Research On Land Use/Land Cover Classification Based On Mixed Pixel Decomposition Decision Tree

Remote sensing technology is an important means of land use / land cover information extraction and classification,the selection of classification method is an important part of land use / cover classification,classification methods directly affect the classification accuracy.Remote sensing image classification method is almost the majority of pixel based classification,the influence of mixed pixel cannot be excluded in the process of classification,especially for the low resolution image,the existence of mixed pixels is reduced to a great extent the classification accuracy;however,using of high-resolution imagery for visual interpretation costs too much and can not be applied to a wide region,how to combine with the decomposition of mixed pixels classification method,so as to improve the classification accuracy is worth exploring.Based on summary and analysis of relevant research results at home and abroad,the GF-1 PMS/Landsat-8 OLI image data as the main data source,to the city drinking water source-Yunlong Reservoir watershed for experimentation area,to explore a suitable for low resolution images based on land use / mixed pixel decomposition decision tree cover classification method.The main research contents and conclusions are as follows:(1)Data preparation and preprocessing.Complete boundary of Yunlong Reservoir Watershed extraction and GF-1 PMS/Landsat-8 OLI image preprocessing,a high number of data of land use / cover types of artificial visual interpretation.The corresponding vegetation index(NDVI,PVI,RVI,EVI,DVI),slope and slope data were extracted to establish the decision tree feature data set.(2)Linear spectral unmixing.Using field collection types of spectral data to create typical research area spectral library,its application in linear mixed pixel decomposition,improve each unmixing abundance data' precision(Arboreal forest abundance,sparse shrub abundance,High albedo abundance,Grassland abundance,Water abundance,Arable land(including crops)abundance,Arable land(no crops)abundance,Low albedo abundance,Desert and bare surface abundance.The result shows that: Gram-Schmidt(GS)image fusion and construction to improve the efficiency and accuracy of data unmixing endmember selection from many levels,which fully constrained least squares unmixing is a good solution to the unconstrained linear spectral unmixing abundance information extracted from the negative situation the RMSE error is controlled at about 0.174913,improve the accuracy of pixel unmixing and practicability;(3)Selection of training samples.Using image fusion DEM 3D terrain classification and training samples(ROI)to improve the sample separation.The results showed that: 3D Terrain Based on image fusion DEM training sample selection and auxiliary classification to break the traditional mode of training sample selection,not only from the perspective of color synthesis principle and integrated 3D terrain from different angles(overlooking,looking up,head up,side view)selected region of interest,improve the efficiency and accuracy of sample selection,and greatly improve the degree of separation of samples(sample separation is more than 1.9);(4)Land use / land cover classification of research area.The characteristics of data to establish a MNF component,were obtained by decomposition of mixed pixel abundance data and terrain,texture,spectral characteristics,ISODATA and other auxiliary data structure set,QUEST,CRUISE,See5.0/C5.0 algorithm uses decision tree mining rules,and rules of application in Yunlong Reservoir Basin Land Use / cover subdivision.Finally,evaluate the accuracy.Results showed that: Mixed pixel decomposition based on decision tree classification method' accuracy is higher than the Maximum Likelihood Method,Artificial Neural Network,Support Vector Machine(SVM),the classification accuracy is reduced form QUEST ? CRUISE ? See5.0/C5.0 ? SVM ? BP Neural Network ? Maximum Likelihood Method;the QUEST has the highest classification accuracy,Kappa coefficient is 0.95,the overall accuracy is 95.87%;CRUISE 2D Kappa classification accuracy coefficient is 0.89 times 92.14%,the overall accuracy of maximum likelihood classification;the lowest accuracy,Kappa coefficient is 0.76,the overall accuracy is 83.40%,therefore,more suitable for classification of mixed pixel decomposition fusion decision tree in complex terrain,landscape broken area of land use / land cover classification,and has high classification accuracy.Although the mixed pixel decomposition decision tree of land use / land cover classification method' classification accuracy is generally higher,but the method in constructing complex feature data set is relatively complicated,and the implementation algorithm of the classification process will take a long time.It is necessary to do further research on how to optimize the construction of complex feature data sets and improve the classification efficiency of mixed pixel decomposition decision tree.