteo‐hydrological data sets to enhance generalization across diverse catchments. However, this strategy faces two key limitations: (a) TL performance for a specific target may be hindered by irrelevant source inputs and suboptimal fusion of source information; and (b) TL data requirements are demanding. Here, we propose a Multi‐Source Adaptive‐Fusion Transfer Learning (MSAF‐TL) network to address these limitations. MSAF‐TL first extracts independent temporal features from each highly related source catchment and then weights them based on the similarity of their static attributes to those of the target. These weighted features are adaptively fused using a cross‐attention mechanism, strengthening the final target prediction. We test this framework using a large‐sample hydrological data set for Great Britain (CAMELS‐GB). Results show that small sets of highly similar source catchments can match the transferability of large data sets. For a target catchment with only 1 year of training data, MSAF‐TL achieves an NSE of 0.85 using only the top two similar sources, compared with 0.78 when using all 668 CAMELS‐GB catchments. With limited sources, MSAF‐TL consistently outperforms conventional multi‐source TL (MSTL) across varying source quantities and target conditions. This performance gap widens under greater data scarcity, with average NSE improvements over MSTL reaching 22% when only 6 months of target training data are available. Overall, this study demonstrates that MSAF‐TL can achieve more accurate and stable TL performance in data‐scarce regions, and that data requirements can be reduced without compromising performance through source pre‐selection.

Streamflow gauging records remain sparse in many regions, limiting hydrological predictions. Transfer Learning (TL), which uses knowledge from data‑rich sources to improve predictions in data‑scarce targets, has emerged as a promising solution. Most TL applications in hydrology pre-train deep learning (DL) models reply on large meteo‑hydrological datasets to enhance generalization across diverse catchments. However, this strategy faces two key limitations: (i) TL performance for a specific target may be hindered by irrelevant source inputs and suboptimal fusion of source information; and (ii) TL data requirements are demanding. Here, we propose a Multi-source Adaptive-fusion Transfer Learning (MSAF-TL) network to address these limitations. MSAF-TL firstly extracts independent temporal features from each highly-related source catchment and then weights them based on the similarity of their static attributes to those of the target. These weighted features are adaptively fused using a cross-attention mechanism, strengthening the final target prediction. We test this framework using a large-sample hydrological dataset for Great Britain (CAMELS-GB). Results show that small sets of highly similar source catchments can match the transferability of large datasets. For a target catchment with only one year of training data, MSAF‑TL achieves an NSE of 0.85 using only the top two similar sources, compared with 0.78 when using all 668 CAMELS‑GB catchments. With limited sources, MSAF‑TL consistently outperforms conventional multi‑source TL (MSTL) across varying source quantities and target conditions. This performance gap widens under greater data scarcity, with average NSE improvements over MSTL reaching 22% when only six months of target training data are available. Overall, this study demonstrates that MSAF-TL can achieve more accurate and stable TL performance in data-scarce regions, and the data requirements can be reduced without compromising performance through source pre-selection.