Use of Eigenvector-Generated Scatter Plots in Clustering Image Data

Over the last few years we have been analyzing state-of-the-art spectral temporal data of many events. Our goal was to develop specific techniques to classify and identify events based on these measurements. While the techniques evolved from one data type, we focus in this paper on the technique itself and its potential efficacy when applied to other data types. We use a Singular Value Decomposition SVD technique to cluster like events by forming a scatter plot from the first two eigenvectors. An evaluation of this approach using real data as well as simulations is given. A novel technique is introduced to assess cluster stability in the absence of ground truth. Results are presented along with the effects of misalignment of data samples, compression, training sets, and classifiers. The overall methodology is quite powerful and has remarkable noise immunity.