Abstract: A hyperspectral image classification method based on two-step smoothing and feature weighting is proposed to address the problem of various image filtering based spatial spectral joint classification methods being difficult to preserve weak edges of the image while denoising. Firstly, the original hyperspectral image is preprocessed by minimum maximum normalization, and then principal component analysis is used to reduce the dimensionality of the hyperspectral image. Next, using windowed domain transformation recursive filtering to obtain feature images with weakened noise while preserving weak edges, and then smoothing the feature images again through L0 gradient minimization to further suppress noise and enhance edges. After that, each feature image is weighted by variance. Finally, support vector machine is used for classification. Experiments were conducted on two datasets, and the classification accuracy of this method improved by 14.06% and 25.75% compared to spectral feature-based methods, and by 0.76%～4.3% and 1.5%～5.69% compared to various filtering algorithms in this field. Moreover, the classification results better reflect the true land cover categories.

Key words: hyperspectral image classification, principal component analysis, windowed domain transform recursive filtering, L0 gradient minimization smoothing, feature weighting