| Abstract | During the last decade, a number of initiatives have been undertaken to create systematic national and global data sets of processed satellite imagery. An important application of these data is the
derivation of large area (i.e. multi-scene) land cover products. Such products, however, can be expected to exhibit internal variations in information quality for two principal reasons. First, they have been assembled from a multi-temporal mix of
satellite scenes acquired under differing seasonal and atmospheric conditions. Second, intra-product landscape diversity will lead to spatially varying levels of class commission errors. Detailed modelling of these variations with conventional ground
truth is prohibitively expensive and hence an alternative, albeit indirect, accuracy assessment method must be sought, preferably one that provides a measure of classification confidence at the pixel level. In this paper we propose a method for
confidence estimation based on the analysis of classification consistency in regions of overlapping image coverage between Landsat scenes from adjacent orbital paths and rows. We have developed an overall land cover mapping methodology that exploits
consistency evaluation both to improve classification performance during product generation and to conduct post-generation accuracy assessment. This methodology has been implemented within a proto-type mapping system, QUAD-LACC (Guindon, 2002), and
is being employed to derive synoptic land cover products of the Great Lakes watershed from archival Landsat Multi-spectral scanner (MSS) imagery. Our methodology involves an independent clustering and classification of each Landsat scene. The
interpretation quality of each scene is assessed by comparing its classification of pixels in overlap regions with those of its four nearest neighbouring scenes. Consistency statistics are then used both to identify mislabelled clusters and to assign
a measure of classification confidence to each cluster. Finally, the scene-based classifications are 'composited' to generate a final seamless land cover product and an accompanying confidence layer. At the pixel level, this layer quantifies a
cumulative confidence that encapsulates the number of independent label estimates available, their level of agreement and the inherent confidence of their parent clusters. It should be noted that others have suggested using overlap regions for
accuracy characterization, not in classification but rather landscape metric estimation (e.g. Brown et al., 2000). |