142 Seismic Imaging We present here two examples of impressive FWI results, based on marine (Figure 5.15) and land (Figure 5.16) datasets. In the former, the starting model was obtained by a conventional ray-based reflection travel time tomography (Bishop et al., 1985). FWI was then used to refine the model, leading to clear channels in the shallow part (in white in Figure 5.15-b) and to the presence of gas in the deeper part (in black in Figure 5.15-d). It is possible to see gas leakage along the main faults. The second example, in a land acquisition context (Inner Mongolia, China), was more challenging due to the presence of highly energetic surface waves (Baeten et al., 2008; Brossier et al., 2009). Here, they were filtered out in a pre-processing step. The initial velocity was derived from travel time tomography and is mainly a 1D model (not represented here). Specific attention was paid to the preservation of energy in the dataset between 1.5 to 2 Hz: this was a crucial step in the FWI construction of the velocity model. In both cases, the acoustic FWI largely outperforms standard travel time tomography. More work is needed in future to consider higher frequencies and more complex physics. 5.6 Conclusions Full Waveform Inversion is a technique to obtain seismic quantitative images of the subsurface. However, there are a number of difficulties in terms of its applicability. In particular, a low to high frequency strategy may need to be applied. A carefully considered initial model as well as suitable pre-processing steps must be determined. Finally, the multi-parameter estimation (beyond P-wave estimation) is still an active area of research. In future, FWI will hopefully become “full”: currently, windows are applied to the data to select transmitted waves, for example, or to remove surface waves. These waves contain interesting information on the subsurface (Pérez Solano et al., 2014). The question is to know how to efficiently extract it. There have been a number of possible alternatives to FWI, for example the Adaptive Waveform Inversion strategy (Warner and Guasch, 2016). Usually, the objective is to remove the cycle-skipping effects or to be able to consider the reflected waves more easily (as in the Reflection Waveform Inversion approach (Zhou et al., 2015)). Another alternative is to split the problem into two parts: the estimation of the macro-model containing the main structures (migration velocity analysis or tomography), and of the model perturbation (migration) (Symes, 2008; Chauris and Cocher, 2017). Despite larger computation capabilities and memory allocation, FWI can only process limited frequencies in 3D and usually follows a deterministic approach: starting from an initial model, the model is iteratively updated. In future, uncertainties around the final solutions should be properly evaluated, taking into account
RkJQdWJsaXNoZXIy NjA3NzQ=