Extraction of Urban Built-up Land Features
After producing SAVI, MNDWI, and NDBI images, a new image dataset was created, which used these three new images as three bands. The further extraction of urban built-up land was carried out based on this new dataset. The change from an original seven-multispectral-band image into the three-thematic-band image largely reduces correlation among three bands (Table 4). Consequently, three major urban land-use classes, vegetation, water, and built-up land are well separated (Figure 4). Compared with the original image, moreover, spectral signature analysis was also greatly simplified owing to the reducing of banddimensions
(Figure 3 and Figure 5). Three methods were used to extract built-up land features from the new images composed of the three thematic-oriented bands, which are principal components analysis (PCA), logic calculation, and supervised classification methods. PCA is a method that examines principal components eigenvector loadings to decide which of the PC images will concentrate information related directly to the theoretical spectral signatures of specific target materials. The technique is able to predict whether the target material is represented by bright or dark pixels in the relevant PC image according to the magnitude and sign of the eigenvectors. Table 5 describes the PC transformation on the new images based on the covariance matrix and is the base for identifying which PC has the greatest loadings (values) for NDBI band (representing
the built-up land class), but that also has opposite signs (_ or _) with SAVI and MNDWI bands. It is obvious that the built-up lands cannot be identified from PC3 as all three input bands have positive loadings in the two images, and is also difficult to be separated from vegetation in PC1 of the two images because both NDBI and SAVI bands in Quanzhou image have close positive loadings, and the NDBI band of Fuzhou image only has a small loading (0.014). Therefore, the built-up lands can only be mapped by PC2 based on a strong positive or negative loading with an opposite sign from SAVI and MNDWI bands. The negative sign of the loading for NDBI-band in Quanzhou image indicates that the built-up lands will be present by dark pixels. Accordingly, by negating (i.e., multiplying by _1) PC2 image, built-up lands of the image were mapped as bright pixels. Finally, a threshold value (Table 6) was used to extract built-up features from the PC2 image to form a binary-image with the built-up land class assigned a value of 1 and all non-built-up land classes a
value of 0 (Figure 6a and 6f). The second method is an “if-then-else” logic calculation through a band spectral signature analysis. A simple rulebased logic tree is used to segment urban built-up lands
from non-urban built-up lands. Figure 5 illustrates the signatures of the three new bands in the two index-derived images. A distinct feature is that the mean values of band 2 (NDBI-band) of both new images are all greater than those of