GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.2 GNGTS 2024 Integratng cross-gradient joint inversion of ERT and SRT data and unsupervised machine learning on structured meshes incorporatng topography G. Penta de Peppo 1 , M. Cercato 1 , G. De Donno 1 1 “Sapienza” University of Rome – DICEA Introducton Among the geophysical methods applied for Civil and Environmental Engineering, the combined use of electrical resistvity tomography (ERT) and seismic refracton tomography (SRT) has shown to be efectve for investgatng the hydrogeological framework where civil structures and infrastructures are located, since these high-resoluton methods are highly sensitve to diferent physical propertes of the subsurface. Combining them in a coupled inversion scheme can signifcantly improve model reconstructon, so that the resultng subsurface models are more consistent and reliable than those obtained by comparing the results of individual inversions (Doetsch et al., 2010). During the last decades coupled inversions have become increasingly popular, in partcular those involving one or more common additonal terms in the objectve functon to be minimized, even though their practcal implementaton can be complex. The structural gradient-based joint inversion approach is valid when changes in the geophysical propertes are aligned, which is a reasonable assumpton in a wide range of scenarios. The huge number of real-world case studies suggests that the cross-gradient joint inversion introduced by Gallardo and Meju (2004) is presently one of the most robust approaches. Currently, there are many applicatons of cross-gradient joint inversion on structured meshes with fat topography, but such algorithms are usually unable to manage non-fat surfaces. Jordi et al. (2020) developed a novel scheme on unstructured meshes, which can obviously adapt also to not-fat topographies, but for this purpose they modifed the original cross-gradient method, which considers the direct neighbourhood of a single cell. We developed a new cross-gradient joint inversion routne in Python to process apparent resistvity and travel-tme data incorporatng topography, without modifying the original approach from Gallardo and Meju (2004) whose efectveness is well demonstrated. In order to assess the impact of the coupled inversion scheme in both qualitatve and quanttatve terms a new standardizaton of the cross-gradient parameter is proposed (“Normalized cross-gradient”, NCG). Afer the joint inversion procedure, an unsupervised machine learning algorithm is applied for improving the fnal interpretaton by integratng the two output models into a single cross-secton. Our approach is applied to both synthetc and feld examples related to the applicaton of ERT and SRT techniques to Civil and Environmental Engineering.