Fine-tuning long short-term memory models for seamless transition in hydrological modelling: From pre-training to post-application

Pre-trained models like FourCastNet, Pangu and GraphCast have gained popularity in the meteorological field. In hydrology, data-driven rainfall-runoff models based on long short-term memory (LSTM) networks have been successfully applied for various purposes. As large-sample hydrological datasets (e.g., Caravan) continue to grow, it is foreseeable that pre-trained models tailored for hydrology will emerge. These pre-trained models have the potential to bypass the need for training data-driven models from scratch, enabling us to focus more swiftly on customized applications. Additionally, they offer opportunities explore model performance in changing environment, which is also a key consideration when using data-driven models in unseen scenarios. However, the hydrological field has seen limited attempts to employ, transfer, and fine-tune pre-trained models. This study aims to explore the possibility of using fine-tuning techniques to achieve a smooth transition of LSTM-based rainfall-runoff models from pre-training to post-application scenarios. By utilizing ERA5-Land reanalysis precipitation data within the Caravan dataset, we calibrated a pre-trained LSTM model for runoff simulation. Subsequently, we transitioned the model to use near-real-time satellite precipitation estimates as the input, targeting satellite-driven predictions. Our results show that fine-tuning parameters lead to improvements in various metrics, including the Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), and hydrological signature metrics such as high and low flows, compared to outcomes without parameter fine-tuning. Specifically, fine-tuning using locally calibrated models enhanced performance in 73.5% of the basins. In contrast, the results of fine-tuning regional models were mixed; while it benefited 55.1% of the basins, it also led to a deterioration in model performance in 44.9% of cases. This study is a pioneering exploration of the adaptability of LSTM models from pre-training to post-application. It also lays the groundwork for future investigations aimed at enhancing the adaptability of data-driven models to the impacts of changing environment. • We explored the feasibility of fine-tuning pre-trained LSTM-based rainfall-runoff models for hydrological applications. • Local fine-tuning enhanced 73.5% of basins, whereas regional fine-tuning benefited only 55.1%. • The quality of precipitation data significantly affects the performance of both pre-trained and fine-tuned models.