Thematic land-cover map assimilation and synthesis: the case of locating potential bioenergy feedstock in eastern Ontario, Canada

Abstract

Insufficient spatial coverage of existing land-cover data is a common limitation to timely and effective spatial analysis. Achieving spatial completeness of land-cover data is the most challenging for large study areas which straddle ecological or administrative boundaries, and where individuals and agencies lack access to, and the means to process, raw data from which to derive spatially complete land-cover maps. In many cases, various sources of secondary data are available, so that land-cover map assimilation and synthesis can resolve this research problem. The following paper develops a reliable and repeatable framework for assimilating and synthesizing pre-classified data sets. Assimilation is achieved through data reformatting and map legend reconciliation in the context of a specific application. Individual maps are assessed for accuracy at various geographic scales and levels of thematic precision, with an emphasis on the ‘area of overlap’, in order to extract information that guides the synthesis process. The quality of the synthesized land-cover data set is evaluated using advanced accuracy assessment methods, including a measure describing the ‘magnitude of disagreement’. This method is applied to derive a seamless thematic map of the land cover of eastern Ontario from two disparate map series. The importance of assessing data quality throughout the process using multiple reference data sets is highlighted, and limitations of the method are discussed.

Keywords:

• land-cover map
• bioenergy
• data quality
• accuracy assessment
• error magnitude

Acknowledgments

The authors acknowledge and thank the Ontario Graduate Scholarship Program, the Department of Geography at Queen’s University and the ForValueNet NSERC Strategic Network for providing academic funding and support of this research. A thank you is also extended to the anonymous referees for sharing their research wisdom and for valuable insights on preceding drafts of the paper. Responsibility for errors rests solely with the authors.

Notes

1. 1. There is also a burgeoning literature on ‘data fusion’ (for an early review, see Pohl and Van Genderen 1998). These methods/algorithms are designed to synthesize raw remotely sensed data rather than secondary land-cover data, and so they are not considered in this paper.

2. 2. Incidentally, the average difference between the number of samples collected for each class through a random process and what would be expected using a stratified sample is only 3.2% in the OLCDB map and 3.5% in the SOLRIS map, which means that all classes are well represented relative to their map proportion.