Hyperspectral Endmember Extraction Preprocessing Using Combination of Euclidean and Geodesic Distances

摘要

Combination the spatial-contextual information in spectral unmixing as a preprocessing of endmember extraction algorithms (EEAs) has been an important issue in hyperspectral image analysis. Particularly, this paper performs a new preprocessing framework using combination of spectral Geodesic and spatial Euclidean distances prior to classical spectral-based EEAs. It exploits both spatial and spectral features of image pixels in order to look for high spectrally correlated and spatially homogenous regions where pure spectral signatures are more likely to be found. For this purpose, it exerts a new correlation coefficient quantity on spatially homogenous pixels designated by spectral weighting determination and appraising the cluster label of spatial neighbours of pure pixels. The novel preprocessing hampers from useless computation of a great number of mixed pixels executed by EEAs. Additionally, two new spectral Geodesic and spatial Euclidean distances are presented to specify the final mean vector which exploits in correlation coefficient computations. The validation of our preprocessing is deliberated on two real hyperspectral datasets from the viewpoints of RMSE and SAD based errors in comparison with other schemes. Experimental consequences declare that such preprocessing can amend figures of unmixing accuracy without intensifying the complexity and with no requirement of changing EEAs.