Continuous-output neural networks for EM ellipticity pattern recognition

The interpretation of images of magnetic field polarization (ellipticity) acquired with a geophysics electromagnetic (EM) imaging system is treated as a pattern-recognition problem. A continuous-output backpropagation network is presented with images from a target in various locations and is taught to associate the spatial location of the target with the pattern in the image. Five different data representations were examined for training speed, accuracy, and generalization capabilities. The results are shown to be relatively insensitive to network design, but the overall errors decrease as the the size of the input vector decreases. The network located the target in the field data within 3% of the horizontal location and 1.2% of the depth.