Generalized Split-Sample Test Interpretation Using Rainfall Runoff Information Gain

Abstract

Rainfall-runoff conceptual models are used largely for river discharge prediction, for waterworks design, and as support for water quality assessment. The generalized split-sample test (GSST) recently was recommended to analyze rainfall-runoff models’ performance. Moreover, it was found that parameter transfer may be conditioned to the precipitation conditions of the donor and receiver periods. This study focused on the generalized split-sample test results, and analyzed them in terms of the information gain between rainfall and runoff series. This issue was not considered before in GSST interpretation. Six small to moderate-sized basins (50–500 km2) in northern Tunisia were studied using the daily bucket with a bottom hole (BBH) model and the GSST calibration-validation approach. The mean absolute error and the Nash–Sutcliffe efficiency (NSE) were adopted to quantify model performance. The analysis suggests that the mean information gain (MIG) may be an indicator of the explored hydrological conditions of the assessment periods. In addition, results show that validation periods characterized by high MIG improved robustness, displaying low standard deviation of monthly NSE and enhanced accuracy, as shown by mean monthly NSE. The study of the effects of underlying physiographic factors suggests that the transfer from period to period is likely to be more robust for moderate-size basins than for small basins and that basin steepness tends to decrease the robustness of the transfer.