Skip to Main Content
Compressing hyperspectral data such that class discrimination is maintained is a difficult task. In a supervised classification scenario, one has hope of meeting this goal since all information needed to classify the data is present. The challenge then is to determine which subset of features are important for classification and which are not. We summarize our generalized relevance learning vector quantization-improved (GRLVQI) , based on Hammer and Villmann's GRLVQ , for the joint classification and feature extraction of hyperspectral data. Our previous results show exceptional classification and feature reduction results on 23 and 35 classes of a real hyperspectral data set , . We present a scalable architecture for GRLVQI targeted for hardware implementation enabling hyperspectral- based real-time decision making functions. Implementing the GRLVQI system with floating point gate array technology likely provides a means of developing a partial reconfigurable GRLVQI system that allows one add or remove classes on-the-fly without adversely affecting the current state of the classifier or feature extractor.