Evaluation of machine learning approaches for predicting streamflow metrics across the conterminous United States
Few regional or national scale studies have evaluated machine learning approaches for predicting streamflow metrics at ungaged locations. Most such studies are limited by the number of dimensions of the streamflow regime investigated. This study, in contrast, provides a comprehensive evaluation of the streamflow regime based on three widely available machine learning approaches (support vector regression, random forest, and cubist regression) and on multiple linear regression to predict 106 natural streamflow metrics at ungaged locations. This evaluation is done for 545 streamgages across the northwest United States for recurrence-interval flood metrics and for 1,851 sites in the conterminous United States for non-flood metrics. The results indicate that for flood metrics, predictions by cubist regression and support vector regressions have substantially less error than the other approaches. For all the remaining streamflow metrics, random forest models outperform the other methods.
|Evaluation of machine learning approaches for predicting streamflow metrics across the conterminous United States
|Ken Eng, David M. Wolock
|USGS Numbered Series
|Scientific Investigations Report
|USGS Publications Warehouse
|WMA - Integrated Modeling and Prediction Division