Enhancing marine magnetic anomaly interpretation with anisotropic diffusion and deep transfer learning.

Area of Science:

• Geophysics
• Marine Geology
• Deep Learning Applications

Background:

• Linear magnetic anomalies (LMAs) are crucial for understanding oceanic crust formation and evolution.
• Interpreting LMAs is challenging due to sparse, irregular ship tracks, leading to subjectivity and time consumption.
• Existing methods struggle with discontinuities in marine magnetic data.

Purpose of the Study:

• To minimize subjectivity and accelerate the interpretation of marine magnetic anomalies using deep learning (DL).
• To address challenges of data discontinuity and limited labeled datasets in DL applications for marine magnetics.

Main Methods:

• Implemented deep learning models, including standard convolutional neural networks (CNN) and transfer learning.
• Utilized anisotropic diffusion to enhance LMA continuity along local orientations.
• Investigated the effectiveness of transfer learning with and without anisotropic diffusion.

Main Results:

• Transfer learning significantly improved prediction accuracy compared to standard CNNs.
• Anisotropic diffusion further enhanced prediction accuracy when used with transfer learning.
• The best-performing model was applied to data from Shatsky Rise and Azores Plateau, successfully identifying LMAs from spreading ridge volcanism.

Conclusions:

• Deep learning, particularly transfer learning combined with anisotropic diffusion, offers a robust solution for interpreting marine magnetic anomalies.
• This approach enhances accuracy and efficiency, overcoming limitations of traditional methods.
• The model successfully identified volcanic features and correlated nonlinear anomalies with complex tectonic settings.