Early satellite systems with multispectral resolutions on the order of 20m (SPOT 1 4) or 30m (Landsat 4 7) have been able to obtain imagery of sufficient resolution to effectively classify areas with large contiguous areas of information classes, i.e. scenes of dense forest vegetation or cropland. However, accurate classifications of urban areas, where features on the ground change meter by meter, have been impossible to achieve with such lower resolution systems. Satellite systems like Ikonos and QuickBird with resolutions in the panchromatic band at 1 meter and resolutions in the multispectral range of 4 meters produce images of high enough resolution that classification of urban scenes can be done accurately.
[...] Viewing the resulting image (Figure the vast majority of the water and shadow pixels were indeed preserved but a host of asphalt pavement pixels were included in the image as well. Figure 4. Image showing pixels with Red < 230 and NIR < 300; water, shadow, and asphalt pixels visible. (ERDAS Imagine) After checking the DN values of the included asphalt pixels, it was clear that the asphalt pixels' DN values were well within the spectral range of the water and shadow pixels intended to be included in the image. [...]
[...] Pixels at these locations could be separated from the rest of the image as water and problems with the classification would not be nearly as formidable as with other methods. This method would have much the same effect that the vector masking approach would have by avoiding spectral properties as much as possible in removing water pixels. It would also have similar problems, however, in that it would be nearly impossible to remove all water pixels from the image based on DEM characteristics since the resolution of the DEM would most likely be much less than that of the QuickBird image. [...]
[...] (ERDAS Imagine) Also visible in Figure 6 are small bodies of water, east of Lake Wingra, which are associated with the lake but weren't included in the vector layer. The smaller lakes near the western edge and southern extent of the QuickBird scene were also not included in the vector layer. This was a considerable problem since the pixels in these water bodies and other water pixels that weren't included in the vector layer would make the vector masking approach effectively useless because the same errors of commission would occur just as they had with the first approach. [...]
[...] A bilinear interpolation produced a georectified version of the QuickBird image which could then be reprojected into WTM coordinates and overlain with the vector layer containing the lake polygons. ERDAS Imagine did not have WTM coordinates in its database of coordinate systems for reprojection so the WTM 83 parameters had to be entered manually. The parameters are summarized in Table 1. Projection Spheroid Datum Scale Factor at Central Meridian Longitude of Central Meridian Latitude of Origin Table 1. Wisconsin Transverse Mercator Parameters. [...]
[...] It was decided that an unsupervised ISODATA classification procedure should be used since a supervised classification procedure would take far too long considering the variety of different land cover types and features present in the image. The number of spectral classes outputted by the unsupervised classification was set to be 100 in consideration of the scene's spectral diversity. The spectral classes would then be merged into a handful of information classes. Some of the difficulties with classifying a QuickBird image result from its high resolution. [...]
Online readingwith our online reader
Content validatedby our reading committee