Analysis I: RGB Images


For this study, two types of processing were performed. The first was the extraction of imaging spectrograph data at three different wavelengths, which was used to produce a red-green-blue (RGB) composite image. The second analysis procedure involved using images from the spectrographs at several wavelengths in an attempt to perorm a supervised spectral classification in order to unambiguously identify features within a scene. Before both of these could be attempted, some preprocessing must be done to extract images of a specific wavelength from the spectral imager data.

Preprocessing

Since the imaging spectrometers use video-rate CCD detectors, the raw spectra are stored initially as either digital video files or on a casette tape. The BigSpex data, on digital video tape, were digitized via video capture software.


SWF file Figure 4. An example of video signal from the BigSpex imaging spectrograph (right panel). The sequence of images from the webcam (1 second time resolution) is shown on the left. (N.B. Image quality reduced for portability)

A demonstration of what data acquisition would look like appears in Figure 4. Assuming the webcam and spectrograph are co aligned, the video signal shown in the right-hand panel corresponds to a narrow slice through the center of the webcam image. Note that the vertical (i.e. colinear with the slit) field of view for the spectrograph is less than the webcam. Spatial structures colinear with the slit appear as vertical structure, while any relative spectral variation will show up as horizontal structure, which we expect to be fairly smoothly varying.


Figure 5. Keogram at 557.0-nm of the a section of the first 2000m pass during flight #1.

To produce an image at one wavelength from the video data, a vertical slice is extracted for the entire duration of the video, producing what is known as a keogram. Figure 5 is a keogram for the green wavelength 557.0-nm. Images of any wavelength imaged by the detector are produced by extracting the appropriate column of pixels from the video data. These images were created using the YaPlaySpex application written by Fred Sigernes.

Red-Green-Blue (RGB) Composite Images

Once images of specific wavelengths have been created from the spectrograph video data, it is easy to select several images to create a red-green-blue (RGB) composite image. This is most simply done using an application called Image Calculator written by K. Heia at Fiskeriforskning division of Norut.


Figure 6. Image Calculator used to create a RGB composite image.

Using this application, one selects images to represent the red, green and blue colors in the image. Figure 7 below shows individual images at 480.0-, 557.0- and 630.0-nm and the resulting RGB composite.


Figure 7. RGB composite image formed from three images of wavelengths 480.0-, 557.0- and 630.0-nm

Close inspection of the RGB composite does show a slight separation in the colors. For example, where there is a discrete feature common to all three images (e.g. a black shadow on snow or the edge of a building), the individual colors can be separated by a small amount, as in this blown up portion of Figure 7. Here a blue-green edge on the left side of this group of buildings is observed; on the right side the shadows have a reddish edge. The reason this occurs is owing to the ability of the YaPlaySpex software to extract the correct pixel column from every frame. Simply put, when the operating system is using a large amount of resources, the YaPlaySpex software occasionally drops individual frames, resulting in images which may not be the correct horizontal length. A possible solution to this problem is to configure the software to extract pixel columns of user-defined wavelengths simultaneously from each frame. That way, if frames are dropped, it affects all the images, rendering the differences between them to be much smaller.


Figure 8. Animation (SWF) showing the variation of magnification along the slit direction with wavelength.

Finally, it is also important to note that the magnification in the direction parallel to the slit is not uniform over the entire field of view of the detector. That is, the vertical magnification decreases with increasing wavelength, as illustrated in Figure 8. Therefore, as the spectral images are presented here, there is not a 1:1 correspondence between a pixel p(x,y,&lambda 1) and another pixel in a different spectral image p(x,y,&lambda 2). Although no attempt to quantify or correct for this aberration was attempted, it should be possible to crudely correct for this by performing a vertical scaling operation, the amount of which increases with increasing wavelength. Note that the resulting images would need to be cropped to the original size, as the Image Calculator software needs spectral images of identical size to perform a classification.


(Finished 10.05.05 -JMH)