Unsupervised feature selection (UFS) is a standard approach to reduce the dimensionality of hyperspectral images (HSIs). The main idea in UFS is to define a similarity metric, and select the features minimizing the metric to reduce the data redundancy. In this paper, we proposed a novel criterion for unsupervised dimensionality reduction based on the representation of spectral reflectance to capture dominant reflectance variations. Since capturing all the spectral information from an entire hyperspectral dataset is a time-consuming process, we proposed a heuristic algorithm named Greedy Search for Spectral Representation (GSSR). This algorithm divides the spectrum into spectral regions with less spectral variations and merges them. GSSR, similar to feature selection techniques, preserves the original data from being distorted or compromised by a transformation. We compared the GSSR algorithm with well-known existing algorithms in different experiments using various datasets. Comparison with the best approximation to represent single spectra as well as entire hyperspectral scene revealed that spectral representation is almost the same. The difference between the best spectral representation and the ones provided by GSSR is less than 0.01%; while on average, GSSR is about 660 times faster to represent single spectra and 37 times faster for a complete hyperspectral scene. Five well-known unsupervised dimensionality reduction methods were also implemented and used for comparison analysis. Based on the image classification accuracy over two hyperspectral datasets, the spectral features identified by the proposed criterion improved the classification accuracy as well.

This paper proposes a novel self-learned integrated framework of active learning (AL) and semi-supervised learning (SSL). SSL methods try to estimate a certain semilabel for unlabeled samples. While AL methods select the most informative unlabeled samples for the current classifying model and provide their labels by human expertise. An excessive human-machine interaction is required for labeling the selected instances. Whereas, providing reliable labels is a sensitive, time-consuming and expensive step. In our framework, we try to decrease the required human supervision by incorporating SSL method. In addition, the participation rate of each AL and SSL methods in the framework is adaptive and determined based on the certainty of the classifier at each iteration. The experiments were carried out on Pavia University image data which is an urban scene. The results showed the efficiency and the excellent performance of the proposed method in both terms of accuracy and computational cost.

One of the main steps in hyperspectral image classification is the selection of bands that provide the best separability among classes. It is usually understood that the selected bands for classification must contain a large amount of information, and the correlation among selected bands should be small to avoid redundancy. At the same time for optimal classification, class separability should be at maximum value. The question arises whether the most informative spectral regions are really the same as the most discriminant ones for a given set of classes. Answering the question, we developed a new method named Spectral Region Splitting (SRS) to identify interesting spectral regions. This article concludes that the optimal informative and the optimal separable spectral regions are not identical. Furthermore, the cause of the difference is proven theoretically.