A model on achieving higher performance in the classification of hyperspectral satellite data: a case study on Hyperion data

Abstract

Hyperspectral remote sensing data is characterized by the large number of contiguous spectral bands with narrow bandwidth. Enormous information available in hyperspectral data is quite challenging for classification as compared to multispectral remote sensing data. Most of the widely used conventional'hard'classifiersareproducinginconsistentclas- sification results while employed in classification of hyperspectral data. In this paper, we present an effective hyperspectral classification model for achieving higher accu- racy. The proposed model is characterized by three major components: dimensionality reduction using principal com- ponent analysis (PCA), multiresolution segmentation, and fuzzy membership-based nearest neighbor (NN)-classifica- tion. Here, the bands of the dimensionality-reduced images are represented by the first principal component (PC) of each of the spectral region covered by the hyperspectral sensor. Then, multiresolution segmentation is carried out on these PC composite images based on color and shape homogeneity criterion. The conventional NN-classifier is effectively used by appropriate utilization of fuzzy membership function de- fined on a set of optimal features derived from the segmented image objects. We demonstrate a case study on Hyperion sensor data of Earth Observing-1 (EO-1) satellite. A compar- ative assessment is carried out with other competing tech- niques such as spectral angle mapper (SAM), artificial neural network (ANN), and support vector machine (SVM) on a set of images with different land cover surfaces. The proposed classification model outperforms the existing classification approaches investigated here.