An improved object oriented classification scheme for polarimetric SAR image

The process of object oriented classification is to segment the image first to get different objects, and then classify the objects, the result of segmentation directly affects the accuracy of final classification. For this problem, an improved object oriented classification method for PolSAR image is proposed in this paper: firstly estimate each object by calculating the percentages of different classes of all selected pixels within the object, if all percentages within the object are below a certain threshold, it can be considered that this object has deviation on segmentation, so pixel-based classification is conducted on the object; otherwise, object oriented classification is proceeded on the object, the results of pixel-based and object oriented are combined at last. Improved classifier dynamic selection algorithm is used as the classification algorithm to fulfill the decision-level fusion on three base classifiers with huge diversity including Wishart, kernel-KNN and Wishart-KNN. PolSAR images of AIRSAR and EMISAR are used as experimental data for the classification experiments, and the results show that improved object oriented classification algorithm achieves the highest performance on accuracy, it can take full advantage of merits from pixel-base and object oriented classification algorithms and can be applied well in practice.