A Knowledge-Based Strategy for Interpretation of SWIR Hyperspectral Images of Rocks

Abstract

Strategies to interpret short-wave infrared hyperspectral images of rocks involve the application of analysis and classification steps that guide the extraction of geological and mineralogical information with the aim of creating mineral maps. Pre-existing strategies often rely on the use of statistical measures between reference and image spectra that are scene dependent. Therefore, classification thresholds based on statistical measures to create mineral maps are also scene dependent. This is problematic because thresholds must be adjusted between images to produce mineral maps of the same accuracy. We developed an innovative, knowledge-based strategy to perform mineralogical analyses and create classifications that overcome this problem by using physics-based wavelength positions of absorption features that are invariant between scenes as the main sources of mineral information. The strategy to interpret short-wave infrared hyperspectral images of rocks is implemented using the open source Hyperspectral Python package (HypPy) and demonstrated on a series of hyperspectral images of hydrothermally altered rock samples. The results show how expert knowledge can be embedded into a standardized processing chain to develop reproducible mineral maps without relying on statistical matching criteria.

1. Introduction

Laboratory-based infrared hyperspectral imaging is a fast and non-destructive method to obtain mineralogical information from rock samples [1,2,3,4,5,6]. The continuous coverage allows for the evaluation of spatial patterns, such as microstructures, that emerge as a result of the varying mineralogical composition in hyperspectral images. The level of mineralogical detail that can be derived from an image dependent on several factors, including the spatial resolution of the image, the spectral resolution of the camera, and the wavelength range imaged [7]. The wavelength range imaged by the camera determines the minerals that can be mapped because the wavelength positions of diagnostic absorption features vary between minerals [8,9].

Numerous methods have been developed to extract and interpret mineralogical information from hyperspectral images, and to create mineral maps, e.g., [10,11,12]. Interpretation strategies are defined as the series of steps performed to extract mineralogical information from hyperspectral images by the application of specific analysis, classification, and/or mapping methods. An example of such a strategy is the commonly used ENVI (ENVI is a trademark of NV5 Geospatial Solutions, Inc., Broomfield, CO, USA) hourglass approach [13,14]. This approach involves (i) the identification of mineralogical endmembers (i.e., the unique spectra of minerals present in the image), (ii) matching these endmember spectra to each of the image pixel spectra using statistical measures, (iii) defining thresholds that reflect positive matches to the various endmembers, and (iv) the classification of pixels using the defined thresholds in a mineral map of the identified endmembers. Many commonly used methods for matching the endmember spectra to image pixel spectra include the Spectral Angle Mapper (SAM) [15], Mixture-Tuned Matched Filtering [16], and Linear Spectral Unmixing [17] and, more recently, machine learning algorithms [18]. Various software environments and expert system frameworks have been developed to implement hyperspectral image processing and interpretation strategies, including Tetracorder [19], Material Identification and Characterization Algorithm (MICA) [20], ENVI, The Spectral Geologist (TSG, a trademark of CSIRO Australia), Hyperspectral Python (HypPy) [21], and HyLiTE [22]. More recently, machine learning algorithms have been included in hyperspectral mineral mapping workflows, e.g., [23,24,25,26].

Interpretation strategies that rely on statistics-based algorithms to measure similarities between endmembers and image pixel spectra have several weaknesses. First, the algorithms require the setting of threshold values to define matches between the reference and unknown spectra. The definition of thresholds is a subjective process because the values are influenced by the image characteristics, which depend on factors such as the specifications of the camera used, the applied calibration procedures, and the minerals present in the rock, e.g., [27]. Therefore, the statistics-based thresholds for classifying pixel-spectra into minerals vary between measurement campaigns and different rocks. This means that there are no universal statistical thresholds for the classification of hyperspectral images into mineral maps. The definition of thresholds requires systematic manual interference; this is time consuming and therefore not practical in the case of the classification of large numbers of images of multiple rocks, measured during different campaigns. Secondly, statistical matching algorithms, such as SAM, do not provide information on the specific wavelength range at which the endmember and the pixel spectrum match. The degree of matching is expressed as one value for the match over the entire wavelength interval and often dominated by large spectral trends and features, rather than sharp absorption features. This means that it is impossible to know whether the match is the result of a positive match between the absorption features of the unknown and reference spectra, or a similarity in the overall shape of the curves of the two reflectance spectra.

