Machine learning-enhanced baseflow estimation from digital separation methods: insights from south-western Nigeria

This study evaluates how machine learning (ML) models (Random Forest, XGBoost, KNN, and Linear Regression) can replicate and enhance baseflow estimates generated by the Eckhardt filter, a broadly accepted digital separation method. Employing 40 years of daily streamflow records from six rivers (Ogun, Osun, Owena, Yewa, Oba and Ogbese) in South-Western Nigeria, we contrast ML-derived baseflow with conventional techniques, such as the UKIH and BFI approaches, while setting an Eckhardt-filtered benchmark. The outcomes show that ML models produced sustained high efficiency in replicating the Eckhardt filter output with R2 ranging from 0.988 to 0.995 and lower RMSE across a broad spectrum of hydroclimatic regimes. The feature importance assessment underscored the prevalent importance of lagged discharge, thereby emphasizing the pronounced temporal autocorrelation in baseflow formation. Conventional techniques exhibited inconsistent performance, most conspicuously under highly dynamic flow regimes, whereas ML models generalized satisfactorily across all the catchments. These results highlight the potential of ML techniques for baseflow prediction in tropical, data-limited contexts, where physical in situ observations are limited.