Using deep learning based methods to classify salt bodies in seismic images

Abstract

Identification of salt bodies buried under the earth's surface remains important for hydrocarbon industry as they form traps to hold oil and gas reservoirs in the subsurface. Seismic data interpretation on a large volume of data is labour-intensive and time-consuming task if performed manually. Many efforts have been made in the past to automate or semi-automate the process of extracting the salt bodies from migrated images using many proposed seismic attributes. There remains need of domain experts for picking the right geological features for interpretation and these attributes may not completely interpret the noise-contaminated seismic data. In the recent past, machine-learning algorithms have also been used to identify salt bodies using seismic attributes and they out-perform the existing methods. However, there is still need to have directly assisted processes of seismic data interpretation without the compulsion of providing seismic features. Recent researches have used deep learning methods for seismic facies classification and fault detection, which resulted with better outcomes than other machine learning algorithms. In this document, we propose a novel deep learning method based on U-Net plus Se-ResNet to identify salt bodies in seismic images. This method does not require domain knowledge expert and it is expected to extract required useful features automatically from the seismic data. After 10-fold cross-validation, a promising intersection over union (IOU) score has been achieved.