Transfer Learning for Hydrological Modelling and XAI-Based Physical Consistency Assessment in Reconstructing Streamflow Time Series in Data-Scarce Regions

Accurate streamflow monitoring is essential for water resources management, yet many Brazilian watersheds lack sufficiently long historical records to support effective decision-making. This challenge is particularly critical in the Metropolitan Region of Belo Horizonte (RMBH), which depends on major reservoirs located within its territory – such as Rio Manso, Serra Azul, Vargem das Flores, and the Ibirité (REGAP) reservoir – for industrial and domestic water supply. Several of these strategic systems suffer from limited or inconsistent hydrological monitoring, complicating operational planning, increasing the risk of water shortages and of compromising reservoirs flow outcome capacity. Transfer Learning (TL) with Long Short-Term Memory (LSTM) networks emerges as a promising strategy to overcome this limitation, enabling the development of hydrological models in watersheds with little or no historical data. This study investigates the application of TL to enhance daily streamflow prediction in data-scarce basins of the Metropolitan Region of Belo Horizonte (RMBH), while assessing the optimal length of local streamflow records required to improve hydrological modelling through fine-tuning of a regional TL model. For this, 23 watersheds with similar hydrological behaviour and geomorphological characteristics were previously selected in the RMBH to evaluate the feasibility of reconstructing streamflow time series in data-scarce regions. Satellite-derived products and reanalysis datasets were employed as inputs to overcome limitations in hydrometeorological data availability. Furthermore, eXplainable Artificial Intelligence (XAI) methods are employed to explore the physical feasibility of knowledge transfer, with the potential to identify which watershed attributes – such as drainage area, elevation, soil-moisture dynamics, land-use composition, and climatic seasonality – most strongly influence whether hydrological behaviour learned in source basins can be meaningfully transferred to target basins. Significant performance gains can be achieved with only one to two years of local data, allowing accurate models to be developed rapidly even in newly monitored watersheds. This improves considerably the decision-making in data scarce regions, primarily those ones with some water conflicts. XAI analyses confirmed the physical soundness of the predictions, supporting more reliable streamflow reconstruction. However, further methodological improvements are required, as some watersheds were unable to benefit from transfer learning. Overall, TL represents a powerful direction for streamflow modelling in regions with limited monitoring, while XAI provides a framework to understand the physical consistency of the transferred knowledge and to determine the minimum monitoring effort required to build reliable local models.