Computers and Geosciences 115 (2018) 56–65

Contents lists available at ScienceDirect

Computers and Geosciences

journal homepage: www.elsevier.com/locate/cageo

Research paper

Quantitative X-ray Map Analyser (Q-XRMA): A new GIS-based statistical

approach to Mineral Image Analysis
Gaetano Ortolano a, Roberto Visalli a, *, Gaston Godard b, Rosolino Cirrincione a
a
Department of Biological, Geological and Environmental Sciences, University of Catania, Corso Italia, 98125, Italy
b
Universit
e Paris-Diderot, Institut de Physique du Globe de Paris, UMR CNRS 7154, F-75238, Paris cedex 05, France

A R T I C L E I N F O A B S T R A C T

Keywords: We present a new ArcGIS®-based tool developed in the Python programming language for calibrating EDS/WDS
Numerical petrology X-ray element maps, with the aim of acquiring quantitative information of petrological interest. The calibration
Linear regression procedure is based on a multiple linear regression technique that takes into account interdependence among
Mineral formula quantification elements and is constrained by the stoichiometry of minerals. The procedure requires an appropriate number of
ArcGIS® spot analyses for use as internal standards and provides several test indexes for a rapid check of calibration ac-
Python
curacy. The code is based on an earlier image-processing tool designed primarily for classifying minerals in X-ray
element maps; the original Python code has now been enhanced to yield calibrated maps of mineral end-members
or the chemical parameters of each classified mineral. The semi-automated procedure can be used to extract a
dataset that is automatically stored within queryable tables. As a case study, the software was applied to an
amphibolite-facies garnet-bearing micaschist. The calibrated images obtained for both anhydrous (i.e., garnet and
plagioclase) and hydrous (i.e., biotite) phases show a good fit with corresponding electron microprobe analyses.
This new GIS-based tool package can thus find useful application in petrology and materials science research.
Moreover, the huge quantity of data extracted opens new opportunities for the development of a thin-section
microchemical database that, using a GIS platform, can be linked with other major global geoscience databases.

1. Introduction
Advanced geological and petrographic investigations are increasingly
supported by a collection of quantitative information. In this context,
quantitative X-ray mapping is emerging as an innovative tool that can be
applied in many fields, from basic scientific research to mining industry,
cultural heritage and forensic geology, as demonstrated by the large in-
vestments made by manufacturers of microanalytical devices. However,
in contrast to single spot analyses, X-ray maps are characterised by semi-
quantitative raw data that is not corrected, for instance, for the mean
atomic number, absorption or fluorescence (ZAF) effects (Lanari et al.,
2018 and references therein). In this context, to derive fully quantitative
data on element concentrations, good analytical standardization pro-
cedures are crucial in overcoming the intrinsic limits of X-ray mapping
(Lanari et al., 2018 and references therein), such the previously
mentioned ZAF effects, background noise and possible peak overlap in
energy-dispersive-spectrometer (EDS) X-ray maps, or the volatilisation of
light elements due to prolonged exposure under a finely focused electron
beam.
Taking into careful consideration the problems above, a new per-
forming multiple linear regression technique was applied for the first
time to calibrate the unprocessed wavelength-dispersive-spectrometer
(WDS) or EDS X-ray maps using high-precision spot analyses as inter-
nal standards (De Andrade et al., 2006). The developed procedure, which
largely modifies the pre-existing ArcGIS®-based image processing plat-
form by Ortolano et al. (2014b) (i.e., X-ray Map Analyser), aims to cali-
brate sets of X-ray maps for each classified phase.
In general, image processing in materials science is mainly based on
the multivariate statistical analysis of raw X-ray intensities in EDS/WDS
matrices or of images with multiple raster bands (typically 8-bit images
with 256 intensity levels) recording the distribution of chemical ele-
ments. This is one of the techniques applied routinely in the investigation
of many geo-petrological processes (e.g., Airaghi et al., 2017; Belfiore
et al., 2016; Coutelas et al., 2004; Lanari et al., 2013; Loury et al., 2016;
Marmo et al., 2002; M
eszaros et al., 2016; Ortolano et al., 2014a; Vidal
et al., 2006). The technique has been adopted also thanks to the devel-
opment of semi-automated tools for classifying mineral phases and
quantifying modal parameters from selected thin section micro-domains

* Corresponding author.
E-mail address: rvisalli@unict.it (R. Visalli).