We present a novel strategy for the interpretation of hyperspectral images and the creation of mineral maps that does not depend on statistical relationships or statistical matching criteria. It does not require abundant training points to establish relationships between the reference and the image spectra as in supervised machine learning methods (e.g., [18]). The strategy is knowledge-based and suitable for the interpretation of large hyperspectral image data sets of multiple rocks. The main characteristics are (i) the incorporation of expert knowledge in the various steps in the workflow, (ii) the focus on the analysis of the physics-based wavelength positions of absorption features, (iii) the evaluation of spatial patterns in wavelength images and summary products, and (iv) the automation of parts of the processing and classification steps. The strategy differs from previous strategies because of an extensive exploratory analysis of spatial patterns in many different wavelength images, maps, summary products, and classifications that were created without the use of statistical matching algorithms and which forms the basis for the development of a customized and automated processing chain to create mineral maps.

In this paper, we describe the strengths and weaknesses of the interpretation strategy using an example analysis of the previously collected hydrothermally altered rocks that were imaged with a hyperspectral camera in the short-wave infrared (SWIR) wavelength range at 26 μm pixel size.

2. Materials and Methods

2.1. Image Acquisition and Preprocessing

Hyperspectral images were acquired from slabs of 11 rock samples (see Section 2.8) with a SWIR camera (Specim, model SWIR 3, Serial number SN430024, date 12 November 2015, Oulu, Finland) and OLES Macro lens mounted in a Sisuchema instrument setup (Specim, Spectral Imaging Ltd., Oulu, Finland) from 894 to 2511 nm in 288 bands at 26 μm pixel size. Image sizes were 384 by approximately 1200 pixels, resulting in images of 11 mm wide and approximately 3 cm long.

The raw images were converted to reflectance using the dark current and white reference measurements acquired during the recording of the images. Noise reduction involved the (i) removal of bad pixels that produced anomalous pixel values, (ii) identification and removal of bad bands near the lower (1000 nm) and higher (2500 nm) margins of the measured wavelength range, and (iii) application of a spatial–spectral smoothing filter that averaged five spatial and two spectral neighboring pixel values [21]. The filter was applied to reduce noise in the pixel spectra by a factor of 7 (approximately). Care was taken to ensure that shallow absorption features were not filtered out by the spectral smoothing process. The noise-reduced images are available at [28].

2.2. Interpretation Strategy

The strategy for the interpretation of SWIR hyperspectral images of rocks is presented in Figure 1. Step 1 involves the creation of wavelength maps, summary products, decision tree classifications and the mean spectra of classes. The methods to create these products are described in Section 2.3, Section 2.4, Section 2.5 and Section 2.6, respectively. Step 2 involves the exploratory analysis of the images, maps and spectra, produced in Step 1, in order to determine the variations in spectral features, reflectance spectra and their spatial distribution in the hyperspectral images.

Step 3 involves comparing the images and spectra produced in Step 1 with the complementary mineralogical and chemical analyses acquired by, for example, petrography and electron microprobe. The aim of Step 3 is to identify the different minerals with their chemical variations and their spatial patterns. In Step 4 an improved classification scheme is developed that is specific to the imaged rock samples, based on the exploratory analysis in Step 2 and in comparison with complementary data in Step 3. The processing chain describes the processing and classification steps to create this improved classification and can be automated. The application of the processing chain in Step 5 results in the creation of mineral maps.

