GNGTS 2018 - 37° Convegno Nazionale

GNGTS 2018 S essione 3.1 587 Fig. 3. data, instead, rely on 200 sources and 800 receivers. After the training, we test the GAN on different in-lines and x-lines never used for training. We show the input migrated section (Fig. 1 top) and the corresponding output (Fig. 1 bottom). We choose the L1 norm in order to force the output image to be sparse. The low-resolution artifacts have been almost completely eliminated and the produced image is very similar to the target. Testing 79 images, we obtained an average SNR of 17.15dB and 16.6dB for the outputs obtained with and without L1 constraint, respectively. The computational time required to obtain the output image is way less than a dense RTM image. In the considered example, the time needed for generating a single migrated section with the dense geometry was around 40 minutes; on the other hand, in the proposed pipeline the computation was performed in 2 minutes, almost entirely dedicated to migration with the coarse geometry and only few seconds were needed to generate the GAN output. Reflectivity from migrated images. As a proof-of-concept, we proposed a synthetic example for deconvolution-like problems (deghosting, LS-RTM, etc.). The input I is a standard depth migrated image and the desired output I ref is an image of subsurface’s reflectivity, where the ground truth has been computed from the Sigsbee stratigraphic velocity model ν ( x, z ) as We trained on 194 pairs of patches and then we tested it on a different validation set of 193 pairs. Both training and validation sets were taken from the left-side portion of the model, such that the network can be tested on the unseen right-side model.