An Improved Classification Scheme with Adaptive Region Growing and Wishart Classification Algorithm for Digital Images

This paper proposes a new ARGWishart(Adaptive Region Growing-Wishart) classification algorithm for digital images. It integrates the adaptive region growing algorithm and Wishart maximum likelihood classification algorithm for difficulties that arise from selecting training samples, and instability of the final classification accuracy on PolSAR(Polarimetric Synthetic Aperture Radar) image. At first, the main diagonal elements of the polarimetric coherency matrix are extracted from the image. A few are presentative and clear sample data are selected manually. An improved adaptive region growing algorithm is developed by introducing the OTSU method for the image. After this, we can obtain enhanced training samples by which each class center can be calculated. Then, unknown samples are classified by the Wishart maximum likelihood classification algorithm. Thus, we can finish the classification on the polarimetric SAR image. For comparison, we use PALSAR(Phased Array Type L-Band Synthetic Aperture Radar) and RADARSAT-2 images as data sources and apply SVM(Support Vector Machine), the Wishart supervised classifier, and the ARG-Wishart classifier to conduct comparative research on two experimental plots, namely, Lishui County in Jiangsu Province and Jiangning District in Nanjing City. The results show that the ARG-Wishart classifier can achieve better results than the SVM and the Wishart classifier in terms of overall classification accuracy and Kappa coefficient.