Feature extraction of hyperspectral images with a matting model

Owing to the limitations of the imaging sensor and theoretical aspect, hyperspectral images (HSIs) are contaminated with some unwanted components such as noise and a lack of spatial information. This article proposes a spatial-spectral feature enhancement model to eliminate interference, modify spectral distortion, and increase the useful features. The framework firstly proposes an effective spatial feature-based strategy for selecting a band with the most edge information to serve as alpha channel. Given the alpha channel, the continuous foreground and background are estimated by the closed form solution. Finally, feature-enhanced HSI is obtained by linearly combining the selected band, hyper foreground and background. Experimental results of the ground-based data and remotely sensed data indicate that the proposed feature enhancement algorithm provides effective performance in enhancing spatial-spectral features and reducing noise. Especially, the feature-enhanced data have positive influence on both unmixing and classification.