Land use and cover changes on the Loess Plateau: A comparison of six global or national land use and cover datasets

Abstract

Image classification often produces large deviations between land use and land cover (LULC) datasets and the ‘real’ changes, leading to uncertainty in the results of LULC related assessments and the propagated impacts through modelling. LULC products are widely used as input for various large-scale climatic, ecological and hydrological models, but the accuracy and authenticity associated with data quality are rarely fully considered. In the study, six widely used global or national LULC datasets, MODIS-MCD12Q1, EAS CCI-LC, GlobeLand30, GLASS-GLC, CAS-CLUDs and ChinaCover, are used to assess the consistency and reliability of LULC on the Loess Plateau, where land cover has undergone major changes due to “Grain to Green” Project. Results show that MODIS and GLASS products have low quality, with the overall accuracy of 55.3~58.2% and 34.7~39.4% respectively, and the areal and spatial results cannot reflect the real changes of the Loess Plateau. Large areas of croplands in MODIS-MCD12Q1 are classified as natural grassland. Croplands in GLASS-GLC are overestimated in the central parts of the Loess Plateau. Both of MODIS and GLASS products are hard to separate woodlands from grasslands. ESA CCI-LC has higher classification accuracy (73.9%~74.2%) than the released MODIS and GLASS products. The woodlands in ESA CCI-LC is relatively underestimated than that of CAS-CLUD and ChinaCover, and the conversion feature from cropland to forest and grasses is almost absent on ESA CCI-LC maps. Although GlobeLand30 has a high overall accuracy at 86. 6~86.7%, it is inadequate to get the characteristic of returning of cropland to forest and grasses. The most similar land covers are CAS-CLUDs and ChinaCover, which are considered to have highest classification accuracy ranging from 89.4% to 91.6% and can reflect the actual LULC status and its changes on the Loess Plateau. A blending LULC dataset is developed and the overall accuracies for all classes can be improved by 1.63–7.49%.