Fine Classification For Vegetation In Arid Areas Based On Object-Oriented Method Using Remote Sensing Images

Vegetation,as the most sensitive indicator of ecological environment,is of great significance for ecological protection,planning and construction.Among them,oasis vegetation in arid and semi-arid areas is not only the key link between oasis and desert,but also the direct responder of oasis habitat.Due to its continental arid climate and unique geographical location,xinjiang has bred a large number of typical oasis system,among which the growth type and distribution of vegetation are closely related to the stability and security of oasis ecosystem.Therefore,the study of oasis vegetation in some arid areas of xinjiang is conducive to the fundamental understanding and the existing ecological environment problems solving,it also has a great reference value for oasis sustainable development strategy setting.With the rapid development of a series of high resolution satellites and object-oriented technology,GF-2 satellite has been successfully applied into the research of object-oriented vegetation extraction,Compared with the traditional image-based image classification,objectoriented technology can make full use of image information which contain plenty of spectral,shape and texture characteristics to participate in the classification.However,too many features will reduce the efficiency of vegetation extraction and lead to the decrease of classification accuracy in the case of limited samples.Therefore,feature selection is an inevitable choice to improve the effect of object-oriented classification of high resolution image..In this paper,with the support of GF-2 high-resolution remote sensing and field metrical data,taking the middle and lower reaches of the ulungu river basin in xinjiang as the research area,An object oriented classification framework combining feature selection and machine learning algorithm is proposed to achieve fine extraction of oasis vegetation in arid areas.in other words,Firstly using ReliefF and CFS to filtrate original feature space,and then appling optimize data to C4.5,SVM,RF to construct six kinds of object-oriented model of vegetation classification,and compare their sensitivity and classification accuracy.Finally,using PSO algorithm to optimize the parameters and get the best vegetation classification model in the study area.This method can provide technical support for the first survey of forest and grass germplasm resources and the investigation of germplasm resources in China.The main research content are as follows:(1)Based on GF-2 remote sensing data,a series of preprocessing processes were used to improve the image quality and intelligibility.Then calculating each band to get OIF value,selecting the optimal band combination was NIR-Red-Green(432).Combining with field metrical data,a three-level vegetation classification system which could achieve accurate classification of vegetation species was established in study area.By calculating the JM distance of the four kinds of plantings,it could be seen that the vegetation had good separability.(2)Multi-scale segmentation algorithm and ESP scale evaluation tool were used to obtain the optimal image segmentation scale parameters,Aiming at different features,multilevel segmentation experiments were carried out.Using the membership function,The vegetation object layer was obtained which could provide data support for the subsequent experimental analysis.(3)Through detailed analysis of h.ammodendron,achnatherum splendens,populus euphratica,reed four kinds of vegetations spectral reflection in different bands,the spectral feature,vegetation index feature,shape feature and texture feature of the image object were extracted.Aiming at the time-consuming problem of GF-2 image texture extraction,PCA principal component transformation was performed and 5×5 windows was selected to extract texture information.After transformation,the extraction efficiency of texture feature is greatly improved while the image information is not received.After transformation,the image information is not affected meanwhile the extraction efficiency of texture feature is greatly improved.(4)Using ReliefF and CFS to filtrate the initial feature space,it is found that feature selection is an effective method to improve the classification accuracy and efficiency in object-oriented research.Among them,30 features are optimized by ReliefF algorithm,while 9 features are directly selected by CFS algorithm for subsequent classification,which can effectively filtrate out redundant features,and the screening ability is obviously better than ReliefF algorithm.(5)The optimized data was applied to C4.5,SVM and RF,to construct six kinds of object-oriented vegetation classification models.According to the analysis of different evaluation indexes such as production accuracy and user accuracy,It could be concluded that compared to C4.5 and SVM,the classification accuracy of RF model was performed better in most cases,especially when training samples were small,RF modle was more stable.Among them,the overall classification accuracy of CFS-RF is the highest,and the optimal classification results can be obtained by using the least features,which is more suitable for the accurate extraction of oasis vegetation in Xinjiang.Using PSO parameter optimization,the final classification accuracy of CFS-RF model is 92.35% and the Kappa coefficient is 0.90,which satisfied the demands of vegetation extraction accuracy.