**3. Data processing**

Historically, the first Raman maps were constructed manually by making a map point by point. The sample was then not scanned continuously but was immobile beneath the laser beam during the acquisition of each spectrum (**Figure 4**). These first maps were interesting for the study of homogeneity in materials, but, because of the limited number of spectra associated with this time-consuming method, imaging was not very efficient. With the recent development in fine positioning systems and synchronisation between the positioning system and the CCD detector, automated scanning is now available. This automation allows high resolution

**Figure 4.** Principle of Raman mapping. (a) Laser spot size versus area of interest and different ways of scanning, (b) point by point, (c) continuously and horizontally and (d) continuously and vertically. The spectrum associated to each pixel corresponds to the average spectrum of the different associated phases. The concentration map is thus darker where the vermicular structure is analysed. If the spot size is larger than the pixel size (left row), there is oversampling, and the associated maps are the same whatever the method used. However, if the laser spot size is small in comparison with the pixel size (right row), there is undersampling and the associated maps depend on the type of the scan.

In all scanning methods of imaging (e.g. Raman mapping as well as atomic force microscopy, scanning electron microscopy, etc.), the lines are accumulated one by one and the spatial

mapping point by point as well as by continuous scanning.

168 Raman Spectroscopy and Applications

We have chosen a number of different rocks and minerals to illustrate methods of data processing for improving image quality and for obtaining specific information. Measurements were made mainly on polished thin sections as well as on polished rock surfaces.

#### **3.1. Classical maps**

Raman scans are mostly used to map the composition of a sample over a particular area. The interest of mapping with respect to single spectrum analysis is that it is possible to observe the association and spatial distribution of different phases (see **Figure 5**). These kinds of images are obtained by selecting a peak in the spectrum of each phase and by plotting its intensity in order to obtain an image of its concentration (**Figures 3b** and **5b**), or by associating a colour to a given peak in order to obtain an image in which each phase is represented by a colour (**Figures 3c** and **5d**). However, the quality of the images can be improved by more complex data processing. The use of the peak area instead of the peak intensity leads to slightly better contrast, in particular if the peak is wide (**Figure 5c**). A second step involves extracting the spectrum of each phase and using a correlation process with the global data set to dissociate phases when there is partial overlap of their peaks (**Figure 5d**–**f**). Calculations can be applied to the data set, such as derivatives to eliminate the background or to use the square or the cube of the data set to increase the signal/nose ratio. Finally, more advanced tools can be used, such as principal component analysis, to improve the detection of different phases [9].

**Figure 5.** Raman mapping of silicified microorganisms from the Gunflint formation and average spectrum of carbonaceous matter. (a) Optical microscopy image in transmitted light. Associated Raman images of quartz obtained (b) using the intensity of the main peak of quartz at 465 cm−1 and (c) the area of the same peak. (d) Raman compositional map with quartz in orange and carbonaceous matter in green corresponding to the superimposition of the Raman images obtained using the full spectrum correlation of (e) quartz and (f) carbonaceous matter. Boolean masks (0 in black and 1 in yellow) can be obtained by directly drawing on the image (g) or by using a mathematical threshold filter (h) in order to obtain a high signal/noise ratio average spectrum of the fossilised filament (i).

Mapping also permits by obtaining high quality spectra of a phase. Indeed, it is possible to create a mask (i.e. Boolean image) by directly drawing on an image (**Figure 5g**) or by using thresholds (**Figure 5f**) to select the areas of interest, to multiply it to the data set in order to select only the associated spectra, and finally to average these spectra to obtain the corresponding average spectrum (**Figure 5i**). Different masks can be applied to the same data set. For example, it is possible to use a mask corresponding to the area where the signal of the phase is the highest and a second one corresponding to the areas where the background level is the lowest. The average spectrum thus obtained can be more representative than that obtained in only one point. Of course, the map can also be used to locate the area where the signal is highest before acquiring a single spot spectrum.

### **3.2. Detection and identification of small phases**

are obtained by selecting a peak in the spectrum of each phase and by plotting its intensity in order to obtain an image of its concentration (**Figures 3b** and **5b**), or by associating a colour to a given peak in order to obtain an image in which each phase is represented by a colour (**Figures 3c** and **5d**). However, the quality of the images can be improved by more complex data processing. The use of the peak area instead of the peak intensity leads to slightly better contrast, in particular if the peak is wide (**Figure 5c**). A second step involves extracting the spectrum of each phase and using a correlation process with the global data set to dissociate phases when there is partial overlap of their peaks (**Figure 5d**–**f**). Calculations can be applied to the data set, such as derivatives to eliminate the background or to use the square or the cube of the data set to increase the signal/nose ratio. Finally, more advanced tools can be used, such

170 Raman Spectroscopy and Applications

as principal component analysis, to improve the detection of different phases [9].

