Creation of Digital City Models – From a Single High-Resolution Satellite Image
A 3-D Digital City Model has been created by extracting the relative height of buildings, using single Quick Bird satellite image, collected at an off-nadir viewing angle.
Study area and inputs:
A sample of Quick-Bird imagery acquired on December 28, 2001, having an off-nadir collection angle of 24.5 degrees was procured for this study (Fig. 1). The area chosen was a core city portion of Abu Dhabi, the capital of the United Arab Emirates, covering approximately 2 sq. kms of area. This part of the city has a well-organized network of roads and high-rise buildings. These buildings are distinctly separated from one another.
Fig 1. Quick-Bird Multi-spectral-PAN merged image at 0.67m ground resolution
Methodology and results :
A true representation of the ground is required for obtaining a high success rate in determining Land Use/Land Cover classifications. Most image processing software tools are based on spectral classification, with recent object-oriented classifiers also providing a very good classification of the image. However, if we classify two images of the same area, one an on-nadir collection angle and the other an off-nadir view, the results of the off-nadir view will show an increase in the area of the objects that have a higher relative height. The reason being that these objects expose more of their lateral sides, at the cost of low height features, which get concealed in the shadow.
Keeping this in mind, the authors adopted a semi-automated method of feature extraction to arrive at a true-to-ground classification with maximum relevant information and the least effort. The classes of more relevance (buildings) were extracted manually, and other classes (grass, open and trees) were derived from semi-automatic image classification techniques.
- A base layer was created in which grass, open bare grounds, and trees were classified spectrally and all the remaining features were merged into the class “open area”.
- Road centrelines were captured by placing vertices along road junctions and curvatures. A buffer was defined and created for the roads, in order to closely match their width with the roads on the image. This method of extraction of roads helped in representing roads in the best possible way. An automatic (surface material based) classification would have hidden roads in areas with high-rise buildings, since these buildings would have masked the roads, given the off-nadir collection. This is a common phenomenon in high-density urban areas.
- Building rooftops were captured as polygons in a CAD environment. The location of these polygons were then edited/moved to the actual building base, to replicate the exact ground coverage of each building. The polygons were processed in two different ways.
- The building height computation is based on the premise that the distance between the base and top of the buildings represents the height of the building. To quantify the height, the distance (in map units) between centroids of the building top and base polygons is computed and assigned to the elevation property of polygons. To create a DEM, these polygons are then converted to a raster file, in which, all the non-building pixels inherit a ‘Null’ value, whereas the building pixels contain the ‘Z’ factor (Fig 2).
- The building rooftop polygons are deleted and the base polygons are maintained for processing, as they represent the built up class in the Land Use layer.
4 .The final Land Use layer (Fig. 2) is generated by merging: a. Building base polygons converted to raster, b. Buffered road layer converted to raster, and 3. The open, grass and tree classes resulting from the image processing task.
The generated DEM did not comprise true building heights, but represented a factor that was in proportion to the building heights. If the actual measured height of one of the represented buildings is available in the area of coverage, then by applying a simple multiplication factor to relative height, the true heights of all buildings can be obtained. Although no actual ground measurements have been performed to validate the claim, it is expected that accuracy within 2 m (three pixels on ground for the test image) of actual height can be obtained by using this method.
This mixed method of feature extraction, with a height value assigned to each building, gives a 3-D affect that can be viewed in a fly-through mode. It can also be used for performing line-of-sight or accentuation of signal calculations.
Fig 2: Perspective view of the Land Use layer, draped over the relative building height DEM, the Land Use classes represented are roads, open area, trees, grasslands and buildings.