GNGTS 2018 - 37° Convegno Nazionale

736 GNGTS 2018 S essione 3.3 been tested by changing type, level and window length of the noise affecting MT time series. Furthermore, in order to identify the most reliable apparent resistivity and phase values among the different impedance tensors clusters provided by the SOM analysis for each analyzed period, a selection criterion is provided and tested on synthetic and field MT data. SOM clustering technique. The self-organizing map (SOM) is a neural network widely used for data exploratory analysis (Kohonen, 1998). ASOM consists of a set of nodes (neurons) distributed on a low-dimensional grid, where each node has an associated D-dimensional weight (or prototype) vector, with D the input vectors dimension. A neighborhood relationship links adjacent nodes characterizing the SOM map structure. The latter identifies the local lattice structure and the global map shape. The weight vectors are initialized with random values before the SOM processing that consists of an iterative procedure: at each step, an input vector is compared with all the weight vectors of the nodes by calculating the Euclidean distance between them. The node whose weight vector is closer to the input vector is called Best Matching Unit (BMU) and its weight vector, as well as those of its neighboring nodes, are adjusted to move towards the input vector by using a neighborhood function. The latter is a decreasing function of the distance between BMU and n th node of the grid. This procedure is repeated for all the input vectors to provide a two-dimensional map with new weights for each node. The final result of the SOM analysis is a map in which each node contains a certain number of input vectors with similar features. It is worth noting that, even if SOM clustering has been widely used in the last two decades for different geophysical applications, from meteorology to seismology (e.g., Esposito et al. , 2008; Liu and Weisberg, 2011), its application to MT data is an almost unexplored research field, except for the first attempts proposed by D’Auria et al. (2015) and Carbonari et al. (2017). Application to MT data. SOM clustering of MT data is performed by using as input vectors the normalized impedance tensors obtained after a DWT decomposition of the MT signal. In particular, the latter is firstly decomposed through a Discrete Wavelet procedure (D’Auria et al. , 2015; Carbonari et al. , 2017), then, for each wavelet scale, the impedance tensor is estimated Fig. 1 - Apparent resistivities and phases retrieved from the SOM clustering of the impedance tensors respectively in absence (a) and presence (b) of noise. Green and orange circles indicate apparent resistivity and phase retrieved from, respectively, the xy and yx components of the impedance tensor, while the continuous blue line shows the noise-free synthetic MT curve.