2.3. Wavelength Mapping

Wavelength images and maps were created from the noise-reduced reflectance images. The wavelength mapping method has been described in various papers, e.g., [29,30]. The method was performed in two steps. First, wavelength images were calculated that contained the wavelength positions and depths of absorption features in each pixel spectrum. The wavelength positions of the deepest absorption features in a specified wavelength range were calculated by continuum removal over that wavelength range. The continuum was removed by division in order to negate the effect of albedo differences between image pixels. All local minima were determined on the continuum removed spectrum. In the remainder, only the three deepest local minima were taken into account. Subsequently, a second-order polynomial was fitted through the following three points that form an absorption feature, i.e., the local minimum and the spectrally neighboring points at shorter and longer wavelengths. The lowest point in the fitted parabola represented the wavelength value of the deepest feature. The difference between the reflectance value at this point and the continuum provided the depth of the feature. The same was performed on the second and third deepest absorption features in the specified wavelength range. After that, a wavelength map was created by displaying the wavelength position in color and the feature depth in gray scale in the same map. Wavelength images were calculated over wavelength ranges of 1300–1600 nm, 1650–1850 nm, 1850–2100 nm, and 2100–2400 nm. The three deepest features were calculated for each wavelength range. The values of the wavelength positions and depths of absorption features were analyzed in scatterplots, and as color composite images of the wavelength positions of the three deepest wavelength positions in red, green, and blue, respectively. Wavelength maps were created from the wavelength images with stretching intervals of 1300–1600 nm, 1650–1850 nm, 1850–2100 nm, 2100–2400 nm, and 2185–2225 nm. These wavelength ranges contain complementary information on specific molecular bonds, e.g., in OH, H₂O, AlOH, FeOH, MgOH, SiOH, and C${\mathrm{O}}_3^{2-}$ in lattices of the minerals present in the rocks (see [9] for an overview of wavelength ranges of mineral absorption features, and [31] for the wavelength positions of SiOH bonds). The wavelength maps provided intuitively interpretable visualizations of wavelength positions and depths of mineral absorption features in the rock samples.

2.4. Summary Products

Summary products are herein defined as images that result from mathematical operations on image bands or spectra to indicate specific minerals or parameters, e.g., [32]. Examples of summary products are band ratios and spectral indices. The summary products calculated from the noise-reduced reflectance images in this study are albedo, ferrous drop, illite–kaolinite, and Shannon entropy (

Table 1. Description of summary products. See Appendix A for an overview of the command-line syntax to create the summary products.

Name	Description	Interpretation
Albedo	Mean reflectance value of all bands in a pixel spectrum.	Brightness.
Ferrous drop (fedrop)	Ratio of reflectance bands at 1600 over 1310 nm [33].	Indication of ferrous iron in minerals, e.g., illite and carbonates. High values indicate abundant ferrous iron in the mineral.
Illite crystallinity (illx)	Ratio of the depths of deepest features between 2100–2400 and 1850–2100 nm, i.e., the depths of the Al-OH feature and water features of smectite–illite–muscovite minerals [33].	Indication of the degree of ordering of the mineral, e.g., [34]; High values indicate relatively high degrees of ordering.
Illite–kaolinite (illkaol)	Ratio of reflectance bands at 2164 over 2180 nm. The 2164 nm band is positioned at the second deepest feature of the doublet feature of kaolinite and the 2180 nm band is positioned at the high between the double feature [33].	Indication for the relative amounts of illite (high values) and kaolinite (low values). Note that the values are affected by the type of kaolinite in the rock and the center wavelengths of the bands of the hyperspectral camera used.
Shannon entropy	Measure from information theory: $$H\left(x\right)=-\sum _{i=1}^{n}p\left(x_i\right){\log }_{2}p\left(x_i\right),$$ $H\left(x\right)$ represents the entropy of a reflectance spectrum of n bands, where each reflectance value $\left(x_i\right)$ of each band is subtracted from 1 [35]. The spectrum is converted to density distribution prior to calculating $H\left(x\right)$.	It measures the deviation from a flat horizontal spectrum. A flat spectrum results in highest Shannon entropy values. Spectra with few but deep features produce low values.

