Application of natural language modeling techniques in natural gas segmentation in seismic reflection images

The most common indirect method the Oil and Gas industry uses to survey an area looking for hydrocarbon accumulations is based on the physical principle of seismic reflection. Geoscientists look for sudden signal intensity peaks which may indicate the accumulations. Most machine learning methods that automate this task using seismic reflection data are based on considering whole seismic lines as images. In this work, we propose a method to automate the segmentation of natural gas accumulations by taking into account the temporal nature of the data and turns reflection amplitudes into word-like objects, using a modified version of WordPiece tokenization, and a Robustly Optimized Bidirectional Encoder Representation from Transformers Pretraining Approach (RoBERTa) to segment each seismic trace that forms the image. As a post-processing step, we apply the mathematical morphology techniques of opening and closing to improve the initial segmentation. We also analyze the presence of a seismic imaging problem in the dataset and how it affects the resulting metrics depending on the dataset’s train-test split choice. Lastly, we compare the proposed method against two baseline models present in the literature. Experimental results show that the proposed method generalizes better than the baseline models and is more efficient to segment previously unseen gas accumulations, effectively decreasing the time between the seismic survey (data acquisition) and exploratory drilling phases. It also paves the way to use other methods from Natural Language Processing in geological research and time series tasks in other research areas.