Electromagnetic Data Evaluation Using A Neural Network: Initial Investigation-Underground Storage Tanks

Environmental and engineering studies utilize multiple methods of investigation with an<br>integrated geophysical approach. This complicates data interpretation because while integrating<br>data collected by a single method is no simple task, that of integrating multiple data types is<br>even less so. Traditionally, simple analytical and numerical models have been used to interpret<br>geophysical anomalies. However, interpretive models for the types of objects frequently<br>encountered at environmental restoration sites have not been available. Simulating<br>environmental targets as mineralized ore bodies has been our only available capability. This is<br>neither cost-, nor labor-effective, and is wholly inappropriate for cultural artifacts.<br>Artificial intelligence neural network concepts can be applied to these processes of discriminating<br>anomalies of interest from the host matrix, Neural networks are taught the defining response<br>parameters for a particular archetype. The trained network is then used to interpret geophysical<br>data anomalies relative to the response characteristics of the archetype.<br>For our investigation, frequency-domain electromagnetic (EM) data from known test<br>configurations were used to train a neural network to discriminate underground storage tanks.<br>The trained neural network was then used to “interpret” EM data collected at Hickam Air Force<br>Base, Hawaii. The results were compared with known “ground truth” excavation information<br>to determine the accuracy of the performance of the neural network.<br>The program demonstrated that artificial intelligence concepts can facilitate interpretation of<br>geophysical anomalies. In addition, the neural network offers the potential of truly integrating<br>collected data sets. The current state-of-the-art of integrated interpretation is to overlay separate<br>interpretations of the multiple methods employed and derive a “best fit” conclusion from this<br>collage. The neural network provides a means of integrating the separate data sets in either a<br>multiple (parallel), or in a step-wise (serial) integral response.