The challenging task of automatically detecting gas accumulation from seismic reflection surveys has recently gained a new interest in the emergence of deep learning algorithms. These algorithms have an excellent potential to assist the interpretation of potential gas reservoirs but require a large amount of labeled seismic data. The production of these labeled data, however, requires many hours of expert work. Furthermore, given the variance between different seismic surveys, a solution designed for one survey may not perform well in another, especially when comparing onshore and offshore data. In this work, we propose to use a transfer learning methodology to expand an existing classifier and apply it to a different type of seismic survey. Our base model is a Recurrent Neural Network trained and adjusted to detect bright spots in the F3 3D seismic survey in the offshore North Sea, Netherlands. We present the results of our transfer learning proposal applied to a 2D onshore survey in the Parnaiba Basin, Brazil. All the nines existing gas fields were discovered in the last decade with more than 20,000 km of 2D and 482 km2 of 3D seismic data and 150 drilled wells, among exploration and production. The methodology presented here yielded results evaluated by recall, specificity, accuracy, and AUC indexes. Considering the growth in E&P activities in onshore Brazil, this model could be a tool to reduce geological risk and be considered as one of the criteria in prospect ranking.
Presentation Date: Tuesday, October 13, 2020
Session Start Time: 1:50 PM
Presentation Time: 2:15 PM
Location: Poster Station 10
Presentation Type: Poster
