Earthquake focal mechanisms are derived from moment tensor solutions and provide insights into earthquake rupture characteristics, including strike, dip, and rake angles. Common methods for analysing seismic records (seismograms) include first motion polarity of P-waves, waveform inversion, and amplitude analysis. However, these methods often require significant human intervention. With the increasing availability of strong-motion earthquake records (accelerograms) from diverse epicentral distances, depths, and tectonic regions, there is potential to analyse these records in non-traditional feature spaces—such as shape, texture, and pixel intensity—to predict earthquake fault mechanisms. Artificial intelligence (AI) advancements, especially machine learning algorithms, allow for the analysis of large datasets and the extraction of repeated patterns or behaviours, which could enhance earthquake characterization. While traditional approaches typically rely on time-domain analysis requiring extensive preprocessing, this study aims to develop a non-traditional workflow to classify strong ground motion signals based on their focal mechanism types using machine learning. The proposed workflow involves the following steps: (1) Transforming each signal from the time domain (1D) to the frequency domain via a spectrogram, creating a 2D image representation; (2) Segmenting the spectrogram image to extract shape features; (3) Generating texture- and pixel-intensity-based descriptors. The optimal feature subset is then fed into four machine learning classifiers—random forest, support vector machine, naive Bayes, and an artificial neural network. The method demonstrates efficient classification of seismic records, offering a pathway for automating the analysis and improving overall efficiency.