https://doi.org/10.1016/j.cageo.2018.03.001
Received 23 June 2017; Received in revised form 23 February 2018; Accepted 2 March 2018
Available online 6 March 2018
0098-3004/© 2018 Elsevier Ltd. All rights reserved.
G. Ortolano et al.
(e.g., Friel and Lyman, 2006; Gu, 2003; Lanari and Engi, 2017; Launeau
et al., 1994; Ortolano et al., 2014b; Tarquini and Favalli, 2010; Tinkham
and Ghent, 2005).
Others tools use a very efficient clustering algorithm (i.e., K-means) to
obtain several homogeneous groups of pixels automatically bounded by a
mask (e.g., Cossio and Borghi, 1998; Cossio et al., 2002; Lanari et al.,
2014). In particular, software such as Petromod (Cossio et al., 2002) and
XMapTools (Lanari et al., 2014) treat each of these groups of pixels as an
array of X-ray intensity values representing a single mineral phase, where
the element pixel array is a function of chemical concentration. Such
software use linear regression functions to obtain new sets of calibrated
X-ray images, converting qualitative raster images into calibrated grid
formats (i.e., an array of equally-sized, square grid points storing con-
centrations as absolute numeric values) (Cossio and Borghi, 1998; Cossio
et al., 2002; De Andrade et al., 2006; Lanari et al., 2014; Togami et al.,
2000).
In this context, calibration techniques are obviously a crucial step in
the quantitative analysis of raster imagery. Careful control of the pre-
diction model is required to obtain physically significant results.
The new procedure differs substantially from the previous one (i.e., X-
ray Map Analyser: Ortolano et al., 2014b), which has been completely
revised and now includes: (a) an additional analytical cycle based on a
multiple linear regression that is used to produce X-ray grid images of
each identified mineral phase; (b) a third analytical cycle that uses
calibrated maps to extract the zoning patterns of complex solid-solution
mineral phases. The third cycle is useful in visualising the progressive
changes in end-member concentrations, which are generally related to
multivariant chemical reactions in complex systems.
2. Algorithms
Ortolano et al. (2014b) used the Python programming language to
develop a semi-automated image processing procedure that was imple-
mented in ArcGIS®. This procedure, based on the multivariate statistical
Fig. 1. Simplified flow charts of the GIS-based image-processing procedure.
The procedure is subdivided into three analytical cycles.
57
Computers and Geosciences 115 (2018) 56–65
analysis of X-ray maps, was designed for classifying chemically homo-
geneous zones and extracting quantitative information such as mineral
modes (e.g., Belfiore et al., 2016; Ortolano et al., 2014a). The Python
code of this procedure was implemented in a first analytical cycle (Fig. 1)
used to classify all recognizable phases within the entire selected domain
and in a second cycle for a more in-depth analysis of the previously
identified phases and the detection of any sub-phases such as those
related to mineral zoning. Both cycles use principal component analysis
together with a maximum likelihood classification algorithm to generate
graphical outputs. Although this image processing procedure yields a
significant amount of information, it lacks an algorithm able to transform
the derived raster X-ray image of identified phases into a standard grid
format.
A new Python code was therefore implemented to cyclically calibrate
the sets of X-ray maps for each classified phase using the ArcGIS®
licensed software available on demand at the Geoinformatics and Image
Analysis Lab of the Biological, Geological and Environmental Sciences
Department at the University of Catania.
The second cycle (Fig. 1) has been completely modified. Based on
microprobe spot analyses, it introduces a multiple linear regression
technique to predict element concentration values for each pixel
(expressed as atoms per formula unit, a.p.f.u) within a specific phase.
This is a crucial step for implementing a third analytical cycle (Fig. 1)
used to derive maps of end-member proportions and to highlight
chemical variations within the investigated phases.
The following sections describe the main steps of the new analytical
cycles, focusing on the adopted multiple linear regression technique.
Further specific insights are also provided in the user guide available in
the supplementary material within the software (i.e., Q-XRMA).
2.1. Second cycle: calibration algorithm
We chose a multiple linear regression algorithm (Hosmer and
Lemeshow, 1989; Marquardt, 1980; Neter et al., 1996) because it can
G. Ortolano et al.
take into account natural compositional constraints relating to the stoi-
chiometry of minerals. This algorithm also helps overcome issues relating
to possible correlations (e.g., linear, exponential, logarithmic) between
element concentrations and corresponding pixel intensities. The
element-concentration vs. pixel-value relationship can be affected, for
instance, by: (a) the initial selection of device parameter settings, or (b)
the use of data from different acquisition instruments. The multiple linear
regression can be used to derive, for each pixel, the absolute concentra-
tion of an element (i.e., the dependent variable - DV) in a specific mineral
phase, taking into account the weight of all of the other elements (i.e., the
explanatory variables - EVs) in order to preserve stoichiometry.
One of the main difficulties in adopting a multiple linear regression
approach is finding the correct explanatory variables to explain any
possible interdependent variations in the parameters of interest. To
address this point, the pixel intensities of major elements in common
minerals (i.e., Al, Ca, Fe, K, Mg, Mn, Na, Si, Ti) were selected as
explanatory variables.
The calibration procedure is based on the application of the following
interdependent multiple linear regression equation:
h     mance and explain the variation in the dependent variable. For
255 255 255
ECi F ¼ α1 EGVi F0 þ α2 EGVj F0 þ α3 EGVk F0 þ…
 i
255
þ αn EGVn F0 þ βi þ γ i ; (1)
where:
 EF is a discretized positive value;
 i represents the element of unknown concentration;
 j; k; l…:n represent the other elements;
 ECi F ¼ the absolute and discretized concentrations of element i
calculated for each pixel; this variable is dependent of the explanatory
variables;
