| With the development of science technology, remote sensing image has been significantly improved in spatial resolution, time resolution and spectral resolution. It plays an important role in environmental monitoring, land use, national defense and military fields.High-resolution remote sensing image has obvious structure and texture characteristics, its information is very abundant, but its band is less, which makes it lack of spectral information, and it also has the phenomenon of "different images same object" and "different objects same image ", these factors make the extraction of image information becomes more complicated and difficult.For remote sensing image classification, fuzzy method has been shown its advantages in analyzing mixed pixels and promoting the accuracy of classification. As the world's leading image analysis software, the basic idea of eCognition is object-oriented fuzzy classification. Meanwhile, texture as an important feature of the image reflects the information of spatial structure, it is also an important factor that can not be ignored during classification. Based on these considerations, in order to fully understand the object-oriented classification techniques and the idea of fuzzy classification, this paper take four QuickBird images which have been preprocessed as the experimental datas, do some research about fuzzy classification based on pixel and object-oriented fuzzy classification so as to obtain the classification experimental results which based on Fuzzy logic.Experimental work including:(1) Use eCognition software as the experimental platform, use fuzzy logic algorithms make the object-oriented image classification experiment under the rules set mode and Quickmap mode respectively. Research the multi-scale segmentation, spectral difference segmentation and edge detection algorithm, experimental contents including the establishment of feature space, membership function classification, nearest neighbor classifier and Box classifiers. The overall classification accuracy of experiment 1 is 95.68%, Kappa coefficient is 0.9407, which is better than the maximum likelihood classification results; (2) Make texture analysis experiments under MATLAB, In three different windows (3×3,5×5,7×7), use GLCM calculate the energy, entropy, correlation, contrast and homogeneity from 0°,45°,90°,135°directions, averaging the texture measures in four directions as the final results. By comparison, the homogeneity measure in 5×5 window reflects the image texture better;(3) According to the characteristics of remote sensing data, design a fuzzy classification system, use the image's mean red, mean green, mean blue and texture information as input variables, set membership functions and fuzzy rules, complete fuzzy inference and defuzzification, realize fuzzy classification by MATLAB program and obtained good results. For the extraction of waters in experiment 1, the producer accuracy can reach to 95.24%, user accuracy is 100.00%, and Kappa coefficient is 0.9693;(4) Get a series of conclusions about experiments, comparison of results and different experimental methods by analyzing, discussing and summarizing the four experiments' result. These conclusions have reference value for the practitioners and scholars who will use object-oriented fuzzy classification or fuzzy classification based on pixel;(5) By the comparation of classification experiments which completed under eCognition software, ERDAS software and MATLAB, we can draw some conclusions:(a) For the high-resolution remote sensing image information extraction, multiresoluation segmentation has more strong advantages than other segmentation algorithms; (b) texture analysis has become an important way to improve the classification accuracy; (c) fuzzy classification method based on pixel or object-oriented has great value in dealing with the boundary problems, and uncertain problems about pixels, region and block; (d) Fuzzy logic theory has a long-term significance in the application of remote sensing image classification. |
PDF Full Text Request