Review of hyperspectral remote sensing image classification

Abstract

Studies on hyperspectral remote sensing image classification have developed rapidly with the progress of related disciplines, including pattern recognition, machine learning, and remote sensing technology. This review generates a systematic summary and conducts a comprehensive evaluation of the advancements in current techniques for hyperspectral remote sensing image classification. Based on an overview of different classification schemes, we examine the recent progress in per-pixel classification algorithms for hyperspectral images from six aspects, namely, new classifier design (e. g., kernel-based methods), feature mining, spectral spatial classification, active and semisupervised learning, sparse representation for classification, and multiple classifier systems. Future research directions are discussed as well. On the one hand, new theories and methods of machine learning should be introduced continuously into hyperspectral image classification. Moreover, multisource data and multidimensional feature spaces may improve the accuracy, generalization capability, and automation degree of a classifier. On the other hand, new classification methods should be designed in consideration of practical requirements to meet the needs of real applications and to emphasize the advantages of fine spectra in hyperspectral remote sensing.