Abstract: [Objective] A feature selection method is proposed for improving the accuracy of land cover classification, which considers the combination of separable distance of sample probability and band correlation coefficient.[Method] The several derived vegetation indices and texture characteristics used were extracted from Landsat-8 OLI data in Jiangle county of Fujian Province. Then the optimal features were identified by traditional feature selection method and improved method, respectively. By comparing the separable values of vegetation types based on the best features, the accuracy of feature selection based on two methods was determined. The different scenarios based on primary spectral data, selected vegetation indices and textural images were classified by support vector machine classification algorithm to explore selected features on improving the land cover classes.[Result] The improved separability method can more accurately select the features with higher discrimination while avoiding the selection of redundant bands. For vegetation index and texture features, a single feature cannot maximize the separability of vegetation classes while two feature combinations can significantly improve vegetation separability. Compared with the other texture features in the same window sizes, the ratio vegetation index and the texture features based on contrast, variance and the second moment with small window sizes had a better performance in improving the vegetation classification accuracy. The combinations of optimal vegetation indices as extra bands into OLI multi-spectral bands did not significantly improve overall classification performance (OCA). The combination of textural images and primary spectral bands improved the OCA, which was especially valuable for improving vegetation classification accuracy. The combination of both vegetation indices and textural images with multi-spectral bands provided the best classification performance. The overall classification accuracy and overall Kappa coefficient (OKA) were increased by 7.41% and 8.5%, respectively.[Conclusion] The feature selection method based on improved transformation divergence can balance the separability among all classes, better to avoid redundant features. So, the selected features by the improved method can increase accuracy of specified classes in the study area.

Key words: OLI remote sensing image, optimal vegetation index, optimal texture feature, improved divergence analysis, support vector machine (SVM) classification