Electromagnetic Induction (EMI) methods are successfully used for terrestrial unexploded ordnance (UXO) classification and have shown potential for the detection and classification of ordnance in marine settings. However, there are additional complicating factors for EMI methods in a marine setting: sensor positioning is less accurate than in a terrestrial setting and the conductive seawater and sediments have an EMI response that needs to be accounted for in data processing. Machine learning (ML), and in particular, neural networks, have been demonstrated to be effective tools that complement, or sometimes replace, physics-driven data processing techniques. This project aims to develop a combination of ML and geophysical inversion technologies for the classification of UXO from marine EMI data. The specific objectives are to:

  1. Develop a neural-network based approach for the classification of ordnance directly from EMI data, assuming that the background responses have been removed; this approach will be an alternative to the library-based inversion approach typically applied to terrestrial data and will enable the project team to examine strategies for making data processing and classification more robust to background responses and positional errors.
  2. Design a computationally efficient approach for estimating and removing the EMI response of the background and sediments from marine EMI data and use this to test how inaccuracies in the removal of the background response impact the ML approach in (1).
  3. Examine network architectures and input features, derived from the measured data, that are tolerant of positional errors for classifications from the ML approach in (1).

Technical Approach

This research focuses on development of computational methods and algorithms for marine EMI data. The project team will examine ML and deterministic, physics-driven methods in order to identify where ML can be an effective tool to either complement or replace deterministic approaches for classification of ordnance from EMI data.

For the first objective, the project team will build upon the early-stage work that demonstrates the potential for using convolutional neural networks (CNN) to classify ordnance directly from terrestrial EMI data. Work thus far has been primarily on synthetically generated data, and the next step is to apply this approach to field data sets. The project team will begin by examining well-characterized terrestrial sites so that they can assess the impact of a range of environmental factors and noise sources, and develop strategies that are robust to a range of these factors.

To address the second objective, the project team will use and further advance the open-source inversion software, SimPEG, for estimating a one-dimensional (1D) conductivity model from EMI data. The project team will compare two approaches: (a) a standard 1D inversion that treats each sounding independently and (b) a spatially constrained approach that promotes continuity of structures between individual soundings so that a smooth three-dimensional model is obtained. To assess the effectiveness of each, the project team will use synthetic datasets and apply the ML approach developed in Objective 1.

For the final objective, the project team will attempt to design strategies for the CNN approach in Objective 1 that are robust to errors in sensor positioning. They will begin by testing how the CNN performs with increasing positional uncertainty. They will then explore if using different input features, for example, ratios of measured data improves classifications.


The current approach for working with marine EM data involves multiple steps to deal with instrument orientation, conductivity background and final classification. ML has the potential to address each of these challenges individually or perhaps jointly, but research in this area is still in early stages. The research will illuminate where, and under what conditions, an ML approach might aid or replace a deterministic approach currently used. It will therefore provide a strong foundation for future development of ML approaches to finding ordnance items in a marine environment.