A fully convolutional neural network to classify UXOs using multicomponent electromagnetic induction data
Unexploded ordnance (UXO) clearance relies on electromagnetic induction (EMI) methods to detect and classify buried and submerged metallic objects, but the standard classification workflow requires physics-based inversion and manual expert review. We present a fully convolutional neural network (CNN) that classifies UXO directly from multicomponent time-domain EMI data, bypassing the inversion step entirely. The CNN uses a segmentation-like architecture that preserves the spatial dimensions of the input and outputs a high-resolution classification map suitable for generating a diglist. Two CNNs are used in sequence: the first detects anomalies due to metallic objects and characterizes the spatially correlated noise present in the preprocessed field data; the second performs multi-class classification using synthetic training data augmented with that noise. A physics-based parameterization of the clutter class is used to train the network to discriminate clutter from UXO across a broad range of object types. We validate this approach on two field datasets acquired with the UltraTEMA-4 system at the Sequim Bay marine test site. On a calibration grid with 21 seeded objects, the CNN detects all UXOs and discriminates two out of four clutter objects. On an independent blind grid with 35 seeded objects (17 UXOs, 18 clutter), the CNN detects and classifies all UXOs and achieves a false-positive count comparable to that of the physics-based inversion approach. We see the two methods as complementary: once trained, the CNN classifies rapidly and with minimal supervision, making it well suited as a fast first-pass tool or as an independent check on physics-based results.