Improved seismic imaging based on resolution enhancement and pattern recognition

In later years the petroleum industry has moved into exploring fields with more complex geological structures. Examples can be salt deposits in the Nordkapp basin in the Norwegian Barents Sea, carbonate reservoirs in the Middle East and presalt discoverie s in the Santos basin in Brazil. Thus there is a strong need to further develop techniques employed within seismic processing and imaging to better handle such complex scenarios. The project will develop and investigate different techniques with a potenti al of giving sharper images of the subsurface. Parts of the activity will focus on improvement of the signal-to-noise ratio by processing the seismic data in a way that honours the shape of the geological layers more accurately. As part of such a processi ng strategy, the construction of the subsurface velocity field can be more optimized. Accurate velocity model building is crucial in case of complex geology in order to constrain and optimize the final seismic image. By combining techniques from the image analysis and pattern recognition community, complex geological features can be more efficiently detected and extracted from the data and thus make iterative velocity model building less dependent on manual interpretation. To further enhance the sharpness or resolution of the final seismic image, newer developments from the signal processing community will be combined with more conventional wave-based seismic techniques. The project is a collaboration between the Center for Imaging, University of Oslo and the partners Cepetro/Unicamp, Brazil and PGS Geophysical AS, and involves two PhD-candidates. An exchange program of PhD-students and academic staff between the universities in Norway and Brazil is also included. The project has studied imaging of salt structures using deep learning to get a realistic 3D model of the structures. Deep learning is a machine learning method which has grown popular over the last five year and revolutionalized tasks like face recognition or image retrieval. We have used deep learning to train algorithms to recognize seismic textures with and without salt. Based on this, the algorithm can make automatic models of salt based on seismic data. We compared automatic models to manual interpretations, and the automatic models fit well with manual interpretations.

Prosjektet var blant de første til å ta i bruk dyp læring innen automatisk tolning av seismiske data. Stipendiat Anders Waldeland har publisert dette i The Leading Edge, blitt invitert til en rekke firma i bransjen for å holde foredrag om dette, og spilt inn en E-lecture for European Association for Geoscientists and Engineers (publiseres høsten 2018). Så langt har Anders Waldeland disputert (mars 2018), mens Hao Zhao forventes å levere sin avhandling i 2018.

The project is of cross-disciplinary nature and represents a combination of petroleum geophysics (seismic processing and imaging) and image analysis/pattern recognition. Two PhD candidates will be attached to the project. It can be subdivided into four we ll defined work packages: - WP1: Enhancement of the signal-to-noise level of pre-stack seismic data employing true amplitude type of CRS technique (including the use of slope and curvature attributes). - WP2: Estimation of migration velocity fields in tim e employing CRS attributes and mapping of such velocities to depth using IIP-tomography as a constraint. - WP3: Depth-velocity model building in case of complex geological structures like salt bodies. The initial depth velocity model from the mapping tech nique described in WP2 will be iteratively updated employing PSDM and (semi)automatic detection of complex structures using pattern recognition/texture analysis. - WP4: Development super-resolution type of PSDM based on the combined concept of Fresnel ape rture and MUSIC.