<p>This data is based on GLC_FCS10 global data and superimposed on the Xinjiang data extracted from the administrative boundary mask of the Xinjiang Uygur Autonomous Region in 2024. The land use types are divided into 10 categories and 30 categories: cultivated land, forest, shrub, grassland, tundra, wetland, impermeable surface, bare land, water, and permanent ice and snow. The data format is grid (.tif format), and the data spatial resolution is 10 meters. The data coordinate system is GCS_WGS_1984, and the data year is 2023. Two forms of data are provided this time. One is stored in the form of 5°×5° tiles with latitude and longitude according to the original data without masking (stored in the folder GLC_FCS10maps_2023_5x5_XJ), and the other is stored in the form of 5°×5° tiles after masking (stored in the folder GLC_FCS10maps_2023_5x5_XJ_Clip). The layer color file (GLC_FCS10maps_colormap.lyr), the full text of the data paper and the GLC_FCS10 user manual are also provided. </p>
The global data is data released on the zenodo platform by Liu Liangyun, a researcher at the Institute of Aerospace Information Innovation of China Academy of Sciences. The dataset is referred to as GLC_FCS10 for short. The dataset integrates Sentinel-1 and Sentinel-2 time series images and is built with the help of the Google Earth Engine platform. The overall accuracy reaches 83.16% globally and 85.09% in the United States, which is significantly superior to many publicly released 10-meter products. Compared with mainstream 10/30-meter products, GLC_FCS10 shows obvious advantages in terms of classification richness, spatial detail expression capabilities and classification accuracy.
The production of the GLC_FCS10 dataset adopts a hierarchical, multi-source fusion fine processing framework: First, based on multi-source prior products (Such as GISD30 impervious surface, GWL_FCS30D wetland, etc.) High-confidence training samples are generated through spatio-temporal consistency analysis and spatial filtering, and the spectral mass center method is used to down-scale the 30m sample to 10m resolution, and the samples are assigned according to hierarchical strategies (The impermeable surface is equally distributed/the natural land type is distributed in proportion to the area); Secondly, the time-series percentile characteristics of Sentinel-2 optical images are integrated (including NDVI and other indices) and texture features, polarization characteristics of Sentinel-1 radar data, and ASTER GDEM terrain parameters to build a multi-source feature set; Then a three-level classification process is used-first distinguishing impervious surfaces from natural landmarks (Random forest model combined with morphological optimization), re-isolating wetlands (Independent modeling of coastal and inland areas), finally refined classification of the remaining 20 types of natural features and local adaptive modeling through 983 5°×5° grids (Including neighborhood 3×3 grid samples) to solve spatial heterogeneity; finally, global verification (56,121 sampling points, OA=83.16%) and third-party verification (U.S. LCMAP dataset, OA=85.09%) confirmed that its accuracy exceeds mainstream products, especially in wetland identification (F1=66.63%) and cultivated land classification.
The overall accuracy of GLC_FCS10 reaches 83.16%, the kappa coefficient is 0.789 globally, and the overall accuracy in the United States is 85.09%. Compared with five released 10-meter or 30-meter land cover products, it also shows that GLC_FCS10 has higher accuracy.
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
10.5281/zenodo.14729664