From:
Imaging Spectroscopy:
Earth and Planetary Remote Sensing with the
USGS Tetracorder and Expert Systems

Tables 0-8 present the details of the rules for each material. As discussed above, the spectral identifications are partitioned into spectral regions (called groups) where spectral features are compared to make an identification. For example, clays, carbonates, and other materials containing light elements have vibrational overtones and combination features in the 1.5 to 2.5-m spectral region (with most diagnostic features in the 2-2.5 region). Fe²⁺ and Fe³⁺-bearing minerals have broad electronic absorptions in the ultraviolet to 1.5-m region. Thus these two regions are separated for analysis, but the boundaries can be crossed whenever diagnostic information occurs in another region, as in the case of jarosite or alunite (Figure 4). The following is a general overview of what is in each table and why.

A.1 Table Description

Each table consists of 11 columns.

Column 1) The first column is a description of the material (e.g. H₂O ice). Also in this column is sometimes a keyword used for Not features. For example, the first entry in Table 0 includes the keyword "GREENVEG." In other entries (e.g. entry 3: H₂O-Ice) this would include the keyword NOTGREENVEG to indicate that a feature matching GREENVEG should not be present (exactly which feature is defined in column 8).

Column 2) Second column is the spectral group.

Column 3) Third column is the material number, which runs sequentially through all the tables (250 reference materials/spectra through Table 7).

Column 4) The fourth column is the spectral feature number for that sample.

Column 5) The fifth column is the feature type: Dia = diagnostic, Opt = optional, and Not = a Not feature.

Column 6) The left continuum interval.

Column 7) The right continuum interval. The feature is defined to exist between the left and right continuum intervals. It does not matter that the feature center is not defined because the analysis system matches total feature shape between and including the continuum, not simply its position.

Column 8) Spectral continuum and Not constraints, including the definition of Not features, continuum thresholds or limits, and continuum slopes. Examine material 1, feature 1 (first entry in Table 0): ct 0.05 means the continuum at the feature center (derived by the analysis algorithms) must be greater than 0.05 reflectance. If the reflectance is below this level, the spectrum is too low to robustly detect this material (Note: this level is spectrometer signal-to-noise dependent). The sequence rct/lct > 1.0 means the right continuum divided by the left continuum must be greater than one (a positive slope). Figure 10b shows that the typical vegetation spectrum has a large increase in reflectance from 0.5 to 0.7 micron. The threshold constraints can also include an upper limit (this can be useful in mid-infrared emittance spectra).

A Not feature, for example in the H₂O ice spectrum (material 3, feature 3 in Table 0 = "1 0.5a 0.5"), is defined as Not the GREENVEG material. (NOTGREENVEG=material 1, 1 = feature 1 of the green vegetation spectrum) and it can have a depth of no more than 0.5 (the a means absolute) and the correlation coefficient must be at least 0.5. The Not feature can also be relative in depth to another feature. For example, examine the montmorillonite entry, material 126, feature 2, Table 2 (the entry "2 0.12r 0.3"): the montmorillonite should not have a muscovite feature 2 (the 2.3-m feature) with a strength greater than 0.12 times the depth of the montmorillonite feature 1 (0.12r1 is 0.12 times the derived depth of feature 1) and the fit to that muscovite feature can be no greater than 0.3 (if it is higher than 0.3 that feature is detected). Relative depths allow tracking of feature strengths as abundance varies from spectrum to spectrum. These entries begin to show the complexity that we have found necessary to do the identifications and thus built into the rules. If any constraint is violated, and it is a diagnostic feature, that material can't be present. If a Not feature is found, the material of interest is ruled out. If the feature is optional, its presence is set to zero (the depth and fit values are set to zero). The feature correlation coefficients are weighted by the relative areas of the features in the reference spectra and a weighted correlation coefficient and weighted depth computed.

Column 9) The correlation coefficient (fit) thresholds. Any fit below the value given rejects the material bearing that feature. The feature fitting routine computes the weighted depth, weighted fit (least squares correlation coefficient), and the weighted fit times depth as output (weighting described in paper I for when multiple features are used). The thresholds apply to the weighted fits and depths, and weighted fits times depths derived in the previous columns. There are several ways to constrain the thresholds:

FITALL means the fit threshold applies to all output, including the band depths. If a fit falls below the threshold, all output for that material is set to zero.

DEPTHALL means apply the band depth threshold to all output, including the fits. DEPTH or FIT thresholds applies only to the one output parameter.

The DEFAULTs are described at the end of each table. DEPTH-FIT means apply the threshold on the fit to the depth output.

FDALL applies the weighted fit times depth thresholds to all output, and

FD-DEPTH applies the depth threshold to the fit*depth output. These constraints prevent the algorithm from choosing a match in the presence of high noise relative to the feature strengths.

Global parameters are defined in Table 8.

Column 10) Describes what action might be taken. For example, Table 3 describes searching for the presence of vegetation. If it is found, the algorithm next executes special cases (1, 2, 3, 4, and 5) to further analyze the vegetation spectra.

Column 11) The identification of the spectrum (file and record number where the spectrum can be found). The splib04 spectra are from the Clark et al. (1993a) spectral library; others will be included in the next library, currently being assembled.

Appendix A, Table 0

Group 0 Reference Materials