Canadian Forest Service Publications
An approach for edge matching large-area satellite image classifications. 2007. Wulder, M.A.; Han, T.; White, J.C.; Butson, C.R.; Hall, R.J. Canadian Journal of Remote Sensing 33(4): 266-277.
Issued by: Pacific Forestry Centre
Catalog ID: 27427
CFS Availability: PDF (download)
Large area land cover mapping based on remotely sensed data often requires combining individual, or large groups, of classified images to produce final map products. Operational and logistic considerations are typically confronted when classifying medium spatial resolution satellite imagery (i.e. Landsat), with the mapping typically partitioned for logistical purposes based upon spectral, ecological, or political considerations, or combinations thereof. At the locations where logistically based production zones join, visual discontinuities can emerge. Transparent and systematic approaches for addressing the discontinuities are desired for the Earth Observation for Sustainable Development of Forests (EOSD) project. This large-area land cover mapping project is producing map products for Canada’s forested ecozones. A distributed implementation plan, largely based upon grouping provincial and territorial political units, was followed for production. Within each production zone scene-to-scene discontinuities are rare, typically related to image acquisition date and related phenological state. In contrast at the production zone boundaries discontinuities can emerge due to differences in support data available, or more commonly due to differences in the attribution of density classes. Of the over 475 scenes classified, it is projected that less than 30 (about 6.3% of total) will require processing to minimize the cross-boundary discontinuity.
Options for mitigating the discontinuities are described and demonstrated in the context of different scenarios of overlap found along the EOSD production zone boundaries (complete overlap, partial overlap, and no overlap) using two subsets of a Landsat scene along the shared provincial border between British Columbia and Alberta, Canada. Analysis of image gradients provides a quantitative basis for identification of discontinuities and also relates the results of the likelihood based re-labelling process. Through this process only density descriptors of cover-types are altered, largely maintaining classification integrity. The process as presented is generic and is suitable for addressing edge discontinuities that can emerge when undertaking a large-area land cover classification project.