Bayesian-hierarchical SAR classifier

This research combines three existing statistical tools (optimal Bayesian techniques, maximum entropy density estimation, and hierarchical classification techniques) to develop a practical, robust technique for classifying short vegetation, which we call the Bayesian-hierarchical classifier. We apply this technique to real SAR data, using it to identify five types of short vegetation. It yields high accuracies across an entire growing season, despite the natural fluctuations of vegetation dielectric constant and structural characteristics due to changes in each plant type as it matures. We also show the results of two existing classification techniques, the ISOCLUS unsupervised clustering technique, and a simple maximum likelihood estimator with Gaussian assumptions, on the same data set used to develop the Bayesian-hierarchical classifier. All classifications and comparisons are done on the basis of independent testing and training data, an important criterion to ensure accurate evaluation under real world conditions. The Bayesian-hierarchical classifier has an average accuracy of 95% on the diagonals of the confusion matrix, improving over the other techniques by between 10-20%