Subsurface Object Detection Method using Blind Source Separation Algorithm

• A novel non-invasive framework for subsurface object detection using acoustic signals. • FastICA-based Blind Source Separation enhances signal clarity in noisy environments. • Spectral features (MFCCs, Chroma, Mel coefficients) reveal consistent object-induced shifts. • Achieved 91.67 % classification accuracy across soil, rice husk, and sand terrains. • Demonstrated strong generalization and reliability under variable field conditions. Accurate detection of underground anomalies such as voids, sinkholes, and buried cavities is critical for mitigating geohazards and protecting infrastructure. Traditional imaging techniques often suffer from resolution degradation and signal contamination in heterogeneous soil environments. This study presents a novel acoustic-based object-truthing framework that integrates Blind Source Separation (BSS) using Fast Independent Component Analysis (FastICA) to isolate target-relevant signals from noisy mixtures. A custom-built sensing platform was deployed across three terrain types such as soil, rice husk, and sand—to capture subsurface acoustic responses. Key spectral features, including Mel-Frequency Cepstral Coefficients (MFCCs), Chroma Vectors, and Mel Spectrogram Coefficients, were extracted to characterize frequency content, harmonic structure, and perceptual energy distribution. Controlled trials revealed consistent spectral shifts due to buried hollow spheres, with MFCC means increasing from –298.77 to –223.70, Chroma means decreasing from 0.5782 to 0.1296, and Mel coefficients inverting from –51.45 to 6.435, confirming strong object-induced resonance effects. The integrated classification model achieved an accuracy of 91.67 %, with F1 scores ranging from 0.90 to 0.93 across all classes. Additional metrics, including Matthews Correlation Coefficient (0.8761) and Cohen’s Kappa (0.8750), validated the model’s reliability. Post-separation signal-to-noise ratio (SNR) improved by up to 18 dB, enabling robust detection even in acoustically challenging conditions. These findings demonstrate that the proposed BSS–ML pipeline offers a cost-efficient, interpretable, and terrain-resilient solution for subsurface anomaly detection in geophysical applications.