Estimation of the prevalence of non-perennial rivers and streams in anthropogenically altered river basins by random Forest modeling: A case study for the Yellow River basin

• Actual prevalence of non-perennial reaches can be estimated by Random Forests. • Human impact predictors need to be included in anthropogenically altered basins. • Available global-scale and local data for predictors can be easily combined. • Our approach can be applied to other basins with daily streamflow observations. Whether or not rivers and streams sometimes stop flowing affects biodiversity and ecosystem services. Up to now, large-scale estimation of the prevalence of non-perennial reaches (NPRs) was mostly limited to the prediction of naturalized streamflow intermittence. Using a Random Forest (RF) modeling approach, we present the first high-resolution estimate of the actual prevalence of NPRs in a large river basin where the streamflow is strongly impacted by humans. Using daily time series of observed streamflow at 224 gauging stations, we developed and analyzed four RF variants with different combinations of predictor variables that included seven human impact (HI) predictors to estimate the prevalence of NPRs in the Yellow River basin. The HI predictors quantify the relation of mean annual streamflow to the storage capacity of artificial reservoirs and sectoral water abstractions from groundwater and surface water. Predictor importance and partial dependence plots were derived to understand the relation between the predictor variables and intermittence status. With a balanced accuracy (BACC) of 0.89, a sensitivity of 0.85, and a specificity of 0.93, the RF variant with all 19 predictors achieved the highest performance. A total of 35 % of the 38,871 reaches with a network length of 65,450 km were identified as NPRs. We found that it was necessary to include HI predictors in the modeling of NPR prevalence in the Yellow River basin as the variant without them obtained a balanced accuracy of only 0.78, overestimating NPRs as compared to the observations and resulting in an estimated 51 % of NPRs in the basin. The other two model variants with HI predictors but different other predictors lead to NPR fractions of 27 % and 45 %, indicating a very high sensitivity of estimated NPR prevalence to the selection of predictors. Our study shows how globally available predictors can be combined with local daily data streamflow observations and local predictors to estimate NPR prevalence in river basins that are strongly affected by humans.