Rainfall-Runoff Simulation in Ungauged Tributary Streams Using Drainage Area Ratio-Based Multivariate Adaptive Regression Spline and Random Forest Hybrid Models

Abstract

For various reasons, it is not always possible to obtain adequate and reliable long-term streamflow records in a river basin. It is known that streamflow records are even shorter when the stations located on tributary channels are of the interest. Hence, it is necessary to develop dependable streamflow estimation models for the tributary streams that play a key role in the micro-hydrology of the basin. In this study, rainfall-runoff models are developed to estimate the daily streamflow in ungauged tributary streams. Precipitation and streamflow in the most similar gauging station on the main channel and lagged values up to three days before on the same tributary station are used as the input variables of the allocated models. To select the most similar gauging station, a similarity index criterion is developed and used in the analysis. Then, two scenarios based on the streamflow or the corresponding set of direct runoff and base-flow in the same station are used. By applying multivariate adaptive regression spline (MARS) and random forest (RF) methods, several rainfall-runoff models are developed and evaluated based on determination coefficient, mean absolute percentage error, root mean square error, relative peak flow, scatter plot and time series plot. Alternatively, the MARS and RF models are combined with a drainage area ratio (DAR) model to produce the DAR-MARS and DAR-RF models. It is concluded that the direct runoff in the mainstream is more effective on the streamflow of the tributary station, while the integration of models with DAR enhanced the capabilities of the models in estimation of extreme values in the streamflow time series.