**Figure 5.** Raman mapping of silicified microorganisms from the Gunflint formation and average spectrum of carbonaceous matter. (a) Optical microscopy image in transmitted light. Associated Raman images of quartz obtained (b) using the intensity of the main peak of quartz at 465 cm−1 and (c) the area of the same peak. (d) Raman compositional map with quartz in orange and carbonaceous matter in green corresponding to the superimposition of the Raman images obtained using the full spectrum correlation of (e) quartz and (f) carbonaceous matter. Boolean masks (0 in black and 1 in yellow) can be obtained by directly drawing on the image (g) or by using a mathematical threshold filter (h) in order

Mapping also permits by obtaining high quality spectra of a phase. Indeed, it is possible to create a mask (i.e. Boolean image) by directly drawing on an image (**Figure 5g**) or by using thresholds (**Figure 5f**) to select the areas of interest, to multiply it to the data set in order to select only the associated spectra, and finally to average these spectra to obtain the corresponding average spectrum (**Figure 5i**). Different masks can be applied to the same data set. For example, it is possible to use a mask corresponding to the area where the signal of the phase is the highest and a second one corresponding to the areas where the background level is the lowest. The average spectrum thus obtained can be more representative than that

to obtain a high signal/noise ratio average spectrum of the fossilised filament (i).

Although the spatial resolution of Raman spectroscopy is the same as that of optical microscopy, each pixel of an image is not just a colour but corresponds to a spectrum. Raman spectroscopy may thus permit identification of very small grains that are difficult to identify using classical optical microscopy (e.g. the pyrite grain shown by the black arrow in **Figure 6c** or the grain of SiC in **Figure 9d**). A mineral phase may also be invisible in optical microscopy, and Raman mapping is then the only way to detect it, as shown in **Figure 6**. Finally, at larger scales, mapping can be a good means to localise phases present in minor or trace amounts in a sample.

**Figure 6.** Optical and associated Raman maps of (a–c) a polished thin section of a hydrothermal vein from the quarry of Neuvial, Mazerier, Allier, France and (d–f) a thin section of the ~800 My-old Draken Formation, Svalbard (see [10] for more information). The Raman map (c) shows the presence of carbonaceous matter (green), orthoclase (blue) and a pyrite grain (pink and black arrow) in the quartz matrix (orange). The carbonaceous matter and the pyrite grain were invisible in the optical microscopy image (a), and the orthoclase was too small to be identified by polarised optical microscopy (b). The optical image (d) shows a fossil planktonic microorganism. The presence of opal (in yellow) and anatase (in red) is highlighted in the composition (e) and opal (f) Raman maps, whereas they are invisible in optical microscopy. The quartz matrix is in orange in (e). Image size 30 × 30 μm2 .

#### **3.3. Identification of spectrally close phases**

Not only is the spatial resolution better than mapping but also the spectral resolution. Indeed, due to the high number of spectra accumulated for each image, very small spectral changes can be statistically detected. **Figure 7** shows the position of the centre of mass of the main peak of a quartz spectrum during analysis of crushed samples of various grain sizes (see Ref. [1] for more information). Although the resolution of the Raman spectrometer is ~1 cm−1, the image highlights peak shifts lower than 0.5 cm−1. Such small variations would not be identifiable using single spectrum analysis, while they are statistically obvious in this image.

**Figure 7.** Raman analysis of quartz powders of four different grain sizes (decreasing grain size from left to right) showing the shift of the main peak at ~465 cm−1 towards the lower values with the decreasing grain size (see Ref. [1] for more information).

A similar approach can be helpful for mineralogy because it is often difficult to distinguish minerals having relatively similar Raman spectra, for example, carbonates [11]. However, by using peak position and the full width of the peaks at half maximum (FWHM) or the background level, it is possible to distinguish and identify different phases in the images. **Figure 8** shows different Raman images obtained from the same data set of carbonate deposits in a vesicle in basalt from Svalbard (see Refs. [12] and [13] for more information). The different carbonate phases can be distinguished because of the variation in the ratio Mg/Ca during deposition, resulting in phases ranging from dolomite to magnesite.

**Figure 8.** (a) Optical image and Raman maps of (b) the position, (c) the FWHM of the main peak position around 1000 cm−1, (d) and (e) of the second carbonate peak around 320 cm−1 and (f) the background level illustrating different phases in carbonate deposited in a vesicle formed in basalt from Svalbard (see Ref. [13] for more information). The corresponding areas can be selected (g) in order to obtain the associated average spectra (h).

