SVM-based hyperspectral image classification using intrinsic dimension

Hyperspectral images currently have potential applications in many scientific areas due to their high spectral resolution and consequently their good information contents. Nevertheless, extracting suitable and adequate features from this data is crucial for any analysis and especially for the classification algorithms. To overcome this issue, dimension reduction techniques are proposed and have showed a direct effect on improving the classifier efficiency of hyperspectral images. One common approach to decreasing the dimensionality is feature extraction by considering the intrinsic dimensionality of the data. In this paper, using support vector machines (SVMs), we propose a procedure that improves feature extraction methods through estimating the appropriate dimension of the transformation space. Due to the nature of the SVM classifier that can efficiently handle the high-dimensional data, our procedure tunes and optimizes the input space data to achieve better output performance. This study presents an efficient method for classifying hyperspectral images with SVMs by considering a suitable dimension for extracted features. The results reveal the superiority of the implemented method to improve the overall accuracy of classified imagery by a margin of nearly 10 %. © 2013, Saudi Society for Geosciences.