LARS Tech Report Number
060772
Abstract
The standard procedure at LARS for classification analysis of aircraft scanner data has been to training samples (fields) based on the ground truth data and to use these directly or in a form subdivided via clustering as training for the computer. In this process the major emphasis is on defining class statistics which describe actual materials of interest as designated by the user. The approach assumes these materials are spectrally separable and classification is performed to test if this is true. In the approach described here the multispectral data is first clustered to determine what spectrally separable groups exist in the data. After these separable groups are found they are related to their physical meaning or "ground truth". This approach is in a sense the reverse of ground truth oriented training procedure referred to above.
There are two requirements often cited for multispectral pattern recognition to be useful for a particular material in the scene. First it must be spectrally separable from all the others in the scene and, two, it must be of informational value. In the existing approach the features which are of informational value are first defined. In the cluster oriented approach the spectrally separable classes are found first and time consuming training and test classification analysis to determine the spectral separability of unseparable materials is avoided.
Date of this Version
January 1972