GNGTS 2017 - 36° Convegno Nazionale

554 GNGTS 2017 S essione 3.1 good reconstruction of the medium scale geological characteristics in the central part of the model, such as the thrust structures and the salt bodies. Differently, due to the poor seismic illumination, the quality of the reconstruction decreases moving toward the leftmost and deeper edges of the model, where the predicted velocities remain very similar to the starting models. At a first glance, Fig. 1i shows that the estimated Vp / Vs ratio contains some high-frequency characteristics that are not present in the starting Vp / Vs model. At the large scale of Fig. 1, seems that having realistic P- and S-wave velocity starting models, and thus a realistic Vp / Vs ratio, allows for an accurate reconstruction of the subsurface Vp and Vs fields. Fig. 2a shows an example of observed shot gather filtered below 60 Hz. Figs. 2b to 2d show the corresponding shot computed on the starting model, on the final predicted model, and the differences between observed and the predicted shots, respectively. In Fig. 2b note that due to the long-wavelength structure, only diving waves generate on the starting model. The fair match between observed and predicted data demonstrates that the inversion actually converges toward a minimum of the misfit function. Indeed, even though some reflections are unpredicted and the correct amplitudes of the events are sometimes mispredicted, the diving waves are totally recovered and the overall energy of the waveform differences is small. Figs. 3a and 3b offer a closer look at the final predictions that allows for a more quantitative assessment of their features. In this case the results are compared along two velocity columns of the Marmousi-2 model located at 10935 m and 15000 m from the leftmost edge of the model. We note that the estimated Vp model fairly reproduces the velocity variations and the velocity contrasts of the true model. Differently, the estimated Vs field underpredicts the velocity contrasts that characterize the true model. In other words, the estimated shear-wave velocity is mainly a low-pass filtered version of the true one. But more importantly the comparison of the estimated and the true Vp / Vs ratios, demonstrate that the EFWI has not been able to retrieve reliable estimations of this elastic parameter. In particular, if we focus our attention on Fig. 2 - Synthetic seismic data computed on the true model (a), on the starting model (b), and on the predicted model (c). (d) Differences for a single shot gather between the observed seismic data (a) and the predicted data (c).