ALBERT

Description: Due to the still enormous burden of unexploded ordnance (UXO) in the subsurface worldwide, the safe recovery of a wide variety of buried weapons and ammunition requires efficient and reliable detection methods. Using a deep learning approach applied to magnetic field data distributed areal along the surface, we aim to achieve a more accurate localization of UXO and small magnetically effective objects in general by detecting the specific signature of their magnetic anomaly. To investigate the applicability of this approach, we developed a deep convolutional neural network that performs image segmentation in different potential measurement scenarios. In this process, the sought small-scale target signals should be distinguished from different background fields containing, e.g., several types of noise. For this purpose, extensive synthetic data sets were generated first using numerical simulations of the magnetic dipole fields of multiple objects. The resulting multi-dipole scenarios and corresponding masks are then passed to the network, which is trained on a test and validation set to produce a representative model for the trained simulation examples. At the end of the training process, this model is supposed to be able to predict yet unknown examples from an evaluation set. Subsequently, the prediction quality of the resulting model needs to be analyzed to fine-tune the parameters of the assumed network architecture or even the architecture itself. The poster we present deals with the generation and preprocessing of appropriate training data, the applied network architecture, and preliminary results of the first evaluation stages.

Description: poster