ABSTRACT
This chapter describes a new approach for converting raster remotely sensed imagery to the vector data model. Most techniques for converting from the raster to the vector data model amount to threading vector boundaries between pixels representing different classes. The new approach, which comprises two stages, allows the threading of vector boundaries through the original image pixels, thereby facilitating sub-pixel geometric accuracy. The first stage is concerned with estimating the proportions of specific classes that the pixel may represent. Several techniques are available for this including mixture modelling, neural networks, and fuzzy c-means classification. In the second stage, an algorithm is implemented to determine where the relative proportions of each class occur within each pixel. The algorithm works by assuming spatial dependence within and between pixels. The pixels of the image are divided into several smaller units, amounting to an increase in spatial resolution, and the land cover is allocated to the smaller cells within the larger pixels so as to maximize spatial dependence. The resulting map allows vector boundaries to be threaded between the smaller units. The approach presented raises some important issues for further research.
