Automatic Earthquake Mechanism Classification based on Wrapper and Shallow Learning

In this work, we propose an automatic multiclass classification method using metaheuristic-based wrapper strategies and shallow learning classifiers to maximize the primary focal mechanism classification in seismic motion data. The proposed method was trained and validated on a public seismic motion database, after transforming the raw signals into numerical feature vectors. The best classification scheme was formed using the wrapper method with a genetic algorithm approach and a naive Bayes-based fitness function, combined with a seven-nearest neighbors classifier. This scheme achieved a successful area under the receiver operating characteristic curve score of 0.807 and 0.940 for the training and test stages, respectively. These results corroborate the effective reduction of the original feature space from 25 to 12 features while maximizing the classification performance of three seismic activity classes: strike-slip, reverse-oblique, and normal-oblique. Among the three analyzed groups of features, the shape-based set stands as the most discriminant due to the geometric pattern differentiation in the spectrogram image. The promising results obtained allow the proposed method to be considered a powerful tool for monitoring primary earthquake focal mechanisms.