Comparison Between Genetic Programming and Neural Network in Classification of Buried Unexploded Ordnance (UXO) Targets

摘要

Ground penetrating radar (GPR) has been utilized for detection and classification of unexploded ordnance (UXO) for both civilian and military purposes for many years [1]. UXO classification processes of the GPR technology often involve complex qualitative features such as 2D scattering image and are performed subjectively by human operators. Thus, inconsistent and subjective classification performance associated with human factor such as fatigue, memory fading, learning capability or complex features are clearly inevitable.[2] In order to overcome this issue, an automatic, objective classification method that does not require a trained operator is essential. Artificial intelligence (AI) technologies, such as Neural Network (NN) and Fuzzy system, have been investigated and applied to develop autonomous classification algorithms for UXO and land mine detection and showed promising results. [2,3] Recently, genetic programming (GP), which is relatively new method of the AI techniques, has been introduced for classification [4]. A preliminary comparison between the effectiveness of the NN and the GP techniques in a classification point of view has been conducted using images of alphabet characters [5]. From this study, GP showed better performance than NN in the various levels of problem difficulty. GP also provided robustness to untrained data, which caused difficulties in the case of the NN. In this paper, we present the results of our next step effort in comparison of classification performances between the NN and the GP techniques based on the simulated scattering patterns of UXO-like object and non-UXO objects. For this comparative study, 2 dimensional scattering images from one UXO target and four non-UXO objects were generated by numerical simulation tool (FEKO). For non-UXO objects, the most challenging targets to discriminate from UXO, since all these objects produce resonance signal as UXO-like targets do [6], were selected. Classification performances of both techniques (NN vs. GP) in different level of noise and in the case of presence of untrained data were examined and the results and observations are discussed.