<title>Unsupervised/supervised hybrid networks for identification of TSS-1 satellite</title>

Neural networks have potential advantages such as real-time operation and robustness based on their parallel structure, self-organization, fuzziness, and particularly their adaptive learning ability. A single neural network is useful for identification of objects. To carry out identifying complex objects, however, it is necessary to consider hybrid architectures of two or more networks, which offer some degrees of improvement in performances. In this paper, neural learning techniques, the self-organizing feature mapping (SOFM), and learning vector quantization (LVQ2) have been applied to the automatic target recognition problem in the presence of a satellite object with high level noises. SOFM, unsupervised learning captures the homogeneity within-class characteristics; whereas LVQ2, supervised learning captures the heterogeneity of between-class.