| Abstract | Land cover is a fundamental parameter in the study of biosphere-atmosphere interactions and the assessment of long-term environmental change, as well as for resource management and policy formulation.
In the past, it has been difficult to produce land cover maps over large areas which are thematically consistent and which portray land cover distribution at one point in time. This situation persisted through the 1980s when high resolution satellite
data were available, for reasons of cost and the impossibility of obtaining complete cloud-free coverage during one growing season/snowfree period. Recent work in the processing and analysis of medium resolution optical data from the Advanced Very
High Resolution Radiometer (AVHRR) indicates that this type of data has high potential for classifying land cover to a fairly high accuracy provided the spatial resolution (1km) is adequate for the application of interest. In this paper, we present
the methodology and initial results of a project designed to classify the land cover of Canada as it was in 1993. To obtain the maximum possible amount of information from the satellite data, three original channels (red, near infrared, thermal
infrared) and a vegetation index obtained at 10-day intervals during the growing season are used. Various clustering approaches were examined with the goal of developing an objective and repeatable method. The results obtained so far indicate high
sensitivity of the data to land cover variations, especially in transition areas (grassland to boreal forest, boreal forest to tundra) and in areas of mixed coverage (e.g., cropland and deciduous forest; bare and wetlands in tundra). Class labelling
and accuracy assessment for huge territories such as landmass of Canada pose special challenges; these issues will be discussed in the paper. |