 α1; 2; 3; …n ¼ the calculated coefficients of the regression between
element concentrations measured through spot analysis and the cor-
responding pixel intensity values in the original X-ray raster images;
255 255 255
 EGVi F0 ; EGVj F0 … EGVn F0 ¼ the discrete pixel intensity values of
the i, j, k, …, nth, single band element maps; these are independent
explanatory variables;
 βi ¼ value of the intercept for element i, which is the expected value
of the dependent variable (ECi F) if all the independent (explanatory)
variables (GV) are zero;
 γ i ¼ random error for element i, defined as the difference between
measured and predicted values.
The adopted calibration procedure also yields several statistical pa-
rameters used to assess model performance and evaluate the reliability of
predicted results. They can be summarised as follows:
a) Regression coefficients (αn): the coefficient calculated for each
explanatory variable (EGVn F) indicates the type of correlation (i.e.,
negative, positive or no correlation) between explanatory and
dependent variables. For instance, to predict the Ca content in
plagioclase (ECCa F), one would expect a negative correlation (i.e.,
αNa < 0) with the Na pixel values (EGVNa F). If the calibration pro-
cedure yields an unexpected positive correlation for the above-
mentioned relationship, one must proceed with great caution.
b) Probability: this parameter is used to check whether the regression
coefficient of each EV is near or equal to zero. The smaller the
probability value, the lower the likelihood that the coefficient is equal
to zero, and the higher the correlation between dependent and
explanatory variables. This low probability implies that those specific
coefficients are relevant for the regression model. In contrast, EVs
with coefficients showing high probability values should be removed
from the regression equation. In our procedure, when the probability
is lower than 0.01, the value on the regression report is marked with
58
Computers and Geosciences 115 (2018) 56–65
an asterisk as a reminder that the coefficient of that specific EV is
consistent with the prediction model. Some explanatory variables
could be considered good predictors also from a theoretical point of
view (e.g., for preserving stoichiometry), even though they yield
regression coefficients with high probability values. This assumption
is statistically valid, given the consistency among other test indexes
(i.e., Joint F-Statistic and Joint Wald Statistic), which highlight the
mutual interdependence relationship between the regression
coefficients.
c) Variance inflation factor (VIF): the VIF is a control parameter that
depicts the weight of redundancy among EVs. It is an index that
measures how much the multicollinearity (i.e., a situation in which
two or more explanatory variables in a multiple regression model
show linear correlation) affects the variance values. If the explanatory
variables have VIF values larger than about 7.5, the variance increases
due to collinearity and the model is not accurate. As a rule of thumb,
EVs should be removed from the regression model.
d) R-Squared (R2) and Corrected Akaike's Information Criterion (AICc): the
R2 values, ranging from 0.0 to 1.0, are measures of model perfor-
instance, an R2 value of 0.8 means that the EVs explain 80 percent of
the variation in the dependent variable. The AICc is used to compare
different models for the same dependent variable (e.g., increasing or
decreasing the number of spot analysis). In comparing different
calibration models, the model with the lowest AICc is to be preferred.
e) Joint F-Statistic and Joint Wald Statistic: these test indexes are measures
of the overall statistical significance of the model, since they assess
regression coefficient relationships simultaneously. As a result, these
tests measure the statistical meaning of the multiple linear regression
models, even when one or more of the single coefficients are statis-
tically meaningless (i.e., with a probability value > 0.01). In partic-
ular, these tests compare a model with no explanatory variables (i.e.,
an intercept-only model) to the user-specified model (i.e., a multiple
linear regression model). When these tests yield values lower than
0.01 (i.e., marked with an asterisk on the regression report), the user-
specified model provides a better fit than the intercept-only model
and, consequently, the set of explanatory variables improves the fit.
When such a result is achieved, the applied multiple regression
models can be considered reliable if the value associated with the
Koenker (BP) statistic test (see below) is greater than 0.01.
f) The Koenker (BP) Statistic test: this is used to determine whether the
EVs in the model have a consistent correlation with the dependent
variable (DV) across the entire study area. If test values are lower than
0.01, the EVs vs. DV correlations are non-stationary or hetero-
scedastic (i.e., characterised by an anomalous value distribution).
Heteroscedasticity provides information about the error between data
sets, such as the observed and predicted values for a specific DV, and
can be considered a measurement of deviation with respect to the
average. When an asterisk marks the Koenker (BP) test, there is a
heteroscedastic distribution. As the Q-XRMA software writes test and
regression results in specific folders and tables, the heteroscedastic/
homoscedastic condition can be easily verified by constructing scatter
plots between estimated/predicted values vs. residuals (i.e., the
observed dependent variable values minus the predicted ones).
g) The Jarque-Bera statistic test: this is used to assess whether the re-
siduals show a normal Gaussian distribution. When this test yields a
value lower than 0.01, the residuals are not normally distributed,
indicating a likely bias that can influence the reliability of predictions.
This bias could be due, for instance, to the presence of outliers in the
spot analyses. The latter can affect the modelled regression relation-
ship by producing a non-best fit and a biased regression coefficient. A
possible solution is to remove the outliers and start the procedure
again.
During the calibration procedure, the results of each of these tests are
displayed on the screen, and values lower than 0.01 are marked with an
G. Ortolano et al. Computers and Geosciences 115 (2018) 56–65

