Item

Abstract

Identification of microearthquakes at source depth holds significant importance in the field of microearthquake monitoring. Taking 256 microearthquake events (1.5 < ML < 4) in Changning Shale gas exploration area in the south of Sichuan Basin as the engineering background, this paper introduced a method of extracting six feature sets and 6 × 24 feature parameters, which are derived from microearthquake waveform in time and frequency domains based on Empirical Mode Decomposition and Hilbert Transform. The feature importance ranking and 22 key feature parameters closely related to source depth information were obtained using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms. In addition, principal component analysis (PCA) was used to reduce dimensionality and reconstruct the feature space. The classification performance of multiple algorithms, including XGBoost, Support vector machine (SVM), Logistic Regression (LR), K-Nearest (KN), RF, and Decision Tree (DT) models, was compared. The results show that both the 22-dimensional feature parameters and the feature space reconstructed by PCA can effectively distinguish shallow events with source depths less than 1 km from deep events with source depths greater than 6 km. Using the evaluation indicators of receiver operating characteristic, sensitivity, and specificity, it is believed that XGBoost, SVM, and RF classifiers outperform LR, KN, and DT in identifying source depth. Among them, XGBoost classifiers are the least affected by random sampling and changes in sample proportion. The machine learning technology used in this study can effectively perform automatic source depth classification on seismic signals.