Global land cover map validation, comparison and integration for different user communities

Observation of global-scale land cover is of importance to international initiatives, governments, and scientific communities that endeavour to understand and monitor changes affecting the environment. Various applications such as climate models, ecosystem modelling and hydrological models use a number of global land cover (GLC) maps that were produced from different initiatives. The users have different requirements regarding spatial, temporal and thematic aspects of GLC maps as well as their accuracy. For example, climate modellers typically use GLC maps at 1km spatial resolution or coarser whereas this resolution is too coarse for land change science studies to detect small-scale changes. Furthermore, to determine the fitness of GLC maps for certain applications, map accuracy assessments need to consider the perspectives of the users as confusion between certain classes can have a strong impact on specific applications whereas for other applications they are inconsequential. Therefore, generation and assessment of GLC maps needs to account for different user requirements and perspectives. This PhD research aimed to account for different user requirements in assessing, comparing and as well as improving GLC maps.

Firstly, the characteristics of current GLC reference datasets that have been used for calibration and validation of GLC maps were reviewed and analysed. Findings revealed varying GLC reference dataset suitability levels depending on the reference data characteristics, user requirements and target maps. Nonetheless, several datasets (LC-CCI, GOFC-GOLD, FAO-FRA and Geo-Wiki) were identified as generally being suitable for re-use for multiple user groups. This highlights the potentiality of GLC reference datasets for multiple uses and public access of existing reference datasets in improving the usability of the datasets outside their intended use.

Secondly, a comparative assessment of thematic accuracies of GLC maps based on an existing reference dataset was conducted. The Globcover-2005 reference dataset was processed to assess and compare Globcover, LC-CCI and MODIS maps for the year 2005. These maps were evaluated from the perspective of several user applications using a weighted accuracy assessment procedure. Overall accuracies of the maps ranged between 61.3 ± 1.5% and 71.4 ± 1.3%. The overall weighted accuracy varied between 80-92% for the considered applications. The latter accuracy is higher because confusions between some classes were deemed inconsequential for the applications considered. To determine fitness of use of GLC maps, accuracy of GLC maps should be assessed per application; there is no single-figure accuracy estimate expressing map fitness for all purposes.

Thirdly, this research assesses the spatial accuracy of Globcover-2009, Land Cover-CCI-2010, MODIS-2010 and Globeland30 in Africa using publicly available GLC reference datasets. Spatial accuracy was modelled by the spatial autocorrelation structures of the local correspondence between map and reference data. Created correspondence maps showed spatial patterns indicating zonal differences in the degree with which different GLC maps matched the reference data. The results showed the potentiality of integrating current GLC maps along with reference data to create an improved GLC map. Different integration methods including geostatistical approaches were tested and assessed by cross-validation. The integration methods based on geostatistical approach resulted in 4.5%–13% higher correspondence with the reference LC than any of the input GLC maps. An improved GLC map was presented based on the based integration method. This GLC map has 10% higher global correspondence with reference LC than the individual input maps.

 

Figure 1. The integrated GLC map

Lastly, the thematic requirements of different GLC map users was addressed and a  concept of producing GLC maps with user-specific legends based on area fraction maps of LC classes is proposed. It is demonstrated by creating GLC maps with user-specific legends from the perspectives of land system modelling and biodiversity assessments. This PhD research demonstrates the importance of accounting for the requirements and perspectives of user applications in validating, comparing and improving GLC maps. The work also includes improving the efficient use of existing GLC reference datasets, comparative accuracy assessment of GLC maps using both the design based and model based approaches as well as presenting an integration method to improve current GLC maps to better meet different application needs.