Regionalised Fine-Tuning of LSTMs for Streamflow Prediction in Ungauged Catchments

Accurate streamflow prediction in ungauged basins remains one of the central challenges in hydrology. While deep learning models, particularly long short-term memory (LSTM) networks, have demonstrated strong performance in large-sample rainfall-runoff modelling, current continental-scale implementations are spatially agnostic. They rely on static catchment attributes to encode regional differences, but do not explicitly exploit spatially structured hydrological similarity, limiting their ability to represent local hydro-climatic and geological controls in ungauged settings. In contrast, traditional regionalisation approaches explicitly transfer information from spatially or hydrologically similar donor catchments.

We introduce Regionalised Fine-Tuning (ReFT), a method that integrates donor-based regionalisation LSTM by spatially weighting the fine-tuning of a pretrained continental model. Starting from a global model trained across all gauged basins, ReFT adapts the model for each ungauged target catchment by minimising a distance-weighted loss over neighbouring donor catchments. This transforms spatially agnostic LSTMs into spatially aware regional specialists while preserving the broad process representations learned from large-sample training.

The method is evaluated using 222 Australian catchments from CAMELS-AUS under a spatial out-of-sample cross-validation framework. Two LSTM variants are tested: a climate-only model and a hybrid model that also uses runoff from the Australian Water Resources Assessment - Landscape (AWRA-L) land surface model as a dynamic predictor. ReFT-LSTMs are compared against a continental LSTM without fine-tuning, a regionalised GR4J conceptual model, and AWRA-L.

Results show that ReFT produces systematic improvements across most of the Nash–Sutcliffe Efficiency distribution, with median and upper-quantile skill exceeding both the continental LSTM and traditional benchmarks. In most catchments, ReFT-LSTMs outperform both AWRA-L and regionalised GR4J, demonstrating that spatially weighted fine-tuning can effectively substitute for local calibration data in ungauged basins. Remaining weaknesses are confined to specific challenging regions, where conceptual model constraints remain advantageous. Overall, ReFT provides a principled pathway to reconcile large-sample deep learning with regional hydrological realism for PUB applications.