We generated an annual fractional tree cover dataset (GLOBMAP) in China from 2000 to 2022 with a resolution of 250 m using the MODIS land surface reflectance products. The dataset is generated using a feedforward neural network. A total of 2000 samples are automatically extracted from the very high spatial resolution imagery on Google Earth through an auto-segmentation algorithm to calibrate and validate the model, with 80% for training and the remaining 20% for validation. Seven highly discriminative input features are extracted from the annual series of MODIS data, including values in the red (B01) and short wavelength infra-red 2.1 (SWIR 2.1: B07) bands for three key time phases, which are the day of year (DOY) of the maximum NDVI during the starting month (near 136~167), the medium month (near 197~228) and the ending month (near 259~289) of the growing season, and the yearly minimum NDVI. The root mean square error (RMSE) and Mean Absolute Error (MAE) of the dataset are 11.78% and 7.39%, respectively. The dataset provides fractional tree cover maps in China from 2000 to 2022. The data are projected by WGS 1984 and are stored with GeoTIFF format. The files are designated with the naming structure “GLOBMAP_YYYY_TreeCover_China.tif”, where “YYYY” represents the year of the data. The valid range of tree cover value is 0-100, with a scale factor of 1.0 and the unit of percentage (%).