A discriminative-generative approach to the characterization of subsurface contaminant source zones

Abstract

Large-scale contamination of ground water due to improper disposal of hazardous chemicals poses a global threat to drinking water supplies. Effective restoration and remediation of such sites relies upon a knowledge of the contaminant's distribution within the subsurface. Obtaining a detailed map of the existing distribution is usually not feasible; rather partial knowledge in terms of certain metrics that characterize the distribution has recently been shown to be sufficient for planning and monitoring remediation strategies. In this work we explore the prediction of a representative metric based upon down-gradient concentration profiles using a classification framework where each class represents a particular sub-range of the metric. Initial experiments show that our proposed model can be used effectively for predicting the metric.

A mixture of experts based discretization approach for characterizing subsurface contaminant source zones

“…We have used the MoE model to estimate a suitable discretization of the metric and in this work we have specifically shown that it can be effectively used for predicting the percentage-mass of DNAPL in pools from the down-gradient concentration profiles. Comparing to our preliminary work [2] where we used Gaussians to model the bin distribution, we get similar accuracies and bin ranges but the overlap between bins is considerably smaller. In the future we would like to compare the effectiveness of this approach against more sophisticated methods that explicitly force bins to be disjoint and also measure its performance against regression based on the traditional MoE model.…”

“…Similarly, the corresponding metric values for that group define the extent of the bin in metric space. In previous work we modeled bin distributions using Gaussians, which can lead to large overlap [2]. So, to obtain bins in metric space that minimally overlap, we model the distribution of metric values in the k-th bin using the following function:…”

“…Work in [1] suggests that knowledge of metrics summarizing the pool and ganglia distribution may be sufficient for predicting the performance of a remediation strategy. Except for our preliminary work [2] there has not been much research on methods for deterimining such metrics from data available in the field.…”