). The Shannon entropy was used because of its property of highlighting the spectral contrast between image pixels. The illite crystallinity (Table 1) was calculated from the absorption feature depths in wavelength images between 1850–2100 nm and 2100–2400 nm.

2.5. Decision Trees

Decision trees were applied in order to convert the wavelength images into classified images and to slice all summary products into classes according to predefined, generically applicable thresholds. Decision trees consist of a series of binary nodes with Boolean statements that are true or false [36]. An example of such a Boolean statement is ”the wavelength position of the deepest feature is between 2190 and 2200 nm.” This statement is either true or false depending on the values in the wavelength image. By connecting multiple binary nodes, a decision tree is formed that classifies all pixels based on the expressions in the nodes. A major advantage of the classification of wavelength images and summary products using invariant thresholds in the decision trees is that the resulting images contain the same sets of classes between images. This makes the classes comparable between images and across different acquisitions. A previous example of the use of a decision tree to classify a spectral index map to map mineralogy is provided by [37].

The decision trees in this study were built using experience with the interpretation of mineral reflectance spectra. This includes knowledge of the wavelength positions and depths of absorption features in the spectra and the range of values of summary products calculated from the spectra. This knowledge was used to determine the thresholds used in the Boolean nodes of the decision trees. Although the setting of the thresholds introduced a degree of subjectivity in the classification process, it is based on the choices made by experts and is consistent between images. The decision trees $dt\_2100\_2400$, $dt\_albedo$, $dt\_fedrop$, $dt\_illcryst$ and $dt\_illkaol$ (Table 2) were created for the exploratory analysis and comparison with other mineral and chemical analyses in steps 2 and 3 in Figure 1.

Decision tree $dt\_mineral\_map$ (Table 2, Figure 2) resulted from the improved classification scheme in Step 4 and was applied in the processing chain in Step 5 to create mineral maps (Figure 1). The depth and wavelength position of absorption features between 2100 and 2400 nm and illite crystallinity were used for the classification. The decision tree classified the images into illite–muscovite minerals of different composition and crystallinity, kaolinite, and different types of chlorite. The application of this tree involved the (i) selection of pixels with a minimum feature depth of 0.05 reflectance in the rectangular box at the upper left of the decision tree in Figure 2 (pixels with shallower depth end up in the box below are named “aspectral”), (ii) subdivision of pixels with absorption feature larger than 5% along right-hand branches of the three of the wavelength of the deepest absorption features (W1) at 2180 nm, 2200 nm, 2210 nm, 2220 nm, 2240 nm, 2260 nm, 2300 nm, and 2360 nm (along the second column in Figure 2), and (iii) further subdivision of the pixels by slicing the wavelength positions of second deepest absorption features (W2) at 2160 nm, 2180 nm, 2340 nm, and and 2400 nm, and crystallinity values (Ix) at 2 and 3.25, along the third to sixth column. In the resulting classification, all image pixels are assigned to one of the classes in the decision tree and labeled accordingly. Classes that were labeled ”other” indicate minerals not present in the rock sample set or present in very small quantities. Note that the mineral classes represent either pure minerals or mixtures of minerals.

The decision trees were also applied to reference spectra from spectral libraries to test the performance of decision trees and to verify if the classifications produced the required results.

Table 2. Description of decision trees used to classify wavelength images (Section 2.3) and summary products in Table 1. See Figure 2 and Appendix C for the design of the decision trees. Thresholds for slicing and classification are based on experience with the interpretation of mineral spectra.

