A new feature selection method for segmentation of VHR satellite image

The goal of this paper is a new feature selection method for segmentation of very high resolution satellite images. We propose a reasonable number of feature types based on spatial and spectral features containing 1^st order and 2^nd order statistics, Gabor filter and two spectral indices. Having 227 generated features for each pixel, the appropriate and essential features are selected by three steps procedure using a training image. We applied the selected features for segmentation of the images captured by Quickbird satellite. Furthermore, we compared the result of our proposed method with well known feature selection methods. Using different segmentation evaluation measures, our comparison show the efficiency of the proposed method in providing better segmentation.