Estimation of suspended sediment concentration and yield using linear models, random forests and quantile regression forests

For sediment yield estimation, intermittent measurements of suspended sediment concentration (SSC) have to be interpolated to derive a continuous sedigraph. Traditionally, sediment rating curves (SRCs) based on univariate linear regression of discharge and SSC (or the logarithms thereof) are used but alternative approaches (e.g. fuzzy logic, artificial neural networks, etc.) exist. This paper presents a comparison of the applicability of traditional SRCs, generalized linear models (GLMs) and non-parametric regression using Random Forests (RF) and Quantile Regression Forests (QRF) applied to a dataset of SSC obtained for four subcatchments (0.08, 41, 145 and 445 km²) in the Central Spanish Pyrenees. The observed SSCs are highly variable and range over six orders of magnitude. For these data, traditional SRCs performed inadequately due to the over-simplification of relating SSC solely to discharge. Instead, the multitude of acting processes required more flexibility to model these nonlinear relationships. Thus, alternative advanced machine learning techniques that have been successfully applied in other disciplines were tested. GLMs provide the option of including other relevant process variables (e.g. rainfall intensities and temporal information) but require the selection of the most appropriate predictors. For the given datasets, the investigated variable selection methods produced inconsistent results. All proposed GLMs showed an inferior performance, whereas RF and QRF proved to be very robust and performed favourably for reproducing sediment dynamics. QRF additionally provides estimates on the accuracy of the predictions and thus allows the assessment of uncertainties in the estimated sediment yield that is not commonly found in other methods. The capabilities of RF and QRF concerning the interpretation of predictor effects are also outlined.