Combine three 10m GLC products with OSM data to generate reliable training labels, and use these labels along with 1m resolution images from Google Earth to train the proposed framework. This framework solves the label noise caused by resolution mismatch between images and labels by combining resolution preserving backbone, weakly supervised module, and self supervised loss function, thereby automatically improving VHR land cover results without the need for manual annotation. Based on large-scale storage and computing servers, the 73.25TB dataset was processed to obtain SinoLC-1 covering approximately 9.6 million square kilometers throughout China.

    The SinoLC-1 product is validated using a visual interpretation validation set consisting of over 100000 random samples and a statistical validation set collected from official land survey reports provided by the Chinese government. The verification results indicate that the overall accuracy of SinoLC-1 is 73.61%, with a kappa coefficient of 0.6595. The validation of various provinces and regions further demonstrates the accuracy of this dataset throughout China. In addition, statistical verification results indicate that SinoLC-1 is consistent with the official survey report, with an overall misestimation rate of 6.4%. In addition, SinoLC-1 was compared with five other widely used GLC products. The results indicate that SinoLC-1 has the highest spatial resolution and the finest landscape details.