Spatial scaling issues in the production of land-cover maps using satellite remote sensing data

Land-cover is an important data source both in its own right and as a surrogate for many environmental variables. Remote sensing is the only viable means by which such data can be acquired at regional and global scales. To produce accurate land-cover maps, the appropriate number and nature of classes must be identified. It is expected that the type of classes that can be discriminated will alter as the spatial scale (resolution) of the data changes. This hypothesis is examined in three experiments performed on satellite sensor data of three sites (SW Niger, N Norfolk and NW England). The first experiment examines the spatial structure of raw image data for evidence of domains of scale in the corresponding scenes. The results suggest that domains of scale may exist over the range of spatial scales (resolutions) studied (20m-25m to 12km), each of which may demand a separate taxonomy. The second experiment examines thematic maps (c.f. raw images) for further evidence by analysing the region size distributions, and changes in class proportions and landscape ecology indices as data are degraded from fine to coarse spatial scales. The results provide only weak confirmation of the trends identified in the raw imagery. The third experiment examines how the nature of 'pure' classes identified at some fine spatial scale changes as the data are degraded and the pixels become increasingly mixed. The detected spectral responses will change and, as a consequence, new class definitions are required to maintain an accurate representation of the land-cover. The novel elements of this thesis include: (i) an objective comparison of several measures used to quantify spatial structure in raw image and thematic map data; and (ii) the relationship between land- cover classes and the spatial resolution is examined using a new technique based on clustering in class proportion space.