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

## TL;DR

This paper introduces a method combining anisotropic diffusion and deep transfer learning to improve the interpretation of marine magnetic anomalies, reducing subjectivity and increasing accuracy.

## Contribution

The novel approach integrates anisotropic diffusion with deep transfer learning to address data sparsity and irregularity in marine magnetic anomaly interpretation.

## Key findings

- Anisotropic diffusion improved the continuity of magnetic anomalies, enhancing deep learning model performance.
- Deep transfer learning significantly increased prediction accuracy compared to standard CNN models.
- The method successfully identified linear magnetic anomalies in the Shatsky Rise and Azores Plateau regions.

## Abstract

Linear magnetic anomalies (LMAs) in marine settings provide important clues about formation and evolution of the oceanic crust. LMA interpretation can be challenging as it relies on visual inspection of spatial patterns of magnetic anomalies, which may not be well defined due to sparse and irregular ship tracks. Interpreting such magnetic anomalies based on human perception is inherently subjective as well as time-consuming. We aim to minimize subjectivity and speed up the interpretation by using deep learning (DL). Two significant challenges arise when applying DL to marine magnetic anomalies. First, the anomalies may exhibit discontinuities due to sparse and highly irregular tracklines. Second, the quantity of labeled marine magnetic data maps is very limited. For the first challenge, we employed anisotropic diffusion to smooth LMA along the local orientations, thereby enhancing the continuity. For the second challenge, we investigated deep transfer learning. We implemented three different DL models, namely, standard convolutional neural network (CNN), transfer learning without anisotropic diffusion and with it. When applied to a test dataset consisting of magnetic anomalies from East Pacific Rise and Reykjanes, CNN without transfer learning achieved moderate accuracy. With transfer learning, the prediction accuracy improved substantially. When anisotropically diffused marine magnetic anomalies were used as input, the prediction accuracy reached an even higher level. We applied our best-performing deep learning model—transfer learning combined with anisotropic diffusion—to marine magnetic anomalies from the Shatsky Rise region in the western Pacific and the Azores Plateau region in the northern Atlantic. Predictions at both areas exhibit LMAs resulting from spreading ridge volcanism. Some of the nonlinear predictions are due to poor data coverage, while others are correlated with complex tectonics such as tectonic reorganization, fracture zones, faults, etc.

The online version contains supplementary material available at 10.1038/s41598-025-30926-1.