Effect of Data Augmentation Method in Applied Science Data-Based Salt Area Estimation with U-Net

Abstract

Oil and natural gas rank first as energy inputs worldwide. Other subsurface resources, such as salt, provide clues to obtaining these natural resources. Salt accumulation areas are subsurface resources used to locate oil and gas fields. Seismic images, which are geological data, provide information for locating underground resources. Manual interpretation of these images requires expert knowledge and experience. This time-consuming and laborious method is also limited by the fact that it cannot be replicated. Deep learning is a very successful method for image segmentation in recent years. Automating the detection of subsurface reserves in seismic images using artificial intelligence methods reduces time, cost and workload factors. In this study, we aim to identify salt areas using U-net architecture on the salt identification challenge shared by TGS (the world’s leading geoscience data company) Salt Identification Challenge on kaggle.com. In addition, the effect of data augmentation methods on the designed system is investigated. The data set used in the system consists of seismic images that are combined together for automatic detection of salt mass. The study aims to obtain the highest accuracy and the lowest error rate to detect salt areas from seismic images. As a result of the study, the IoU (Intersection over Union) value of the system designed without data augmentation method is 0.9390, while the IoU value of the system designed using data augmentation method is 0.9445.