Landmine detection and classification through hybrid neural networks and fuzzy set

Unlike conventional weapons, the buried landmine, if not removed, remains a threat. The clearance of anti-personnel landmine has become an important issue in military, economic, and especially humanitarian concerns worldwide. Traditionally, the landmine clearance is conducted by human expert or animal. With the advance of technology, various detection and classification strategies and sensors have been proposed to provide better solutions. The research conducted in the areas of landmine detection is voluminous. This work seeks to focus on improving the detection and classification of landmine by applying neural networks and fuzzy set theory. The phase information in scattering parameters dataset measured by separated aperture sensor and ignored in previous research is shown to be valuable with proper selection of neural network model capable of handling complex-valued data. By using the complex-operator and complex chain rule, a simplified derivation of the back-propagation learning method is constructed. A two-stage hybrid neural network architecture, combining a complex-valued feedforward neural network and a self-organizing feature map, is proposed to classify the types of landmine by integrating fuzzy information at pixel level. From the proposed architecture, a new technique using self-organizing feature map to elicit fuzzy membership function for classification from data with labelling information is discovered.