2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Semantic Segmentation Pipeline for Seismic Data


Seismic reflection surveying is the most used method to obtain subsurface information in the O&G exploration industry. Through this data, one may obtain structural and stratigraphic geometric features and potential hydrocarbon deposit locations. Even though it is paramount, seismic data interpretation is an extraordinarily time-consuming and human-intensive task, mainly due to the ever-larger volumes of seismic data and the geological complexity present in the study areas. Aiming to address this issue, computer-aid systems assisting geoscientists to interpret this large and complex data in a faster and more accurate manner represent vital importance for the development of the exploration industry. As presented in recent work, deep learning techniques have been helping human-centric tasks in seismic stratigraphic mappings, such as the classification of lithologies and facies, segmentation of salt bodies, tracking of faults and even the semantic segmentation of 3D seismic data. In this work, we present the evolution of our semantic segmentation model presented at AAPG2018. We simplified the development pipeline and improved the efficiency of training, obtaining an improvement in the accuracy of the models and speeding the process. Moreover, our model can now capture thinner layers of strata with higher accuracy. The new pipeline to produce a segmentation model comprises three parts, instead of six: (1) process the post-stacked data generating training, validation and testing sets; (2) create the segmentation model and train the system; (3) deploy the model for the regular usage. We removed the dataset generation for classification, pre-model training and transfer learning from the classifier. However, we can still apply transfer learning to speed training up. In our experiments, we used two public datasets: Netherlands Offshore in F3, and Penobscot from Scotia basin comprising nine and seven horizons, respectively. The new pipeline produced models that able to achieve more than 95% of pixel accuracy using less than 2% of data for training, and our qualitative results showed that the model could provide a mask very close to the actual interpretation.