Decision Tree	Resulting Classified Image	Description
$dt\_2100\_2400$	wave2100–2400_class	Classification based on depth of deepest and wavelength positions of first and second deepest absorption features in wavelength image between 2100 and 2400 nm (see Figure A2).
$dt\_albedo$	albedo_class	Slicing of albedo image at thresholds: 0.25, 0.38 and 0.50 (see Figure A3).
$dt\_fedrop$	fedrop_class	Slicing of ferrous drop (fedrop) image at thresholds: 1.1, 1.2, 1.3, 1.4 and 1.5 (see Figure A4).
$dt\_illcryst$	illx_class	Slicing of illite crystallinity (illx) image at thresholds: 0.25, 0.33, 0.5, 1, 2, 3 and 4 (see Figure A5).
$dt\_ill\_kaol$	illkaol_class	Slicing of illite over kaolinite (illkaol) image at thresholds: 0.95, 0.97, 0.99, 1.0, 1.01, 1.03 and 1.05 (see Figure A6).
$dt\_mineral\_map$	mineral_map	Classification using depth and wavelength positions of deepest features in the wavelength images between 2100 and 2400 nm and the illite crystallinity image. Customized for the rock sample set in this study (see Figure 2).

Table 2. Description of decision trees used to classify wavelength images (Section 2.3) and summary products in Table 1. See Figure 2 and Appendix C for the design of the decision trees. Thresholds for slicing and classification are based on experience with the interpretation of mineral spectra.

Decision Tree	Resulting Classified Image	Description
$dt\_2100\_2400$	wave2100–2400_class	Classification based on depth of deepest and wavelength positions of first and second deepest absorption features in wavelength image between 2100 and 2400 nm (see Figure A2).
$dt\_albedo$	albedo_class	Slicing of albedo image at thresholds: 0.25, 0.38 and 0.50 (see Figure A3).
$dt\_fedrop$	fedrop_class	Slicing of ferrous drop (fedrop) image at thresholds: 1.1, 1.2, 1.3, 1.4 and 1.5 (see Figure A4).
$dt\_illcryst$	illx_class	Slicing of illite crystallinity (illx) image at thresholds: 0.25, 0.33, 0.5, 1, 2, 3 and 4 (see Figure A5).
$dt\_ill\_kaol$	illkaol_class	Slicing of illite over kaolinite (illkaol) image at thresholds: 0.95, 0.97, 0.99, 1.0, 1.01, 1.03 and 1.05 (see Figure A6).
$dt\_mineral\_map$	mineral_map	Classification using depth and wavelength positions of deepest features in the wavelength images between 2100 and 2400 nm and the illite crystallinity image. Customized for the rock sample set in this study (see Figure 2).

Figure 2. Design of decision tree $dt\_mineral\_map$ (Table 2) developed in Step 4 (Figure 1) for the classification of the hyperspectral images of the eleven rock samples in this study. Input images are depth of deepest absorption feature between 2100 and 2400 nm (D1) and wavelength position of deepest (W1) and second deepest (W2) features and illite crystallinity (Ix). Down pointing arrows arrows present a “False” (F) to the binary statements, horizontal arrows present a “True” (T). The class “other” represents classes that did not occur in the rock sample set or occurred in very small quantities. The colors are similar to those in the resulting mineral map. ill = illite, musc = muscovite, chlt = chlorite, unspec = unspecified, lx = low crystallinity, hx = high crystallinity, lw = long wavelength, sw = short wavelength.

Figure 2. Design of decision tree $dt\_mineral\_map$ (Table 2) developed in Step 4 (Figure 1) for the classification of the hyperspectral images of the eleven rock samples in this study. Input images are depth of deepest absorption feature between 2100 and 2400 nm (D1) and wavelength position of deepest (W1) and second deepest (W2) features and illite crystallinity (Ix). Down pointing arrows arrows present a “False” (F) to the binary statements, horizontal arrows present a “True” (T). The class “other” represents classes that did not occur in the rock sample set or occurred in very small quantities. The colors are similar to those in the resulting mineral map. ill = illite, musc = muscovite, chlt = chlorite, unspec = unspecified, lx = low crystallinity, hx = high crystallinity, lw = long wavelength, sw = short wavelength.

