The 30m annual land cover dataset in Loess Plateau from 2013 to 2023

We created the first annual Loess Plateau Land Cover Dataset (LLCD) based on a deep learning framework. First, we selected a widely recognized and verified multi-source dataset, automatically extracted three commonly confirmed stable samples from them, and combined them with our visual interpretation samples to form the standard training labels for this study. At the same time, the two types of data were used as noise labels to assist the training of the model. The preprocessed images, standard labels, and noise labels were combined into training pairs to jointly train the proposed TSFNet model.Compared with existing land cover products, LLCD has better performance in the overall accuracy of the Loess Plateau and the classification of easily confused land types. Based on the quantitative evaluation of 5400 independent validation samples, LLCD's OA reached up to 92.59%, and the average F1score of the more difficult to identify types such as cultivated land, grassland, and bare land was greater than 91%.