Hyperspectral Images Classification by Combination of Spatial Features Based on Local Surface Fitting and Spectral Features

Hyperspectral sensors are important tools in monitoring the phenomena of the Earth due to the acquisition of a large number of spectral bands. Hyperspectral image classification is one of the most important fields of hyperspectral data processing, and so far there have been many attempts to increase its accuracy. Spatial features are important due to their ability to increase classification accuracy. In the present paper, a new method is proposed for the spatial features generation of hyperspectral images based on local surface fitting technique. In this method, a surface is fitted to the gray level intensity of the image in the local window around each pixel, and the fitted coefficients, the coefficients of the first and second fundamental forms, curvatures, divergence of the gradient, the area of ​​the gray level intensity of the image and the volume enclosed below the surface are produced in the various window sizes as spatial features. Proposed spatial features stacked with spectral features and form the spectral-spatial vector. this rich spatial-spectral vector is classified with K-nearest neighbor and support vector machine classifiers. The experiments of this paper that are conducted on two real hyperspectral images in agricultural and urban areas show the superiority of the proposed method. The final results show that the overall accuracy of the proposed method in the best case is 7%  higher than other competitor methods.