asterisk. This warning allows the user to quickly check the reliability of  If XpLAB  0
the model. Model results are saved to specific folders as output tables.
ðXpLAB  Xmc LAB Þ  l 
XpGIS ¼ * Xm GIS  Xmr GIS  þ Xmc GIS (2a)
jðXml LAB  Xmr LAB Þj
2.2. Second cycle: map calibration (data setting and outputs)

In order to preserve the interdependence relationship between

explanatory variables (i.e., the pixel values of selected major elements),  If XpLAB < 0
the calibration procedure considers one phase at a time, previously ðXmc LAB  XpLAB Þ  l 
classified during the first cycle of analysis (see above). To run the cali- XpGIS ¼ * Xm GIS  Xmr GIS  þ Xmc GIS (2b)
jðXml LAB  Xmr LAB Þj
bration procedure, all selected EVs are required as input (i.e., Al, Ca, Fe,
K, Mg, Mn, Na, Si, Ti X-ray maps), and two spreadsheets must be
compiled as well (supplementary material within the software – Q-
XRMA). This operation must be repeated whenever a new mineralogical  If YpLAB  0
phase is investigated.
The first spreadsheet is used to ensure the correct location of internal ðYpLAB  Ymc LAB Þ  t 
YpGIS ¼ * Ym GIS  Ymb GIS  þ Ymc GIS (2c)
standards within the map (i.e., Device vs. ArcGIS® coordinate conversion jðYmt LAB  Ymb LAB Þj
tool). It converts the original device coordinates for each spot into the
ArcGIS® coordinate system, using the following equations:

Fig. 2. Editing step of the ArcGIS® point shapefile location.

