Título: Seismic Facies Segmentation Using Convolutional Neural Networks
Autores: Jonh Brian Lemos, Lorena da Silva Oliveira Santos & Alexsandro Guerra Cerqueira
Resumo: Seismic facies segmentation is a crucial task for hydrocarbon exploration and reservoir characterization. However, due to the high amount of data involved in geophysics, manual seismic labeling is very time-consuming and sometimes prone to errors due to its subjective nature. Artificial intelligence algorithms, especially deep learning techniques, have had a significant impact on seismic data analysis. Nevertheless, many techniques require large amounts of labeled data, often becoming unfeasible in the context of seismic segmentation. This study proposes a workflow that uses only 10% of the data for training, which is generally the amount labeled in real-world analysis, to predict the remaining sections. We employed a U-Net architecture to train two models on the Parihaka 3D seismic survey, one for the inline orientation and another for the crossline orientation. The data processing included normalization, patching, and the merging of the two prediction cubes based on the maximum probability of their classes. The proposed method achieved an accuracy of 95% and a Mean IoU of 87% on the merged cube. These results demonstrate that the U-Net architecture was able to extract key features from seismic data and provide accurate labels. This approach represents a practical and efficient tool to accelerate the seismic labeling process while generating reliable results.
Palavras-chave: Seismic Facies Segmentation; Seismic Data; Convolution Neural Networks; U-Net.
Páginas: 6
Código DOI: 10.21528/CBIC2025-1191628
Artigo em PDF: CBIC_2025_paper1191628.pdf
Arquivo BibTeX:
CBIC_2025_1191628.bib