Transfer Learning for Streamflow Modelling Among Sub-Basins of the Brazilian Semi-Arid Region

Le résumé fourni par la source

Streamflow forecasting is an essential component of effective water resources management, particularly in regions highly vulnerable to extreme hydroclimatic events, such as the Brazilian semi-arid region, which is characterized by pronounced spatial and temporal variability of precipitation, frequent droughts, and occasional flood events. The scarcity, irregularity, and limited duration of hydrological data in many watersheds of this region pose significant challenges to traditional hydrological modeling approaches, restricting the ability to make informed decisions in water resources planning and operational management. In recent years, machine learning–based models, particularly Long Short-Term Memory (LSTM) recurrent neural networks, have shown considerable potential for streamflow modelling due to their ability to capture complex nonlinear relationships and long-term temporal dependencies between precipitation, catchment storage, and runoff generation processes. However, the modelling performance is highly dependent on the availability of extensive and continuous historical records, which limits their direct applicability in data-scarce watersheds. In this context, transfer learning has emerged as a promising strategy to overcome these limitations by enabling the transfer of knowledge learned in well-monitored source sub-basins to improve predictions in target watersheds with limited data availability. This study aims to evaluate the transferability of deep learning models for streamflow modelling among watersheds of the Brazilian semi-arid region, considering different scenarios of data availability. The study also seeks to identify the main physical and hydrological parameters that influence both the performance and transferability of the models. LSTM models were initially pre-trained on watersheds with longer historical records and subsequently fine-tuned for watersheds with varying levels of available local data. Performance evaluation, conducted using widely adopted hydrological metrics, demonstrated that knowledge transfer is effective, allowing significant gains in predictive accuracy even when local datasets are limited. Furthermore, it was observed that certain hydrological and physiographic attributes exert a direct influence on the models’ ability to generalize to new basins. The application of eXplainable Artificial Intelligence (XAI) techniques further reinforced the physical consistency of the streamflow modelling, enhancing both interpretability and reliability of the results. Overall, the use of transfer learning proved to be a highly promising strategy for improving hydrological modelling in data-scarce semi-arid regions, reducing dependence on long-term monitoring, supporting more effective water resources management, and contributing to risk mitigation and sustainability in these vulnerable environments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.