Spectral Comparison Using k-Means Clustering

There is a growing number of infrared (IR) spectral signature data in the scientific community gathered from a variety of sensors using a variety of collection techniques. As the quantity of collected data grows, automated solutions for searching and matching signatures need to be developed. When searching and matching signatures, reducing computational complexity and increasing matching accuracy are essential. We present a signature classification method via k-means clustering using a novel application of spectral angle mapping to efficiently determine spectral similarity. We evaluate the method against spectral data in the "SigDB" spectral analysis software application developed by the Johns Hopkins University Applied Physics Laboratory (JHU/APL). The key component to this approach is the set of characteristic functions used to map signatures' similarity into a spatial representation. Existing methods used to autonomously identify and classify IR spectral data include spectral angle mapping and key feature detection. Spectral mapping is computationally slow due to the need for direct individual comparison, and key feature detection improves computation time but is limited by the specific features selected for comparison. The accuracy and computation time of the spectral cluster classification method is evaluated against spectral angle mapping and visual analyses on the ASTER NASA spectral library. The goal of this method is to improve both the accuracy and speed of classifying newly collected unlabeled spectra. We find that the proposed method of scoring signatures offers a speed increase of three orders of magnitude in comparing spectra at the expense of a high false positive rate, suitable for use as a first-pass filter. We further find that the k-means cluster-based classification is highly sensitive to the selection of initial cluster centroids, and offer alternative solutions to use with our scoring method.