The points close to grain boundaries or fractures should be removed.

Fig. 3. Regression report of the calibration procedure displayed by the software.

Red-delimited areas show the results of the statistical tests executed for iron calibration within garnet, used to evaluate the accuracy of the calibration model (see text
for details). (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

59
G. Ortolano et al.
 If YpLAB < 0
ðYmc LAB  YpLAB Þ  t 
YpGIS ¼ * Ym GIS  Ymb GIS  þ Ymc GIS (2d)
jðYmt LAB  Ymb LAB Þj
where XpGIS is the calculated ArcGIS® X coordinate of the spots, XpLAB is
the equivalent analytical device X coordinate, XmLAB and XmGIS are,
respectively, the device X coordinate of the X-ray map and its equivalent
on the ArcGIS® reference system, whereas apices l; c and r stand for,
respectively, the bottom left corner, the centre and the bottom right
corner of the map along the X axis (i.e., Eq. (2a, b)). The same applies to
the Y coordinate (i.e., Eq. (2c, d)) where b; c and t stand for respectively
the bottom left corner, the centre and the top left corner of the X-ray map
along the Y axis.
The coordinate conversion spreadsheet can be used in two different
cases: (a) when X-ray maps and the analysis of internal standards have
been completed using the same device; (b) when X-ray maps and the
analysis of internal standard have been completed using different de-
vices. In the first case, the device coordinates of the X-ray maps and their
equivalents in the ArcGIS® coordinate system are required as input in-
formation, along with the device coordinates of each single internal
standard. The ArcGIS® X–Y coordinates can be easily obtained using the
extent information in the layer properties menu of the map loaded into
Fig. 4. Window shell of the third analytical cycle.
Green-delimited area displays the mineral phases already identified and classified during the first analytical cycle. These phases are used as masks to derive calibrated
maps. Such maps can display either mineral end-members (blue-delimited area; first computational mode) or arithmetic combinations of elements (red-delimited area;
second computational mode). (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)
60
Computers and Geosciences 115 (2018) 56–65
the ArcGIS® project. In the second case, one must provide: (a) the device
X–Y coordinates of at least two spot analyses and their equivalent values
in the ArcGIS® coordinate system; (b) the sizes of the X-ray maps; (c) the
ArcGIS® coordinates of the X-ray maps; and (d) the device coordinates of
each single spot analysis. After entering all the input values, the ArcGIS®
X–Y coordinates that are automatically calculated for each single spot
analysis can be subsequently inserted into the second spreadsheet.
The second spreadsheet (PtAnls.xlsx) contains information regarding
the sample location, mineral phases, sub-phases of the mineral investi-
gated, number of analyses, chemical compositions and the ArcGIS® X–Y
coordinates. The software uses the coordinates to automatically create
the point shapefile required to complete the calibration steps. To avoid
unexpected calibration results, the points projected by the software must
not be located close to other mineral phases or fractures; such points
should be removed from the point shapefile (Fig. 2).
Once the input files have been set correctly, it is possible to run the X-
ray map calibration procedure. During this procedure, the software ap-
plies statistical tests to assess the statistical meaningfulness of each
calibration model. Fig. 3 show an example of iron calibration within
garnet; the red areas highlight statistical test results (e.g., VIF, Joint F and
Wald-Statistics, Koenker and Jarque-Bera statistics) which suggest that
the applied calibration model can be considered reliable for the inves-
tigated dependent variable (CFe).
G. Ortolano et al. Computers and Geosciences 115 (2018) 56–65

