A Pixon-based hyperspectral image segmentation method used for remote sensing data classification

Image segmentation plays a key role in remote sensing especially as a preprocessing step for further applications such as classification. The data dimensionality and high spectral resolution of the images make it more challenging to precisely segment and consequently classify the Hyperspectral data. It may convince us to utilize the spatial features as well as the spectral characteristics of data to gain better classification results. In this paper we propose a Pixon-based image segmentation technique. We also apply a PDE-based smoothing algorithm to construct larger segments which are more homogenous. The resulted segment maps are then fed into the SVM classifier and the final thematic class maps are produced. The results of applying the proposed method on well-known Hyperspectral datasets imply that using the gained segments instead of pixels in the classification step leads to a considerable compression ratio as well as significant improvements in the classification accuracy and validity.