Segmentation-aided classification of hyperspectral data using spatial dependency of spectral bands

Classifying every pixel of a hyperspectral image with a certain land-cover type is the cornerstone of hyperspectral image analysis. In the present study a segmentation-aided methodology for the spectral-spatial classification of hyperspectral data is proposed. It considers the spatial dependence of the spectral bands, deals with the curse of dimensionality and handles the spectral variability. A local spatial regularization of spectral information is used, in order to derive an informative joint spectral-spatial representation of the data. A contiguity-based segmentation algorithm is formulated, in order to build the object-wise texture that can aid classifier learning. The hybrid use of the segmentation texture is evaluated in both pre-processing (i.e. selecting representative pixels to learn the classifier) and post-processing (i.e. refining predicted labels and removing possible outlier classifications). The experiments performed with the proposed methodology provide encouraging results, also compared to several recent state-of-the-art approaches.