An evaluation of per-parcel land cover mapping using maximum likelihood class probabilities

Abstract

The Centre for Ecology and Hydrology (CEH) has recently developed a per-parcel classification procedure which integrates remotely sensed imagery with digital cartography. The aim of the work reported here was to compare the per-parcel approach with a conventional per-pixel classification through examination of maximum likelihood class probabilities. An Airborne Thematic Mapper (ATM) image with 1.25 m spatial resolution was used as a source of training data. The ATM image was spatially degraded to 10 m resolution to form the remotely sensed input and the study area was subdivided into land parcels by manual digitizing. The per-parcel approach extracted raster data from a core area described by a shrunken version of a land parcel, thus eliminating mixed boundary pixels. Boundary pixels were shown to be more variable than the core pixels and their removal improved classification confidence. The per-parcel approach calculated the mean spectral response of the extracted core pixels. This was shown to remove within-parcel variation and improve classification confidence in comparison to per-pixel class allocations. A parcel-based representation was shown to be most appropriate for mapping agricultural land cover in comparison to semi-natural areas, because agricultural landscapes have an inherent parcel structure. When land cover is heterogeneous, as in many semi-natural areas, a per-pixel classification would appear to be more appropriate. A hybrid classifier, which could switch between per-parcel and per-pixel mapping, was suggested as a powerful land cover mapping tool.