Abstract
Although Hyperspectral technique was successful in natural environment and resources, it was used less in urban-related studies. This study assesses the feasibility of Hyperspectral technique in urban remote sensing. Data was acquired by the Pushbroom Hyperspectral technique Imager (PHI) at Shahe county, north of Beijing, China in June 1997. Because of the spectral complexity of urban and man-made objects, we proffer a method of classification, hierarchical and mask technique to classify the different urban objects. The method was based on relative comparison of spectral features between class pairs. The study shows that Hyperspectral data can represent the minor differences of urban objects, and can be used for the step by step classification of vegetation, water body, coal area, road and buildings with different classification of vegetation, water body, coal area, road and buildings with different roof materials in urban area.
1. Introduction
In the VIS-NIR region, urban objects have sihgnificatn spectral fingerprints, this enables spectral recognition and extraction of different terrestrial and classes (Ben-Dor 1999, Liu 1999). Due to the influence of human beings, the spectral characters are very complex, therefore, it is difficult to classify the land cover types by traditional classification techniques. The disadvantages of traditional on e layer classification techniques lies in that, only one set of possible features can be selected, and pixels of each class need to be compared with all other classes before it can be identified as a definite class. Moreover, the features used by the classification methods are selected according to the mean distance between all class pairs, this do not guarantee the optimal discrimination and recognition of each class.
In pattern recognition, classification by comparison between twp classes is the simplest problem. It provides optimal classification environment for the class pair. Land cover types have intrinsic hierarchical structure which were used by the decision tree classifiers. Primary problems considered by decision tree classifiers are to be define the classes which need to be discriminated by each node, how to control class overlaps and how to decide the number of tree layers (Safacia, 1990). This method can be used in multi-spectral and Hyperspectral image classification, where the rank of classes are defined by their spectral behavior.
Jia (1998) proffered a progressive binary decision tree classifier. The advantage of this method is that the tree structure is simple which makes the software development more general. It considers only one class pair at each node in a layer. This enables the appropriate features to be selected according to the particular pair as well as decrease the dimension of the data. This is important for Hyperspectral data analysis. The defect is that only at the last layer can all the classes be discriminated; non is classified at the middle layers. For M class problem it needs M-1 layers and M(M-1) nodes. The tree structure may grow very large if the class number is big.
In reality, some classes, such as vegetation, has spectral character that is very different from others, so it is possible to separate them by only one node in a layer. Some classes have similar character that can be treated as a general class, and can be discriminated by a node from others. Therefore, we conceive a strategy, hierarchical and masking method, to classify the different spectral classes of urban man-made objects. Compared with the method of Jia, the number of nodes and layers decrease obviously while the advantages of decision tree classifier can be preserved.
2. Data Representation
In accordance with the importance of urban survey, a studying project has been conducted in June, 1997 in Shahe, north of Beijing, China. A Chinese developed Pushbroom Hyperspectral Imager (PHI) has been used for the Hyperspectral data acquisition which was fixed on a 2-D stabilized platform in Y-5 airplane. Flight altitude was 1000 meter and GPS has been used for navigation and positioning. PHI is a CCD imaging spectrometer, which has 244 spectral bands from visible to near infrared spectral region (400-850 nm). The basic technical data of PHI are shown in table 1. in this flight , 15 channels have been selected for the data acquisition (table 2). Its pixel size is about 1.5m*1.5m. Overall 4 flight lines have been executed to cover this area.
Table 1: Technical Specification of the PHI
| Spectral Range (m) |
0.40~0.85 |
| Number of Bands |
244 |
| Spectral Sampling |
1.8nm |
| Spectral Resolution |
<5nm |
| Field of View |
210 |
| Spatial Sampling |
376 pixel/line |
| Spatial Resolution |
1.5mrad |
| Digitization |
12bit |
| Data Rate |
7.2 M sample/Sec |
| Frame Rate |
60Fr/Sec (Maximum) |
Table 2 Selected Channels Drom PHI in Beihai Flight
| Band no. |
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
| Wavelength (nm) |
456 |
501.5 |
531.8 |
577.2 |
607.5 |
637.9 |
668.2 |
683.3 |
| Band no. |
9 |
10 |
11 |
12 |
13 |
14 |
15 |
|
| Wavelength (nm) |
698.5 |
713.6 |
728.8 |
744 |
759.1 |
804.6 |
850 |
|
One of the four lines, which cover the most typical urban objects, was selected. The terrestrial cover types in this area were decided through image interpretation and ground survey. Land cover types in this area are consisted of vegetation, water body, roads and roofs with different materials. The overall 17 classes are listed in Table 3. in order to get rid of the influence of atmosphere and the sun spectral radiance effect, data were preprocessed by a generally used normalization method, internal average relative reflectance (IARR) calibration to convert it into relative reflectance.
Table 3. The classes found from PH1 in Beihai Flight
| 1 Vegetation |
2 While roof |
| 3 Water body |
4 Coat |
| 5 Grey tile roof |
6 Concrete roof |
| 7 Asbestine tile roof |
8 Asphalt road |
| 9 High way |
10 Oil paper roof |
| 11 Railway |
12 Unknow roof1 |
| 13 Unknow roof2 |
14 Cyan tile roof |
| 15 Metal jar |
16 Red tile roof |
| 17 Greylish red tile roof |
|
