Hyperspectral feature selection for forest classification

Hyperspectral imagery contains many correlated bands, which not only increase computing complexity but also degrade classification accuracy if not enough training data are available [David A. Landgrebe, 2003]. To mitigate this problem, linear image transformations, such as principle components analysis (PCA), minimum noise fraction (MNF), and canonical analysis, are employed in hyperspectral data applications. Though these transformations are very effective for data decorrelation and feature extraction, they do not maintain the physical meaning of the original spectral bands, for each band of the transformations is a linear combination of the original spectral bands. In this study, we propose two methods of hyperspectral feature selection as means of data dimension reduction. These methods calculate band significance in terms of the eigenvalue contribution of each original band to the transformed image, which is termed the "band score." The band scores enable band ranking. Feature selection is, therefore, achieved by selecting the highest ranked bands. In addition, the band ranking also indicates the preferred band locations across the wavelength range for specific applications. This information benefits future imaging sensor development. In this study, both low and high altitude AVIRIS datasets acquired over the Greater Victoria Watershed (GVWD), British Columbia, Canada, are used for assessing feature selection methods. From each dataset a set of thirteen bands is selected. The classification results with the selected bands are compared with those obtained from all-band MNF and canonical transformations.