A New Method for Contextual Classification of Polarimetric SAR Data Based on Combining SVM and MRF

Forest species classification has become increasingly important due to changing environmental conditions and the need to sustainably manage forest resources. They have an important role in the global hydrological and biochemical cycles. Remote sensing has been shown to be valuable for forestry applications such as monitoring, forest inventory, biophysical variable estimation, species discrimination or classification. In this regard, polarimetric synthetic aperture radar (PolSAR) data can provide information on the structure of forests. Moreover, recent studies show that image classification techniques, which use both scattering and spatial information, are more suitable, effective, and robust than those that use only spectral information. The main motivation of this paper is presenting an appropriate contextual classifier for PolSAR data.
Relating the posterior Markov random field (MRF) energy function to the support vector machine (SVM) classifier, the proposed method takes advantage of both the parametric and nonparametric classifiers which efficiently combines SVM and Wishart classifiers. The proposed contextual image classifier adopts the ICM approach to converge to a local minimum and represent a good tradeoff between classification accuracy and computation burden. In proposed method, the computational cost of the training stage is exactly similar to that of training stage in SVM. However, for classifying test pixels, computational cost of the proposed method is the sum of those of SVM, Wishart, and also MRF.
The method allows taking into account various types of features. In particular, the features obtained directly from original data, the features which are derived using the well-known decomposition methods, and the SAR discriminators are used as input features. Moreover, the covariance matrix of the PolSAR data and the Wishart distribution are used to compute the MRF energy function. The proposed method modifies the decision function and the constraints of SVM based on the integration of contextual information. Selection of the appropriate features and optimization of requiring parameters perform simultaneously using genetic algorithm (GA).
In this study, two Radarsat-2 polarimetric images acquired in the leaf-off and leaf-on seasons are used from a forest area. A total of six classes, including white pine, red pine, poplar, and red oak are cinsidered. Two other classes in this study are water area and ground vegetation region.
In this study, the selection of the appropriate features and optimization of requiring parameters are performed simultaneously using GA. Then, the proposed algorithm is compared with the Wishart, Wishart-MRF (WMRF), and SVM as the baseline classifiers. Comparison of the accuracy of the proposed method with baseline methods is performed. The results show that this algorithm allowed approximately 16%, 11%, and 7% increases in overall accuracy with respect to the Wishart, WMRF, and SVM classifiers, respectively. Moreover, proposed method allows 25.29%, 19.74%, and 11.76% increase in average accuracies of forest species with respect to the Wishart, WMRF, and SVM classifiers, respectively. This demonstrates the efficiency of the proposed method for classification of forest species. Also, the results show that, incorporating contextual information into proposed method significantly improves the spatial regularity of the classification results and reduce the sensitivity to noise or speckle.