Transfer Learning Applied to Seismic Images Classification

Abstract

Seismic imaging is an essential tool in oil and gas (O&G) exploration since it provides information about subsurface geometry and structural features. Also, by gathering other sources of data, it is possible to identify relevant localities of hydrocarbon accumulation. Current petroleum exploration demands analysis and interpretation of large volumes of seismic data in strict deadlines. Therefore, computational systems that assist the expert in classifying subsurface features aiming to speed up the analysis process are paramount to the industry development.

The growing popularity of deep learning inspired scientists to apply such methods to seismic data in real data sets. Although this technique has shown good results, a well-known issue in deep learning systems is the difficulty to find a good starting point to adjust model's parameters. A poor initialization may lead to longer training sessions or to the inability of finding a solution. To address the initialization problem, we propose the use of transfer learning to set a good starting point to the parameters of our model. The idea behind this technique is to use previous knowledge obtained from one task in another. In our approach, to train a convolutional neural network (CNN) for a new data set, we initialize the model using the values of the parameters from a CNN trained with another seismic cube.

We conducted two main experiments using real seismic data sets from Scotia and Central-Graben basins. The first one was designed to verify if it is possible to train a model with a highly limited number of examples using previous knowledge. We proposed the second experiment to check whether training information from one cube would be useful to set a good starting point for the new model. The results showed performance improvements, even using cubes from regions that are not geologically similar.

AAPG Datapages/Search and Discovery Article #90323 ©2018 AAPG Annual Convention and Exhibition, Salt Lake City, Utah, May 20-23, 2018