Electromagnetic induction (EMI) methods are commonly used to classify unexploded ordnance (UXO) in both terrestrial and marine settings. Modern time-domain systems used for classification are multicomponent which means they acquire many transmitter-receiver pairs at multiple time-channels.

Traditionally, classification is done using a physics-based inversion approach where polarizability curves are estimated from the EMI data. These curves are then compared with those in a library

to look for a match based on some misfit. In this work, the project team developed a convolutional neural network (CNN) that classifies UXO directly from EMI data. Analogous to an image segmentation problem, the CNN outputs a classification map that preserves the spatial dimensions of the input. In this way, the CNN produces high-resolution results and can handle the multiple transmitter-receiver pairs and the per-line acquisition of multicomponent systems. The project team trains the CNN using synthetic data generated with a dipole forward model considering relevant UXO and clutter objects. A careful design of the clutter classes is needed to maximize clutter discrimination. A physics-based parameterization of the clutter class is used to maximize clutter discrimination. The approach was tested using field data acquired with the UltraTEMA-4 system in the Sequim Bay marine test site.

It was found that the CNN results can be affected by spatially and temporally correlated noise remaining in the preprocessed data. Including this systematic noise in the training dataset was crucial to improving the classification results for the field data. To tackle this issue, a two-step workflow was used. First, a CNN was trained to detect metallic objects in field data. From this, patches of data that do not contain metallic objects were extracted and used to generate a new training data set where patches containing only "background signal" were added to the synthetic data. A second CNN was trained with these data to perform the classification. Using this workflow, classification results for the field data show that the approach detects all UXOs and classifies more than 90% as the correct type while also discriminating 70% of the clutter.

A key advantage of the CNN is that, once trained, it may be used to provide real-time classification results on the field.