Using Multiple Alternative Models to Evaluate Groundwater Model

Here we test model discrimination criteria through comparisons with cross-validation experiments. Three common model discrimination criteria are considered [AICc, BIC, and KIC (as used, KIC is equivalent to MLBMA)] along with two other statistics [the calculated error variance and the weighted sum of squared residuals]. The study is conducted using the computer codes UCODE_2005 and MMA and a MODFLOW model of the Maggia Valley ground-water system in Southern Switzerland. The tests focus on prediction of eight heads and three flows midway along the valley where ecological consequences and, therefore, model precision are of great concern. Sixty-four alternative models were designed deterministically and differ in how the river, recharge, bedrock topography, and hydraulic conductivity are represented. Computationally frugal linear (local) sensitivity analysis statistics made it easy to obtain substantial insight into the many alternative models. For example, leverage statistics were used to measure the importance of the observations omitted for the cross-validation experiments. Results include: (1) None of the model discrimination criteria consistently identified the most accurate modes, where accuracy was evaluated using measured values and cross-validation. (2) The most significant model improvements occurred with introduction of spatially distributed recharge and improved bedrock topography. (3) The simplest models poorly represented the system in the area of interest.