PolSAR Terrain Classification Based On Ensemble Learning And Feature Selection

With the continuous development of Remote Sensing, the requirement of the performance of Remote Sensing detection becomes increasingly demanding. With more complex issues to be addressed, the traditional radars can not fit for the demand of higher performance. Polarimetric synthetic aperture radar(Pol SAR) has been launched as a new radar which can provide more terrain scattering information. Because Pol SAR can provide more information than SAR, it becomes a popular research in the field of remote sensing in recent decades. Moreover, Pol SAR image classification is considered as a typical application of Pol SAR data processing, which can use samples, features and classifiers to separate differen terrain types. Currently, Pol SAR image classification has been successfully applied to many practical problems.For Pol SAR image classification, samples, features and classifiers are three improtant factors which affect the final classification result. Pol SAR image data contains more information and more features than SAR image data. However, it brings more difficult to choose proper samples and features and also gives more burden on computation complexity. The samples and features with redundant, similar or harmful information will jepardise the final classification result. In order to get better image classification results, this paper concentrates on samples selection, features selection and classifiers designing. We combine some classical algorithms, such as ensemble classifiers and Fisher linear discriminant together, which are shown as follows:1. For the supervised PolSAR image classification, the selection of the training samples may affect the final classification results and there are no specific rules about how to select proper training samples at present. Hence, in order to effectively deal with that problem, a novel supervised polarimetric synthetic aperture radar image classification method is proposed based on pruning ensemble classifiers, by which the selection of training samples are converted into the selection of classifiers. In the proposed method, firstly, the training samples are randomly obtained from each class of Pol SAR image, and they are divided into several blocks which are respectively used as the training subsets to produce individual classifiers of an ensemble. Then the training subsets which are harmful for Pol SAR image classification are wiped off via eliminating theindividual classifiers corresponding to them based on a pruning ensemble classifiers algorithm. In particular, matching pursuit optimization ensemble classifiers algorithm is employed to prune classifiers of an ensemble, and support vector machine is used as the basic classifier algorithm. The experimental results demonstrate that the proposed method obtains better classification results compared against others methods.2. Aiming to deal with the problem that for Pol SAR images the pixels in the same class may have different appearance because of the topographical slopes and the radar look angle. To improve the image classification result, two supervised polarimetric synthetic aperture radar image classification methods are proposed based on Naive Bayes Combination(NBC) and Weighted Majority Vote(WMV) ensemble. In the first proposed method, the Naive Bayes Combination is adopted to learn on different training samples to get classification surfaces in order to improve the classification results. Firstly, we extract some features and choose some pixels as the original training samples for the classification, and randomly divide the training samples into several training sample subsets. After that, the frame of Naive Bayes combination is obtained based on the training sample subsets. Finally, Naive Bayes Combination gives the final classification results. Support vector machine is used as the basic classifier algorithm for constructing the Naive Bayes Combination. For the proposed Pol SAR image classification method based on Weighted Majority Vote ensemble, Firstly, the features are extracted from the Pol SAR data, and several groups of pixels in one class are chosen as the training sample subsets. After that, the component classifiers learn on different training samples give the predictive labels for the pixels, and the weights are calculated on these labels. Finally, the predictive labels are combined together to get the final classification result. The experimental results demonstrate the effectiveness of the two proposed methods on AIRSAR and Radarsat-2 data.3. How to use the features is very crucial for the PolSAR image classification, however, there are still no specific rules for it. For solving the aforementioned problem, a supervised Polarimetric SAR image classification method based on weighted ensemble via 0-1 matrix decomposition is proposed. This proposed method adopt matrix decomposition ensemble to learn on different feature subsets to get coefficients to improve the final classification results. Firstly, we extract some features from Pol SAR data as initial feature group and randomly divided it into several feature subsets. After that, according to ensemble algorithm to get the different weights based on the featuresubsets, small coefficients are assigned to bad classification results to decrease the harmful affection of some features. The final classification result is achieved by combining the results together. The experimental results of L-band and C-band Pol SAR data demonstrate the effectiveness of the proposed method.4. A feature selection method and a feature weighting method via Fisher linear discriminant for polarimetric synthetic aperture radar(Pol SAR) image classification have been proposed. Appropriate feature selection is essential for obtaining accurate classifications, but so far has remained an open research problem. We propose two supervised three-component feature selection methods based on the Fisher linear discriminant. Fisher linear discriminant method is used to calculate paramemters for the features. Then these paramethers are modified according to a three-component scattering power decomposition model, combining both physical and statistical scattering characteristics to adapt them for the particular scattering mechanisms inherent in Pol SAR data and assigned to the coherency matrix to enhance the discriminating ability of the features. Freeman decomposition and Wishart classifier are used to classify the Pol SAR image. The effectiveness of the proposed methods are demonstrated by experiments using NASA/JPL AIRSAR L-band and CSA Radarsat-2 C-band Pol SAR images of the San Francisco area.5. An improved three-component model-based scattering decomposition method by implementing the new volume scattering model for the polarimetric SAR image is presented. The modified volume model which is suit for the target decomposition of Pol SAR data is used in the proposed method aiming to reduce the emergence of negative powers in the Freeman decomposition. The experimental results of NASA/JPL AIRSAR L-band, CSA Radarsat-2 C-band and DLR ESAR L-band data illustrate that the proposed method can effectively reduce the number of negative powers, which demonstrate the effectiveness of the proposed method on Pol SAR data.