Extraction and Recognition of Urban Objects
by Hyperspectral Remote Sensing
3. Methodology
First, we should analyze the Spectral reflectance of the classes. The relative spectra reflectance derived by IARR were plotted in figure 1 (Note that the spectrum of white roof is much higher than that of other classes, therefore its curve exceeds the selected limit of y-axis). From Fig. 1 we found that different classes have different channels that allow particular class to be separated from other classes. For examples, vegetation has unique spectral response in red and near infrared channels that can be described by the well-known index NDVI, and though water body and coal have similar spectral response, they exhibit different changing trend at channel 9 and 10. These differences in channels and changing fashion require different feature set and different classifiers to reach optimal classification.
Figure 1: Sketch map of classification at a particular node
With these considerations in mind, we proffer a strategy of classification, hierarchical and masking method, to classify the classes step by step. See Fig. 2 for the sketch of classification at a particular node. At each node of the decision tree, relative comparison was made to separate only one general class into two or several subgroups, each subgroup represents one definite class or one general class that consisted of several definite classes. A mask was made for each of the subgroup to screen it out from other subgroups. If the subgroup is a general class, it was processed at a node in the succeeding layer. One of the advantages of the strategy is that, at each node, features can be selected according to the classes that should be separated. Usually, the features can be selected the most appropriate features which allow the relative comparison. The other advantage of this strategy is that at each node, we can process the data confined by the mask by some image processing methods to emphasizes the spectral difference, and use different classifiers. This makes the advantages of the decision tree more clear and powerful. Actually if we use the same feature set and same classifier, the classifier result are the same as the single layered classification method. After all the classes have been separated, we can composite them into a unique classification map.
Figure 2: Classification flow chart
Using this method in our study site, we can extract the classes defined in Table3 step by step. Figure 3 gives the classification flow chart. At node N1, vegetation was separated first according to NDVI. NDVI was computed using channel 8 and 13. Because this feature is very strong in urban area, it induce very little error. At node N2, the remaining part can be separated into 5 subgroups according to the spectral difference from channel 5 to 8. Class 2 and 5, White roof and Grey tile roof can be separated from node N2. A parallelepiped classifier can be used (Richards, 1994). The water body in the scene are heavily polluted and appears as dark as coal. Their channel reflectance are very similar except the minor different trends at channel 9 and 10. Simple band operation such as subtract 9 from 10 can make them separable at node N3. Red and grayish red tile roofs have prominent increasing tendency from channel 5 to channel 9, and have different reflectance value between this region, so they can be separated from class 8 to 15 at node N5 and separated with each other at node N6. Subgroups that need to be further classified at node N4, N5 contains the classes with very similar Spectral response, unsupervised classifier cane be used to classify them.
Figure 3 Relative reflectance of classes
