Enhancing seismic activity classification in augmented soil gas radon time series data through computational intelligence techniques

Abstract

The classification of seismic activities with anomalous radon time series (ARTS) is promising but hindered by unbalanced data. Rare events like earthquakes make standard classification methods less effective. For such types of problems supervised classification algorithms do not work effectively since they are designed to learn on balanced datasets. Specifically, different techniques viz. Synthetic Minority Over-sampling Technique (SMOTE), Density-Based Synthetic Minority Over-sampling Technique (DBSMOTE), and Adaptive Synthetic Sampling (ADASYN) have been used to address the class imbalance problem inherited within the original soil gas radon time series. Numerous machine and deep learning methods are considered to be the prominent classifiers, including logistic regression, support vector machines (SVM), XGBoost, long-short-term memory (LSTM), and convolutional neural networks (CNN). The investigation has been focused on optimizing the classification of imbalanced datasets through the selection of oversampling methods and learning algorithms. This study indicates that window size and overlapping parameters greatly affect model performance, particularly in discerning seismic events, with larger windows presenting challenges. SVM stands out as a relatively accurate classifier, consistently achieving competitive AUC values without overfitting. Data augmentation techniques show mixed effects, underscoring the need for careful selection based on dataset characteristics. This study suggests that SVM with oversampling methods offers an effective approach for seismic anomaly classification in earthquake prediction using soil radon gas time series data. The dataset for this study comprises soil radon and thoron gas concentration time series along with meteorological parameters spanning 14 months. Four seismic events were recorded during the data collection period. during the whole study period.