Three-Dimensional Gravity Inversion Based on Attention Feature Fusion.

Area of Science:

• Geophysics
• Deep Learning Applications
• Geological Surveying

Background:

• Three-dimensional gravity inversion estimates subsurface anomaly sources using surface gravity data.
• Deep learning (DL) methods are increasingly applied to 3D gravity inversion, showing promising results.
• Existing DL methods can suffer from feature loss and limited vertical resolution.

Purpose of the Study:

• To propose a novel 3D gravity inversion method based on the U-Net architecture with an attention feature fusion mechanism.
• To enhance the vertical resolution and mitigate the skin effect in 3D gravity inversion.
• To evaluate the performance, anti-noise ability, and generalization of the proposed method.

Main Methods:

• Utilizing the U-Net network as a foundational framework for 3D gravity inversion.
• Implementing an attention feature fusion module to aggregate global and local semantic features, preserving feature information.
• Connecting encoder and decoder features via long hops to integrate coarse-grained and fine-grained semantic information.

Main Results:

• The proposed attention-based U-Net method achieves higher vertical resolution compared to standard U-Net.
• The attention feature fusion module is identified as crucial for improving inversion accuracy and vertical resolution.
• The network demonstrates strong anti-noise capabilities and good generalization performance in experiments.
• Application to the San Nicolas deposit in Mexico accurately predicted sulfur deposit location and shape, aligning with geological data.

Conclusions:

• The developed attention feature fusion U-Net method offers a significant advancement in 3D gravity inversion.
• This approach effectively addresses limitations of previous DL methods, particularly regarding vertical resolution and skin effect.
• The method shows practical utility and reliability for real-world geological exploration, as demonstrated by the Mexican deposit case study.