Hyperspectral Spatial-Spectral Feature Classification Based on Adequate Adaptive Segmentation

This paper proposes some novel classification scheme based on adaptive spatial vicinity for hyperspectral remote sensed images. Different segmentation methods such as Robust Color Morphological Gradient (RCMG), Expectation Maximization (EM) and Recursive Hierarchical Segmentation (RHSEG) have been generalized to hyperspectral image analysis and their extensions; Hyperspectral Robust Color Morphological Gradient (HRCMG), Adequate Expectation Maximization (AEM) and Hyperspectral Recursive Hierarchical Image Segmentation (HRHSEG) were introduced and applied in the empirical implementation. Experiments were based on two available hyperspectral data sets (Indiana Pines and Hekla). Experimental results were compared with three analysis measurements (overall accuracy, average accuracy and Kappa factor) as well as their classification maps with pixelwise methods and some previous spatial-spectral approaches such as EMP and ECHO. All of the quantitate quality measures of proposed method were better than other reviewed approaches, but the classification map of proposed approach is so artificial in some cases. The novel segmentation methods (HRCMG, AEM and HRHSEG) are applied, and the accuracy was improved in compare with elder schemes, when the median voting scheme is employed