Authors: Amin Aghababaei, Xueyi Li, Norm Jones, and Gustavious Williams – Brigham Young University; Eniola Webster-Esho, Prabhakar Clement – The University of Alabama; Ryan Van Der Heijden, Donna Rizzo – University of Vermont
Title: Nationwide Identification of Baseflow Dominant Periods: Integrating Manual Expertise into Machine Learning
Abstract: Hydrograph separation is a challenging topic in hydrology. Several separation models, based on graphical or digital filters methods, have been introduced to separate total flow into quick flow or baseflow components. However, while these models generally work well for flood conditions, they are not designed for base-flow dominant conditions. Base-flow dominant conditions are influenced by groundwater and drought, and processes are often regional. The concept of baseflow dominant flows in low-flow conditions is different from separating peak flow from base flow. These low-flow conditions have unique characteristics that require specialized modeling approaches.
To address this new condition, we selected about 200 gauges, evenly distributed across the United States streams and manually identified and labeled baseflow dominant periods. We visually inspected the streamflow hydrographs and identified periods of low flow with relatively slow changes that are likely attributed to baseflow. This resulted in a comprehensive dataset, spanning diverse geographical regions, with baseflow periods and values labeled that can be used for developing a more accurate and generalized model. We used these labeled data, to train a supervised classification model with an accuracy of 91 percent. The high accuracy of this model demonstrates its identifying baseflow dominant periods that can be used to study the interaction of groundwater and flow, or the impact of various drought processes on stream baseflow.