Hyperspectral Analysis Example of Urban Data

The figures here show an example analysis of an airborne hyperspectral data flightline over the Washington DC Mall. Hyperspectral sensors gather data in a large number of spectral bands (a few 10's to several hundred). In this case there were 210 bands in the 0.4 to 2.4 µm region of the visible and infrared spectrum. This data set contains 1208 scan lines with 307 pixels in each scan line. It totals approximately 150 Megabytes. With data that complex, one might expect a rather complex analysis process, however, it has been possible to find quite simple and inexpensive means to do so. The steps used and the time needed on a personal computer for this analysis are listed in the following table and are briefly describe below.

Operation	CPU Time (sec.)	Analyst Time
Display Image	18
Define Classes		< 20 min
Feature Extraction	12
Reformat	67
Initial Classification	34
Inspect and Add 2 Training Fields		5 minutes
Final Classification	33
Total	164 sec = 2.7 min.	25 minutes

Define Classes. A software application program called MultiSpec, available to anyone at no cost from http://dynamo.ecn.purdue.edu/~biehl/MultiSpec/, is used. The first step is to present to the analyst a view of the data set in image form so that training samples, examples of each class desired in the final thematic map, can be marked. A simulated color infrared photograph form is convenient for this purpose; to do so, bands 60, 27, and 17 are used in MultiSpec for the red, green, and blue colors, respectively. The result is shown at left above.

Feature Extraction. After designating an initial set of training areas, a feature extraction algorithm is applied to determine a feature subspace that is optimal for discriminating between the specific classes defined. The algorithm used is called Discriminate Analysis Feature Extraction (DAFE). The result is a linear combination of the original 210 bands to form 210 new bands that automatically occur in descending order of their value for producing an effective discrimination. From the MultiSpec output, it is seen that the first nine of these new features will be adequate for successfully discriminating between the classes.

Reformatting. The new features defined above are used to create a 9 band data set consisting of the first nine of the new features, thus reducing the dimensionality of the data set from 210 to 9.

Initial Classification. Having defined the classes and the features, next an initial classification is carried out. An algorithm in MultiSpec called ECHO (Extraction and Classification of Homogeneous Objects) is used. This algorithm is a maximum likelihood classifier that first segments the scene into spectrally homogeneous objects. It then classifies the objects.

Finalize Training. An inspection of the initial classification result indicates that some improvement in the set of classes is called for. To do so, two additional training fields were selected and added to the training set.

Final Classification. The data were again classified using the new training set. The result is shown at right above. Note that the procedure used here does not require complex preprocessing such as correction for atmospheric effects, absolute calibration, transformation from radiance to reflectance, etc.

Because of its dimensionality, hyperspectral data potentially provides the capability to discriminate between nearly any set of classes. Signal processing research has shown that, of all the variables to the data analysis process, the most important one is the size and quality of the classifier training set. There are a number of steps in addition to those used above that could be taken to further polish the result at right above, but this current result appears to be satisfactory for many practical circumstances.

Research of this nature is conducted in the Purdue Electrical and Computer Engineering School's Multispectral Image Processing Laboratory (MIP Lab). For further information about hyperspectral data analysis, one may consult the documentation page of the above URL or contact Professor David Landgrebe at landgreb@ecn.purdue.edu.