Once the calibration of the selected elements is complete, new cali-
brated maps are produced. In turn, these may be used as input files for the
third analytical cycle (Fig. 1) used to derive mineral end-member maps.
2.3. Third cycle: map of mineral end-member fractions
The main advantage of the calibration procedure is that it can
combine the calibrated elemental maps to produce maps of mineral end-
members fractions (and other compositional variables such as XMg). To
this end, the software can be used to complete a third analytical cycle in
two different computational modes. The first mode (blue area in Fig. 4)
yields different end-member maps for a set of common minerals, using a
specific phase identified earlier as a mask file (green area in Fig. 4). The
second mode (red square in Fig. 4) enables the manual selection of ele-
ments to be input into the raster calculator, which formulates and exe-
cutes map algebra expressions to create specific end-member maps of
mineral phases detected during the first cycle (green area in Fig. 4). In the
latter case, the user must define the name of the output map, the number
of elements to be used and the correct expression for deriving graphical
outputs.
In both computational modes, the graphical outputs and the related
normalised end-member concentration for each pixel are stored within
queryable database format tables.
3. Case study
As a case study, we investigated a garnet-plagioclase-biotite-bearing
microdomain in a metapelite sample (see Fig. 5). In order to test the
applied methodology accurately, we used input data from two different
laboratories, taking care to correctly locate spot analyses within the pixel
matrices of the corresponding EDS X-ray maps.
Spot analyses used as internal standards to calibrate images were
performed using a Cameca SX-FIVE electron microprobe (EMP) equipped
Environmental Sciences. Operating conditions were: 20-kV accelerating
potential and 8-nA beam current; a dwell time of 500 ms per 128 frames,
corresponding to an average acquisition time of 22 h and 40 min; a res-
olution of 1200  900 pixels, corresponding to a square pixel with a side
of about 2 μm.
As a preliminary operation at the beginning of the first cycle (Fig. 1), a
focal median filter was applied to the native X-ray raster images with the
aim of reducing the background noise common to EDS images (see
Ortolano et al., 2014b, for details).
Garnet, plagioclase and biotite were selected as target minerals due to
their grain size and chemical composition distributions, which was
considered particularly appropriate for highlighting the results of our
procedure.
3.1. Garnet
With the aim of highlighting the maximum range of compositional
variability in the selected zoned garnet porphyroblast, 95 spot analyses
(Table S2a in supplementary material) were acquired along a core to rim
transect and used as internal standards for the calibration procedure. The
spreadsheet “Device vs. ArcGIS®” (see description in x2.2) was used to
convert the coordinates.
The sum of cations occupying sites X, Y and Z (Table 1), calculated for
each pixel, is in agreement with the structural formula of garnet (i.e.,
X3Y2[ZO4]3) and is reported in Supplementary Table S3a. This finding
was further confirmed by the statistically meaningful results for each
considered element, which were consistent with the test indexes
Table 1
Garnet structural formula on the basis of 12 oxygen with relative standard
deviation.
Garnet core Garnet mantle Garnet Rim
(18 881) (72 428) (164 859)

with five WDS spectrometers at the Camparis laboratory in Paris. Oper- Element Average St.Dev. Average St.Dev. Average St.Dev.
ating conditions were: ~1-μm beam spot, 15-kV accelerating voltage, and Si 2.98 0.006 2.97 0.006 2.97 0.006
10-nA beam current. Natural minerals and synthetic oxides were used as Ti 0.01 0.000 0.00 0.000 0.00 0.001
standards. Adopted mineral abbreviations are from Siivola and Schmid Al 1.97 0.008 1.97 0.007 1.98 0.006
(2007), whereas mineral formulae were calculated using MINPET 2.02 Fe 1.84 0.128 2.28 0.105 2.49 0.075
Mg 0.08 0.043 0.22 0.044 0.31 0.000
software (Richard, 1995). All mineral analyses and structural formulae Mn 0.39 0.054 0.23 0.058 0.12 0.052
are reported in the supplementary material (Table S1). EDS elemental Ca 0.71 0.126 0.33 0.102 0.14 0.072
maps, provided in the Q-XRMA software supplementary material, were Na 0.00 0.005 0.00 0.003 0.00 0.002
acquired using a Tescan Vega-LMU scanning electron microscope K 0.00 0.000 0.00 0.000 0.00 0.040
equipped with an EDAX Neptune XM4-60 micro-analyser characterised XAlm 0.60 0.048 0.74 0.034 0.81 0.024
by an ultra-thin Be window coupled with an EDAX wavelength-dispersive XGrs 0.24 0.046 0.11 0.033 0.05 0.023
low-energy X-ray spectrometer calibrated for light elements, located at XPrp 0.02 0.015 0.07 0.014 0.10 0.013
XSps 0.13 0.019 0.07 0.019 0.04 0.017
the University of Catania, Department of Biological, Geological and

Fig. 5. Classification of the mineral phases.

(a)mineralogical distribution map resulting from the classification; (b) histogram of the mineral modal percentages.

61
G. Ortolano et al. Computers and Geosciences 115 (2018) 56–65

calculated automatically by the procedure (Table S4a in supplementary
material). The results of the calibration process for divalent cations (i.e.,
Ca, Fe, Mg and Mn maps) are shown in Fig. 6a, where it is possible to
observe the concentric zonation in garnet.
The third analytical cycle was then executed to derive end-member
maps of almandine, grossular, pyrope and spessartine (Fig. 6b), which
were used to extract chemical information from the calibrated outputs
(Fig. 6c).
3.2. Plagioclase
Following the same procedure used for garnet, elemental maps of
plagioclase were calibrated using 20 spot analyses (Table S2b in sup-
plementary material) of different grains to highlight and quantify
compositional variations possible linked to different growth stages. The
resulting stoichiometry (Table 2) is consistent with the structural formula
of plagioclase (i.e., X[Z4O8]: Table S3b), as further supported by the
reliable test index results for the calibration procedure (Table S4b in
supplementary material). The calibrated map of calcium (Fig. 7a) high-
lights a higher Ca content in grains enclosed within garnet (i.e.,
maximum value of 0.44 a.p.f.u), whereas the calibrated map of sodium
(Fig. 7a) shows a higher Na content in the matrix grains (i.e., maximum
value of 0.92 a.p.f.u.), thereby allowing the identification of different
plagioclase generations. The third analytical cycle was therefore used to
derive albite and anorthite component maps (Fig. 7b). These two outputs
were used to classify different generations of plagioclase, as shown on the
ternary feldspar classification diagram in Fig. 7c. The large number of
calibrated pixels improved the compositional characterisation of the
mineral phase by increasing the amount of available compositional data
(i.e., 170 854 calibrated pixels: Fig. 7c), thereby providing chemical in-
formation also in areas not directly covered by EMP spot analyses.

