Deep Learning-based Model for Automatic Salt Rock Segmentation

In places where petroleum and natural gas accumulate, a large number of salt layer deposits are likely to form under the surface of the earth. Locations of petroleum and natural gas can be found through precise positioning. Salt rock areas are traditionally located by experts through annotations on seismic images from professional equipment. However, manual labeling is a tedious and lengthy process, and is not objective. The inaccurate judgment of the location of a salt body will create hidden safety hazards. For a more accurate and automatic process, a salt rock segmentation method based on a U-Net model and deep supervision is proposed, using Kaggle platform data provided by the TGS-NOPEC Geophysical Company (TGS). Based on the data, single model precision of 87.32% mAP is obtained by training the model directly. Using transfer learning, ResNeSt loaded with a pretrained model is used as the backbone network of the encoder. To further improve the accuracy, some modules are added to the decoder. A series of experiments are conducted using a standardized method, whose results show that the proposed model delivers higher accuracy than some state-of-the-art models do.