Accuracy Analysis of Hyperspectral Imagery Classification Using Level Sets

Image classification is an important task in the remote sensing field. In a previous study, the authors presented a semi-automated supervised level set-based hyperspectral image segmentation algorithm (LSHSA) (Ball and Bruce, 2005). The LSHSA method used specialized speed functions created using pixel similarity and class discriminator functions. The pixel similarity function was based on an exponential term using three of the data bands with equal contributions from each band. The class discriminator functions had experimentally determined thresholds that were based on the training data, and were used to stop the segmentation at natural boundaries. This procedure is modified by using stepwise Fisher’s linear discriminant analysis (FLDA) with a receiver operating characteristics (ROC) decision metric to determine the best bands for class separation. Four new speed functions are proposed and investigated, based on an identity matrix, the covariance matrix, FLDA weighting coefficients, and the Fisher’s between-class covariance matrix. A HYDICE 191 band hyperspectral image of the Washington D.C. Mall area is used to validate the algorithm (Ball and Bruce, 2005). The classes segmented are grass, water, trees, paths, shadows and buildings. The results of the proposed algorithm are compared to the results from the previous study using the LSHSA and the maximum likelihood (ML) classifier. The results show that the new method provides better results than the previous LSHSA and ML. The main contribution of this paper is a new best bands-based speed function for segmenting hyperspectral imagery using the level set methodology with improved results over the original LSHSA.