A data-driven analysis of frequent patterns and variable importance for streamflow trend attribution

Abstract

Identifying key driving forces for streamflow variation is essential for improving sustainable water resource management in terms of understanding how changes in the watershed translate to changes in streamflow. In this study, the relationships between trends in total annual streamflow and trends in watershed characteristics across the contiguous U.S. during 1981-2016 are investigated with data from 2,621 USGS gages. The regions of homogeneous hydrologic change, i.e. watersheds that are undergoing similar statistically significant streamflow trends, are delineated and frequent pattern mining (i.e. Apriori algorithm) and variable importance (i.e. Random Forest) are used to derive the key driving forces for these regions. As expected, the trends in streamflow are highly associated with the trends of precipitation. In contrast, the influences of anthropogenic factors vary substantially across regions. Particularly, the influence of water use change tends to be significant in the regions dominated by agricultural land, e.g. Dakotas. The importance of land use change is highlighted in the regions with relatively large forest coverage, e.g. Northeast. However, these important identified water use changes are not frequently associated with the increasing streamflow in sub-regions, e.g. Great Lakes, and thus the significance of the water use impacts are site-specific. Therefore, the changes in climate and land use are frequently and importantly identified together in the sub-regions with increasing streamflow, which can be collectively used to discover the major causes of the streamflow trends in those regions. Although the impacts of changing water use are highlighted in the Southwest, climate trends are primarily responsible for the decreasing streamflow.