‘Single-Family Housing (less structured)’ (D) is also located in the southern and western parts of the study area (Fig. 3). These areas began as small settlements around railway stations. In the 1910s, the areas grew into large areas with single-family houses, and several areas separated from the city. These independent areas were planned so they did not significantly intervene with the often hilly terrain (hence, ‘less structured’). These areas primarily consist of single-family houses, which are not necessarily made of wood,and gardens (Utställningsförslag 1997, 1997).
‘Narrow Block Housing’ (E) is found in the southern and western parts of the study area (Fig. 3) and is primarily composed of blocks of flats that are 3–4 storeys high and 7–10 m wide. Inspired by the functionalistic movement, these buildings appeared during the 1930s; they are often interspersed in the natural landscape with preserved slopes and other terrain features (Andersson, 1998).
A cloud-free SPOT5 satellite scene captured on 4 June 2008 with a resolution of 10 10 m was used in this study. Lantmäteriet, the Swedish mapping, cadastral and land registration authority,orthorectified the data and used 10–20 ground control points (GCPs) for each SPOT5 scene. The GCPs were selected from orthophotos with a 1 1 m resolution, and the elevation information was taken from a 50 50 m digital elevation model. The root mean square value was expected to be 2–5 m. The study area covered 7 6 km of the satellite scene (Fig. 5).
In the WICS procedure, the input data, consisting of a multispectral satellite scene, must be reduced to a number of input
classes (step 1 in Fig. 1). Therefore, a subset of the satellite image,including all four spectral bands, corresponding to the study area was processed using a per-pixel-based iterative self-organising data analysis tool (ISODATA) in ArcGIS Spatial Analyst. The ISODATA tool assigns a number, in this case 20, of cluster vectors in the dimensional space created by the four spectral bands. The pixels are then assigned to the nearest cluster vector in Euclidian terms, and new cluster vectors are calculated based on the position of the assigned pixels. The procedure is repeated until very few pixels change clusters. Therefore, this procedure is very similar to the aforementioned K-mean clustering, except that ISODATA allows for different numbers of clusters (merging). The benefit of using this method is that it is possible to reduce the data complexity whilst retaining as much of the variance as possible. The result is the reduction of the dataset containing four spectral bands into a dataset containing only 20 classes, as shown in Fig. 5. The number 20 is based on a limitation of the WICS method: more