Detecção e delimitação de gás natural em imagens sísmicas Usando MLP-Mixer e U-Net

• Autor(a) principal: CIPRIANO, Carolina Lima Saraiva
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/4618
• topic: hidrocarbonetos;; imagens sísmicas;; aprendizado de máquina; MLP- Mixer;; U-Net.; hydrocarbons;; seismic images;; machine learning;; MLP-Mixer;; U-Net.; Ciência da Computação

Energy use is essential for carrying out various activities, from the simplest to the most complex. Hydrocarbons are one of the main energy sources currently used and originate from the decomposition of organic compounds in underground or marine geological layers. To identify these layers and detect the extent of hydrocarbon deposits, it is necessary to carry out a seismic survey. Seismic data acquired through the seismic reflection method are important for prospecting hydrocarbons, but the techniques for analyzing these data are complex for specialists and are subject to errors and divergences. In addition, hydrocarbon reservoirs can have different sizes and shapes, ranging from large and continuous to small and fragmented. The growth of deep learning has greatly emphasized segmentation, classification and object detection tasks in images from different areas. Consequently, the use of machine learning on seismic data is also growing. For this reason, this work proposes an automatic method of detection and delimitation of natural gas regions in 2D seismic images using the neural network MLP-Mixer and U-Net. The MLP-Mixer is used to detect regions of interest that may contain gas, reducing the number of false positives and facilitating gas delimitation. U-Net is used for the gas region delimitation task, which is the most complex step. Despite the natural imbalance between the number of gas and non-gas samples, U-Net had a satisfactory performance in the proposed method. The results obtained detecting of hydrocarbons were competitive, with an accuracy of 99.6% for inline seismic sections and 99.55% for crossline seismic sections and a specificity of 99.79% for inline seismic sections and 99.73% for crossline seismic sections.