An application of divergence measurement using transformed video data

Publisher Website
Google Scholar
Abstract
Multispectral aerial video data (0.42–0.43 μm, 0.52–0.55 μm, 0.64–0.67 μm, and 0.85–0.89 μm) with 0.13 meter resolution were collected over test plots of cotton, sorghum, cantaloupe, soil, pigweed, and johnson‐grass on May 31 and July 24, 1983 near Weslaco, Texas. These data were transformed into four principal component (PC) bands for each date and used to classify the six features. The classification accuracy of individual features and associated omission/commission errors were determined. Classification accuracy characteristics associated with different numbers of PC bands were analyzed using divergence measurements derived from class training statistics. Evaluation of PC transformed data divergence indicated that: (1) divergence identified the same number and type of PC bands needed for successful feature discrimination as was determined through classification, (2) correlations between omission/commission misclassification errors and divergence values averaged ‐0.92 for all classifications conducted, (3) correlations between classification accuracy and divergence averaged 0.85 for all classifications conducted. Divergence measurements derived from transformed data are potentially valuable as a guide for feature selection in classification.

This publication has 2 references indexed in Scilit: