Improving predictions of hydrological low-flow indices in ungaged basins using machine learning

We compare the ability of eight machine-learning models (elastic net, gradient boosting, kernel-k-nearest neighbors, two variants of support vector machines, M5-cubist, random forest, and a meta-learning ensemble M5-cubist model) and four baseline models (ordinary kriging, a unit area discharge model, and two variants of censored regression) to generate estimates of the annual minimum 7-day mean streamflow with an annual exceedance probability of 90% (7Q10) at 224 unregulated sites in South Carolina, Georgia, and Alabama, USA. The machine-learning models produced substantially lower cross validation errors compared to the baseline models. The meta-learning M5-cubist model had the lowest root-mean-squared-error of 26.72 cubic feet per second. Partial dependence plots show that 7Q10s are likely moderated by late summer and early fall precipitation and the infiltration capacity of basin soils.