Note: since this article was first published, Spectral AnalyzerTM added the capability for using ASTER, SPOT, Landsat Surface Reflectance and EO-1 ALI multispectral data.
Spectral analysis using Landsat multispectral data is one of the most powerful tools for exploring and characterizing many aspects of the earth's surface. Applications of spectral analysis include mineral exploration, vegetation analysis, land use, forest health and many other applications. PANCROMA offers a powerful tool for conducting such analyses called the Landsat Spectral AnalyzerTM (LSA). False color band combinations and vegetation indices are well-known methods for using multispectral data to extract information from the earth's surface. The PANCROMA LSA uses all of the Landsat spectral bands for extracting such information. The LSA is used in conjunction with databases such as the USGS Digital Spectral Library. A Digital Spectrum (DS) is a plot of reflectance versus wavelength. Reflectance values typically vary between zero (complete absorption) and unity (complete reflectance). Instruments used to record such spectra are calibrated against special card standards to make sure that this is so. Digital spectra are recorded for minerals, flora, and various man-made substances as well as snow and ice. The individual spectra contained in the databases constitute an electromagnetic "fingerprint". Matching remotely sensed data to a DS can give indications of the presence of the target minerals or substances. The DS for the mineral Muscovite is shown below.
The PANCROMA LSA is a semi-automated utility for searching Landsat multispectral data for spectral fingerprints. The user sets the search criteria for each multispectral channel. The LSA then interrogates the six Landsat multispectral channels to determine which if any pixels match the fingerprint data. It then displays each "hit" as a red pixel on a gray band image. An example will make the use of this utility clearer. This example will look for something not very exotic - Lawn Grass. We will use a subset of a Row 23 Path 39 Landsat scene of Louisiana near the town of Abbeyville. Lake Peigneur is at the lower right. I subsetted the scene for faster processing, although you should have no problem processing full Landsat scenes with computers with more than 2GB of RAM. The band 123 (BGR) image is shown below.
I then pulled a spectrum for Lawn Grass from the USGS library. The spectrum is shown below.
The next step is to overlay the Landsat channels over the spectrum. I did this in PowerPoint® although you can use any convenient graphics editor.
I then struck horizontal lines that marked the intersection of the Landsat bands with the spectral curve. My strategy was to require that the reflectance be "at least" as large as the peaks and "no more than" for the valleys. The next step will show how the LSA allows such discrimination.
A note of caution: there are five "Lawn Grass" spectra listed and six "Grass" spectra, each different from the others to a greater or lesser degree. I chose one arbitrarily to see what I would get. In reality, library spectra may be useful but it is best not to place blind faith in them, as we shall soon see. The best spectra are the ones that come from a reliable source directly from your target area. You should at least read the Documentation that accompanies each USGS spectrum.
Now we are ready to scan our data. Select 'File' | 'Open and open six Landsat band files, band1 (Blue), band2 (Green), band3 (Red), band4 (NIR) band 5 (SWIR) and band 7 (LWIR). Now select 'Spectral Analysis' | 'Landsat Spectral Analyzer' | 'Six File Method'. The Spectral Criteria Form will be displayed as shown below.
There are six track bar sliders corresponding to the six input Landsat bands. Beneath each one is a radio button that lets you select between 'Must Exceed' and 'Less Than' as level qualifiers. Each multispectral band block also has an 'Include' check box. Un-checking this box excludes the band from the analysis. At the bottom of the screen is a check box that lets you lock all the sliders and another check box that allows you to save your settings between runs. Note that I have input the selection criteria from my channel plot.
When you click 'OK', a second data entry screen appears. These are the usual reflectance computation criteria. Since the spectra are given as reflectances, the band digital numbers (DNs) must be converted. This will be done in the same way as described in Section 49 Landsat Top of Atmosphere Reflectance. Make sure that the 'Render Color Image' check box is checked (this is the default). When you click "OK', the LSA will convert the DNs to TOA reflectances. It will report the maximum and minimum reflectance values to the Main Window data screen. Note that some reflectance values may be less than zero. This of course does not make sense. PANCROMA uses the method proscribed in the Landsat Data Users Handbook directly without biasing the computed values to make them all positive or zero. Since the magnitude of any negative reflectances is usually very small, this is not a problem.
PANCROMA will mark each pixel that matches the comparison criteria by coloring it red, and displaying it on the band1 grayscale image. The results of my first run were exactly... nothing. No hits at all. I inspected the maximum reflectance value for the band4 channel and noticed that it was around 0.8. I concluded that there might be a standard error between the measured reflectances and the spectrum. I made a second run decreasing all channel values by 0.2 units and again got no hits. I then decreased all channels by another 0.1 unit. This yielded the image shown below. (The image is shown at half scale so that it fits on the page conveniently.)
Comparison of the marked plots with the RGB image shown that the scan has apparently marked the plowed fields, not the lawn grass. I "ground truthed" the image using a high resolution aerial photograph and this seemed to confirm my suspicion. I also included a band 145 false color composite, commonly used to discriminate vegetation for comparison.
Although the run was not successful in discriminating Lawn Grass, a few conclusions can be drawn:
The method did seem to be a fairly sensitive discriminator for the plowed fields
More experimentation with the selection criteria might find the grassy fields as intended
The technique appears to be more sensitive and more positive than either false color composites or classification
It is clear why more multispectral bands are better, and why hyperspectral data is the best for this purpose. It is also clear why 11 bit or 16 bit data is better than 8 bit data, as the former yields more resolving power for applications like this
The PANCROMA LSA is a powerful tool for exploring and characterizing geological and agricultural aspects of the earth's surface with many possible useful applications. However like most tools, it still relies on the expertise and skill of the user to produce the best results. This example showed how it is possible to discriminate plowed fields in an image. Although this particular task was relatively simple, the method is equally applicable to cases where the discrimination is not so obvious, as in mineral exploration.