Exploring inductive transfer approaches for machine learning based rainfall-runoff modelling

Abstract

Machine learning has become an increasingly important approach for rainfall–runoff prediction, where accurate modelling is often constrained by limited training data and the need for robust generalisation. Inductive transfer techniques offer a promising means of improving model performance by leveraging knowledge across related tasks. We investigate three forms of inductive transfer for rainfall–runoff modelling: data augmentation, task augmentation, and additional input. These techniques are implemented within Long Short-Term Memory (LSTM) networks using catchment attributes from 1981 to 2020. Experiments use catchments from the Caravan, CAMELS-US, and Caravan-GRDC datasets. Performance is assessed with Nash-Sutcliffe Efficiency (NSE), Pearson's correlation coefficient (r), Kling–Gupta Efficiency (KGE), Fractional High–Flow Bias (FHV), and Fractional Low-Flow Bias (FLV). Data augmentation improves the median NSE by up to 3% in low-data regimes, whereas its benefits become marginal when larger training datasets are available. Across the global-scale evaluation, inductive transfer techniques achieved median NSE values between 0.73 and 0.87, consistently outperforming the baseline LSTM while improving correlation, overall hydrological performance, and flow bias metrics. These findings demonstrate that inductive transfer is an effective strategy for enhancing rainfall–runoff prediction, particularly in data-scarce settings, while task augmentation and additional input provide the most consistent and robust gains across diverse hydrological conditions.