A study on vegetation cover extraction using a Wishart H-α classifier based on fully polarimetric Radarsat-2 data

This article investigates the potential of vegetation cover extraction using fully polarimetric Radarsat-2 data based on the Wishart H-α classification method. In our research, the polarimetric entropy (H) and scattering angle (α) obtained by Cloude–Pottier decomposition technique are used to classify land cover, and the classification results are then used as the initial classification of the Wishart classifier. The classes of the Wishart H-α classification method are associated with vegetation based on the scattering mechanism of vegetation. In our experiment, a fully polarimetric Radarsat-2 image located in the northeast of China, acquired on 4 August 2012, is used to extract the information of vegetation cover, and the corresponding Lansat-7 image is used to verify the classification results. The overall accuracies with different multi-looks window size (3 × 3, 5 × 5, and 7 × 7) are 54.0%, 88.0%, and 73.1%, respectively. The difference of the classification accuracies is influenced by various scattering mechanisms of the lake area. After multi-look processing by a 3 × 3 or 7 × 7 window, the scattering mechanism of the lake should be Bragg scattering rather than fully or partly high-entropy multi-scattering in the Wishart H-α classification. The results show that the water (Bragg scattering) is misclassified as vegetation (high-entropy multi-scattering) by the Wishart H-α classification in C band. The misclassification is caused by the Wishart H-α classifier and the scattering mechanism of the shallow in the study area. Consequently, a modified Wishart H-α classifier that keeps Bragg scattering from high-entropy multi-scattering in each iteration is proposed. Moreover, the analysis on the features of the shallow indicates that the polarimetric entropy increases with the change of roughness. Therefore, the partial lake area presenting high-entropy multi-scattering mechanism is not recognized as vegetation. The final overall accuracies are 81.7%, 88.0%, and 88.0% with the 3 × 3, 5 × 5, and 7 × 7 multi-look window, respectively.