全球尺度多源土地覆被数据融合与评价研究

精确的全球及区域尺度土地覆被遥感分类数据是全球变化、陆地表层过程模拟、生态文明建设及区域可持续发展等研究的重要基础数据。本文以5套全球土地覆被数据集GLCC、UMD、GLC2000、MODIS LC、GlobCover为研究对象,结合MODIS VCF、MODIS Cropland Probability以及AVHRR CFTC数据集,设计一种基于模糊逻辑思想的证据融合方法实现上述多源土地覆被信息的决策融合,生成一套依据植物功能型分类的全球1 km土地覆被融合数据SYNLCover。结果显示,与5套源土地覆被数据集相比：① 在总体一致性精度上,SYNLCover的8个生物形态类型和12个目标类型的平均总体一致性精度最高,分别约为65.6%和59.4%,其次依次是MODIS LC、GLC2000、GLCC和GlobCover,UMD的最低,分别约为48.9%和42.6%,而且SYNLCover与5套源土地覆被数据集两两相比的总体一致性都是最好的;② 在类型一致性精度上,除灌丛类型外,SYNLCover中包括森林、草地、耕地、湿地、水体、城镇建筑和其他7种生物形态类型,以及森林类型的5种叶属性的平均一致性精度也是最高的,如其他类型的平均一致性精度可达67.73%;③ 除灌丛和湿地类型外,SYNLCover的其余6种生物形态类型的平均一致性精度均比其在5套源数据中相应的一致性精度的最大值提高了10%~15%左右;森林类型的5种叶属性的一致性精度也提高了约10%。SYNLCover分类精度的提高反映了本研究设计的多源数据融合方法的可行性和有效性。

Data fusion and accuracy evaluation of multi-source global land cover datasets

Accurate global and regional land cover classification datasets based on remote sensing are of fundamental importance in research on global changes, land surface process modeling, ecological progress, and regional sustainable development and so on. The overall objective of this study is to present a decision-fuse method that integrates existing multi-source land cover information into a 'best-estimate' dataset using fuzzy logic. Combined with another three global datasets, i.e., MODIS VCF (Vegetation Continuous Field), MODIS Cropland Probability, and AVHRR CFTC (Continuous Fields of Tree Cover), this method is applied to five global land cover datasets (GLCC, UMD, GLC2000, MODIS LC, and GlobCover) to generate a new 1-km global land cover product SYNLCover with desired legends, which are properly defined in terms of plant functional types. Pixel-based comparisons among these six global land cover datasets are performed, and results reveal that compared with five original global land cover datasets: (1) In terms of map-specific consistency, overall consistencies of both eight life forms and twelve objective legends of SYNLCover are the highest, accounting for about 65.6% and 59.4%, respectively; followed by the accuracy of MODIS LC, GLC2000, GLCC, and GlobCover in a descending order, and the lowest map-specific consistencies of life forms and objective legends are separately 48.9% and 42.6% in UMD. Besides, among all dataset pairs, SYNLCover agrees best with each original land cover dataset regarding the occurrences of life forms and leaf attributes. (2) In terms of class-specific consistency, it is suggested that SYNLCover gets the highest average class consistencies for all the five leaf attributes, as well as major life forms except Shrubland, among which the consistency for Others in SYNLCover is up to 67.73%. (3) For Trees, Grassland, Cropland, Water, Urban and built-up and Others, SYNLCover shows particular improved average class-consistencies by about 10% to 15% over the maximum consistency of original datasets, and the consistencies of five leaf attributes in SYNLCover also increases by about 10%. This study indicates a successful integration of multi-source land cover information into a new refined dataset with improved characteristics scientifically.