Improving seismic signal classification of different ground activities with advanced AI and signal processing techniques

This study presents a comparative framework for seismic signal classification based on adaptive decomposition and Hilbert spectral analysis. A self-recorded vibration dataset containing approximately 4,500 one-second segments of pedestrian, bicycle, and vehicle activities was acquired at a 1 kHz sampling rate using a buried geophone sensor. The raw signals were preprocessed through direct-current removal, fourth-order Butterworth band-pass filtering (1–50 Hz), and min–max normalization to mitigate environmental and instrumental noise. Three decomposition methods—Variational Mode Decomposition (VMD), Empirical Mode Decomposition (EMD), and Matching Pursuit Decomposition (MPD) were independently applied to obtain narrow-band oscillatory components. The Hilbert Transform (HT) was subsequently employed to derive marginal-spectral features, including spectral energy, spectral entropy, and dominant frequency, preserving a clear physical correspondence to source–soil coupling characteristics. These features were used to train multiple classifiers, including Support Vector Machine, Logistic Regression, k-Nearest Neighbors, Gradient Boosting, Random Forest, and Extreme Gradient Boosting (XGBoost), under group-aware cross-validation to avoid trial-level data leakage. Experimental evaluation demonstrates that VMD-HT features provide the highest discriminative performance, achieving a test accuracy of 91.4% and a macro-F1 score of 0.89 using the Random Forest classifier, followed by XGBoost with 90.3% accuracy. The EMD-HT framework attained 87.5% accuracy (macro-F1 = 0.84), while MPD-HT achieved 84.2% accuracy (macro-F1 = 0.81). Ensemble classifiers consistently outperformed linear models, confirming the advantage of nonlinear decision boundaries for decomposed spectral representations. The observed results indicate that VMD yields superior mode separation and spectral compactness compared with EMD and MPD, leading to improved feature stability and reduced overlap across classes. The proposed framework thus offers a robust and physically interpretable approach for ground-vibration classification, with potential applications in intrusion detection, structural health monitoring, and environmental sensing.