Salt Deposits Identification Using Enhanced U-net-based Mixed Attention Loss Function for Semantic Segmentation

Numerous areas of Earth’s surface with huge accumulations of oil and gas have high salt deposits under surface. Determining such deposits assists several countries in enhancing the storage capacity of petroleum reserves and exploring new ones. However, determining such a deposit is challenging because expert seismic images need human interpretation of salt bodies, which causes biased and greatly variable translations. This research proposes the Enhanced U-net with Mixed Attention Loss Function (EU-net-based MALF) to automatically and accurately identify whether or not the seismic images have salt deposit regions. Initially, the images are gathered from Tomlinson Geophysical Services (TGS) dataset to evaluate the model performance. Then, the nearest neighbor interpolation method and min-max normalization are employed for image resizing and normalizing. At last, the EU-net-MALF is performed for semantic segmentation. The EU-net-MALF achieves a better accuracy of 98.26% compared to existing methods like fully convolutional DenseNet, U-Net and deep supervision, and Interactive Salt Segmentation based on Convolutional Neural Network and GraphCut (SaltISCG).