Abstract: This paper describes the production of China land cover data for version 1.0 Global Map with the emphasis on the developed land cover classification methodology-a combination of unsupervised image clustering and supervised decision tree classification. The 23 MODIS/TERRA 16 d composites (MOD43B4 product) acquired over the whole year of 2003 are the main remote sensing data sources. Firstly,the composite in the period of August 13 to 28 which corresponds to the greenest image of the year was segmented with unsupervised clustering algorithm. The output class number is set to 60 in order to keep any subtle difference between different land cover types. Further clustering is performed to the time average of four composites starting on July 12 and ending on Sept. 13 with output class numbers reduced to 40. The polygons output from the two-round clustering are then classified into 16 land cover types using a decision tree classifier. The MODIS spectral reflectance and two indexes of multiple MODIS composites consist of the variables of the decision tree. The accuracy of the final classification result was evaluated by comparison to existing 1∶500 000 China land use map and 1∶1 000 000 China vegetation map and other high resolution images. This evaluation indicates the overall accuracy of China land cover product is 89.14%. At the same time,22 filed sites were selected around Beijing as check points to compare the classification result with the ground truth,which suggests a similar evaluation of the accuracy.