GNGTS 2021 - Atti del 39° Convegno Nazionale

501 GNGTS 2021 S essione 3.3 With the application of ensemble learning, we get two different predictions: one on the sin- gle trace and the other on its inverted version in time, then combining them with the geometric mean, as it gives better results than, for instance, the arithmetic mean. The prediction is given as a probability set that associates a probability value to each point: the value indicates the probability of the point to be a reflector, i.e. to belong to a reflecting sur- face. A threshold above which a point is labeled as a reflector can be set. The optimum threshold is estimated by evaluating the number of points classified as reflec- tors vs the threshold. We perform this task by using the algorithm described in Satopaa et al. (2019). The threshold is set at the sharp inflection point clearly visible in the resulting curve, thus limiting the subjectivity of the choice. Results The first test is the application of this algorithm to a synthetic dataset, the Marmousi model (Martin, 2004). Figure 1 shows part of the data with automatic the prediction superimposed. Prediction is quite accurate and horizons are consistent in 2-D even we remark that, as previously discussed, the prediction is fully 1-D and each extracted point is totally independent from all the other. Fig. 1 - Example of automatic extraction of horizons from the Marmousi Dataset: horizons are marked in red and superimposed on input data. In order to test the proposed methodology on field data we use a 2-D marine seismic section of the WS10 exploration project, obtained in autumn 2010 in the west Mediterranean Sea by the Istituto Nazionale di Oceanografia e Geofisica Sperimentale (OGS), which also performed the data processing (Geletti et al., 2014). The selected portion of the seismic section images a rifted margin of the eastern Sardo-Provençal Basin characterized by a faulted salt dome and a portion of an almost undisturbed sedimentary sequence (Figure 2). For such reason, the analyzed data represent an interesting and complex test for the proposed procedure. The same data is used to tested an independent picking strategy, which can be analyzed for comparison (Forte et al., 2016). The NN is able to properly extract all the main horizons, both where they are sub-horizontal (i.e. in the shallow part) and where they exhibit a relevant dip (i.e. along the flanks of the salt dome). As desired, horizons interrupt at the fault location, while, correctly, no horizons are de- tected inside the salt dome.