A study is described in which optimal textural and spectral features are selected for scene segmentation. A set of 46 textural features and 3 spectral features were available for image classification. A method was developed which used a thresholded separability measure to select the best features for scene segmentation. The measure was based on the Mahalanobis distance between class means. The optimal feature selection process was applied to a variety of images and classification results using 4 features ranged from 91% to 97% with independent data sets. The use of the thresholded Mahalanobis-like distance for optimal feature selection was compared to the more common thresholded divergence separability measure and was found to choose features which were equally good for classification. The Mahalanobis-like measure had the additional advantage of using only 1/6 the time needed to calculate the divergence measure.
Library of Congress Subject Headings
Remote sensing--Data processing; Image processing--Digital techniques; Imaging systems--Image quality
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Rosenblum, Wendy, "Optimal selection of textural and spectral features for scene segmentation" (1990). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus