The lack of annotation data is an important factor restricting the research on large-scale medium-resolution land cover classification mapping. In order to solve this problem, this paper proposes a multi-scale land cover annotation data generation method, which uses a 10 m resolution land cover product (10LC) to correct the 30 m land cover product (30LC) and generate a 30 m land cover annotation dataset (30LC dataset). In this method, the 3 × 3 window slide of the 10 m resolution product is used to match the 30 m land cover product by pixel, and the 30 m land cover annotation dataset is generated according to the maximum number of categories in the window, which effectively uses the recognition ability of the 10 m resolution product on the ground object boundary and mixed ground objects, overcomes the deficiency of 30 m resolution data detail feature extraction, and also reduces the noise impact when using a single product. Based on the 30 m land cover annotation dataset and Landsat8 imagery, this paper uses the ASPP_U-Net model to carry out land cover classification mapping in the Lancang-Mekong region. The experimental results show that the overall accuracy of land cover classification in the Lancang-Mekong region is 84.28%. Compared with directly using 30 m FROM-GLC2017 as the annotation data, the accuracy is improved by 12%.