A number of classification algorithms have been proposed for PolSAR data. There are basically three approaches: (1) algorithms based on a statistical model, (2) algorithms based on the scattering mechanism of electromagnetic waves, and (3) knowledge-based algorithms. In the first category, classification is done based on the specified probability density function. The second approaches, the classification of PolSAR images are based on some form of target decomposition theories. Third approaches included two steps. First, extraction of knowledge from PolSAR data, and second application of this knowledge to the classification of other pixels. In these approaches it is possible to include scattering model results and common knowledge about the targets.
The main purpose of the proposed method is the knowledge-based and object-based classification of PolSAR data. To improve the classification results, the contextual information should be considered for incorporation into the classifiers. For this, the object-oriented package eCognition was used to implement the object-oriented image analysis of PolSAR images. The multi-resolution segmentation module was used to perform object delineation in this study. Proposed method can apply prior knowledge, expert knowledge and data knowledge in the process of classification. A combination of Support Vector Machine and Decision Tree (SVM-DT) was presented for fusion of three level knowledge. The SVM based binary decision tree architecture takes advantage of both the efficient computation of the decision tree architecture and the high classification accuracy of SVMs. The SVM-DT architecture was designed to provide superior multi-class classification performance. Utilizing this architecture, N-1 SVMs needed to be trained for an N class problem, this can lead to a dramatic improvement in recognition speed when addressing problems with big number of classes. Since DTs are often constructed using a portion of the training patterns to accomplish individual classifications at each node, the node classifiers should be robust in the presence of “bad” samples or outliers. The SVM’s remarkable performance with regard to sparse and noisy data makes them suitable for binary classification trees.
The incorporation of prior knowledge into SVMs is the key element that allows to increase the performance in many applications. In this paper the prior knowledge was used for compensating the unbalanced data in SVM classification. Furthermore, the expert knowledge provides information for designing a decision tree and a feature selection for classification algorithm. Also, the data knowledge was used in various classification steps e.g. features for SVM classification and feature selection. A Radarsat-2 image of the Petawawa forest area including six land cover classes: red oak (Or), white pine (Pw), black spruce (Sb), urban (Ur), water (Wa), and ground vegetation (GV) was chosen for this study. In this research, six experiments were provided for evaluating three level knowledge effect: Wishart, SVM, SVM-DT, object-based SVM-DT, object-based SVM-DT and feature selection, adding prior knowledge to object-based SVM-DT and feature selection. The results show the positive effect to forest classes for adding various knowledge to the classification. Although, the effect of some knowledge on other classes is non-positive. Eventually, the overall accuracy of the proposed method is 87.36. The proposed algorithm outperformed the Wishart classifier by 15% and SVM classifier by 9%.