Aprendizado profundo para identificação de litologias em rochas carbonáticas do pré-sal

Abstract

The characterization of pre-salt reservoirs is still a challenge in the oil industry due to the geological peculiarities and the heterogeneity of carbonate rocks. These difficulties encourage the development of new methods, such as computed tomography of the rocks for image generation. As with characterization, new methods have been developed in the computational field, one of which is deep learning, which is considered state of the art in several tasks, with emphasis on computer vision. With this, it is presented, in this dissertation, its application for the identification of patterns in microtomography images of cilindric rock samples (plug) for the lithological classification of rock samples. Two convolutional neural network models were proposed and implemented, a base model and this same model modified by adding a layer of spatial pyramid pooling (pooling strategy), where both were compared with the use of the technique of transfer learning with the model Inception v3. As a reference for these models, classic methods for extracting characteristics from images were used as input to a multilayer perceptron. The dataset of interest has samples of 60 plugs made available by Petrobras, which are distributed in three different classes, labeled by specialists. Each plug has 100 images available for use, thus summing up to 6,000 images. The results obtained showed that the Inception v3 model surpassed, on average, the other models, reaching more than 80% of average accuracy. With that, a workflow was developed to automate and accelerate the lithological classification process of pre-salt carbonate rocks.