GNGTS 2018 - 37° Convegno Nazionale

584 GNGTS 2018 S essione 3.1 A GENERATIVE-ADVERSARIAL NETWORK FOR PROCESSING SEISMIC IMAGES F. Picetti, V. Lipari, P. Bestagini, S. Tubaro Dipartimento di Elettronica, Informazione e Bioingegneria - Politecnico di Milano, Italy Background and motivations Applications in hydrocarbons exploration and reservoir characterization require a subsurface mapping at increasingly higher resolution and higher fidelity. Moreover, as for hydrocarbon exploration, the areas of interest are more and more complex to analyze. This, along with the increasing amount of data, determined a high demand for advanced and efficient seismic imaging methodologies. In recent years, machine learning techniques have started to be investigated to face some of these issues; supervised classification methods have been increasingly explored as a helping tool for the interpretation step (Hall 2016, Bestagini 2017). As a matter of fact, the recent advancements brought by convolutional neural networks (CNNs) have greatly impacted the whole signal and image processing community. Among the different architectures, generative adversarial networks (GANs) emerged as a promising approach for problems that need some form of regularization that is not easy to express through simple modeling (Goodfellow 2014). However, most of the components of seismic imaging workflows are large scale ill-posed inverse problems, rather than classification ones. Traditionally addressed with model-based analytical approaches, recently some of imaging tasks have been solved through data-driven deep learning techniques (Lucas 2018). While leading to state-of-the-art results in computer vision, image processing and various related field, deep learning has barely started to be studied for inverse imaging problems (Al-Regib 2018). McCann (2017) provides a review of recent applications of convolutional neural networks for biomedical imaging problems. Recently, Araya-Polo (2018) proposed a deep learning strategy for seismic velocity model building. In this work we introduce a possible way of using deep learning for seismic imaging applications. We propose to use a GAN as a tool for processing seismic images obtained via Reverse Time Migration (RTM). We formulate the problem as the estimation of a post processing operator that can be learned through a training phase to tackle different problems. Specifically, the GAN is fed with pairs of images composed of input images and desired output images depending on the target application (e.g., deconvolved images, Least Squares RTM images, deghosted images, etc.). The used CNN architecture builds upon the recently proposed pix2pix GAN (Isola 2017). Its potential is shown by tailoring the proposed architecture to two different proofs-of-concept on synthetic migrated data. The first application is data interpolation in the image space: we aim at recovering an image obtained with a dense source-receiver acquisition geometry from an image migrated with a very coarse acquisition geometry. The example we show is obtained on several 2D migrated sections of the SEG/EAGE Overthrust velocity model. The second application is deconvolution. We train our CNN on a portion of the well-known SMAART JV Sigsbee velocity model, to transform the migrated image on the corresponding reflectivity section (obtained from the stratigraphic velocity model). Then we predict the reflectivity from the remaining part of the migrated image. Preliminary results confirm the positive impact that deep learning can have in seismic image processing in the future. GAN for seismic image processing. In this section we introduce the way we cast seismic image processing problem in the GAN framework. The goal of the proposed method is to build a processing block that takes a migrated imageas input and produces an image Î . During training, the machine is fed with a set of K pairs of migrated images I ( k ) and the