Application of an Artificial Neural Network for Airborne Magnetic Data Discrimination
Jeannie Norton, Battelle, Oak Ridge, Tennessee
Jacob Sheehan*, Zonge International, Denver, Colorado
Les Beard*, Zonge International, Tucson, Arizona
Paper – [pdf] UXO_SAGEEP_ANNs_for_ordnance_LPB_2011-UXO
The objective of this work was to develop and apply an artificial neural network (ANN) to discriminate ordnance from non-ordnance based on input derived from airborne vertical magnetic gradient data. The project assessed whether more output classifications are beneficial
and/or required for effective and consistent discrimination, e.g. large ordnance, small ordnance, scrap, and geology. While it may ultimately be possible to determine ordnance type using ANNs, for this project we got best results with the funda-mental classification scheme: UXO or not UXO.
Artificial neural networks (ANNs) are computational constructs that attempt to mimic the workings of the human brain with respect to the brain’s ability to detect patterns. ANNs can be trained to determine which class an object belongs based on selected inputs. The inputs are weighted according to their relevance in determining the class of an object. This process is shown in Figure 1 for a simple one layer neural network. The set of weighting values (w) is determined by setting up a series of linear equations in which the inputs and outputs are known, and then finding the solution for w in a least squares sense. Once the weights have been established, the ANN can be tested on other data to see if the predicted output matches the known data class (Lawrence, 1994).
In our specific case, we wanted to predict whether the source of a magnetic anomaly was produced by ordnance. The output classes fell into one of two categories. Either the item of interest was or was not unexploded ordnance. This latter category included geology, fragments from exploded ordnance, and non-ordnance items, for example: tools, bailing wire, automobile parts. The class of ordnance only included ordnance items that were 40mm or larger. The input consisted of anomaly amplitudes and parameters derived from least squares inversion of the magnetic data for a single dipole source—e.g., dipole magnetic moment, depth of burial of the source, orientation of the source dipole, and the least squares goodness of fit.