Abstract: Corn field remote sensing images have a mass of endmember spectral variability and complexity, that results in the bad classification of planting area. A corn field of remote sensing image classification method based on segmentation-derived regions and feature likeness is proposed. First, principal component analysis (PCA) is used to extract the first principal component from the fusion image which is fused by the panchromatic and multi-spectral image, to acquire the monochromatic image I which contains rich information. Then, do a Watershed segmentation to I, we can get a graph of a split target area. Then build characteristic group which is composed of texture, brightness and contour feature likeness. At last Based on the principle of random forests, extract the corn target using the characteristic group. With the testing using GF-1 satellite remote sensing data and the results comparison analysis of the support vector machine (SVM), neural network algorithm and maximum likelihood algorithm, it shows that the classification accuracy of this method is superior to other algorithms.

Key words: endmember spectral, segmentation-derived regions, feature likeness