Multi-level Segmentation Classification Method Research Based On Remote-sensing Image

The object-oriented method is advanced of the pixel-based method.This method has been used widely as an advanced method in recent years,especially of the multi-scale classification.This paper employed this method because it can give a good result accuracy.Classified the images with the methods of multi-scales combined with nearest neighbor classification and threshold classification respectively and comparative the classification accuracy.Another,the segmentation parameters have significant influence on the result with the experiments procedure.But,there are any document literature of scientific study on the segmentation parameters.Therefore,this study design a part to discuss the segmentation parameters with the result.The segmentation parameters is a key procedure of the segmentation result in the object-oriented classification,further effect the result of the classification.Segmentation evaluation function is a significant standard of the segmentation quality.Scale,shape and compactness are the parameters,the combination of the different levels of the three parameters may lead to the different result,which could evaluate the quality of the segmentation.We improve the approach based on the segmentation evaluation function,and dig into the affectation of the area factor.Four segmentation evaluation function had been used to analyze the effect of the three parameters,scale,shape and compactness based on Statistics.The Landsat TM and GF-1 remote sensing images as the sample of the experience,which was selected from the county,with 10 pieces of each kind of images.It turned out that: First,the segmentation scale is the most important parameter to the result and the shape is heavier than the compactness.Second,the high quality of the segmentation ask for the value of the shape is small and the compactness is big.Third,the area could improve the stability of the segmentation evaluation function.Forth,the proposed method correspond to the existed method and it can evaluate the segmentation.Fifth,the different resolution have the same effect on the selection of segmentation parameters.For the problem of limitation of the mainly subjective or lacking of widely used of the methods of mult-scale object-oriented classification,the method of multi-scale classification for the research.The data source was Landsat-8 OLI images of Milin county in Tabet.First,the optimal segmentations of multi-scale segmentation should be confirmed.This paper proposed the function model between the accuracy of multi-scale classification and segmentation scales.Then,classified the images with the methods of multi-scales combined with nearest neighbor classification and threshold classification respectively.The result showed that the segmentation scales was 190,150,100,60,and the accuracy of multi-scale segmentation is higher than the single one.Others,the accuracy of nearest neighbor classification with the multi-scale is higher than the threshold classification's,the accuracy was 0.86 and 0.72,and the kappa was 0.72 and 0.69.The function model of optimal segmentation scale is a method can be used widely with scientific theoretical.The method of nearest neighbor classification and threshold classification with multi-scale provide a foundation for the change of forest.Finally,analysis the vegetation changes of the Milin by NDVI.The vegetation type of Milin was extracted with the method of object-oriented.Then,the result were calculate by the spatial analysis and obtained the transition matrix.