3.3. Biotite

Lastly, to test the calibration procedure on a hydrous mineralogical

phase, 22 randomly distributed spot analyses from several biotite crystals
(Table S2c in supplementary material) were used to set up the calibration
models correctly. As in the case of the anhydrous phases, the weight of
each cation obtained for each pixel through the map calibration pro-
cedure (Table 3) appears to be consistent with the structural formula of
biotite (i.e., WY3[Z4O10](OH)2: Table S3c in supplementary material).
Low Joint test index values (see x2.1) together with the satisfactory re-
sults of other statistical tests such as the variance inflation factor (VIF),
the Koenker (BP) and the Jarque-Bera tests, confirmed the accuracy of
the calibration models used for the single investigated elements (see
Table S4c in supplementary material). The calibrated maps of divalent
cations (i.e., Fe and Mg) occupying the Y site in biotite are shown in
Fig. 8a, where it is possible to observe a higher iron content when in
contact with garnet. This increase in iron within biotite grains close to
garnet was probably caused by diffusive exchange of iron and magnesium
between the two mineralogical phases, which led to an increase in
magnesium within garnet. Using the raster calculator (i.e., second
computational mode of the third cycle; see description in x2.3) biotite
XMg values (Fig. 8b) were subsequently determined to highlight possible

Table 2
Plagioclase structural formula on the basis of 8 oxygen with relative standard
deviation.
Plagioclase 1 (24 825) Plagioclase 2 (131 329)

Element Average St.Dev. Average St.Dev.

Si 2.70 0.074 2.75 0.066

Ti 0.00 0.000 0.00 0.000
Al 1.28 0.063 1.27 0.056
Fe 0.00 0.000 0.00 0.000
Mg 0.00 0.000 0.00 0.000
Mn 0.00 0.000 0.00 0.000
Ca 0.30 0.039 0.25 0.030
Na 0.62 0.063 0.66 0.060
Fig. 6. Outputs of the second and third analytical cycles for garnet. K 0.01 0.006 0.00 0.005
(a) Ca, Fe, Mg and Mn calibrated images expressed in a.p.f.u.; (b) almandine,
grossular, pyrope and spessartine component maps expressed as mole percent- XAb 0.68 0.036 0.72 0.034
XAn 0.32 0.036 0.28 0.034
age; (c) extraction of end-member fractions from the calibrated maps along
XOr 0.00 0.000 0.00 0.000
a segment.

62
G. Ortolano et al. Computers and Geosciences 115 (2018) 56–65

Fig. 7. Outputs of the second and third analytical cycles for plagioclase.
(a) Ca and Na calibrated images expressed in a.p.f.u.; (b) albite and anorthite
component maps expressed as mole percentage; (c) pixels composition of
plagioclase obtained from the calibrated maps reported in the feldspar
ternary diagram.

chemical variations linked to different metamorphic evolutionary stages.

To this end, the compositions obtained from the pixels calibrated for AlIV
and XFe were plotted on the biotite classification diagram (Fig. 8c). The
significantly larger chemical dataset obtained through the calibration
procedure (i.e., 89 017 calibrated pixels), confirms the rather heteroge-
neous composition of biotite (i.e., AlIV: 2.42–2.63; XFe: 0.56–0.63) within
the biotite classification plot.

4. Conclusions

Q-XRMA is a greatly enhanced version of the GIS-based image-pro-

cessing procedure by Ortolano et al. (2014b), the Python code structure
of which has been modified. The new procedure is divided into three
sequential cycles that allow: (a) the mineral classification of the inves-
tigated microdomain; (b) the identification of sub-phases within previ-
ously classified mineral phases; (c) calculation of element concentrations
by calibrating the native qualitative X-ray maps using internal standards;
(d) the distribution of end-members proportions within solid-solution
mineral phases to be shown.
The main aim of this new procedure is the calibration of elemental
maps, expressed in a.p.f.u, computed for previously-classified mineral
phases, so as to obtain several types of mineral end-member maps. EDS or
WDS X-ray maps are calibrated using internal standards analysed with
the same or another device.
Before starting the calibration procedure, a series of control stages
must be completed. First, it is best to acquire the maps before collecting
the spot analysis data array in order to minimise light element volatili-
sation; in the case of EDS X-ray maps, the possible occurrence of over-
lapping element energy peaks should be ascertained. After this, a
significant number of WDS spot analyses (e.g., about 20) is selected for
each single detected mineral phases or sub-phases, taking care to have a
stoichiometrically consistent dataset and carefully associating each spot
with the corresponding map pixel. The latter is completed using a
spreadsheet that converts the device coordinates into ArcGIS®
coordinates.
Data input reliability, a prerequisite for good calibration, was here
assessed through a multiple linear regression technique. This technique
also allows a posteriori statistical assessment of results and is used to
check the accuracy of the statistical models used to calculate a specific
element.
This technique mirrors the natural interdependence of element

Table 3
Biotite structural formula on the basis of 11 oxygen with relative standard
deviation.
Biotite 1 (53 195) Biotite 2 (35 821)

Element Average St.Dev. Average St.Dev.

