Classificação de padrões em imagens sísmicas utilizando inteligência artificial

Abstract

The classification of regions most likely to accumulate hydrocarbons is a procedure that involves a specialized analysis of geophysical and geological data from sedimentary basins. Part of the analysis of these data is performed through the interpretation of data obtained with the seismic reflection method, being a step that requires a considerable amount of time, in addition to being a laborious task, even for an experienced interpreter. Detecting areas conducive to the accumulation of hydrocarbons ("plays and leads") from the point of view of computer vision is an emerging theme that demands enormous challenges. The objective of this work was to evaluate alternative approaches for the automatic classification of regions that present the possibility of accumulation of hydrocarbons, using machine learning techniques to identify patterns in seismic images. In this sense, Artificial Neural Networks (RNA), Convolutional Neural Networks (CNN) and semantic segmentation with a U-Net architecture. A database of seismic images from the Sergipe-Alagoas Basin (northeastern Brazil) was used as input images for training, validation and testing. Performance indicators such as accuracy, precision, recall, F1-Score, and IoU were used to assess the network during the training and validation phase. The results were quite satisfactory, mainly involving CNN and U-Net, and the latter showed a more significant result.