Target spectrum based feature selection (TSFS): A new method based on chain coding for target detection problems

One of the controversial issues in hyperspectral remote sensing methods for target detection is whether the feature selection will be useful. Generally, feature selection methods are divided into variance based and wavelength based methods. Variance based feature selection methods like information-theory-based methods may eliminate the distinctive features because the distinctive features probably are not statistical principal components. Distinctive features are crucial to distinguish target from background and are maintained in wavelength based methods which concentrate on wavelength information. However, beside to wavelength-based information, target spectrum fluctuations are also critical for target detection. In addition, the wavelength based methods are often time consuming iterative methods with high computational cost. This paper introduces a new feature selection method considering hyperspectral target spectrum. The proposed algorithm has been developed based on Chain coding idea. We proposed Chain Filtering, Chain Encoding, and Chain Statistics as filter, embedded, and wrapper feature selection methods. In this paper, Chain filtering, Chain statistics, and Chain encoding approaches have been compared with different types of feature selection methods such as Principle Component Analysis (PCA) and Minimum Noise Fraction (MNF). Numerical tests have been executed using 4 datasets including Cuprite Nevada and Jasper Ridge datasets from AVIRIS, Botswana, and local datasets of Isfahan province from Hyperion sensor and using Constrained Energy Minimization (CEM) target detection. The results show the accuracy of target detection applying the proposed feature selection method increases from 85% to about 92% for Kaolinite and from 77% to 96% for Buddingtonite in comparison with PCA for Cuprite dataset. Furthermore, the increments more than 5% and 17.5% will be achieved in comparison with MNF, respectively. The results have shown that not only the proposed method overcomes the accuracy decrement issue of feature selection in target detection, but also it improves target detection accuracy by eliminating non-informative features for target detection applications. So feature selection will be an efficient tool for target detection if the applied feature selection method picks out the distinctive features well.