2.6. Mean Spectra of Classes

For each class in the classified images, the mean reflectance spectrum for that class was calculated, as well as the number of pixels in that class. The mean spectra were used in exploratory analyses, the comparison with other chemical and mineral analyses, and the development of the improved classification scheme (steps 2, 3 and 4 of Figure 1) to interpret the mineralogy of the class from which it was derived. The percentage of pixels per class was useful to quantify the mineralogical composition and to filter out small classes. Small classes may represent noise (as shown for example, by erroneous pixel spectra) or mineral variations that are not of interest. They can be removed from the analysis or merged with other classes, respectively. Care must be taken not to remove small, but geologically significant, classes.

2.7. HypPy Software

In the interpretation strategy, the processing and classification steps run entirely on routines in the Hyperspectral Python (HypPy) software, version 3. HypPy is open-source software and can be operated with a GUI and from the command line in GNU/Linux and Windows environments [21]. The creation of images and spectra in Step 1 and the application of the processing chain in Step 5 to create the “final” mineral maps (Figure 1) were automated using HypPy software. Automation was achieved by executing command line statements of HypPy (see Appendix A) in a series of GNU/Linux shell scripts.

2.8. Test Sample Set and Validation

The test sample set consists of eleven rock samples. There were collected from hydrothermally altered sections of the Archean Soanesville greenstone belt in the Pilbara craton in Australia collected along a transect crosscutting different alteration zones that showed a variety of minerals and microstructures. The test sample set exhibits a diversity of SWIR active minerals of varying chemical composition (detectable as minor shifts in wavelength positions of absorption features), hull shapes, and overlapping features that are challenging to differentiate using statistics-based mapping methods. Furthermore, the petrography, mineralogy, and chemistry of the samples were analyzed and described in detail by [38,39,40]. Ten of the samples are hydrothermally altered volcanics of andesitic to basaltic composition (see

Table 3. Summary of the geological characteristics of samples used in this study as determined by petrography, whole rock geochemistry, microprobe, and field spectrometer analysis [38,39].

ID	Sample	Description
1	P2003	Weakly sericite altered and silicified muddy chert.
2	P2004	Deuterically altered, silicified, seriticized (Al-rich), xenocrystic phenocrystic andesite.
3	P2005	Deuterically altered, silicified, seriticized (Al-rich), phenocrystic andesite.
4	P2006	Deuterically altered, silicified, seriticized (Al-rich), weakly phenocrystic andesite.
5	P2007	Deuterically altered, silicified, seriticized (Al-rich), weakly phenocrystic quenched andesite.
6	P2008	Deuterically altered, silicified, seriticized (Al-poor), weakly phenocrystic andesite.
7	P2009a	Deuterically altered, silicified, seriticized (Al-poor), weakly xenocrystic amygdaloidal basalt. Contains aproximately 15% kaolinite in amygdales.
8	P2010	Deuterically altered, silicified, seriticized (Al-poor), weakly xenocrystic weakly amygdaloidal basalt.
9	P2012	Deuterically altered, silicified, ferruginous, chloritised basalt.
10	P2013	Deuterically altered, silicified, ferruginous, chloritised (pyroxene-bearing) basalt.
11	P2014	Deuterically altered, silicified, chloritised amygdaloidal andesite.

). Petrographic analyses confirmed that the samples could be described as fine-grained and silicified in nature and composed predominantly of quartz, white micas, and chlorite. Because of the fine-grained nature of the samples, microprobe analysis was performed to investigate the differences between the white mica minerals in the samples. The results showed that the wavelength positions of the white micas were correlated with the Al content of the white micas. The availability of detailed chemical and mineral analyses made it possible to compare and validate the mineral maps produced from the hyperspectral images with known data sets. Three of the samples were chlorite–quartz altered and six were white mica altered. The white-mica-altered samples were divided into two groups, namely, (1) the relatively Al-rich white mica group that was apparent from the relatively short wavelength position of the 2200 nm absorption feature, and (2), relatively Al-poor with a longer wavelength position of the 2200 nm feature. One of the rock samples contained kaolinite, likely the result of weathering, and one sample represents a silicified sediment (chert).

