Improved scatter-matrix-based feature selection method for vegetation classification of hyperspectral image

Abstract: Due to the advances in hyperspectral sensor technology, hyperspectral images have gained a great attention in the precision agriculture. Compared to multispectral images, e.g., Landsat TM (thematic mapper) and MODIS (moderate-resolution imaging spectroradiometer) images, hyperspectral images have higher spectral resolution and provide more contiguous spectrum. Thus, hyperspectral images are expected to have good capability in quantifying vegetation biophysical and biochemical attributes which can reflect crop growth status and guide site-specific agricultural management. In practical applications, vegetation classification is usually required to be conducted first and then the vegetation of interest is discriminated from others. It is easy to distinguish vegetated areas from other surface types by setting the threshold of normalized difference vegetation index (NDVI). As to the discrimination of different vegetation types using hyperspectral image, it is a typical hyperspectral image classification problem. The scatter-matrix-based class separability measure is often favored and chosen as a selection criterion in feature selection due to its simplicity and robustness. The scatter-matrix-based class separability measure is constructed by using 2 of 3 scatter matrices which are within-class scatter matrix, between-class scatter matrix and total scatter matrix. Traditionally, these scatter matrices are calculated from the perspective of all classes. However, direct optimization of this measure tends to select a set of discriminative but mutually redundant features, which restricts the improvement of classification accuracy. In order to avoid selecting mutually redundant features as much as possible, this study proposes an improved scatter-matrix-based feature selection method, which tries to calculate scatter-matrix-based class separability values for each pair of classes and then takes the average of all the pairwise class separability values as the final selection criterion. Feature selection is performed by maximizing the criterion using sequential floating forward search (SFFS). In order to verify whether the proposed feature selection method could well avoid selecting mutually redundant features, the mean square correlation coefficients were calculated for the proposed feature selection method and the conventional scatter-matrix-based feature selection method, and a quantitative comparison was conducted. The classification accuracy of the proposed method was compared with that of several representative feature selection methods that were respectively based on MI (mutual information) and class separability measure. The experiments and comparative analyses were conducted with a widely used hyperspectral image, which was collected over the agriculture area in northwestern Indiana, USA (United States of America) by the AVIRIS (Airborne Visible / Infrared Imaging Spectrometer). The experimental results indicated that: (1) The proposed feature selection method could better alleviate the problem of selecting mutually redundant features, compared to the feature selection method of conventional scatter-matrix-based class separability measure; (2) Compared with the MI-based feature selection methods, the scatter-matrix-based feature selection methods generally got higher classification accuracies, and especially the proposed feature selection method produced the highest classification accuracies on both data sets (87.2% and 90.1%) for vegetation classification of hyperspectral image. The comparative experiments on the classification of a typical agricultural hyperspectral image demonstrate the effectiveness of the proposed feature selection method in the vegetation classification of hyperspectral image.