Abstract
Geographic object-based image analysis (GEOBIA) has become an area of increasing research interest in the field of remote sensing. Land Cover Map 2000 (LCM2000) of the United Kingdom represents an early use of such techniques to create a national mapping product. The work described here makes use of the object-level metadata created in the production process of LCM2000 to assess uncertainty within the database. Many mapping applications using GEOBIA techniques use segmentation algorithms to generate parcels from imagery that have a similar spectral response. These parcels are then classified to create the final map product. Here we show that by recording detailed metadata at the object level the final map can still be analysed further to assess the uncertainty in the classification of the parcel, or to re-evaluate parcels or groups of parcels for a new system of classification, by developing criteria for classification. This can be done not only using the parcels own metadata but also that from the immediate locality. Breaking down the spatial unit of the parcel, using a region-growing technique, prevents the user from being constrained by the parcels once they have been generated and potentially increases the power of any uncertainty analysis using the available metadata in a potentially more informative way. The example presented here uses these techniques to assess the potential extent of woodland within the area of the National Forest. The technique shows considerable promise in analysing greater complexity in land cover than is possible from a standard land cover map. Object-level metadata can be used to provide an analysis of uncertainty in classification and re-evaluate one parcel in the light of the properties of its neighbours. This can identify the areas where classification uncertainty may be problematic to a user and also the areas that could be more accurately described using mixed classes to represent the heterogeneity at the location.
Acknowledgments
The authors would like to thank the University of Leicester for funding this work and the Centre for Ecology and Hydrology and the National Forest Company, for supplying datasets, with a special thanks to Annette McGrath at the National Forest Company for her support. Also thanks go to Lex Comber for helpful discussions in the early stages of this work and supplying data from his own research.