Computational and I/O Cost

The computational cost of the plane-wave RTM is about $\frac{31\times 3}{515} \approx \frac{1}{5}$ of that for conventional shot-domain RTM. Each iteration of the LSRTM is assumed to cost twice that of the RTM method. So, for 30 iterations, the computational cost of the plane-wave LSRTM is about 12 times that of the conventional RTM. The computational cost of the dynamic plane-wave LSRTM is only about 40% of that for conventional RTM. The drawback is that CIGs are not available for velocity analysis and the convergence is lessened because at every iteration the problem is redefined with a new encoding function. For the I/O costs, conventional RTM inputs 515 shots with 480 traces each, and plane-wave migration only needs to read 31 plane-wave gathers with 1260 traces each. Hence, the I/O cost of plane-wave migration and plane-wave least-squares migration is $\frac{31*1260}{515*480}=0.15$ that of conventional RTM, if all the data can be stored in the physical memory, so it might be more suitable for GPU calculations. Table 2 shows the comparison of different methods in terms of computational and I/O cost.

Table 3.2: LSRTM and RTM computational cost, I/O expense, image quality and sensitivity to errors in the migration velocity for the field data example with a marine streamer acquisition geometry.

	RTM	Shot-domain LSRTM	Plane-wave RTM	Plane-wave LSRTM	Dynamic LSRTM
Computation Cost	1	60	0.2	12	0.4
I/O Cost	1	1	0.15	0.15	0.15
Image Quality	good	highest	better	highest	better
Sensitivity to Velocity Errors	good	fair	good	good	good