#### **3.4. Exotic uses**

As noted above, the collected signal depends on different parameters that can be used to make images. More complex maps can thus be obtained by applying different calculations to the images; for instance, the ratio of peak intensity can distinguish changes in crystal orientation, as shown in **Figure 9**. The image of background intensity is also a particularly interesting example. We have seen above that it can be used to make masks, but it can also be used to show variations in composition or to distinguish relatively similar phases (**Figure 8**). The Raman signal can also be modified by defects in the sample, such as cracks or fluid inclusions, as shown in **Figure 9**. Finally, the signal is also associated with fluorescence and could thus give some information about the trace element composition of a mineral.

**Figure 7.** Raman analysis of quartz powders of four different grain sizes (decreasing grain size from left to right) showing the shift of the main peak at ~465 cm−1 towards the lower values with the decreasing grain size (see Ref. [1] for

A similar approach can be helpful for mineralogy because it is often difficult to distinguish minerals having relatively similar Raman spectra, for example, carbonates [11]. However, by using peak position and the full width of the peaks at half maximum (FWHM) or the background level, it is possible to distinguish and identify different phases in the images. **Figure 8** shows different Raman images obtained from the same data set of carbonate deposits in a vesicle in basalt from Svalbard (see Refs. [12] and [13] for more information). The different carbonate phases can be distinguished because of the variation in the ratio Mg/Ca during

**Figure 8.** (a) Optical image and Raman maps of (b) the position, (c) the FWHM of the main peak position around 1000 cm−1, (d) and (e) of the second carbonate peak around 320 cm−1 and (f) the background level illustrating different phases in carbonate deposited in a vesicle formed in basalt from Svalbard (see Ref. [13] for more information). The corre-

As noted above, the collected signal depends on different parameters that can be used to make images. More complex maps can thus be obtained by applying different calculations to the images; for instance, the ratio of peak intensity can distinguish changes in crystal orientation, as shown in **Figure 9**. The image of background intensity is also a particularly interesting

deposition, resulting in phases ranging from dolomite to magnesite.

sponding areas can be selected (g) in order to obtain the associated average spectra (h).

more information).

172 Raman Spectroscopy and Applications

**3.4. Exotic uses**

**Figure 9.** (a) Optical image in polarised/analysed transmitted light microscopy and (b) associated compositional Raman map of a polished thin section of hydrothermal quartz vein (Barberton, South Africa). The image of the intensity ratio of the peak located at 465 cm−1 over the peak located at 120 cm−1 permits highlighting differences in crystalline orientation. (c) Optical image and (d) associated compositional Raman map of a polished thin section of a pearl showing aragonite in purple and vaterite in dark blue. The background intensity image is used to show the vaterite (see Ref. [14] for more information). A small grain of SiC, coming from the polishing process, is also detected (in light blue in d). (e) Optical image and (f) associated compositional Raman map of a chert sample having undergone atmospheric entry (analogue meteorite, see Ref. [15] for more information). The quartz is in orange and the carbonaceous matter in green. The background intensity is displayed in dark blue and highlights cracks in the sample due to stresses related to atmospheric entry.

To conclude, various processing methods applied to Raman mapping data set can highlight particular properties of the mineral phases in a scanned sample. Further examples of such processing include making several scans of the same area using different environmental parameters (different laser wavelengths, external stress, temperature, etc.).

#### **3.5. Statistical treatment**

In all the above maps, the increase in information due to the statistics of a great number of spectra is implemented visually by plotting the appropriate classic spectral parameters. When the components of the sample are unknown, the reference spectra are not available, and the spectral sources are not well identified from spectrum to spectrum, or to extract very fine spectral modification of a single compound, mathematical treatment based on the statistical structure of the hyperspectral data can be implemented. Nowadays, numerous mathematical treatments exist and are tested by spectroscopists in various kinds of applications [9, 16, 17]. As an illustration, **Figure 10** shows the principal component analysis extraction of spectral sources of interest in the Raman mapping of a uranium dioxide ceramic and the corresponding

**Figure 10.** Principal component analysis of a Raman mapping of a uranium dioxide ceramic. At right, the spectral sources with their percentages and at left, the corresponding maps [9].

map [9]. The percentages to the right are the percentage of data variance coming from the corresponding spectral source. The maps show that only the first four sources have any significance, the fifth being only noise. Analysis of the information given by the spectral sources is complex. An initial analysis demonstrates the fact that different types of maps highlight either the grains or the grain borders in the ceramic (see Ref. [9] for more details).
