A new methodology is presented for classifying remotely-sensed imagery. This technique is meant to be locally-adaptive, supports non-gaussian statistics, and still allows one to generate an automatic classification. The new methodology requires training data, just as the standard technique does, but it uses an unsupervised technique (ISODATA) with which to first classify the data. The clusters from the unsupervised step are used with the training data in a supervised classification to get the mapping from cluster to class. Often, the statistics of the classification procedure are ill-conditioned for large feature spaces, and so this new methodology is designed for multi-step classifications. The idea is for the analyst to break up the classification into two or more steps where more general classes are separated first. The automated procedure then determines which subset of all the features are necessary at each step of the process. At the moment this is implemented using an exhaustive search stategy, but other methods are possible and will be explored. The resulting classification reports which channels were important at each stage of the classification process, thus automating the first step in understanding how and why the classification process works. In combination with a simple, unsupervised segmentation algorithm, which is also presented, this technique is then applied to SIR-C/X-SAR data.
展开▼
