Technical note: Does Multiple Basin Training Strategy Guarantee Superior Machine Learning Performance for Streamflow Predictions in Gaged Basins?

Abstract. In recent years, machine learning (ML) has witnessed growing prominence and popularity in hydrological science, offering convenience and ease of use without requiring extensive hydrological expertise or the complexity associated with process-based models. There exists debate regarding optimal training approaches, with some researchers advocating for multi-basin training while questioning the validity of single-basin approaches. This study examines the relationship between training dataset size (number of basins) and model performance. Through comparative analysis, we found that increasing the number of basins for ML training does not necessarily guarantee improved performance of the trained ML model. Specifically, the state-of-the-art global ML (G model) trained by Google with nearly 6,000 global basins underperforms compared to regional ML models trained with hundreds of basins in contiguous US and Great Britain regions for predicting streamflow in both gauged and ungauged basins. Furthermore, we compared the G model with our single-basin (S) ML models, trained for 609 global locations individually, and found that the G model does not consistently outperform S models, as results show S models outperforming the G model in 46 % of case studies. Therefore, the training approach should not be a criterion for judging model validity; instead, the focus should be on the trained model's performance.