Hyperclass: An Unsupervised Classification Method for Hyperspectral Remotely Sensed Data

Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)



First Advisor

Temple Fay

Advisor Department



Hyperspectral remotely sensed data sets can contain hundreds of spectral bands of data. Multispectral data sets, on the other hand, contain less than ten spectral bands of data. Traditional multispectral classification tools don't perform well on hyperspectral data sets because of the computational intensity required in using all spectral bands. In order to be effective, the number of bands needs to be reduced while maintaining information content in the data. This research proposes a group of processes that perform quick and efficient unsupervised classification of hyperspectral data. These processes perform noise band detection, dimensionality reduction and clustering/classification of the data. The quality of the results of hyperspectral data classification can be affected by noise. Noise can be due to either sensor malfunction or poor signal to noise ratio. A noise detection process was developed that indicates which bands contain too much noise to be effective as input to a classification process. Neural network and deterministic Principal Component Analysis (PCA) techniques were compared to determine which method should be used for hyperspectral data dimensionality reduction. The Iterative Self Organizing Data (Isodata) and the Self Organizing Feature Map (Sofm) clustering techniques were used to cluster and classify the data. The Sofm was extended to incorporate several different cluster competition measurement functions. The Sofm was also extended to perform automatic merging of clusters during the final classification phase. All these functions were implemented in software and were integrated into one governing process called HyperClass. Landsat, Rdacs and Aviris data sets were analyzed using the HyperClass process. The processing completed quickly and produced meaningful clustering of data. The data from a soils agriculture project was analyzed using the HyperClass process. The results showed that HyperClass is a very effective tool for hyperspectral data classification.