Evaluating least absolute deviation regression as an inverse model in groundwater flow calibration

Abstract

The automated inverse procedure built into the US Geological Survey MODFLOW groundwater model works well with single layer geologic systems; however, it does not work well with multi-layer systems. This dissertation provides the Least sum Absolute Deviation regression and Expectation Maximization procedures. These new FORTRAN procedures were added to the parameter estimation source code of the MODFLOW groundwater computer model. The resulting inverse MODFLOW model can now calculate the hydraulic conductivity for multi-layer systems. The model was applied to Walnut Creek watershed in Iowa. Hydraulic conductivity parameters computed with the inverse watershed model using the Least sum Absolute Deviation regression and the Expectation Maximization procedures agree with published data. A forward groundwater model of the larger encompassing Upper Skunk River basin in Iowa was created. The hydraulic conductivity parameters computed by both Walnut Creek inverse models were used to create two forward basin models. Using the hydraulic conductivity from the Least sum Absolute Deviation regression and Expectation Maximization methods on the Upper Skunk River basin produced model heads for Walnut Creek that were within two meters of observed heads. The MODFLOW model predicted a confined aquifer system while the Least sum Absolute Deviation regression and Expectation Maximization model predicted an unconfined system. Literature documents Walnut Creek as an unconfined system. This modified MODFLOW model is a better tool for modeling multi-layer groundwater aquifers.