Analysis on spectral unmixing based on manifold learning

The main study on spectral unmixing is to develop a regression between mixed spectral features of main land-cover types and their responding fractional cover. Studying on in situ spectral reflectance data, based on one of the best known algorithms of manifold learning, locally linear embedding (LLE), a new modeling method named constrained least squares locally linear weighted regression(CLS-LLWR) was proposed. Spectral reflectance of four kinds of the mixed land-cover types in different percentages was measured and preliminarily analyzed. The model CLS-LLWR was verified by predicting fractional cover of main land- cover types. Compared with principal component regression (PCR) and partial least squares regression(PLSR), through comparison and analysis of the standard error of prediction(SE), the result shows that the CLS-LLWR has better predictability. This study indicates that manifold study has the potential for the information extraction of mixed land cover types in hyperspectral image.