Magnetic anomalies characterization: Deep learning and explainability. (December 2022)

• Record Type: Journal Article
• Authors: Cárdenas, J.; Denis, C.; Mousannif, H.; Camerlynck, C.; Florsch, N.
• Is Part Of: Computers & geosciences. Volume 169(2022)
• Publication Date: 2022-12
• Subjects: Geophysics -- Deep learning -- Convolutional neural networks -- Magnetic methods -- Explainability -- Unexploded ordnance; CNNs Convolutional Neural Network -- YOLO You Only Look Once -- MSE Mean Square Error -- R&D Research and Development -- IoU Interception over Union -- UXO Unexploded Ordnance -- EMI Electromagnetic Induction -- TDEM Time Domain Electromagnetic; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2022.105227
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: In recent years, deep learning methods have shown great promise in the field of geophysics, especially for seismic interpretation. However, there is very little information with regard to its application in the field of magnetic methods. Our research introduces the use of convolutional neural networks for the characterization of magnetic anomalies. The models developed allow the localization of magnetic dipoles, including counting the number of dipoles, their geographical position, and the prediction of their parameters (magnetic moment, depth, and declination). To go even further, we applied visualization tools to understand our model's predictions and its working principle. The Grad-CAM tool improved prediction performance by identifying several layers that had no influence on the prediction and the t-SNE tool confirmed the strong capacity of our model to differentiate between different parameter combinations. Then, we tested our model with real data to establish its limitations and application domain. Results demonstrate that our model detects dipolar anomalies in a real magnetic map even after learning from a synthetic database with a lower complexity, which indicates a significant generalization capability. We also noticed that it is unable to identify dipole anomalies of shapes and sizes different from those considered for the creation of the synthetic database. Finally, the perspectives for this work consist of creating a more complex database to approach the complexity traditionally observed in magnetic maps, using real data from multiple acquisition campaigns, and other applications with alternative geophysical methods. Highlights: YOLO and DenseNet combination allows locating synthetic dipoles, as well as predicting their parameters. Noisy data up to a certain level of complexity do not affect the characterization of synthetic dipoles. Grad-CAM tool identifies unused layers and neurons. t-SNE tool displays an idea of the logic behind the predictions. The complexity of the synthetic database allows our neural network to perform well on real data.