A Multi-feature Fusion Method For Ecological Land Class Extraction In The Yellow River Source Region

Remote sensing technology is an important means to obtain land use information,and in the acquisition of land use information,the classification method of remote sensing image is a key step of research,select a good classification method can greatly improve the monitoring ability of land use.The traditional classification method based on pixels,not reasonable use of all the feature information of images effectively,it only pay attention to the spectral information and ignore the features in the image space and texture information,after reading a large number of domestic and foreign literature and analyzing contrast,decided to sentry 2satellite image data as data sources,to national park water spring park as the research area of three rivers’ sources,This paper discusses the suitability and accuracy of object-oriented decision tree classification research method in alpine wetland ecosystem.Multi-scale segmentation is selected as the image segmentation method,and the optimal segmentation scale and other segmentation parameters are obtained through several experiments,so that the segmented objects can meet the requirements of classification.Image feature extraction aspect comprehensive extracted the sentry 2 Band2-Band8 A a total of eight original image bands,extract the normalized difference vegetation index NDVI,NDWI normalized water index,chosen by the gray level co-occurrence matrix texture features calculated average,homogeneity,second moment and the similarity,standard deviation,correlation,contrast and entropy,The auxiliary feature information is selected from DEM digital elevation model and fine-grained spectral information calculated based on hyperspectral restoration experiment,which is a binary image.A total of 20 feature indices construct CART decision tree rules and participate in classification.Finally,the accuracy evaluation is carried out,and the conclusions are as follows:（1）According to multiple test analysis and comparison,the segmentation scale of the study area is 45,the shape weight is set to 0.1,and the compactness is set to 0.25.The segmented image effect meets the requirements.（2）For the rule construction of decision tree classification algorithm,the feature index of the rule construction includes texture feature index,spectral feature index,fine-grained spectral information and DEM digital elevation data set for auxiliary classification.This rule construction method is improved and optimized compared with other decision tree rule construction methods.（3）In the third group of experiments,the spectral features,texture features and auxiliary features are all combined together,and the Kappa coefficient obtained based on the classification of CART decision tree is92.51% and the overall accuracy is 94.01%.Compared with the Kappa coefficient 89.50% in Experiment 2,the overall accuracy is 91.61%.In experiment 1,Kappa coefficient 87.25% and overall accuracy 89.79%,as well as Kappa coefficient 76.32% and overall accuracy 80.81% of the traditional maximum likelihood classification method have been significantly improved.It shows that the object-oriented decision tree classification method based on this feature combination is an effective method for alpine wetland ecosystem classification.（4）In the object-oriented decision tree classification method and the maximum likelihood classification experiment,the accuracy of grassland is insufficient compared with other ground objects,which may be due to the occurrence of mixing between forest land and grassland.