Submitted to: International Journal of Remote Sensing
Publication Type: Peer reviewed journal
Publication Acceptance Date: 3/8/2007
Publication Date: 7/4/2007
Citation: Miao, X., Ge, S., Gong, P., Anderson, G.L., Carruthers, R.I. 2007. Applying Class-Based Feature Extractopm Approaches for Supervised Classification of Hyperspectral Imagery. International Journal of Remote Sensing. 33 (3): 162-175. Interpretive Summary: Applying Class-based Feature Extraction Approaches for Supervised Classification of Hyperspectral Imagery Hyperspectral image analysis is a complex method of assessing reflectance patterns from aerial remote sensed data, in our case of invasive plant species. Data interpretation is often difficult as the resulting data comes in complex blocks of information representing wide areas in space and multiple channels of wavelenth reflectance values from 400 to 1000 nanometers. Complex mathematics are often necessary to evaluate these data and to make decisions on how to best characterize targets on the ground. This paper discusses detailed methods of data analysis and suggests new and improved ways of assessing hyperspectral images. These methods apply class-based feature extraction methods to help in classify hyperspectral imagery with the outcome being more accurate assessment and differentiation of on-the-ground targets.
Technical Abstract: Land cover classes in the hyperspectral imagery can be roughly modelled as Gaussian pancakes floating in the sparse hyperspace. We try to overcome the curse of dimension by applying class-based feature extraction approaches and compressing the Gaussian pancakes into the corresponding lower dimensional subspaces. Each pixel can then be labelled accordingly based on the conventional classifiers. Principal components analysis (PCA), probablistic principal components analysis (PPCA) and probablistic factor analysis (PFA) are three feature extraction approaches. We apply the corresponding class-based version of PCA, PPCA, and PFA algorithms to find the lower dimensional class manifolds in the training stage, then project each pixel onto these lower dimensional manifolds respectively and assign the class label according to the maximum likelihood decision rule. These class-based feature extraction classifiers prove more efficient than the conventional techniques. Results from simulations and the classification of a CASI 2 hyperspectral image classification show that these algorithms can be extended effectively to other hyperspectral data applications.