Deep semi-supervised learning using generative adversarial networks for automated seismic facies classification of mass transport complex

Geological and geophysical interpretation is characterised by large and localised datasets that are extremely expensive to acquire. There are clear advantages in applying deep learning techniques to such datasets, but this requires a large amount of suitable data for effective training. Creation of training data can be time-consuming, but novel countermeasures such as Generative Adversarial Networks (GANs) are an enticing alternative with the potential to alleviate this issue by providing synthetic data that is representative of the real data. This paper details an investigation of the potential of GANs for the augmentation of labelled seismic facies from a mass transport complex in training a convolutional neural network (CNN) for facies classification. The study adopts a specific GAN approach, known as the Conditional Style-Based Logo GAN (LoGANv2), because of its capability to generate conditioned output with improved training stability. By using LoGANv2 to synthesise examples that mimic the behaviour of the real data, based on a 3D seismic dataset from the North Carnarvon Basin in Australia, the accuracy of the traditional CNN for facies classification improved from 38% (the benchmark result where no augmentation was applied) to 52%, 61%, 94% and 74% for four trials with different degrees of augmentation. The results show that increasing training size, either through manual annotation (trial 2) or standard manipulations such as image rotation (trial 3 and trial 4), improves the performance of LoGANv2 and hence the CNN classification. The experiments indicate that although standard manipulation can increase the diversity of the training dataset, the balance between diversity and consistency within the datasets is also important, and this should be optimized for LoGANv2 to achieve better performance. In addition, we explored the adaptability of the proposed LoGANv2-CNN approach to unlabelled samples from another survey to demonstrate the robustness and flexibility of the model.