Comparison of Classifiers for Radar Emitter Type Identification

ARTMAP neural network cla.c=;sifiers arc considered for the identification of radar emitter types from their waveform parameters. These classifiers can represent radar emitter type classes \vith one or more prototypes, perform on-line incremental learning to account for novelty encountered in the field, and process radar pulse streams at high speed, making them attractive for real-time applications such a ... s electronic support measures (ESM). The performance of four ARTMAP variantsARTEMAP (Stage 1), ARTMAP~IC 1 fuzzy ARTMAP and Gaussian ARTMAP is assessed with radar data gathered in the field. The k nearest neighbor (kNN) and radial basis function (RDF) cla.<;Sifi.ers are used for reference. Simulation results indicate that fuzzy ARTl\·'lAP and Gaussian ARTMAP achieve an average classification rate consistently higher than that of the other ARTMAP cla.<>sifiers 1 and comparable to that of kNN and RBF. ART-EMAP1 ARTMAP-IC and fuzzy ARTMAP require fewer training epochs than Gaussian ARTMAP and RBF, and substantially fewer prototype vectors (thus, smaller physical memory requirements and faster fielded performance) than Gaussian ARTMAP, RI3F and kNN. Overall, fuzzy ART MAP performs at lca.<;t as well a.<;; the other classifiers in both accuracy and computational complexity, and better than each of them in at least one of these aspects of performance. Incorporation into fuzzy ARTMAP of the MTfeature of ARTMAP-IC is found to be essential for convergence during on-line training with this data set.