The near-surface velocity reversal and its detection via unsupervised machine learning
In land seismic data processing, picking the first arrivals and imaging the near-surface velocity structures are important tasks. However, in many areas, the near-surface weathering layer includes high-velocity reversals, causing the first arrivals to exhibit shingling effects, which are difficult for picking at the far offset. We have used an acoustic full-waveform modeling method in a multilayered half-space to simulate first arrivals with the velocity reversal. Numerical tests indicate that under certain conditions, shingling occurs if the seismic wave propagates through a thin velocity reversal layer embedded in the shallow structures. Detection of shingling is essential for the selection of valid near-surface imaging solutions, such as first-arrival refraction, or waveform solutions for the appropriate areas. We find that an automated detection scheme that uses unsupervised machine learning can help identify the velocity reversal. We test the method on synthetic and real data, and the testing shows that the automated detection result matches our visual judgment well. After the automated detection, appropriate inversion approaches can be applied to corresponding areas.