Date of Award:

5-1979

Document Type:

Dissertation

Degree Name:

Doctor of Philosophy (PhD)

Department:

Electrical and Computer Engineering

Committee Chair(s)

Clair L. Wyatt

Committee

Clair L. Wyatt

Committee

Doran Baker

Committee

David R. Anderson

Committee

James C. Bezdek

Committee

Ronald Thurgood

Abstract

The feasibility of an airborne system to census deer populations using remote sensing techniques is evaluated in this research. The analysis is based upon the study of spectral signatures of deer and various background objects. The deer detection problem is modeled as a multispectral pattern recognition problem where the classification is performed in a multidimensional feature space. A generalized scene, of which the winter range may be considered typical is used in the classification. It includes the deer, evergreen trees, sagebrush, and dry brush against a snow background. The signature stability study led to the conclusion that the direction-angles of the scene-objects are relatively stable and are representative of the objects while the magnitude of the signature vector is more representative of the ambient illumination. Therefore, the classifier design is based upon direction-angles as the multispectral parameter. A general inter-class feature selection procedure, based upon clustering technique, was formulated and used to select best feature sets for the analysis. Bayesian classifiers were designed and tested with a variety of scene composition. The classification performance was estimated by considering the a priori probabilities of "not-deer" objects sufficiently high to eliminate the misclassification of not-deer objects into deer class. The conclusions of this study are: (1) two features are adequate for almost error-free detection of deer in a snow and/or juniper background. (2) Four features are required for the five-class typical generalized scene to obtain a deer detection accuracy of about 60 percent. If only three features are used, the accuracy drops to about 50 percent. (3) Four features provided an accuracy of about 90 percent when the scene is free of dry brush. If only three features are used, the accuracy drops to about 80 percent. The three- or four-feature classifier was shown to be capable of providing useful deer census data in an operational system when combined with appropriate spatial classification techniques. The wavelengths used for the four-feature set are 0.672, 0.764, 0.981, and 0.725 µm. The three-feature set excludes the 0.725 µm wavelength.

Download

Share

COinS

Copyright for this work is retained by the student. If you have any questions regarding the inclusion of this work in the Digital Commons, please email us at .