GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.1 GNGTS 2024 Electromagnetc Inducton Data Inversion: Realistc Prior Models and Spatal Regularizaton with Respect to Known Structures N. Zaru 1 , M. Rossi 2 , G. Vacca 1 , G. Vignoli 1,3 1 DICAAR - University of Cagliari, Italy 2 Engineering Geology Division – Lund University, Sweden 3 Near Surface Land and Marine Geology Department - GEUS, Denmark Introducton Geophysical techniques, including electrical resistvity tomography (ERT) and electromagnetc inducton (EMI), play a crucial role in non-invasive subsurface characterizaton. While ERT involves galvanic contact with the ground of arrays of electrodes, EMI enables contactless measurements and is frequently employed for (quasi-)3D resistvity mapping, partcularly in large-scale surveys, ofen carried out with helicopter-mounted systems (Karshakov, 2017; Yin, 2007). Calibraton diferences between the two methods pose challenges, as ERT relies on internal resistances, while EMI demands metculous adjustments for absolute measurements (Finco, 2023; Minsley, 2012). Both ERT and EMI inversions are ill-posed problems. The selecton of a unique and stable soluton relies on prior knowledge, incorporated either as regularizaton term (deterministc approaches) or as prior distributon (probabilistc strategies) (Tarantola, 1982; Zhdanov, 2002). The present study focuses on enhancing 1D EMI inversion by making use of: 1) realistc prior samples, ensuring compatbility with subsurface expectatons, and 2) spatal constraints with geologically meaningful features (e.g., from 2D ERT sectons or their geological interpretatons). We verifed the performance of the proposed approach in a feld test by comparing our results against ground- penetratng radar (GPR) measurements. Methodology Within this research, a strategic shif was undertaken to signifcantly enhance computatonal efciency. The proposed approach consists of formulatng the prior informaton in terms of an ensemble consistng of 1D resistvity models and the corresponding forward response computed for the model . The core principle underpinning this approach involves an exploraton of the soluton space for the inverse problem using a substantal number of pre- computed couples, typically in the range of 10 5 - 10 6 samples. Consequently, the challenge no longer revolves around the contnual generaton of a soluton, which is computatonally intensive, but instead centers on the search for a suitable one from the pre-existng array of possibilites. When correctly implemented, partcularly by harnessing the computatonal power of Graphics Processing Units (GPU), this method demonstrates a speed advantage of at least two orders of magnitude over traditonal deterministc algorithms (McLachlanet al. 2021). The discriminaton criterion is given by the chi-squared value, defned as P m d = F ( m ) m ( d , m )