GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 - POSTER GNGTS 2023 Dataset_1 Dataset_2 U CNN D CNN U CNN D CNN U Human 5204 53 2902 113 D Human 39 3115 64 993 Accuracy 98.9% 95.7% Tab. 2 – Performance on the test sets of the CNN model trained without time-shift for the centered time windows of the two test sets. Confusion Matrix is shown, where “U” and “D” indicate Up and Down, respectively, of the first-motion polarity determined by the CNN model or human (U CNN means the CNN model determines the polarity to be Up). Accuracies relative to the different datasets are shown in the last row. Conclusions In this study, we developed the CFM convolutional neural network for the automatic determination of P-wave first-motion polarities. The work was carried out in two stages. In the first we focused on the datasets preparation. A first dataset was compiled using part of the INSTANCE catalog. We devised a selection procedure of seismic traces using SOM and PCA to select only suitable traces from this catalog. We used this dataset to train and test the network. We used a second dataset of waveforms recorded at Mt. Pollino to further test our network. We manually picked the waveforms of the second dataset assigning also the first-motion polarity to each of them. In the second stage, we developed the network, training it using a stochastic optimization algorithm with adaptive steps size (ADAM). We attest that the network is able to generalize and predict polarities of waveforms never presented before, in fact we achieve an accuracy of about 99% and of about 96% on two different test sets. For this reason, we believe it can be exploited as an automatic procedure to ease analysts’ load. In addition, we addressed the issue of uncertainty in arrival times and tried to devise a strategy that could improve performance in this case. We performed a second training, simulating uncertainty in P-wave arrival times in training data. We noted the network is able to gain robustness to these uncertainties by including a time-shift in the training set. References Bishop, C. M. and Nasrabadi, N. M.; 2006. Pattern recognition and machine learning (Vol. 4, No. 4, p. 738). New York: Springer. Goodfellow, I., Bengio, Y. and Courville, A.; 2016. Deep Learning. MIT Press. Kohonen T.; 2013. Essentials of the self-organizing map, Neural Networks pag. 52-65.