This thesis develops four skeletonization methods for seismic inversion, seismic imaging, and GPS marker detection to improve both their computational efficiency and accuracy. The first two improve the accuracy of the final inverted images by novel skeletonized inversion methods. The third one improves the quality of seismic imaging by employing skeletonized preconditional operators. The fourth one uses skeletonized data for machine learning (ML) identification of GPS markers in drone photos.
(1) To obtain a good starting model for anisotropic full waveform inversion (FWI), the simultaneous inversion of anisotropic parameters vp0 and epsilon is initially performed using the wave-equation traveltime inversion (WT) method. Then a transmission+reflection wave-equation traveltime and waveform inversion (WTW) method is presented for a vertical transverse isotropic (VTI) medium where both traveltimes and waveforms are inverted for the velocity model.
(2) To mitigate the amplitude mismatch problem, multiscale phase inversion (MPI) is presented where the magnitude spectra of the predicted data are replaced by those of the observed data. Moreover, the data are integrated N times in the time domain to boost the low-frequency components. In this case, the skeletonized data are traces with the substituted magnitude spectra so that only the recorded phase data need to be inverted.
(3) I have developed a velocity-independent workflow for reconstructing a high-quality zero-offset reflection section from prestack data with a deblurring filter. In this case the Hessian inverse is approximated by its skeletonized representation, also known as the deblurring operator.
(4) The GPS markers are only about 0.5*0.5 m^2 in size and are difficult to detect manually in the drone images. The marker has a unique hourglass shape and its color is dark. To take advantage of these features, superpixels are used as the skeletonized representations of the targets. Then a superpixel-based classification method is applied to the aerial images.