GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.2 GNGTS 2024 Identfcaton of archaeological features using Machine Learning techniques applied to electromagnetc data A. Capozzoli 1 , E. Piegari 1 , F. Cella 2 , M. La Manna 1 , V. Paolet 1 1 Dipartmento di Scienze della Terra, dell’Ambiente e delle Risorse, Università degli Studi di Napoli Federico II, Naples, Italy 2 Dipartmento di Scienze e Tecnologie, Università di Camerino, Camerino, Italy Introducton Over the past decades, geophysical investgatons in the archaeological feld have been recognized as an important tool for identfying target areas and planning dig works (e.g., Cella et al., 2015; Di Maio et al., 2016; El Qadi et al., 2019 and references therein). Indeed, they allow exploraton of evermore extended surfaces with a signifcant reducton in costs associated with excavaton campaigns. One of the geophysical prospectng methods with the highest values of the beneft- cost rato is the frequency domain electromagnetc method (FDEM). This method allows obtaining quick informaton on both the electrical and magnetc propertes of the investgated volumes from the measurements of the two components of the EM feld. However, correlatons between the maps of such components are commonly found only by visual inspecton (e.g., Everet, 2013). The aim of the present study is to realize integrated maps of the two EM components in the atempt to obtain automatc identfcaton of the target areas. Material and methods We analyze the electromagnetc (EM) data acquired in 2012 at the archaeological site of Torre Galli (Vibo Valenta, Calabria, Italy) where a magnetc study and two digs were previously conducted (Cella & Fedi, 2015). We apply the K-Means clustering algorithm to three FDEM dataset, one for each of the operatng frequencies (5 kHz, 10 kHz and 15 kHz), in order to retrieve informaton about potental areas of interest at diferent depths. The K-means algorithm is an unsupervised Machine Learning technique grouping data with similar characteristcs in a predefned number of clusters (e.g., Bhatacharya, 2021). We use two approaches to identfy the optmal number of clusters for the specifc problem. Specifcally, we compute the Silhouete coefcient and apply the Elbow method to the parttons of data in the parameter space defned by the two components of the EM feld. Then, we map in the real space the optmal clusters obtained for each analysed datasets and compare our results with the outcome of magnetc prospecton.