This algorithm is an extension of the normal TSA algorithm (Berman et al 2011) working in the SWIR. It uses some scalar information derived from selected features in both the SWIR and the TIR to make a more informed choice from the list of candidate mixtures that are produced by the Subset Selection (Miller 2002) process that is used by TSA.

 Subset Selection produces lists of likely mineral assemblages that explain the observed SWIR spectrum reasonably well. In this context “explain” means that when the spectra of the minerals composing the assemblage are used in the appropriate mathematical model, the modelled spectrum fits the observed spectrum reasonably well. TSA uses the residual errors of these fits (called the SRSS) to select between candidate assemblages. However it is sometimes the case that some of the best-fitting assemblages have quite similar SRSS values and it is useful to add other information to resolve this non-uniqness.

 The example below illustrates the process. The plot at the top is a normal TSA output and the one below is from TSA+ on the same data. This hole was chosen because it was clear from the raw spectra that TSA was missing considerable amounts of chlorite at various places in the hole. The extra information that was added in this case was the presence or absence of the FeOH absorption feature at ~2250 nm. When FeOH bearing minerals do not dominate the assemblage, this feature can be quite weak (thus having little effect on the SRSS) but it is a very good indication of their presence. Here TSA+ examined the low-SRSS candidate assemblages and selected the one that was consistent with the presence or absence of the FeOH feature, even though it may not have been the assemblage with the lowest SRSS. The result is considerably improved.

In addition to the FeOH scalar, TSA+ uses scalars for prehnite (~1470 nm), amphibole (~2390 nm), epidote (~1550 nm) and sulphate (~1760 nm). In general, these scalars seem to make fewer changes to the TSA result than the FeOH scalar, perhaps because TSA is already performing reasonably well for these minerals.


This algorithm has been developed as a replacement for sTSAT, TSG’s system-level TSA for TIR data. sTSAT was the first unmixing algorithm developed for HyLogging data and it operates in a stand-alone, sample-by-sample mode without reference to results from the VNIR and SWIR regions. In contrast jCLST interprets TIR data using the results from a modified TSAT, TSA+ and from scalars operating on selected features the VNIR and the TIR. It uses this extra information to constrain the minerals that are allowed into the mixing model for the unknown sample. The set of minerals to be used in the model is called the Sample Restricted Mineral Set (SRMS).

The mixing model selects all the spectra from the TIR Reference Set that are associated with the SRMS and offers them to a Constrained Least Squares (CLS) procedure (Often called  Non-negative Least Squares or NNLS, Lawson and Hansen, 1995) This fits all spectra that are offered to it, subject to the constraint that they are included with non-negative weights. CLS is solved using a quadratic programming algorithm (Nocedal and Wright, 1999).

The plots below for a hole from Olympic Dam in SA illustrate a case where jCLST has had a major impact on the TIR interpretation. The SWIR data (here from sTSAS) shows two domains the first aspectral/w-mica and the second w-mica/chlorite. The centre plot (from sTSAT) shows serpentine and talc (here shown in dark blue as Other-MgOH) in place of the chlorite, clearly a major error. The lowest plot from jCLST shows a much more reasonable set of minerals where the SWIR results have constrained the TIR interpretation.



These automatic interpretations must be regarded as a starting point for further analysis because:

1.      Only common minerals are used

The suite of minerals offered for inclusion in the unmixing models is restricted. This because the software slows dramatically when it is offered many alternatives and because there are less errors of commission that include minerals that are unlikely to be present anyway. The minerals offered to the unmixing algorithms to make these results are shown opposite. So ONLY these minerals can be found at this stage.

2.      The modelling is not perfect.

Accurate results depend on automatically  finding the minerals that are present in the sample and then  estimating their proportions by inverting a linear model. Neither of these steps is foolproof and there are usually many ways a geologist with some knowledge of the area can improve the interpretation. 

3.    The data does not always define a unique  solution.

This means that some spectra can be equally well modelled by two (or sometimes more!) different mineral assemblages. See the example below.

Active Minerals for TSA+ and jCLST
AMPHIBOLE Actinolite, AmphiboleML48, Edenite, Hornblende, Riebeckite, Tremolite
PLAGIOCLASE Albite, Andesine, Anorthite, Bytownite, Labradorite, Oligoclase
GARNET Almandine, Andradite, Grossular, Spessartine
SULPHATE Alunite-K, Alunite-NH, Alunite-Na, Gypsum, Jarosite
CARBONATE Ankerite, Calcite, Dolomite, Magnesite, Siderite
PYROXENE Augite, Diopside, Enstatite, Hedenbergite
DARK-MICA Biotite, Phlogopite
CHLORITE Chlorite-Fe, Chlorite-FeMg, Chlorite-Mg
OTHER-ALOH Diaspore, Gibbsite, Prehnite, Pyrophyllite, Topaz
KAOLIN Dickite, Kaolinite-PX, Kaolinite-WX, Nacrite
EPIDOTE Epidote, Zoisite
K-FELDSPAR Microcline, Orthoclase
SMECTITE Montmorillonite, Nontronite, Saponite
WHITE-MICA Muscovite, Paragonite, Phengite
OLIVINE Olivine-Fe, Olivine-Mg
SILICA Opal, Quartz
TOURMALINE  Rubellite, Tourmaline, Tourmaline-Fe
OXIDE Hematite, Goethite, Rutile
SERPENTINE Antigorite, Lizardite

