Optimal Feature Selection for the Classification of Hyperspectral Imagery Using Adaptive Spectral–Spatial Clustering

Hyperspectral images captured through the hyperspectral sensors play an imperative part in remote sensing applications in the present context. Unlike traditional images sensed with few bands in the visible spectrum, the hyperspectral (HS) images are obtained with hundreds of spectral band ranges from infrared to ultraviolet regions. Because of its vast spatial and spectral data, it requires an extensive computational system for processing and its hidden features are needed to be unveiled in an effective manner specifically for the classification of HS imagery. This approach exploits the high spectral band correlation and rich spatial information of the HS images for the generation of feature vectors. To attain optimal feature space for the best probable classification, an adaptive approach is incorporated to adaptively choose spectral–spatial features for feature selection to classify the pixels effectively. Furthermore, the HS image encompasses several bands including noisy bands. To categorize the images with great accuracy, it is suggested to eradicate the noisy bands whilst retaining the informative bands. In this research, an adaptive spectral–spatial feature selection scheme is proposed for HS images where the extremely correlated representative bands are considered for analysis with uncorrelated and noisy spectral bands are judiciously discarded during its classification process. This hybrid approach not merely diminishes the computational time and also improves the general classification accuracy significantly. The empirical result displays that the proposed work surpasses the conventional approach of HS image classification systems.