Comparison and integration of feature reduction methods for land cover classification with RapidEye imagery

Feature reduction (FR) methods can effectively reduce the feature set and improve the accuracy for land cover classification (LCC) using high resolution remote sensing data with high dimensional or strongly correlated feature sets. However, FR methods have rarely been applied for LCC in arid regions with complex geographic environments, especially for the integration of feature selection (FS) and feature extraction (FE) methods. This study investigated the comparison and integration of FR methods for LCC in part of Dunhuang Basin, northwestern China, which is a typical inland arid region and groundwater-dependent ecosystems. Five spectral bands and 9 vegetation indices features that derived from RapidEye satellite imagery were used with support vector machines algorithm. Two wrapper FS methods, based on random forest algorithm (varSelRF and Boruta packages in R software), were used. Three FE methods (principal component analysis, PCA; independent component analysis, ICA; and minimum noise fraction transformation, MNF), were employed to extract a reduced number of reconstructed new features. Integration of varSelRF and PCA methods (varSelRF-PCA) was attempted. All 14 features were relevant, indicated by Boruta method; only 6 features, including the red-edge band selected by the varSelRF module, had higher importance. All the five FR methods could improve classification accuracy, but only varSelRF achieved significant improvement. The varSelRF outperformed the FE methods, followed by MNF, PCA, and ICA. The proposed varSelRF-PCA model significantly improved classification accuracy and outperformed all the FS or FE methods.