Development of Semi supervised Classifier for SAR Image Pattern Recognition

In this paper, new semi-supervised classification method has been developed for polarimetric SAR images. The developed method has been used for classification of various land covers such as tree, grass and land. The present work is primarily devised upon the fusion of unsupervised learning ($k$-means) and supervised learning classifier (support vector machine). Our proposed method requires less training data samples as compared to the supervised method (SVM) alone and yields significant accuracy as 95.833% using 65 labelled training data samples. Further, the analysis of the relationship between accuracy and a various number of labelled training data samples have been observed for the selection of the optimal range of accuracy. This analysis shows that the accuracy of more than 90% can be achieved even with the low training data-set. Finally, the visual analysis has been done, which also supports the classification results from the computational analysis.