Intrinsic Dimensionality Estimation in Hyperspectral Images

By development of remote sensing sensors, hyperspectral remote sensing images are now widely available for monitoring the earth’s surface by using high spectral resolution and dimension. However, this large dimension not only increases computational complexity but also degrades classification accuracy. Dimensionality reduction is a major issue to improve the efficiency of the classifiers in hyperspectral images. The common ways for dimensional reduction is feature extraction. Ideally, the reduced representation has a dimensionality that corresponds to the intrinsic dimensionality of the data. There are a wide range of methods in intrinsic dimensionality estimation and dimensionality reduction of hyperspectral Images in literatures. In this paper, we discuss and compare five intrinsic dimensionality estimation (IDE) techniques for hyperspectral dimensionality reduction. We investigate the performance of these techniques for IDE on hyperspectral images, and compare their performances for supervised image classification purpose. These techniques include; Eigen value estimator (EV), Maximum likelihood estimator (ML), Correlation dimension estimator (CD), Packing number estimator (PN) and geodesic minimum spanning tree (GMST) estimator. The K-Nearest Neighbor (KNN) classifier used for supervised image classification. The variety of the distance metric was used and compared in this classifier. The most useful and practical methods for reduction of dimensionality Principal Component Analysis (PCA) and Independent Components Analysis (ICA) were used in output of these IDE techniques. This study presented a review and comparative study of techniques in IDE. Then feature bands used in supervised classification with variety in parameters usage.