The information on land cover at national scales is critical for addressing a range of problems, including the climate change, biodiversity conservation, ecosystem assessment, and environmental modeling. In view of the problems of the existed global land cover products and the deficiency of current data fusion methods, this study aims to develop a general framework for building a hybrid land cover map by the synergistic combination of a number of land-cover classifications with different legends and spatial resolutions based on Dempster Shafer theory. With the validation of GLOBCOVER, MODIS, GLC2000 and GLCNMO in regional and category scale, the results showed that MODIS had the best consistency with the referenced data, followed by GLC2000, and the GLCNMO and GLOBCOVER had the lower consistency with the referenced data. The validated products and reference data had some categorical confusions which mainly occurred in forest, grass, shrub and cropland, especially between shrub and other categories. So shrub had the worst classification precision. Confusions demonstrated the conspicuous regional characteristics, for example, in northeast, Tibet alpine zone and the southeast zone, the confusions mainly occurred in cropland and forest, grass and shrub, cropland and grass respectively. Based those experiences, the author computed the different category weight for four land cover products using the analytic hierarchy process, which will quantify the contribution in the merging process, and completed the land cover category transformation between four land cover products through the LCCS land cover system with eight indexes of the vegetation or no-vegetation, terrestrial or water, cultivated or natural, life type, leaf type and phenomena. A multi-source integrated land cover map was generated based on the Dempster-Shafer evidence theory. Based on the volunteered data from GEOWIKI project which was a validated program for global land cover products, the forest inventory data and cross-validation method, the author evaluated the fusion result, which showed that not only in overall accuracy but also in classification accuracy, the fusion map had an apparent improvement than original land cover products. For the GEOWIKI validation, the fusion map has the highest producer accuracy in forest, grassland, cropland and bare land, but the shrub classification accuracy is lower than that with GLCNMO; The permanent ice classification accuracy is lower than that with MODIS, the water classification accuracy is lower than that with MODIS and GLC2000. For the user accuracy, all the fusion has the highest accuracy except the water and permanent ice. This is because that in the validated category point, the shrub accounts for only 3.29%, the water accounts for only 0.78%, and the permanent ice accounts for 0.87%. For the forest data, the deviations of forest area percentage in all validated regions are all less than 5%. This demonstrates the fusion map is better for the category area scale, which explains that the evidence theory should absorb every land cover product's advantage in the fusion process and make the data complementary. The uncertainty analysis about fusion results show that the overall uncertainty is in a low range about 0-0.3, and this area covers about 94.5 percent of the total study area. For example, the areas with uncertainty value in 0-0.1 mainly is located in northwest, northeast, north china, Sichuan basin and south in Taiwan province. The area with uncertainty value greater than 0.3 is scarce but aggregated, such as the areas with uncertainty value in 0.3-0.4, mainly located in north center of Inner Mongolia, north of Xinjiang, the land cover in this region are grass land and bare land. The most uncertain area is located in northeast of Gansu province, Ningxia Hui Autonomous Region and northeast of Shaanxi province, and these regions' land covers are mainly grassland, cropland and bare land, which illustrate a conspicuous landscape heterogeneity characteristic.
