Inversion of Magnetic Anomaly Based on Cross Attention Transformer
Three-dimensional magnetic anomaly inversion is regarded as one of the most effective methods for accurately retrieving subsurface magnetization distributions. However, existing deep learning methods for magnetic anomaly inversion suffer from issues such as the lack of accuracy in some model structu...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10973049/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Three-dimensional magnetic anomaly inversion is regarded as one of the most effective methods for accurately retrieving subsurface magnetization distributions. However, existing deep learning methods for magnetic anomaly inversion suffer from issues such as the lack of accuracy in some model structures, poor boundary details, and the skin effect. To address this technical challenge, we propose a magnetic anomaly inversion method based on Transformer architectures, with constraints from magnetic anomaly measurement data. Our method employs a hierarchical encoder-decoder network constructed with Transformer Blocks and introduces three key innovations: 1) We propose a Transformer Block based on cross-attention mechanism. Leveraging this mechanism, the Transformer Block can extract features from both magnetic anomaly and magnetic gradient anomaly data, thereby significantly enhancing the accuracy of boundary detection. 2) We propose a learnable Multi-Scale Feature Fusion Module. This module is devised to integrate the multi-scale features from each stage of the encoder, facilitating the decoder to achieve high-precision inversion. 3) We propose a forward constraint loss function. During network training, this loss function ensures that the inversion results adhere to geophysical principles. This methodology not only elevates the inversion accuracy but also effectively alleviates the skin effect. Experimental results show that, compared to other methods, our approach can accurately reconstruct the shape and location of the magnetization model, improve structural accuracy, enhance boundary details, and reduce the skin effect. Furthermore, the method was applied to magnetic anomaly data from a region in Tianjin, China, successfully predicting the distribution of magnetically related pipeline. This demonstrates its potential as a valuable tool for magnetic anomaly inversion. |
|---|---|
| ISSN: | 2169-3536 |