| dc.description.abstract | The detection of geological structures in seismic data is of great importance for applications concerning storage and extraction of resources from the subsurface reservoirs, such as oil and gas exploration and production, CO2 and waste storage, geothermal resources. However, interpretation of seismic data is a time-consuming task. In this thesis, we present an automated approach using light U-net, which is a 3D convolutional neural networks (CNNs) algorithm to identify geological structures such as faults in 3D seismic cubes. Delineating faults from seismic images is a key step for seismic structural interpretation, characterization of reservoirs and well positioning.
In conventional methods for seismic interpretation, faults are picked as discontinuities along seismic reflectors and/or are identified by measuring attributes that estimate discontinuities (e.g. coherence attributes) in seismic data. In this study, we perform fault detection as a binary image segmentation problem through labeling a 3D seismic image with one for fault and zero for no-fault. We have implemented an effective image-to-image fault segmentation using a supervised, fully convolutional neural network. We automatically generate various amounts of 3D seismic images and corresponding binary fault labeling images to train the network, which is shown to be sufficient to train a thriving network of fault segments. Since a binary fault image is highly imbalanced between zero (no-fault) and one (fault), we use a class-balanced binary cross-entropy loss function to change the imbalance so that the network is not trained or gathered to predict only zeros (no-fault). After training several 3D seismic data sets, the network automatically learns to calculate the rich and correct features that are critical for fault detection. Our results from using this approach on multiple 3D seismic surveys from Horda Platfrom and Stord Basin indicate that the neural network can predict faults from 3D seismic images much more accurately and efficiently than conventional methods.
Further, we analyzed frequency of minor faults in 3D around two major faults in Horda Platform, Vette and Øygarden faults. Using this analysis, we investigated how damage zone architecture of these two major faults changes along different scanlines with respect to time and different inlines. Our results show that more minor faults are observed around the Vette Fault, indicating a larger damage zone around this fault. This damage zone is expected to extend southward as minor faults are distributed over a wide region on 1300-1500 inlines. Whereas the damage zone around Øygarden Fault includes a few to no minor faults in seismic scale on most of the inlines. | |