Research On The Spectrum Automatic Extracting Based On Decision Tree Model Of Karst Rocky Desertification

Information extraction and classification of remote sensing technology have been important elements in the remote sensing area.Traditional classification methods,supervised and non-supervised classification as well as the artificial neural network classification,expert system classification and other new methods are based on spectral features in application of remote sensing classification.However,the image itself as a result of the existence of "same object with different spectra,different objects with same spectrum" phenomenon, which rely solely on spectral features of surface features will make errors occur more frequently as well as the situation of leakage.Many studies have shown that spectral information combined with images and other supporting information, can greatly improve the classification accuracy.At present,the multi-spectral remote sensing image has a large amount of information,such as the texture features of information,geometric knowledge, neighboring relations between features,which can be assisted extracting image information features.Texture is important information in remote sensing.It not only reflects the statistical characteristics of gray-scale images,but also reflects the relationship between feature spaces.Visual interpretation is one of the important symbols.Analysis methods combining texture and spectral characteristics can improve the accuracy of the image spectral information (Chenyun,Daijinfang,Lijunjie,2008).Decision tree as a knowledge-based classification applied to the field of remote sensing gradually.Selecting threshold variable and node is the key feature of the extraction of information.The article is trying to modeling Yachi Bijie as the case study,based on the analysis of spectral characteristics of features as well as the decision tree classification method which based on CART algorithm,using the spectral characteristics values of the image,NDVI values, DEM and the principal component PCI,PC2 for the test variables,mining and determining the rules of decision tree nodes from the training sample data,using integration of remote sensing spectral imaging features and auxiliary data to establish the information automatically extracted decision tree model in the study area of rocky desertification,to extract rocky desertification information at different levels from images,in order to complete the rocky desertification classification of remote sensing image in Yachi demonstration zone of Bijie city multi-spectral demonstration zone.Use the GPS points to verify the accuracy of classification results,and compare it with the maximum likelihood methods of supervised classification (Maximum Likelihood Classification,MLC).The results show that the overall classification accuracy based on CART decision tree is 91.4%,8.3%higher than the maximum likelihood classification,Kappa coefficient was 0.893,0.105 more than the maximum likelihood classification,showing that the use the CART decision tree algorithm to build rules of classification is reasonable.It can be fast,efficient access to a large number of classification rules,And the algorithm complexity low,high efficiency,reflected the knowledge-based classification of remote sensing images of rocky desertification in karst regions in the advantages of automatic extraction of information...