Large-scale baseflow index prediction using hydrological modelling, linear and multilevel regression approaches

Baseflow is critical for water balance budget, water resources management, and environmental evaluation. Prediction of baseflow index (BFI), the ratio of baseflow to total streamflow, has a great significance in unravelling the baseflow characteristics for large scale trajectory. Therefore, this study compares BFI predictive performance derived from a new multilevel regression approach along with two other commonly used approaches: hydrological modelling (SIMHYD, a simplified version of the HYDROLOG model, and Xinanjiang model), and linear regression (traditional linear regression, and alternative traditional regression considers the second-order interaction). The multilevel regression approach does not only group the catchments into the four climate zones (arid, tropics, equiseasonal and winter rainfall), but also considers inter-catchment and inter-climate zone variances. Likewise, calibration and two regionalisation techniques namely spatial proximity and integrated similarity are used to obtain the BFI from hydrological modelling approach. Correspondingly, the traditional linear regression technique estimates BFI establishing linear regressions between catchment attributes and four climate zones. Then, all the three approaches are evaluated against combined average estimation from four well-parameterised baseflow separation methods (Lyne-Hollick (LH), United Kingdom Institute of Hydrology (UKIH), Chapman-Maxwell (CM) and Eckhardt (ECK)) at 596 catchments across Australia for 1980-2012. The findings show that the multilevel regression has greatly improved the performance of BFI prediction in comparison to other methods. In particular, the two calibrated and regionalised hydrological models perform worst in predicting BFI with a Nash-Sutcliffe Efficiency (NSE) of -8.44 and -2.58 along with an absolute percent bias (PBIAS) of 81% and 146% (overestimation of baseflow), respectively. However, the traditional linear regression remains in intermediate position with the NSE of 0.57 and bias of 25. In addition, alternative traditional regression also shows very close proximity. In contrast, the multilevel regression approach shows the best performance with the NSE of 0.75 and bias of 19%. The study also demonstrates that the multilevel regression approach can improve BFI prediction, and shows potential for being used in the prediction of other hydrological signatures in large-scale