Boundary Extraction Of Forest Managers Small Class Based On High-resolution Remote Sensing Images

Forest manager small class is the basic unit of forest resources survey and management, and reasonable correct small class division is one of the basic work of forest resources monitoring.Traditional forest zoning approach was artificial "painting on the slope" of the small class boundary, but it was a large amount of work, subjectivity and a large error. Based on high-resolution remote sensing image, automatically extracting forest small class boundary with support vector machine theory is expected to improve efficiency and accuracy of forest small class division.In this paper, with GF-2 remote sensing image in 2015, image segmentation is on the basis of NDVI extraction and principal component analysis. As the sample of the treated actual small class boundaries, based on SVM method, radial basis function selected to categorize images and extract the small class boundaries. The evaluation index which was the area relative error, the forest small class boundary extracted was evaluated the accuracy. The main results were as follows:(1) For the extraction of the forest small class boundaries of the larger canopy density, image segmentation should be appropriate under segmentation. The best segmentation scale combination of image segmentation is segmentation scale 85, merger scale 95.(2) Forest small class boundary were classified and extracted by the use of SVM linear kernel, polynomial kernel function, radial basis kernel function and sigmoid kernel function. Classification accuracy of four kinds of kernel function was less, and area relative error of classification and extracting the small class boundary by radial basis kernel function is minimized, the highest classification accuracy.(3) SVM method compared with three methods of the maximum likelihood method, neural network method and mahalanobis distance method, the smallest area relative error, the highest accuracy of forest small class boundary extraction of 87.69%. Forest small class boundary extraction accuracy of Maximum likelihood after the SVM method is 87.46%.(4) Small class area relative error of as the samples of actual small class boundary of SVM classification extraction is 12.31%, classification accuracy of 87.69%. Small class area relative error of on the image to select the feature samples of SVM classification extraction is 42.65%, classification accuracy of 57.35%, and 30.34% of the former higher than the latter accuracy.