Spatial-Spectral Hyper Spectral Classification Based on Statistical Dependence between Adjacent Pixels

This paper contributes some spectral-spatial classification methodologies and techniques based on spatial homogeneous regionsfor hyper spectral remotely sensed images. These techniques mainly focus on adequate object segmentation and simultaneous combination of spectral and spatial information. Different segmentation methods such as hyper spectral Robust Color Morphological Gradient (HRCMG), Adequate Expectation Maximization (AEM) and hyper spectral Recursive Hierarchical Image Segmentation (HRHSEG) were introduced and applied in the empirical implementations. This paper also contributes integration of the local weighted Markov Random Fields (MRF) on SVM framework for hyper spectral spectral-spatial classification. Using marginal weighting function in the MRF energy function, which preserves the edge of regions, is a new approach. In this paper, merits and issues of different proposed techniques are examined and compared as well as their classification maps with some known spectral-spatial methods for four real hyper spectral images. Experimental results illustrate our proposed approaches including higher accuracies when compared with elder schemes.