| Abstrakt: |
Carbonate rocks are classified based on many proposals, according to the analysis objective. Pre-Salt carbonate oil reservoirs in the Brazilian marginal basins are being evaluated considering depositional textures characteristics for its lithofacies. Considering the advances in image classification using deep learning algorithms, carbonate lithofacies classification models were trained and evaluated using Convolutional Neural Networks. They classify optical microscopy thin section images obtained from sidewall core samples. A relatively small training dataset composed of 642 instances has been modified and enlarged by a series of augmentation techniques. They include different sorts of spectral and geometric transformations on the original images. Synthetic images have also been created using Generative Adversarial Networks, in order to expand the original training dataset and reduce data imbalance between six classes: stromatolite, spherulitite, laminite, grainstone, dolomite and silexite. While training data has been selected from samples of one well in a specific reservoir, test data has been gathered from two other adjacent wells from the same geological context. Results showed that, by gradually applying augmentation techniques on the training dataset, models enhance their generalization capabilities evaluated by accuracy and by test dataset. The model trained based on the original dataset, with no augmentations, has an accuracy of only 42.24%. As augmentations are implemented, this accuracy rises, reaching 86.34% on the model trained using all the techniques. It shows that it is possible to obtain useful models based on smaller and imbalanced datasets. These results also validate the data-centric approach that has been chosen for the creation of these models. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink