Estimating streamflow-recession indexes using automated methods with application to groundwater and surface water interaction

Statistical properties of streamflow recession provide evidence of complex hydrologic processes such as groundwater and surface-water interactions. Bingham (1982, 1986) sought regional definition of generalized connectivity between surface water and groundwater through the computation of persistent streamflow-recession indexes during winter low flows and subsequent regionalization using surficial geology. The streamflow-recession indexes were derived through a hands-on graphical method for selected peak flows from over 200 unregulated U.S. Geological Survey streamgages with at least 20 years of daily-mean record in Tennessee and Alabama. A streamflow recession curve, plotted on semilog graph paper, was created starting at peak streamflow after a precipitation event until the line neared asymptotic with the horizontal axis. The number of days (x-axis) required for streamflow to decrease one log10-cycle (y-axis) was the streamflow-recession index for the station expressed in days per log10-cycle decline in flow. The streamflow-recession indexes determined by Bingham were somewhat subjective because they were derived from visual inspection of two to eight ideal streamflow hydrographs from each streamgage. Boundaries for streamflow-recession index regionalization were determined using streamflow hydrographs, surficial geology, and lithologic contacts. Although streamflow-recession index values have been useful in statistical regionalization studies (Bingham, 1982, 1986; Knight and others, 2012), the subjectivity and time-consuming manual method of the approach has made it problematic to compute streamflow-recession indexes for areas outside of Alabama and Tennessee and using current (2018) streamflow data. For this study, the computationally derived streamflow-recession index is referred to as the “geologic factor” or Gfactor. Gfactor is computed by extracting the daily-streamflow mean for the period of record of the streamgage. The daily-streamflow mean is converted from cubic feet per second to log10-cycle change in flow from the previous day with units of days per log10-cycle change in flow. Days with flow less than the previous day are isolated represent a decrease in flow. The days per log10-cycle change of declining streamflow are plotted as a distribution curve. The 90th percentile was selected as the empirical Gfactor for the streamgage after exploratory comparisons to Bingham's earlier work. This method is applied to 312 streamgages and more than 4.8 million days of streamflow at streams in or bordering Tennessee in the streamflow-recession index areas delineated by Bingham (1986). Results from the automated process are compared to the original Gfactor estimates from the streamflow-recession index areas created by Bingham (1986) to assess whether the method is capturing the similar conceptual information about hydrologic processes. The spatial distribution of potential categorical predictor variables or factors are presented. Categorical variables such as soil type, ecoregion, aquifer outcrop, and lithology are subdivided into classes such as individual lithologies and soil types. Proxy watersheds were created to determine predominant categorical variable classes possibly impacting streamflow. Using Kruskal-Wallis one-way analysis, the classes are used as treatments to assess the relative impact of the categorical variables on Gfactors. Four generalized additive models (GAMs) were computed using the categorical variable to estimate the relative impact on Gfactors. Lithology and aquifer rock type were determined to have the greatest impact on Gfactor, and the two were selected to be tandem factor variables were combined in the final GAM. The decade of streamflow data collection, conglomerate lithology, carbonate aquifer rock type, sandstone aquifer rock type, semiconsolidated sand aquifer rock type, and the spatial distribution of the Gfactors are statistically significant with p-values <0.0001. The R-squared (coefficient of determination) for the final GAM is 0.563, which implies 56 percent of the variation in the data is explained, and the residual standard error is 0.149. The final GAM determines a method of regionalizing Gfactors across the study area. The use of an automated process allows the digestion of effectively all daily-mean streamflow data to compute the Gfactor. The use of vastly larger amounts of information than Bingham could process by hand should make it possible to estimate the Gfactor for larger areas as well as for discrete time periods. The approach herein based on declining days for one log10-cycle change and GAMs is expected to be a useful tool to evaluate the extent of connectivity between surface water and groundwater in a basin.