Understanding the Plots

The figure opposite has six columns showing various results from an automatic analysis of the HyLogger data for this hole. The detailed results have been averaged into equal sample intervals down the hole. Depths are in meters.

  1. On the left, labelled TIR (jCLST), are the results of unmixing the HyLogger TIR data. The results are shown in generalized mineral groups and the legend is below. 
  2. Next is a narrow column that, for this hole, is mostly dark blue. This displays a measure of the error in the jCLST modelling process that has been used to estimate the mineral proportions. The colour coding uses blue for low errors and runs through the rainbow to red for poor modelling results.
  3. The third column, labelled SWIR (TSA+), shows the results of unmixing the HyLogger SWIR data. The same generalized mineral groups used for the TIR data apply here.
  4. The next narrow column (here with a few lighter blues) displays the errors in the TSA+ modelling in a similar way to the TIR modelling errors (column 2)
  5. This plot, labelled FeOH, is the wavelength position (in nanometres) of the so-called FeOH absorption that is seen in chlorites, dark micas and epidotes. For chlorites, the exact position of this feature can give an indication of the chemical composition and the bounds (plotted in dotted red lines) indicate the rough Mg/(Mg+Fe) ratio if the feature was due to pure chlorite.
  6. The last plot, labelled AlOH, is the wavelength position (in nanometres) of the so-called AlOH absorption that is seen in white micas, kaolins and montmorillonites. For white micas the exact position of this feature can give an indication of the chemical composition and the bounds indicate the amount of octahedrally coordinated Al. Generally, as the Al content goes down, the ferro-magnesian content goes up.

Why Are the SWIR and TIR Results Different?

There are two reasons. The first, and most important, is that while most minerals respond in the TIR, not all minerals have useful reflectance features in the SWIR. In general, minerals with hydroxyl groups, carbonates, and sulphates have characteristic reflectance features in both the SWIR and the TIR but the three dimensional silicates (e.g. quartz, feldspars, garnets, pyroxenes, olivine) usually only respond in the TIR and are fairly transparent in the SWIR. This means that the plots will always be different when these 3-D silicates are present.

The second reason is that the two wavelength regions sample the rock in different ways. Except in some complicated situations, all the TIR radiation that is reflected comes off the surface, a process called Strong Surface Scattering. However, most of the SWIR radiation that is returned has been inside the sample where it has bounced around for a while before hitting some scatterer that reflects it back towards the sensor. This process is called Volume Scattering because the volume rather than the surface is sampled. Usually, it is the presence of the 3-D silicates (transparent in the SWIR) that enables the radiation to bounce around inside the sample. While it is doing this it can interact with even small amounts of SWIR-active minerals to create a measurable response.

The hole shown here illustrates these effects. The upper 225m is dominated by quartz with small amounts of SWIR-active white-mica, chlorite and carbonate. There is not enough of the SWIR-active minerals on the surface to be seen in the TIR but they are easily detected in the SWIR by virtue of Volume Scattering.

The samples where the HyLogger saw no measurable SWIR-active minerals are called Aspectral and are here presented as Invalid. The Aspectral samples are normalized with the samples that saw SWIR-active minerals to give 100% over the binned interval. From 20 m to 120 m roughly 50% of the samples were Aspectral.

Below that most samples were still overwhelmingly quartz with chlorite dominating the trace component until ~222 m where the rock changes dramatically and becomes dominated by chlorite with small amounts of quartz and feldspar. This sudden change only shows in the TIR and the chlorite is seen in the TIR because it is now on the surface in appreciable quantities and is no longer being sampled by Volume Scattering processes.

From 240 -285 m the SWIR-active minerals are in trace amounts again (lots of Aspectral) and the TIR is dominated by a mafic assemblage.


Non-uniqueness at medium to low abundance can be a problem. In the plots below the same data (plotted in black) is modelled (in red) as a mixture of quartz, albite and white mica or as a mixture of quartz, albite and K-spar. Below the data and the fits are plotted the spectra of the minerals used in the modelling scaled by the weight that has been used in the model. So, on the left hand side, adding the yellow quartz spectrum to the blue albite spectrum and the brown white mica spectrum gives the red modelled spectrum.