Uma Abordagem Utilizando Séries Temporais para Detecção de Gás em Imagens Sísmicas com Transformer

Abstract

The detection of gas reservoirs in seismic images is complex, time consuming and requires specialized professionals for this task. The great challenge lies in the large amount of data to be analyzed and the difficulty and resources needed to prove the existence of a gas reservoir, since it is necessary to drill the well to prove the fact. One option to get around the large amount of data to be analyzed is to resort to the use of artificial intelligence through the implementation of deep neural networks. In this sense, the scarcity of proven correct data impacts the number of annotated samples that we have access to, which significantly hinders the use of computational methods. In order to overcome the aforementioned difficulties, this work proposes the use of a gas reservoir segmentation method from a one-dimensional perspective with time series analysis. Due to the sequential characteristic of seismic data, it is possible to employ a methodology that interprets each seismic trace present in the input images as an isolated instance of the database. In this way, it is possible to increase the number of samples in the data set significantly. To perform the segmentation of the reservoirs, a new architecture was proposed, which consists of modifying a Transformer network architecture so that it is capable of interpreting temporal series of seismic data. The proposed network is of the sequence-to-sequence type, which means that the network processes an input sequence and returns a new sequence containing the segmented gas region. The choice of this method was due to the fact that this architecture is effective in extracting contextual information without the need to use recurrence in its training, which results in a better performance and allows the parallelization of the process, resulting in a great saving of time, both in training and inference. The database used in the experiments is private and was granted by Eneva for the development of this technology. The metrics used to evaluate the results were accuracy, sensitivity, specificity and AUC, where the method obtained 97.16%, 79.61%, 97.47% and 88.54%, respectively.