A comparative analysis of classification methods for hyperspectral images generated with conventional dimension reduction methods

This paper compared performances of classification methods for a hyperspectral image dataset in view of dimensionality reduction (DR). Among conventional DR methods, principal component analysis, maximum noise fraction, and independent component analysis were used for the purpose of dimension reduction. The study was conducted using these DR techniques on a real hyperspectral image, an AVIRIS dataset with 224 bands, throughout the experiments. It was observed that DR may have a significant effect on the classification performance. After the DR methods were applied to the image dataset, the extracted reduced bands were used for testing classification performances. Four commonly used classification methods including maximum likelihood, Mahalanobis distance, spectral angle mapper, and support vector machines were used for the classification of dimensionally reduced and original (not reduced) images to test whether classification accuracies differed significantly for these images. Experiments reported in this study indicate that second-order statistics-based DR methods could efficiently reduce the dimensionality of the hyperspectral dataset. The maximum likelihood method generally performed better than the other classifiers while the spectral angle mapper exhibited significantly lower accuracies in the experiments. Inclusion of a texture measure in classification improved the classification results in a significant manner.