Electromagnetic Exploration Data Transformation into Geoelectrical Models through Clustering and Rescaling Apparent Resistivity Curves

The Magnetotelluric (MT) method is a passive geophysical technique used to study the Earth's subsurface by measuring variations in natural electromagnetic fields. From this geophysical method is possible to compute the apparent resistivity data of a multi-layered subsurface, that reflect at each depth, the impact of a sequence of layers with varying resistivities above the measurement point. This Thesis is a prosecution of the work done by C. Hernandez, who described a method for rescaling 1D MT apparent resistivity data into layered subsurface resistivity models using a relationship called depth/pseudo-depth rescaling function. The added value of this study lies in applying clustering algorithms (the three employed are k-means, CURE, and OPTICS) to the apparent resistivity data trying to reduce the number of rescaling functions needed (and therefore the number of resistivity models required through inversion processes) for the rescaling process, using one rescaling function for each cluster. In particular, clustering algorithms were employed in a multidimensional space, testing as clustering criteria, combinations of mathematical parameters, related to a physical meaning, that describe the apparent resistivity curves. In the over 30 tests performed on randomly generated synthetic data, various iterations were carried out on Phyton scripts, to determine the most reliable clustering algorithm and the best combination of parameters. These results showed which combination lead , for each cluster, to the minimum error between the rescaled models using a single rescaling function per cluster (using one at a time and calculating the errors) and the rescaled models using the own rescaling function for each resistivity data. The findings on synthetic data, showed the benefit of clustering, in fact, it emerged that knowing less than 5% of resistivity models in a dataset, and applying the rescaling process using one depth/pseudo-depth rescaling function per cluster, yields very low errors compared to rescaling all the data using a random model without considering clustering processes. Once the conditions were defined, this model was also tested on a real dataset (COPROD2) consisting of 35 apparent resistivity data. After clustering the data and rescaling them using a single 'reference' depth/pseudo-depth rescaling function per cluster, the results confirmed the presence of the conductive bodies expected from the inverted models.