Self-Attention Fully Convolutional DenseNets for Automatic Salt Segmentation.

Area of Science:

• Geophysics
• Structural Geology
• Machine Learning

Background:

• 3-D salt segmentation is crucial for seismic exploration and structural geology.
• Manual salt boundary picking is time-consuming for large seismic datasets.
• Accurate segmentation impacts velocity modeling, seismic migration, and full waveform inversion.

Purpose of the Study:

• To develop a generalized, automated 3-D salt segmentation framework.
• To improve the efficiency and accuracy of salt body identification in seismic data.
• To demonstrate the model's robustness and generalizability across different datasets.

Main Methods:

• A fully convolutional DenseNet architecture was developed for automatic salt segmentation.
• A squeeze-and-excitation network was integrated as a self-attention mechanism.
• The framework employs supervised learning and transfer learning techniques.

Main Results:

• The proposed framework achieved robust performance on the Kaggle TGS salt segmentation dataset.
• Transfer learning enabled effective application to a new dataset (3-D SEAM model) with minimal training data.
• Satisfactory results were obtained, demonstrating the model's generalization ability.

Conclusions:

• The developed deep learning framework offers an efficient and accurate solution for 3-D salt segmentation.
• Transfer learning enhances the model's adaptability to new, unseen seismic datasets.
• This automated approach significantly reduces manual workload in geophysical data analysis.