Controlled-source seismic data are used to investigate the Earth's subsurface structures. A problem of the seismic data recorded in the field experiments is that the data can be noisy, especially in the far offsets as the signal-to-noise ratio (SNR) diminishes with offset between the source and receivers. Data with low SNR can heavily degrade the quality of the seismic image, or even lead to a failure of the imaging.
Supervirtual Interferometry (SVI) is a robust and effective technique to significantly improve the SNR of seismic data. This dissertation proposed novel ideas of SVI from three aspects so that SVI has wider and more practical applications. (1) SVI is extended to 3D data, which allows SVI to be applicable for industrial-scale 3D datasets. (2) Windowing the first arrivals, which is a prerequisite of SVI, is automated for noisy seismic data using two machine learning methods: CNN and DBSCAN. Auto-windowed SVI can be used to benefit first break automatic picking by improving the accuracy of the auto-picking results for noisy data. (3) Reflection SVI is developed to enhance the SNR of far-offset reflections. This technique can significantly improve the accuracy of the picked stacking velocity and the quality of the image for poststack migration.
This dissertation also presents a practice of seismic imaging for the Olduvai Basin in Tanzania. The seismic results with refraction tomography and poststack migration delineate the geometry of the basin floor. The maximum basin thickness is around 405 m, which has an important palaeoantropological meaning that hominin palaeoenvironments in the Olduvai Basin possibly date back to 4 Ma, almost doubling the age estimated from the previous coring content.