A Neural Network Approach for Quality Control and Interpretation of Single-Station Microtremor Measurements

ABSTRACT

Single-station ambient noise measurements, analyzed through the horizontal-to-vertical (H/V) spectral ratio, are widely applied in seismic hazard assessment and in stratigraphic reconstructions of the subsurface. In mining exploration and other applied geophysics contexts, large-scale surveys often involve hundreds of measurements to support preliminary bedrock mapping. Under these conditions, there is a strong need for tools that can rapidly evaluate the quality of collected data and guide users—expert or not—in their interpretation. To address this need, we developed and trained a supervised feed-forward neural network to classify different sections of microtremor spectra into seven categories: normal, 1D resonance, 2D or 3D resonance, artifact, velocity inversion, data requiring cleaning, and malfunctioning sensor. The training dataset comprised 3000 examples from over 20 yr of field measurements in diverse geological and environmental settings, processed with consistent windowing, smoothing, and spectral normalization procedures. By analyzing both the individual spectral components and the H/V curve within selected frequency intervals, the neural net produces a probability-based classification for each interval, which can be aggregated into a synthetic quality assessment—accept, reject, or reanalyze—based on predefined criteria. Testing on unseen data yielded an accuracy of about 80%–82%, with the main classification challenge being the distinction between 2D/3D resonances and artifacts due to their spectral similarities. The neural net proved robust to changes in instrument type, analysis window length, and smoothing parameters. This approach offers an alternative to fixed-threshold guidelines such as the SESAME (2004) criteria.