Parsimonious statistical learning models for low-flow estimation

Abstract. Statistical learning methods offer a promising approach for low-flow regionalization. We examine seven statistical models (Lasso, linear, and nonlinear-model-based boosting, sparse partial least squares, principal component regression, random forest, support vector regression) the prediction of winter summer low flow based on hydrologically diverse dataset 260 catchments in Austria. In order to produce models, we adapt recursive feature elimination variable preselection propose using three different ranking (conditional Lasso, linear model-based boosting) each models. Results are evaluated characteristic Q95 (Pr(Q>Q95)=0.95) standardized by catchment area repeated nested cross-validation scheme. found generally high accuracy (RCV2 0.66 0.7) 0.83 0.86). The perform similarly or slightly better than top-kriging model that constitutes current benchmark study area. best-performing regression (winter) nonlinear boosting (summer), but exhibit similar accuracy. use can significantly reduce complexity all with only small loss performance. so-obtained more parsimonious thus easier interpret robust when predicting at ungauged sites. A direct comparison reveals processes be sufficiently captured so there is no need complex add effects. When performing regionalization seasonal climate, temporal stratification into flows was shown increase predictive performance offering an alternative grouping recommended otherwise.