Abstract:
One of the primary aims in seismic data processing is to detect geologically important
features in the seismic images. Subsurface features, like horizons, faults,
unconformities, and other events contain valuable spatio-temporal data which might be
of scientific importance as well as commercially significant. Therefore, it is necessary to
precisely and efficiently detect such features in images. However, the usual method for
extraction of features in seismic images requires human interference in various forms
like visual interpretation and labelling of the pixels by an specialist. This is time-taking
and costly. On the other hand, automated approaches for extraction of seismic features
suffer from various pitfalls. Subsurface features are of different shapes and dimensions,
and parametric modelling approaches often do not work well. Furthermore, seismic
datasets are usually high-dimensional, and intricate algorithmic approaches are
expensive from a computational viewpoint. Lastly, seismic images carry high amounts
of noise, and this further decreases the performance of parametric approaches.
In this thesis, Convolutional Networks, a supervised learning technique recently gaining
attention in the field of machine learning for vision perception tasks was investigated to
perform end-to-end automated seismic feature segmentation. Two popular
convolutional network models namely U-NET and Seg-Net are chosen and transformed
to accomplish seismic feature segmentation and the results were analyzed. Our aim
was to identify complex features but as not much complex features were available in the
data. The limitation in applying CNNs to these images is that the data set is not huge
and we perform data augmentation on the seismic images. Various loss functions were
tried with the algorithms and the final result is obtained with binary cross entropy and
categorical cross entropy in the two architectures. The result from experiments are
compared on the basis of different type of feature characteristics that the architectures
perform.
