Abstract:  The microtremor horizontal-to-vertical spectral ratio (HVSR) method is an efficient and non-invasive geophysical technique widely used in urban geological surveys and engineering investigations. However, transient interferences caused by pedestrians and vehicles can distort the shapes of HVSR curves. Existing transient interference elimination methods have limitations: the STA/LTA (short-term-average over long-term-average) method is prone to misjudgment and requires complex parameter tuning; The manual rejection method is inefficient; And the frequency-window based rejection algorithm considers only peak-frequency information. To address these issues, this study proposes a machine-learning-based interference suppression method for microtremor HVSR data. First, curve-shape features are extracted to train a curve-rejection model for identifying and removing HVSR curves that significantly deviate from the mean trend. Then, peak-related features are extracted to train a peak-identification model for recognizing valid resonance peaks within the  curves. Finally, a density-based spatial clustering of applications with noise (DBSCAN) algorithm is applied to cluster and further eliminate curves containing peaks with abnormal frequencies or amplitudes. The curve-rejection and peak-identification models achieve F1 scores of 0.967 and 0.985 on the test set, respectively, demonstrating excellent classification performance. Case studies show that the proposed method exhibits higher stability and accuracy in eliminating abnormal curves than the STA/LTA and frequency-window based rejection methods. The processed HVSR curves display more concentrated spectral distributions, more convergent standard-deviation curves, and clearer, more stable peaks. Moreover, the proposed method achieves efficient automatic processing while maintaining strong consistency with manual rejection results.

Key words: microtremor HVSR method, machine learning, transient interference, rejection, geophysics