Predicting Streamflow Duration From Crowd-Sourced Flow Observations

Abstract

Streamflow duration is important for aquatic ecosystems and assigning stream protection status. This study predicts streamflow duration, represented as the fraction of time with flow each year, using a combination of sensor data and crowd-sourced visual observations for a study area in northern Colorado, USA. We used 11 stream stage sensors and 177 visual monitoring points to examine how frequently streams should be sampled to compute flow fractions accurately. This showed that the number of visual observations needed to compute accurate flow fractions increases with decreasing flow duration. We then developed random forest models to predict mean annual flow fractions using climate, topographic, and land cover predictors and found that snow persistence, summer precipitation, and drainage area were important predictors. Model performance was best when using sites with ≥10 visual observations. Our model predicts that almost all (98%) of streams in the study region are non-perennial, about 10% more than the amount of non-perennial streams in the National Hydrography Dataset. Stream type maps are sensitive to the time period of data collection and to thresholds used to represent perennial versus non-perennial flow. To improve maps of non-perennial streams, we recommend moving beyond categorical classification of streams to a continuous variable like flow fraction. These efforts can be best supported with frequent observations in time that span streams with a wide range of flow fractions and drainage area attributes.