Developing Spatial Re-classification Techniques for Improved Land-use Monitoring Using High Spatial Resolution Images

The reasons for the poor performance of conventional, per-pixel classification algorithms applied to satellite sensor images of urban areas are examined. It is argued that standard algorithms are poorly adapted to distinguish between different urban land-use categories, particularly in high spatial resolution images, due to the complex spatial pattern of spectrally distinct land-cover types in urban areas. Alternative techniques need to be developed which make use of both spectral and spatial information within the scene. This study examines one such technique that attempts to derive information on land use in two stages. Firstly, by performing a low-level segmentation of the image into a few, broad land-cover types. Secondly, by grouping the classified pixels into discrete land-use categories on the basis of the frequency and spatial arrangement of the class labels. The second stage is performed using a procedure developed in this study, referred to as a SPAtial Re-classification Kernel (SPARK). This examines the number of occasions on which different types of land cover are adjacent to one another, on a pixel-by-pixel basis. This information is used to construct an 'adjacency vector' for the central pixel in the kernel. Land use is inferred by comparing the derived adjacency -vectors with those of previously selected sample areas of the candidate land-use categories. Using this technique, an overall accuracy of greater than 85% is obtained for a SPOT-HRV multispectral sub-scene of London.