Feature Extraction of Hyperspectral Images of Weighted Spatial and Spectral Locality Preserving Projection (WSSLPP)

In order to improve the classification accuracy of Hyperspectral images (HSI) and preprocess HSI by effectively using the spatial and spectral information of HIS, a new spatial-spectral feature extraction method, Weighted Spatial and Spectral Locality Preserving Projection (WSSLPP) is proposed in this paper. The HSI was reconstructed combining the physical characters of HSI to avoid the interference of singular point; then the target functions of locality pixel neighbor preserving embedding (LPNPE) and locality preserving projection (LPP) were weighted and summed, thus the spatial and spectral dimension information of HSI was effectively fused to construct the projection matrix. WSSLPP not only keeps the pixel neighborhood in spatial domain, but also keeps the implicit structure of samples in spectral domain, which helps for the HIS classification. The benchmark verification on Indian Pines and PaviU database show that the classification accuracy resulted from WSSLPP algorithm is significantly higher than that from other algorithms, and the overall classification accuracy is 99.00% and 99.50% respectively, effectively improving the HSI classification accuracy.