Segmentação de fácies sísmicas com redes neurais

Abstract

The interpretation of seismic data is important for the characterization of the shape of the sediments in a geological study area. Traditionally, this work is carried out by visually choosing points that represent the limits of seismic facies and executing a tool to make the inference of other limit points. This process requires a lot of manual labor and can allow some facies to go unidentified, making the resulting work less detailed than it could be. With the increase in the use of deep learning focused on image segmentation, its application in helping seismic interpretation can bring gains by decreasing manual work and the time spent when studying a geological area. Thus, in this work we made a study of the application of deep neural networks of the encoder-decode type for the identification of seismic facies separating lines. As a result, we created a neural network called DNFS, which is based on U-Net and StNet, has fewer parameters than these and is aimed at binary segmentation of seismic data. This type of segmentation allowed us to segment an arbitrary number of seismic facies just by focusing on the transition between them. To use binary segmentation we use a simple method of adapting the data sets on which we did the experiments. This adaptation uses black lines between the intersections of the seismic facies and white color for the rest of the labeled image. For the calculation of loss we use a function composed by the linear combination of the cross-entropy and Jaccard loss functions. To optimize the coefficient of the linear combination of the function that weighs the weight of cross-entropy and Jaccard loss in the loss value, we performed several experiments with the result that if the cross-entropy contributes 75 % and Jaccard loss with 25 %, we could obtain predictions with high fidelity of the separating lines between the seismic facies. We also carried out an extensive experimental evaluation and adjustments of the hyper-parameters and compared the results with the base networks U-Net and StNet applied on the same data sets. In the end, we obtained a neural network that can be trained in approximately 15 minutes and offers an index above 95% relative to the IoU metric on the StData-12 and Facies-Mark datasets.