Spectral Analyzer Contour Plot Compression

Multispectral target identification can be a powerful tool for agricultural, geological, and other types of terrestrial object classification. The technique measures the reflectance or radiance spectrum of a target object, and then uses that spectrum to locate objects with similar spectra. The method can be especially effective when the spectrum of the target object can be measured in the same satellite scene where other examples need to be discovered. The PANCROMA™ Point Spectrum Generator™ and Spectral Analyzer™ tools work together to accomplish this. The method has been described in detail in previous white papers in this series. It generally consists of clicking on the known target (or targets) to generate the spectrum (Point Spectrum Generator™) and then matching the spectrum of the known target with other potential examples with similar spectra.

Spectral matches are often accomplished using Euclidean Distance or Spectral Angle plots. These have also been explained in detail in previous articles so I will not repeat them here, except to mention that they consider multispectral reflectances as vectors and use distance or angular metrics to compare image with target reflectances. One problem that you might encounter when using the technique is compression of the color contour display plot caused by outlying "flyer" distance or angle values. Because the color contour plot scale is defined by the maximum and minimum distances/angles in the scene, even one outlier can cause the legitimate values to be compressed to the low end of the color contour plot scale. This can result in sub optimal discrimination among potential matches.

In order to rectify this, the distribution of distances or angles can be clipped. This can have a pronounced beneficial effect on the plot and the ability to identify potential targets. This will be illustrated using a RapidEye® sample data scene.

RapidEye® scenes consist of five multispectral bands: visible blue, green and red plus two near infrared (NIR) bands. A complete description of the data can be obtained from the RapidEye website. The five multispectral bands are usually "bundled" into a single GeoTiff data file containing the five 16-bit bands. PANCROMA™ can unbundle the bands into individual band files. However for spectral analysis, the application can read the bundled GeoTiff file directly.

To use the utility, open your RapidEye® file by selecting 'File' | 'Open'. Then select 'Spectral Analysis' | 'RapidEye Point Spectrum Generator' | 'Five 16-Bit (Bundled) GeoTiff File'. The Point Spectrum Data Entry box will be presented as described in the previous sections. Select the single band or composite bands that you wish to use to display for selecting your target points. You must also input the Solar Elevation Angle and Acquisition Date from the RapidEye® metadata. The reflectances will be computed and the spectrum displayed. As is usual, the reflecances will be saved and automatically inserted into the Spectral Criteria form for subsequent use by the PANCROMA™ Spectral Analyzer utility. A section of sample image 2011-04-02T031508_RE5_3A-NAC_6683405_113276.tif is shown below.

RapidEye® Image;.

The first step will be to click on our target. In this case we will use the sandy beach area in the lee of the harbor jetty. The Point Spectrum plot of band reflectances that I generated using eight sample points is shown below. (The lack of scatter among the points is generally indicative of good match).

Target Point Spectrum.

Next, we will locate all areas that have similar reflectance spectra to our target. Select 'Close Graphics Window and Reset'. Then, select 'File' | 'Open and open the single GeoTiff file that contains the five multispectral bands once again. Now select 'Spectral Analysis' | 'RapidEye Spectral Analyzer' | 'Five (Bundled) 16-Bit GeoTiff Bands' | 'Euclidean Distance'. When the Spectral Criteria form becomes visible, note that the target reflectance spectrum values have been automatically entered into the target spectrum text boxes by the Point Spectrum Generator™. Check the 'Save settings between runs' check box so that the populated reflectances will be retained and not lost when we run. Accept these values by clicking 'OK'. (Note that you will have to enter the Solar Elevation Angle and acquisition date from the RapidEye metadata for both the Point Spectrum Generator and the Spectral Analyzer into the appropriate data input box when prompted). The resulting Euclidean Distance contour plot is shown below.

Euclidean Distance contour plot.

Note that most of the plotted values are red or orange. There are a few yellow and green pixels in the image. However, although the beach target has been correctly identified, the plot has been skewed by a few very large distances elsewhere in the image. To rectify this, we will make another plot and clip the flyers. Without exiting PANCROMA™, select 'Close Graphics Window and Reset' as before. Then, select 'File' | 'Open and open the single GeoTiff file that contains the five multispectral bands once again.

Now select 'Spectral Analysis' | 'RapidEye Spectral Analyzer' | 'Five (Bundled) 16-Bit GeoTiff Bands' | 'Euclidean Distance'. This time, when the Spectral Criteria Form becomes visible, check the 'Set Upper Cutoff' check box at the lower right hand corner in the Cutoff Group Box as shown below. Use the slider to select the clipping level. In this case, all Euclidean Distance values that exceed 0.95 will be set equal to 0.95. (There is an optimal clipping value that will yield the best definition. (Setting too narrow of a clipping value will force another type of contour compression. Look at the color contour scale at the upper left of the image to decide the proper clipping value.)

Spectral Criteria form.

The image below shows the result. It is easy to see that the dynamic range has increased considerably. Many false positives have been eliminated. It is now possible to discern many targets that were washed out in the unclipped image. In addition to the beaches, the roofs of many houses show similar reflectance spectra. This may be because sand/gravel is a component of many roofing systems.

Clipped Euclidean Distance Contour Plot.

Another example from a Landsat 7 scene is shown below. In this case we want to detect subtle differences in the surface reflectance of the ocean surface in order to quantify sediment levels. The first image shows the RGB color contour plot.

The next image shows the Spectral Analyzer™ run, targeting the apparent highest sediment loaded areas. As you can see, there is little discrimination in the ocean areas. In this case, the compression is not caused by outliers, but instead by the narrow range of reflectances present in the ocean areas as compared to the entire scene. In order to force the sensitivity of the analysis, we will clip the majority of the reflectance range. This will cause the land reflectances to be ignored in the analysis, and will spread the ocean reflectances across the contour plot.

The third image shows the same plot with the Upper Cutoff set at 0.08. The variations in surface reflectance resulting from varying degrees of sediment loading are much more apparent.

This white paper has demonstrated how simple clipping of flyer Euclidean Distance values can significantly improve the color contour plot, thereby improving target identification. The process can work equally well for Spectral Angle or Distance Change analysis.