Multi-scale accuracy assessment of land cover datasets based on histo-variograms

Due to the shortage of single-scale accuracy assessment methods, this paper proposes a multi-scale accuracy assessment method based on histo-variograms, which assesses the accuracy of land cover datasets on a pixel and sub-pixel scale. On a pixel scale, a standing-pixel is introduced as the sample tool to accurately assess datasets directly, whereas on a sub-pixel scale, a non-strictly defined standing-pixel and histo-variogram are applied to evaluate the accuracy of the dominant type in a mixed pixel with different areas and spatial structures. Taking as the experimental area a typical region in northern Zhejiang Province, China, and Landsat TM/ETM+ images as the reference data, accuracy assessment experiments of five large-scale land cover datasets, i.e., UMD, IGBP DISCover, MOD12Q1-2001, GLC2000 and GlobCover2009 are carried out. The result shows that the proposed multi-scale accuracy assessment method can provide a comprehensive evaluation of datasets along with abundant multi-scale accuracy information. The accuracy assessment on a pixel scale can eliminate the error caused by spatial disagreement between reference data and datasets, making the result more objective, while the accuracy assessment on a sub-pixel scale can effectively reflect the relationship between the accuracy and feature attributes, which refers to the feature's spatial structure and area.