Improving a streamflow regression model for Wisconsin streams

Streamflows derived from hydrological models are widely used in decision-making processes in a broad array of natural resources applications. With an increase in computational power and data availability, data-driven modeling methods are becoming more powerful and popular. While it is well-recognized that reasonable model uncertainty is important to support good decision-making, there remain substantial challenges in quantifying uncertainty in hydrological models. One challenge is an inequality in data availability. While large amounts of data are available for well-monitored streams, the vast majority of streams globally are ungauged, with very limited or no streamflow monitoring. In this study, I evaluated the accuracy of a mixed-effects model for streamflow (flow-duration curves) across the state of Wisconsin, the Natural Community Model (NCM), trained on continuously monitored streamflow stations. The NCM is used as the basis for scientific studies and management decisions in Wisconsin, but uncertainty in the NCM has not been quantified yet, and performance has not been assessed formally except at continuously monitored streamflow stations. There are about 4,000 streamflow monitoring stations in Wisconsin, but about 3,500 have fewer than 5 sporadic streamflow measurements. I used an index gauge approach to estimate long-term streamflow percentiles (with uncertainty) from short-term or sporadic streamflow monitoring. I then used these estimates to estimate a flow-duration curve for each short-term or sporadic streamflow station (with uncertainty). These flow-duration targets formed the basis for an assessment of NCM accuracy in ungauged streams. I developed a random forest model for NCM error that provides a qualitative understanding of sources of error in the NCM as well as a quantitative way to correct the NCM using information from the sporadic/short-term streamflow stations that could not be included in the original NCM training set. The updated NCM has significantly reduced error, and I defined a reasonable level of uncertainty to be used with the updated NCM in decision-making and research applications.