Si 2.76 0.029 2.73 0.021

Ti 0.14 0.027 0.13 0.025
Al 1.60 0.021 1.59 0.021
Fe 1.48 0.028 1.47 0.027
Mg 0.88 0.044 0.98 0.039
Mn 0.00 0.001 0.00 0.001
Ca 0.00 0.002 0.00 0.002
Na 0.01 0.012 0.01 0.012
K 0.89 0.025 0.88 0.027

(caption on next column) XMg 0.37 0.012 0.40 0.009

63
G. Ortolano et al.
Fig. 8. Outputs of the second and third analytical cycles for biotite.
(a) Fe and Mg calibrated images expressed in a.p.f.u.; (b) XMg component map
expressed as mole fraction; (c) pixels composition of biotite obtained from the
calibrated maps reported within the biotite-classification diagram.
concentrations within a specific mineral and can be used to extract
quantitative chemical information in a few minutes without requiring
additional laboratory analyses.
An amphibolite-facies garnet micaschist characterised by bell-shaped
zoned garnet associated with heterogeneous biotite and plagioclase
served as a case study for the methodology. The stoichiometry of cali-
brated pixels was checked and deemed satisfactory, considering that the
procedure is easy to implement, relatively fast and low-cost.
Several upgrades of the code introduced the following improvements:
(a) better division of outputs results into specific folders according to the
cycle completed, thereby facilitating the classification of files; (b) all
64
Computers and Geosciences 115 (2018) 56–65
output files are saved in database tables that can be queried; (c)
compatibility with ArcGIS® 9.x and later versions; (d) a new round of
analyses can begin without having to exit the programme.
Lastly, this new GIS-based software can be used to perform a semi-
automated sequential image routine applicable in many areas of petro-
logical research. This new procedure minimises bias resulting from user
selection of the spot location, and yields quantitative distribution maps
that can be used to unravel complex or hidden paragenetic equilibria of
petrological interest.
Acknowledgments
The authors are grateful to Michel Fialin and Nicolas Rividi for the
valuable support received during the microanalytical data acquisition at
the Camparis laboratory (Paris). Financial support from FIR 2014 (Geo-
statistics-based automated geochemical mapping of plutonic complexes:
experimentation on the Serre Batholith, central Calabria) – University of
Catania (2_1_04_01_04), is further gratefully acknowledged. Pierre Lanari
and anonymous reviewer meticulous reviews improved the quality of the
original manuscript and gave us many more occasions of discussion. We
are grateful to Arabella Fiona Palladino for the revision of the English.
We thank Gregoire Mariethoz for the editorial handling of the manuscript
and thoughtful suggestions.
Appendix A. Supplementary data
Q-XRMA is free to download, including the sources of the Python
code, which are also available in the supplementary material of the
present work. All modifications and/or enhancing of the code are
welcomed but have to be freely shared with the scientific community.
The software will be soon downloadable also from the forthcoming
dedicated web-site.
Supplementary data related to this article can be found at https://doi.
org/10.1016/j.cageo.2018.03.001.
References
Airaghi, L., Lanari, P., de Sigoyer, J., Guillot, S., 2017. Microstructural vs compositional
preservation and pseudomorphic replacement of muscovite in deformed metapelites
from the Longmen Shan (Sichuan, China). Lithos 282–283, 262–280. https://doi.org/
10.1016/j.lithos.2017.03.013.
Belfiore, C.M., Fichera, G.V., Ortolano, G., Pezzino, A., Visalli, R., Zappala, L., 2016.
Image processing of the pozzolanic reactions in Roman mortars via X-ray Map
Analyser. Microchem. J. 125, 242–253. https://doi.org/10.1016/j.microc.2015.11.
022.
Cossio, R., Borghi, A., 1998. PETROMAP: MS-DOS software package for quantitative
processing of X-ray maps of zoned minerals. Comput. Geosci. 24, 797–803. https://
doi.org/10.1016/S0098-3004(98)00057-0.
Cossio, R., Borghi, A., Ruffini, R., 2002. Quantitative modal determination of geological
samples based on X-ray multielemental map acquisition. Microsc. Microanal. 8,
139–149. https://doi.org/10.1017/S1431927601020062.
Coutelas, A., Godard, G., Blanc, P., Person, A., 2004. Les mortiers hydrauliques: synthese
bibliographique et premiers r
esultats sur des mortiers de Gaulle romaine. Revue
d’Arch
eom
etrie 28, 127–139. https://doi.org/10.3406/arsci.2004.1068.
De Andrade, V., Vidal, O., Lewin, E., O'Brien, P., Agard, P., 2006. Quantification of
electron microprobe compositional maps of rock thin sections: an optimized method
and examples. J. Metamorph. Geol. 24, 655–668. https://doi.org/10.1111/j.1525-
1314.2006.00660.x.
Friel, J.J., Lyman, C.E., 2006. Tutorial Review: X-ray mapping in electron-beam
instruments. Microsc. Microanal. 12, 2–25. https://doi.org/10.1017/
S1431927606060211.
Gu, Y., 2003. Automated scanning electron microscope based mineral liberation analysis:
an introduction to JKMRC/FEI mineral liberation analyser. J. Miner. Mater. Char.
Eng. 2, 33–41. https://doi.org/10.4236/jmmce.2003.21003.
Hosmer, D.W., Lemeshow, S., 1989. Applied Logistic Regression. Wiley, New York.
Lanari, P., Engi, M., 2017. Local bulk composition effects on mineral assemblages. Rev.
Mineral. Geochem. 83, 55–102. https://doi.org/10.2138/rmg.2017.83.3.
Lanari, P., Riel, N., Guillot, S., Vidal, O., Schwartz, S., P^echer, A., Hattori, K.H., 2013.
Deciphering high-pressure metamorphism in collisional context using microprobe
mapping methods: application to the Stak eclogitic massif (northwest Himalaya).
Geology 41 (2), 111–114. https://doi.org/10.1130/G33523.1.
Lanari, P., Vidal, O., De Andrade, V., Dubacq, B., Lewin, E., Grosch, E., Schwartz, S., 2014.
XMapTools: a MATLAB©-based program for electron microprobe X-ray image
processing and geothermobarometry. Comput. Geosci. 62, 227–240. https://doi.org/
10.1016/j.cageo.2013.08.010.
G. Ortolano et al.
Lanari, P., Vho, A., Bovay, T., Airaghi, L., Centrella, S., 2018. Quantitative compositional
mapping of mineral phases by electron probe micro-analyser. Geol. Soci. London
Special Publ. https://dx.doi.org/10.1144/SP478.4.
Launeau, P., Cruden, A.R., Bouchez, J.L., 1994. Mineral recognition in digital images of
rocks; a new approach using multichannel classification. Can. Mineral. 32, 919–933.
Loury, C., Rolland, Y., Cenki-Tok, B., Lanari, P., Guillot, S., 2016. Late Paleozoic evolution
of the South Tien Shan: insights from P–T estimates and allanite geochronology on
retrogressed eclogites (Chatkal range, Kyrgyzstan). J. Geodyn. 96, 62–80. https://dx.
doi.org/10.1016/j.jog.2015.06.005.
Marmo, B.A., Clarke, G.L., Powell, R., 2002. Fractionation of bulk rock composition due to
porphyroblast growth: effects on eclogite facies mineral equilibria, Pam Peninsula,
New Caledonia. J. Metamorph. Geol. 20 (1), 151–165. https://doi.org/10.1046/j.
0263-4929.2001.00346.x.
Marquardt, D.W., 1980. “You should standardize the predictor variables in your
regression models”, discussion of “a critique of some ridge regression methods,” by G.
Smith and F. Campbell. J. Am. Stat. Assoc. 75 (369), 87–91. https://doi.org/10.
1080/01621459.1980.10477430.
M
eszaros, M., Hofmann, B.A., Lanari, P., Korotev, R.L., Gnos, E., Greber, N.D.,
Greenwood, R.C., Jull, A.J.T., Al-Wagdani, K., Mahjoub, A., Al-Solami, A.A.,
Habinullah, A.N., 2016. Petrology and geochemistry of feldspathic impact-melt
breccia Abar al’ Uj 012, the first lunar meteorite from Saudi Arabia. Meteorit. Planet.
Sci. 51 (10), 1830–1848. https://doi.org/10.1111/maps.12693.
Neter, J., Kutner, M., Nachtsheim, C., Wasserman, W., 1996. Applied Linear Statistical
Models, fourth ed. (Irwin, Homewood, IL).
Ortolano, G., Visalli, R., Cirrincione, R., Rebay, G., 2014a. PT-path reconstruction via
unravelling of peculiar zoning pattern in atoll shaped garnets via image assisted
65
Computers and Geosciences 115 (2018) 56–65
analysis: an example from the Santa Lucia del Mela garnet micaschists (Northeastern
Sicily-Italy). Period. Mineral. 83 (2), 257–297. https://dx.doi.org/10.2451/
2014PM0015.
Ortolano, G., Zappala, L., Mazzoleni, P., 2014b. X-ray Map Analyzer: a new ArcGIS® based
tool for the quantitative statistical data handling of X-ray maps (Geo- and material-
science applications). Comput. Geosci. 72, 49–64. https://doi.org/10.1016/j.cageo.
2014.07.006.
Richard, L.R., 1995. MinPet: Mineralogical and Petrological Data Processing System,
Version 2.02. MinPet Geological Software, Qu
ebec, Canada.
Siivola, J., Schmid, R.A., 2007. Systematic Nomenclature for Metamorphic Rocks: List of
Mineral Abbreviations. Recommendations by the IUGS Subcommission on the
Systematics of Metamorphic Rocks. Web version 01.02.2007. https://www.bgs.ac.
uk/scmr/docs/papers/paper_12.pdf.
Tarquini, S., Favalli, M., 2010. A microscopic information system (MIS) for petrographic
analysis. Comput. Geosci. 36, 665–674. https://doi.org/10.1016/j.cageo.2009.09.
017.
Tinkham, D.K., Ghent, E.D., 2005. XRMapAnal: a program for analysis of quantitative X-
ray maps. Am. Mineral. 90, 737–744. https://doi.org/10.2138/am.2005.1483.
Togami, S., Takano, M., Kumazawa, M., 2000. An algorithm for the transformation of XRF
images into mineral-distribution maps. Can. Mineral. 38, 1283–1294.
Vidal, O., De Andrade, V., Lewin, E., Munoz, M., Parra, T., Pascarelli, S., 2006. P–T
deformation Fe3þ/Fe2þ mapping at the thin section scale and comparison with
XANES mapping: application to a garnet-bearing metapelite from the Sambagawa
metamorphic belt (Japan). J. Metamorph. Geol. 24, 669–683. http://doi.org/10.
1111/j.1525-1314.2006.00661.x.