GNGTS 2022 - Atti del 40° Convegno Nazionale

GNGTS 2022 Sessione 3.3 529 to better solve the increase of resistivity in time. The implemented Time-Lapse EB approach starts with two ensembles of prior realisations of ρ 0 and Δρ which are iteratively updated on the basis of the difference between observed and predicted data. The final output is two ensembles of posterior models from which the standard deviation can be assessed. However, when the means are strongly different from one to another, the standard deviations are not easily comparable. To overcome this issue that could affect the uncertainties interpretation, we calculate the coefficient of variation: (3) Where the index i is the i- th cell of the model, σ i is the associated standard deviation, and ρ i is the mean resistivity value of the i- th cell. The outcomes of the implemented algorithm are benchmarked against the results of a cascaded inversion (Miller et al. , 2008), which is a deterministic least-square approach in which the inverted ρ 0 is used as starting model to invert for ρ 1 . Application to field data. The field data that we employ in this work refer to Pillemark landfill monitoring station in Samsø island (Denmark), in which the aim is to supervise the aquifer to assess the pollution risk (Høyer et al. , 2019). The geology of the site consists of a layer of fill over a sand stratum of various thicknesses in which an aquifer is flowing, and a second aquifer underlies a thick layer of moraine clay. Due to the vertical resolution and penetration depth limit, we will focus on the shallow aquifer. The data are acquired using the Terrameter LS instrument employing a Gradient and Dipole-Dipole configuration with 22 electrodes and 4 meters spacing. From the database we select two data ( d 0 and d 1 ) acquired after one year, 27 th March 2016 and 28 th March of 2017 respectively, to attenuate the seasonal effect on resistivity variation (Figure 1a-b). After the processing step the two datasets consist of 336 data points each. The subsurface is discretised with an unstructured mesh employing the pyGIMLI (Rücker et al., 2017) software package and the forward problem is solved by the open-source BERT software (Günther et al., 2006). The Time-Lapse EB algorithm inversion starts with two Fig. 1 - a) Observed data on 27 th March 2016. b) Observed data on 28 th March 2017. c) Predicted data d 0 computed on the posterior mean model ρ 0 and the corresponding rms error. d) Predicted data d 1 computed on the posterior mean model ρ 1 and the corresponding rms error. The black dots represent the points of the pseudosection.