Neural network interpretation of high-frequency electromagnetic ellipticity data

This paper focuses on the detection of three dimensional (3D) small conductive targets from high-frequency electromagnetic ellipticity data using neural networks. For environmental investigations it is necessary to provide as much information on the location of shallow buried conductive objects or the electrical properties of possible contaminants. The networks are trained with one-dimensional (1D) forward models to estimate the resistivity and dielectric constant structure of the ground. The input is given by ellipticity sounding curves from eleven discrete frequencies in binary steps in a range from 32 kHz to 32 MHz. Halfspace and layered earth neural networks will provide reasonable fit to sounding curves even if they are influenced by shallow conductive 3D objects. We show that a detailed inspection of ellipticity profiles over targets such as a 5 m by 3 m aluminum sheet (depth of 1 m), a 55-gallon barrel (depth of 0.63 m), and two metal desk (depth of approximately 1 m) can help to detect these anomalies. Piecewise halfspace neural network are capable of enhancing the anomalies in resistivity depth sections and provide additional information for the detection and possible localization of the object. The visualization of the results is very important since small targets will show up as subtle anomalies. Based on observations of ellipticity sounding curves and profiles we can train a neural network to classify target responses versus background responses for specific sites, assuming that enough soundings are available to train the neural networks.