Discriminating dipole signals from geologically noisy electromagnetic induction data with convolutional neural networks

Historical unexploded ordnance (UXO) contamination is a widespread environmental challenge, leading to human casualties and chemical contamination. Electromagnetic induction (EMI) methods are commonly used to detect unexploded ordnance in both terrestrial and marine settings. Using traditional advanced geophysical classification, UXOs can be discriminated from other metallic clutter via a physics-based inversion approach that matches obtained polarizability curves from EMI data with a library of common UXO polarizabilities. This workflow requires identifying dipoles in the acquired dataset. Non-dipolar anomalies can complicate the identification of targets of interest. Some geological conditions, for example, in areas with strong magnetic soil responses and areas with metallic clutter, make it hard to discriminate between dipolar and non-dipolar anomalies.

In this work, we build on our previously developed convolutional neural network (CNN) that classifies UXO directly from EMI data [1]. Our CNN outputs a probability map that preserves the spatial dimensions of the input. We train the CNN using synthetic data generated with a dipole forward model that considers relevant UXO and clutter objects, and train it to discriminate those dipoles in field data. A key novelty is the interplay between (1) training the CNN to handle the expected noise levels in the field and (2) transferring the CNN to field sites with potential new or "unseen" types of (geological) noise. We demonstrate test procedures required to build trust in machine learning approaches for UXO classification, where false negatives can have a significant impact.

[1] Heagy, L., Lopez-Alvis, J., Oldenburg, D., Song, L.-P., and Billings, S.: Using convolutional neural networks to classify unexploded ordnance from multicomponent electromagnetic induction data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13722, https://doi.org/10.5194/egusphere-egu25-13722, 2025.