| Land cover datasets are basic data for study on global change,while land cover change is an important component of global change.Accurate global land cover datasets provide key basic support data for research on global change and regional sustainable development,such as global climate change,ecosystem assessment,carbon cycle,hydrological cycle,biodiversity conservation.However,the low accuracy of existing global land cover datasets has increased the uncertainty of applications of them.How to retain right information and enhance accuracy of those products at the same time has become a key problem for land cover classification by remote sensing.The Moderate Resolution Imaging Spectroradiometer(MODIS)data has the characteristics of high time resolution,stability of data acquisition,real-time access and so on,which is widely used in the global and regional land cover classification studies.Thus,it has not only theoretical but also practical significance to enhance the accuracy of MODIS global land cover products(MODIS LC).Firstly,the national land cover database of China at scale of 1:250,000 and global validation sample data set were used as reference data to analyze accuracy and confusions reasons of MODIS LC V051 at national,various climatic regions and provincial levels respectively.The analysis results of MODIS LC can help MODIS LC users select data and researchers improve classification accuracy of land cover.Secondly,Henan and Guizhou provinces were selected as study areas for their relatively complete land cover types and representativeness of different landform types(plain,plateau).Multi-source land cover datasets,topography,annual average air temperature,annual average precipitation and spatial distribution data of population were selected to construct the quantitative relationship between the distribution probability of various land cover types and environmental factors.Thirdly,an automatic method of collecting land cover samples based on multisource datasets was designed to obtain geographical knowledge.Finally,a multi-source data fusion framework based on Dempster-Shafer evidence theory was designed to improve the accuracy of MODIS LC classification.The main work and conclusions were as follows:(1)For the first time,both evaluation method of absolute accuracy and that of relative accuracy were used to evaluate and analyze the accuracy of MODIS LC in depth at national,different climatic regions and provincial levels in China.Furthermore,some important factors influencing the accuracy of MODIS LC were also taken into consideration for the categories confusions,including data source,training samples,classification features,land cover complexity and mixed pixels.The experimental results showed that:(a)Spatial consistency between MODIS LC and CHINALC1:25 at different climatic regions of China was between 55.78% and 70.36%,which needed to be improved.(b)Both classification accuracy of different land cover types at same climatic region and classification accuracy of the same land cover type at different climatic regions existed significantly different.(c)The lack of training samples in China had great influence on classification accuracy of MODIS LC.(d)With increase of land cover complexity,classification accuracy of MODIS LC decreased rapidly.(e)Mixed pixels accounted for 42% of the total area in China,and the spatial consistency between MODIS LC and CHINALC1:25 was only 50% in these areas,which also had significant influence on classification accuracy of MODIS LC.(f)In conventional methods,DEM,slope,precipitation and land surface temperature were used as auxiliary features directly in land cover classification process,which could improve the overall accuracy slightly(less than 1%),but couldn't fully utilize the relationship between ancillary data and land cover spatial distribution.(2)In this paper,an automatic method of land cover samples collection based on multi-source datasets was proposed.And on this basis,the invariant features of the land cover data were introduced into land cover accuracy improvement process by the form of sample dataset.Firstly,the national land cover database of China at scale of 1:250,000 and global land cover data(GlobleLand30)at resolution of 30 meters were chosen as data sources for samples collection.Secondly,the initial sample areas were collected based on spatial consistency analysis and heterogeneity analysis.Finally,invalid samples were removed from the initial samples through technology of sample purification.The practical experimental results of this method in Henan and Guizhou provinces showed that:(a)Although each of them had its own strengths and weaknesses in the accuracy of different land cover types,overall accuracy of the classification product by the automatic method of samples collection based on multisource datasets proposed in this paper became higher than that of global land cover product MODIS LC,which was classified based on manual method of samples selection.(b)Compared with the method of samples collection based on single source land cover dataset,the automatic method of samples collection based on multi-source datasets proposed in this study had better classification stability and higher classification feature separability,which can not only improve automatic degree of land cover classification but also enhance accuracy of land cover classification.(3)Ratio of land cover distribution,correlation coefficient and changing trend of curves were used to obtain the qualitative relationship between topographic factors and the distribution of land cover categories.And on this basis,logistic regression model of land cover in the study area was constructed,with “homogeneous” pixels as samples collection areas,while annual average air temperature,annual average precipitation and spatial distribution data of population were selected as independent variables and land cover category as dependent variable.The results showed that the logistic regression model not only revealed the quantitative relation between environmental factors and the probability of land cover distribution but also provided evidence source for accuracy improvement framework based on D-S evidence theory.And this method solved the problems of high cost and regional characteristics of experts' geographical knowledge in remote sensing and also provided a new way to use environmental auxiliary data in the process of land cover classification.(4)Two kinds of accuracy improvement schemes for land cover classification product with multi-source data were designed under support of Dempster-Shafer evidence theory in Henan and Guizhou provinces and the basic probability assignment function was established based on the membership of different land cover types and user's accuracy.In the first scheme,MODIS global land cover product,the probability of land cover distribution obtained based on environmental factors and obtained by the automatic method of land cover samples collection were used as the datasource of Dempster-Shafer evidence theory,and the maximum credibility criterion was used to determine the best land cover type at each location.In the second scheme,the Dempster-Shafer evidence theory was only used to determine the land cover type at those locations where MODIS LC was disagree with reference land cover product while MODIS LC remained unchanged at the agreement cells.The results showed that overall accuracy,user's accuracy,producer's accuracy,Kappa coefficient of improved land cover product became higher than original MODIS LC.In the first scheme,overall accuracy was improved about 5% in Henan province and 2% in Guizhou province respectively.In the second scheme,overall accuracy was improved about 7%-8% in Henan province and 4% in Guizhou province respectively.In conclusions,on the basis of mining geographical knowledge contained in land cover datasets,environmental factors and automatic collected samples were introduced as sources of evidence for accuracy improvement framework based on D-S evidence theory in this dissertation for the first time.Methods and technologies proposed in this dissertation effectively enhanced accuracy of MODIS LC.It is without doubt that this dissertation extended range and technical level of application of environmental auxiliary data,enriched and perfected theory,technology and methodology of accuracy assessment and enhancement for land cover classification by remote sensing. |
PDF Full Text Request