Enhancing the performance of machine learning models for simulation of river flows in data-scarce catchments using bootstrapping and physics-based data augmentation

Accurate prediction of river flows in data-scarce regions is essential for sustainable water resources management under a changing climate, yet limited observations constrain the performance and application of data-driven models. This study systematically evaluates the effectiveness of statistical and physics-based data augmentation strategies for improving the performance of Machine Learning (ML) models in river flow simulation. The evaluation was conducted in two data-scarce Sub-Saharan African catchments with contrasting climates. To address data scarcity, bootstrapping and physics-based data augmentation were applied to expand the training datasets of the Feed Forward Neural Network (FFNN) and a Long Short-Term Memory (LSTM) model. The resulting model performances were benchmarked against the physically based Hydrologic Engineering Center Hydrologic Modeling System (HEC-HMS) model. Comparison of standalone ML models with HEC-HMS reveals data-dependent performance, with HEC-HMS outperforming ML models on very limited datasets. In contrast, ML models achieve superior performance as data availability increases. Integration of ML models with data augmentation approaches indicated improved performance of FFNN and LSTM, particularly when limited training data was available. Limited or comparable performance was observed when longer training datasets were used. The results further indicated that applied data augmentation approaches were independent of model architecture and catchment hydroclimatic conditions. These findings provide practical guidance for water resource engineers and modellers on the model-specific and data-dependent applicability of data augmentation strategies for river flow modelling in data-scarce regions.