| Classifying land-use type,extracting information and monitoring changes by remote sensing date is one of the hottest research field today.As one of the most important forest type,monitoring the dynamic change of non-commercial forest is of great significance to forestry development.Based on GF-1,RapidEye and Spot-5 images of Yanqing District,pix-based and object-oriented classification methods were used to choose the optimal approach for change extraction.On the basis of object-oriented classification,optimal multi-segmentation scale and establishment of rule set were studied to improve the classification accuracy.Furthermore,through analyzing the results of change area,this research would provide tactical reference and suggestion for non-commercial forest monitoring system.Content and main results of the research are as follows:(1)To obtain an optimal method for image enhancement of Spot-5 and GF-1 data,5 frequently-used method were used:hue,saturation,and value(HSV);Brovey transformation;Gram-Schmidt spectral sharpening;Principle Component(PC);and Pansharp transformation.Qualitative and quantitative analyze were used to access the effect and quality of the fusion images.Results showed that diferent method had different impact on the fusion images mainly on brightness,correlation index and other aspects.For Spot-5 image data,Pansharp transformation had the best effect and could be the optimal choice for further classification study.For GF-1 image data,GS spectral sharpening had higher potential for the following classification than the other methods.(2)Based on the fusion images,three pix-based classification methods had been done to extract non?commercial forest.According to accuracy evaluation index of Spot-5,RapidEye and GF-1 classification results,support machine method(SVM)had the highest totally accuracy and Kappa coefficient which were 85.64%,82.72%and 87.19%,0.83,0.80 and 0.85,respectively.Furthermore,the results indicated that SVM showed better effect on dealing with shadow.Considering all the factors,in this experiment,SVM was the optimal method and mahalanobis distance method was the least.(3)Object-oriented classification was applied to three images based on multi-scale segmentation and rule set system.Object-oriented ratio of mean difference to neighbors to standard deviation(RMAS)and mean value index were used to set the optimal segmentation scale.For each image data,three-level multi-scale segmentation hierarchy(Spot-5:560,500,450,RapidEye:400,350,1 80,GF-1:650,560,480)was built and rules were set on each level.In addition,dichotomy model was introduced to object level from pix level and the FC value was added to the rules to realize the object-based classification.The classification accuracy of Spot-5,RapidEye and GF-1 was 87.1%、92.3%and 92.1%,respectively,which was higher than pix-based classification methods and more suitable for extracting changes.(4)Extracted change information by the comparison of classification results,and built the error matrix of non-commercial forest change information.Result showed that,both the extraction accuracy was over 87%,missed rates and error rates were less than 20%.The area of non-commercial forest was increasing during the 2004 to 2015 which was mainly reflect on increasing forest land area and decreasing other forest land and agriculture land.These changes were closely related to the higher awareness of protection of non-commercial forest in our country and kinds of projects and policy. |
PDF Full Text Request