New Insights in Prediction of Daily River Flow Using SWOT Observations and Machine Learning

Accurate river flow estimation is essential for water resources management, yet continuous in situ observations remain scarce across rivers globally and particularly in Asia and Africa. This study explores the potential of integrating SWOT satellite–derived river width and water surface elevation with limited gauging data to reconstruct river flow of China’s vast river network. The study compared four machine learning models for data integration: Random Forest (RF), XGBoost (XGB), Multi-Layer Perceptron (MLP), and Transformer (TF). A temporally ordered five-fold cross-validation framework was used to evaluate both interpolating and extrapolating performance. Under interpolating mode, RF and XGB effectively reproduced the observed hydrographs, capturing flow variability and extremes. Under extrapolation mode, all models show reduced skill due to short record length, seasonal incompleteness, and zero-flow effects, although neural network models exhibited relatively better performance. These results demonstrate a potential solution for river flow gap filling by using SWOT satellite observations after appropriate data processing, however, the approach requires substantially larger and more diverse training datasets for improving extrapolation performance.