CNN Prediction Enhancement by Post-Processing for Hydrocarbon Detection in Seismic Images

• Published in: IEEE access Vol. 8; pp. 120447 - 120455
• Main Authors: Souza, Jose Fabricio L., Santana, Gabriel L., Batista, Leonardo V., Oliveira, Gustavo P., Roemers-Oliveira, Eduardo, Santos, Moises D.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial intelligence; Artificial neural networks; Coders; Computer architecture; Computer vision; Encoder-decoder; Encoders-Decoders; Geology; Hydrocarbons; Image enhancement; Image reconstruction; Image segmentation; Lead; Outliers (statistics); petroleum exploration; Post-processing; segmentation; seismic imaging; Training; Workflow
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3005916

Seismic image interpretation is indispensable for oil and gas industry. Currently, artificial intelligence has been undertaken to increase the level of confidence in exploratory activities. Detecting potentially recoverable hydrocarbon zones (leads) under the viewpoint of computer vision is an emerging problem that demands thorough examination. This paper introduces a processing workflow to recognize geologic leads in seismic images that resorts to encoder-decoder architectures of a convolutional neural network (CNN) accompanied by segmentation maps and post-processing operations. We have used seismic images collected at offshore sites of the Sergipe-Alagoas Basin (northeast of Brazil) as input. After performing a patch-based data augmentation, a total of 29600 patches were achieved. Out of these, 24000 were used for training, 5000 for validation, and 600 for testing. Each image generated for the training set was post-processed through reconstruction, thresholding - binarization and deblurring -, and outlier removal. By using the dice loss function, intersection-over-union index, and relative areal residual computed after intense cross-validation training rounds, we have shown that the accuracy of the network to detect leads was higher than 80%. Furthermore, the validation error limits were found stable within 5% - 10% in all validation rounds, thereby resulting in a fairly accurate prediction of the pre-labelled hydrocarbon spots.