| Abstract | The generation of large area mosaics from multiple satellite images presents unique problems not normally encountered in the fusion of aerial photographs. Low revisit cycles of satellites combined with
high incidence of cloud cover over most of the Earth's land areas necessitate the merging of scenes acquired over a broad temporal window, typically on the scale of months to years. As a result, one must deal with problems of integrating partially
overlapping scenes which exhibit diverse atmospheric and possibly thematic characteristics into a visually seamless image product. Traditionally, a host of empirical and largely cosmetic procedures have been employed to achieve radiometric
continuity. Usually, one scene is selected as a reference and the radiometric scales of all other scenes are modified to match it through, for example, grey level mean/variance normalization. Additional local operations, such as scene blending in
overlap regions, are used to further suppress inter-scene seam visibility. While these approaches have been successful in producing visually pleasing mosaics, they fail to provide quantitative measures of radiometric and underlying thematic
continuity, information which is crucial if useful landscape information is to be derived from this type of image product. In this paper we describe a unified normalization methodology which is based on detailed analyses of grey level scattergrams
derived from pixels of regions of overlap between scenes. Based on a clustering technique in this 2-D radiometric space, we are able to address a number of relevant issues including; - the identification and editing out of disparate
inter-scene information. This 'outlier' data can arise from the presence of cloud cover in one of the scenes or from significant surface cover change.
- the derivation of normalization coefficients from grey level properties of clusters found in
the edited datasets.
- the estimation of inter-scene thematic continuity based on cluster properties such as correlation and orientation in this grey level space.
- the selection of preferred reference scene candidates. These methodologies have
been applied to an example data set consisting of 38 Landsat MSS scenes covering a portion of northern Ontario.
|