Convolutional neural networks for the characterization of magnetic anomalies

Summary During the last decade, deep learning architectures demonstrated to be very promising in geophysics, notably in the field of seismic interpretation. Their great power and flexibility may produce valuable contributions to finding solutions to a diversity of geophysical constraints, for example, the limitations of inversion algorithms or the fine-tuning often required on the different noises that affect the data. In our paper, we use convolutional neural networks to characterize magnetic geological bodies. This characterization consists of counting the number of dipolar magnetic anomalies produced by these bodies and predicting their corresponding parameters. After several experiments with deep learning architectures, we found that the combination of YOLO and DenseNet achieves excellent performance by reaching an accuracy of over 90% in their respective metrics. We also applied visualization tools to explain our results. We managed to visualize the discriminative area of our network with Grad-Cam, and we observed the "logic" of the predictions by using t-SNE. In future studies, we will evaluate the robustness of the proposed approach using real data, for example, in the context of pyrotechnical detection for unexploded ordnance exploration.