Utilizing the global land cover 2000 reference dataset for a comparative accuracy assessment of 1 km global land cover maps

Many investigators use global land cover (GLC) maps for different purposes, such as an input for global climate models. The current<br/>GLC maps used for such purposes are based on different remote sensing data, methodologies and legends. Consequently,<br/>comparison of GLC maps is difficult and information about their relative utility is limited. The objective of this study is to analyse<br/>and compare the thematic accuracies of GLC maps (i.e., IGBP-DISCover, UMD, MODIS, GLC2000 and SYNMAP) at 1 km<br/>resolutions by (a) re-analysing the GLC2000 reference dataset, (b) applying a generalized GLC legend and (c) comparing their<br/>thematic accuracies at different homogeneity levels. The accuracy assessment was based on the GLC2000 reference dataset with<br/>1253 samples that were visually interpreted. The legends of the GLC maps and the reference datasets were harmonized into 11<br/>general land cover classes. There results show that the map accuracy estimates vary up to 10-16% depending on the homogeneity of<br/>the reference point (HRP) for all the GLC maps. An increase of the HRP resulted in higher overall accuracies but reduced accuracy<br/>confidence for the GLC maps due to less number of accountable samples. The overall accuracy of the SYNMAP was the highest at<br/>any HRP level followed by the GLC2000. The overall accuracies of the maps also varied by up to 10% depending on the definition<br/>of agreement between the reference and map categories in heterogeneous landscape. A careful consideration of heterogeneous<br/>landscape is therefore recommended for future accuracy assessments of land cover maps.<br/>* Corresponding author<br/>1. INTRODUCTION<br/>The consistent and continuous observation of land cover is one<br/>of the most important foundations for understanding the Earth’s<br/>environment and ecosystems (Verburg et al., 2011). Currently,<br/>several global land cover datasets (GLC) have been developed<br/>and these datasets are evolving towards higher spatial resolution<br/>(Gong et al., 2013; Mora et al., 2014) . Most GLC maps were<br/>developed by individual groups as one-time efforts and the<br/>subsequent mapping standards reflect the varied interests,<br/>requirements and methodologies of the originating programs<br/>(Herold et al., 2006). These differences of GLC maps and the<br/>effects of their quality on the model outcome are not always<br/>considered when selecting a map as an input for specific<br/>modeling applications (Verburg et al., 2011). Uncertainties of<br/>GLC maps can result in considerable differences in modeling<br/>outcomes (Hibbard et al., 2010; Nakaegawa, 2011; Verburg et<br/>al., 2011).<br/>The accuracies of GLC maps are assessed using independent<br/>validation datasets and regional maps or cross validated against<br/>training datasets. The results of accuracy assessments of<br/>previous maps indicate that overall area-weighted accuracy is<br/>around 70% for the existing GLC maps (Defourny et al., 2012).<br/>However, the use of different approaches in the GLC map<br/>production (e.g., classification scheme, data sources and<br/>algorithms) as well as in validation data collection (e.g.,<br/>sampling scheme, data source and method of reference<br/>classification) raise inconsistency issues and make map<br/>comparisons difficult. Several comparative analyses of land<br/>cover maps were conducted at regional level