Enhancing Seismic Event Classification in Gujarat Through SeisAug-DrivenData Augmentation for Deep Learning

In anticipation to substitute the existing manual and semi-automated methods for classifying three categories of seismic events (quarry blasts, earthquakes, and noise), we have developed three different convolutional neural network (CNN) models. The first CNN model is based on extracting relevant features from seismograms (waveform), second is based on spectrograms (spectrum), and the third model uses a combination of these two respectively. The CNNs were trained using a labeled seismological waveform dataset recorded at a station SUR from GSNet (Gujarat) during the years 2007-2022. Generally, a common limitation in applying any deep learning techniques is the limited labelled dataset. Therefore, we utilised SeisAug, a Data augmentation (DA) python toolkit to address this challenge to significantly mitigate overfitting by increasing the volume of training data and introducing variability, thereby improving the model's performance on unseen data. A total of 3414 x 3 waveforms were extracted from the three categories of seismic events with a uniform data length of 180 s, considering factors such as coda length, which varies with magnitude and epicentral distance. From this dataset, 15% of the data belonging to each category was split for testing and remaining data was augmented using ‘SeisAug’ toolkit and used for training. The waveform model (WF), spectrogram model (SPEC), and combined model (COM) produced accuracies of 95.32%, 93.13%, and 93.96%, respectively. The robustness of the developed models is indicated by high F1-scores (WF > 0.91, SPEC > 0.92, COM > 0.97) and high area under the curve (AUC) values (WF > 0.98, SPEC > 0.93, COM > 0.98). The high F-scores indicate that these models are very well trained and the probability/possibility of false positives or false negatives is minimum. The high AUC indicates that the model performs well across a range of thresholds and can effectively distinguish between different seismic events. Further, these models produced accuracies of >90% and 100% when tested on completely new datasets from SCEDC and Palitana region (Gujarat) respectively.