Assessment of the cropland classifications in four global land cover datasets: A case study of Shaanxi Province, China

Accurate and reliable cropland surface information is of vital importance for agricultural planning and food security monitoring. As several global land cover datasets have been independently released, an inter-comparison of these data products on the classification of cropland is highly needed. This paper presents an assessment of cropland classifications in four global land cover datasets, i.e., moderate resolution imaging spectrometer (MODIS) land cover product, global land cover map of 2009 (GlobCover2009), finer resolution observation and monitoring of global cropland (FROM-GC) and 30-m global land cover dataset (GlobeLand30). The temporal coverage of these four datasets are circa 2010. One of the typical agricultural regions of China, Shaanxi Province, was selected as the study area. The assessment proceeded from three aspects: accuracy, spatial agreement and absolute area. In accuracy assessment, 506 validation samples, which consist of 168 cropland samples and 338 non-cropland ones, were automatically and systematically selected, and manually interpreted by referencing high-resolution images dated from 2009 to 2011 on Google Earth. The results show that the overall accuracy (OA) of four datasets ranges from 61.26 to 80.63%. GlobeLand30 dataset, with the highest accuracy, is the most accurate dataset for cropland classification. The cropland spatial agreement (mainly located in the plain ecotope of Shaanxi) and the non-cropland spatial agreement (sparsely distributed in the south and middle of Shaanxi) of the four datasets only makes up 33.96% of the whole province. FROM-GC and GlobeLand30, obtaining the highest spatial agreement index of 62.40%, have the highest degree of spatial consistency. In terms of the absolute area, MODIS underestimates the cropland area, while GlobCover2009 significantly overestimates it. These findings are of value in revealing to which extent and on which aspect that these global land cover datasets may agree with each other at small scale on each ecotope region. The approaches taken in this study could be used to derive a fused cropland classification dataset.