Seismic fault detection in real data using transfer learning from a convolutional neural network pre-trained with synthetic seismic data

The challenging task of automatic seismic fault detection recently gained in quality with the emergence of deep learning techniques. Those methods successfully take advantage of a large amount of seismic data and have excellent potential for assisted fault interpretation. However, they are computationally expensive and require a considerable effort to build the dataset and tune the models. In this work, we propose to use Transfer Learning techniques to exploit an existing classifier and apply it to other seismic data with little effort. Our base model is a Convolutional Neural Network (CNN) trained and tuned on synthetic seismic data. We present results of Transfer Learning on the Netherland offshore F3 block in the North Sea. The method gives satisfying results using as input a single interpreted section, despite the naturally high imbalance of the labeled classes. The proposed networks are easily tuned and trained in a few minutes on CPU, making the technique suited for practical day-to-day use.