GNGTS 2018 - 37° Convegno Nazionale

GNGTS 2018 S essione 3.3 771 kept constant (1800 kg/m 3 ) while inverting simultaneously for Vp and Vs parameters. The Ricker wavelet stays as the true source up to now. The time sampling is set to 4.8*e-05 in agreement with to the CFL stability condition. A total number of 8600 time steps correspond to an acquisition duration of 0.412 s (the same as the field one). The quasi-Newton technique is selected in the SEM46 code. The gradient is smoothed using a Bessel filter (Trinh et al. , 2017). In Fig. 2a we show the first gradient with respect to Vp parameter (left) and to Vs parameter (right). The small artefact presents in the shallow Vs gradient (Fig. 2a right) may cause the update of the Vs model in the wrong direction. To solve this potential issue we applied depth- variable boundary constraints on model parameters. After performing multi-parameter inversion without any data hierarchy, the Vp model is not improved significantly, although the Vs model is updated in the correct direction (Fig. 2b and 2c). Further, an alternative workflow can be designed, involving integrated data-oriented strategies (by frequency filtering, time windowing and offset selection and weighting) and/or model-oriented strategies by parameter selection. In the frame of the model-oriented strategies, we perform a mono-parametric Vs reconstruction which improves the shallow Vs model down to 2.5 m depth. With this improved Vs and the old Vp from DCs analysis, a multi-parameter inversion is performed again (without any data hierarchy or model depth pre-conditioning). By doing so, the model is updated in the right direction and below 2.5 m depth. In Fig. 3a and 3b we show a 2D vertical section of the improved Vp and Vs models, while in Fig.3c and 3d we show the 1D profiles for the true, initial and final Vp and Vs models in correspondence of the target. In Fig. 3e and 3f we show the comparison of the matching between true data and initial data and, respectively, between true data and final FWI data. The data misfit shows an improvement not only in correspondence of the target but also for far offsets. This preliminary results make us confident that further data-fitting should improve the inversion behaviour by integrating model driven strategies with designed data selection. Conclusions and Perspectives. The results of the FWI tests highlight some key features of the multi-parameter FWI for near surface complex shaped targets: - The Vp and Vs models retrieved from DCs analysis seem to be good candidates as starting models. Nevertheless, a robust FWI workflow cannot be performed blindly from this Fig. 1- a) 2D slices of the synthetic (“true”) Vp and Vs models that mirror the CNR geometry; b) Seismic data corresponding to the “true” model – shot 1; c) Initial Vp and Vs models built from DCs clustering and data transform; d) Seismic data corresponding to the initial model - shot 1.