Direct prediction of spatially and temporally varying physical properties from time-lapse electrical resistance data

Time-lapse applications of electrical methods have grown significantly over the last decade. However, the quantitative interpretation of tomograms in terms of physical properties, such as salinity, temperature or saturation, remains difficult. In many applications, geophysical models are transformed into hydrological models, but this transformation suffers from spatially and temporally varying resolution resulting from the regularization used by the deterministic inversion. In this study, we investigate a prediction-focused approach (PFA) to directly estimate subsurface physical properties with electrical resistance data, circumventing the need for classic tomographic inversions. First, we generate a prior set of resistance data and physical property forecast through hydrogeological and geophysical simulations mimicking the field experiment. We reduce the dimension of both the data and the forecast through principal component analysis in order to keep the most informative part of both sets in a reduced dimension space. Then, we apply canonical correlation analysis to explore the relationship between the data and the forecast in their reduced dimension space. If a linear relationship can be established, the posterior distribution of the forecast can be directly sampled using a Gaussian process regression where the field data scores are the conditioning data. In this paper, we demonstrate PFA for various physical property distributions. We also develop a framework to propagate the estimated noise level in the reduced dimension space. We validate the results by a Monte Carlo study on the posterior distribution and demonstrate that PFA yields accurate uncertainty for the cases studied.