Investigating Deep Learning Knowledge Transfer in Streamflow Prediction From Global to Local Catchment

Abstract Accurate streamflow prediction is critical for flood forecasting and water resource management, particularly in data‐scarce regions. Deep learning models like Long Short‐Term Memory (LSTM) offer a bridge to hydrologic regionalization utilizing climate data and catchment characteristics to improve behavioral insights and constrain predictive uncertainties. Here we evaluate transfer learning (TL) approaches using 441 “donor” basins from regions rich in quality hydrological data (Scotland GB‐SCT; Switzerland CH; and Canada's British Columbia; BC) to pre‐train LSTM runoff models by fine‐tuning in data‐poor areas like CA (36 target basins). Pairing measured streamflow (lagged) records with global climate data (ERA‐5) boosted the explanatory power of LSTM predictions especially in snowmelt and glacier‐influenced basins. A K‐Means clustering algorithm was applied to categorize basins into five hydrologically meaningful Clusters (labeled 1–5) based on catchment attributes. The results show that TL‐LSTM models perform better when pre‐trained using Clusters compared to the locally trained model (NSE = 0.85 and KGE = 0.80). The LSTM model trained on basins Cluster 3 data (LSTM 3 ) most closely resembling those in the target region yielded the most accurate predictions. Fine‐tuning with limited local data substantially improved prediction accuracy in the validation split (blind‐tested and treated as “ungauged”), evidencing that even short‐records of local data can enhance a regional model of hydrological behavior. These results demonstrate that TL‐LSTM can effectively enhance streamflow prediction in data‐scarce regions. These insights advance understanding of cross‐basin generalization and support the development of efficient, scalable modeling strategies for hydrological prediction in regions with limited observational data.