PolSAR image classification based on polarimetric decomposition and ensemble learning

Abstract: An effective PolSAR image classification technology is the basis of the successful application of PolSAR. The advantage of PolSAR data lies in its rich polarimetric information. How to make full use of the polarimetric information of PolSAR data for classification has always been a hot issue in the research of PolSAR image classification. To better maximize the use of the polarimetric information of PolSAR data for classification to further improve classification accuracy, this study proposed a PolSAR image classification method that was based on polarimetric decomposition and ensemble learning. In this study, the study area was located in the south of Changchun, Jilin, China, and a RADARSAT-2 Fine Quad-Pol image was used as data sources. 1) All the polarimetric decomposition methods provided by the PolSARpro_v5.0 software were used to extract polarimetric parameters for classification support. These decomposition methods were the Pauli, Krogager, Huynen, Barnes, Cloude, H/A/α, Freeman2, Freeman3, Yamaguchi, Neumann, Touzi, Holm, and Van Zyl methods. 61 polarimetric parameters were extracted using these polarimetric decomposition methods and then merged to form a multichannel image. 2) The multichannel image was divided into numerous image objects by implementing multiresolution segmentation. 3) Against each polarimetric decomposition method, features were extracted from the multichannel images. 4) The PSO_SVM wrapper algorithm was applied for feature selection. 5) The land-cover classification was performed by a support vector machine (SVM) classifier for each polarimetric decomposition method. Based on the samples in the validation group, against the classification results of each polarimetric decomposition method, a confusion matrix from which 4 statistics (producer's accuracy, user's accuracy, overall accuracy, and Kappa coefficient) were obtained was established for assessing classification accuracy. The 16 polarimetric decomposition methods were ranked in the following descending order according to their Kappa value, including Yamaguchi4, Pauli, H/A/α, Touzi, Neumann, Freeman3, Van Zyl, Yamaguchi3, Freeman2, Krogager, Cloude, Holm1, Holm2, Huynen, Barnes2, and Barnes1. Finally, the differences of all combinations of the 9 individual classifiers (Pauli, H/A/α, Freeman2, Freeman3, Yamaguchi3, Yamaguchi4, Neumann, Touzi, and Van Zyl) with Kappa values greater than 0.60 were evaluated using the entropy measurement method. Among all combinations, the combination consisting of the classification results of 6 polarimetric decomposition methods, namely, Pauli, Freeman3, Yamaguchi4, Neumann, Touzi, and Van Zyl, had the highest entropy value of 0.282 7. According to the criterion in which the greater the entropy value was, the greater the differences were, this combination was selected for integration. The classification results of Pauli, Freeman3, Yamaguchi4, Neumann, Touzi, and Van Zyl were synthesized using parallel integration mode and weighted voting method. After integration, the overall accuracy and the Kappa value were 92.49% and 0.90, respectively, which were 4.49% and 0.06 higher than the overall accuracy and the Kappa value of Yamaguchi 4 with the highest accuracy before integration, respectively. Moreover, the producer's and user's accuracies of land-cover classes improved in general. To verify the effectiveness of the proposed method, this study compared it with other classification methods based on many polarimetric decomposition methods. A relatively low overall accuracy and Kappa value were obtained by the compared method, and their values were 1.75% and 0.02 lower than those of the proposed method, respectively. The comparison results indicated that the proposed method exhibits better performance for the PolSAR data classification. This study would provide a new concept for the classification of PolSAR data. Moreover, it would expand the construction method of the "diversity" in ensemble learning.