Classification of short vegetation using multifrequency SAR

The focus of this investigation is to develop an algorithm which is able to classify ground cover (in this case short vegetation) based on structural characteristics, using remotely sensed radar data. This investigation is an extension of work done in the Radiation Lab and presented in Y. Kouskoulas et al. (1998). The authors are interested in developing a computationally efficient classifier, which works well for generalized, non-Gaussian distributions. They used a supervised approach, and although the technique is general enough to be used with many kinds of data, they trained and tested it with SAR data. The final algorithm uses polarimetric radar data at two frequencies (L and C), extracted from SAR scenes taken by the AirSAR platform during the months of May, June and July in the summer of 1995. Their supervised method constructs closed shells that exist in that multidimensional space and surround their classes. The classifier was trained on one data set, and achieved over 90% accuracy with five classes in classifying the independent testing data set. The results and methodology are compared to an unsupervised method and discussed