Cropland Extraction from Very High Spatial Resolution Satellite Imagery by Object-Based Classification Using Improved Mean Shift and One-Class Support Vector Machines

The issue of cropland extraction from very high spatial resolution (VHR) satellite imagery remains a great challenge. In this paper, an object-based classification method for cropland extraction from VHR satellite imagery is proposed based on the improved mean shift and one-class SVM. After the fused satellite image is transformed by nonnegative matrix factorization into three bands, the improved mean shift is employed to segment the image. Subsequently, the structure lines of each region in the segmented image are detected, and the standard deviations of the directions of the straight lines are calculated. The spectral information and the above derived texture information are selected as features for the following classification. At last, the support vector data description is utilized to recognize the croplands from the segmented image based on only some cropland samples. Three satellite images with different spatial resolutions are employed to test the algorithm, and the results show that our proposed method obtains a higher overall classification accuracy than the eCognition's method does, and its overall classification accuracy is promoted with the increasing of spatial resolution. Another merit of our method is that it needs only the cropland samples, which is time-saving and cost-saving.