Salt Segment Identification in Seismic Images of Earth Surface using Deep Learning Techniques

Abstract

Salt segmentation is the process of identifying whether a subsurface target is salt or not. There are several places on Earth where there are significant amounts of salt as well as oil and gas. For businesses engaged in oil and gas development, finding the exact locations of significant salt deposits is crucial. Also, lands that have been impacted by salt are not useful for farming. The absorption capacity of the plant reduces due to the presence of salt in the soil solution. So, in order to identify the land that contains salt, salt segmentation is being done. The seismic image of a particular pixel is analysed to classify it either as salt or sediment. TGS Salt Identification Challenge dataset is used which consists of 4,000 seismic image patches of size (101x101x3) and corresponding segmentation masks of size (101x101x1) in training set. 18,000 seismic image patches are present in the test set which are used for evaluation of the model. The existing models have less detection rate. So, this study has proposed two models for identifying the salt region with high detection rate. The primary model used here is a combination of UNET with ResNet-18 and ResNet-34. The secondary model achieves segmentation results by ensembling UNET with ResNet-34, VGG16 and Inceptionv3. Using these two models, the salt region can be determined from the seismic data. IoU is used as performance metric in order to evaluate the model. The outcomes demonstrate that the ensemble model outperforms individual network models and achieves better segmentation results.