The results of the exploratory analyses of the hyperspectral data and the comparison with the petrographic, chemical, and mineralogical data of the eleven rock samples are presented in Section 3.1. The mineral map that results from the comparison is presented in Section 3.2.

3. Results

3.1. Exploratory Analyses

Wavelength maps, summary products, and classifications of the 11 rock samples created in Step 1 of the interpretation strategy (Figure 1) are shown in Figure 3, Figure 4 and Figure 5. Wavelength maps calculated from other wavelength ranges are presented in Appendix B.

The albedo images in Figure 3a show the brightness of the different rocks in shades of grey. These images exhibit rock microstructures, such as layering, phenocrysts, and amygdales, as well as artificial surface features such as lapidary saw marks. Variations in brightness in the images are enhanced by linear stretching (±2 standard deviations) according to the image-specific pixel values. The brightness differences between the rocks can be observed in the multi-level thresholded albedo image (Figure 3b) after applying the same thresholds between images. The sliced albedo images show that the quartz–sericite altered rock samples 2 to 8 are brighter than sample 1 (a chert) and the chlorite–quartz altered rock samples 9 to 11.

Figure 3. Wavelength maps, summary products and classifications of the hyperspectral images of the eleven rock samples: (a) Albedo, (b) classified albedo image (albedo_class), (c) wavelength map between 2100 and 2400 nm, (d) wavelength map between 2100 and 2400 nm and stretched between 2185 and 2225 nm. See Table 2 for decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013 and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. med = medium.

Figure 3. Wavelength maps, summary products and classifications of the hyperspectral images of the eleven rock samples: (a) Albedo, (b) classified albedo image (albedo_class), (c) wavelength map between 2100 and 2400 nm, (d) wavelength map between 2100 and 2400 nm and stretched between 2185 and 2225 nm. See Table 2 for decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013 and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. med = medium.

Figure 4. Wavelength images, wavelength maps, and classifications of the hyperspectral images of the eleven rock samples: (a) color composite of wavelength positions of the first, second and third deepest absorption features, respectively w1, w2 and w3, between 2100 and 2400 nm, (b) classified wavelength image between 2100 and 2400 nm (wave2100–2400_class), (c) wavelength map between 1650 and 1850 nm, and (d) classified fedrop image (fedrop_class). See Table 2 for decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013 and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. ill = illite, musc = muscovite, kaol = kaolinite, chlt = chlorite, med = medium.

Figure 4. Wavelength images, wavelength maps, and classifications of the hyperspectral images of the eleven rock samples: (a) color composite of wavelength positions of the first, second and third deepest absorption features, respectively w1, w2 and w3, between 2100 and 2400 nm, (b) classified wavelength image between 2100 and 2400 nm (wave2100–2400_class), (c) wavelength map between 1650 and 1850 nm, and (d) classified fedrop image (fedrop_class). See Table 2 for decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013 and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. ill = illite, musc = muscovite, kaol = kaolinite, chlt = chlorite, med = medium.

Figure 5. Summary products and classifications of the hyperspectral images of the eleven rock samples: (a) classified illite crystallinity image (illx_class), (b) classified illite over kaolinite image (illkaol_class), (c) entropy, and (d) mineral map (mineral_map). See Table 2 for the decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013, and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. smec = smectite, ill = illite, musc = muscovite, kaol = kaolinite, chlt = chlorite, unspec = unspecified, sw = short wavelength, lw = long wavelength, lx = low crystallinity, hx = high crystallinity.

Figure 5. Summary products and classifications of the hyperspectral images of the eleven rock samples: (a) classified illite crystallinity image (illx_class), (b) classified illite over kaolinite image (illkaol_class), (c) entropy, and (d) mineral map (mineral_map). See Table 2 for the decision trees used for classification. Rock samples: 1 = P2003, 2 = P2004, 3 = P2005, 4 = P2006, 5 = P2007, 6 = P2008, 7 = P2009a, 8 = P2010, 9 = P2012, 10 = P2013, and 11 = P2014. See Table 3 for descriptions of the rock samples. The width of the images is 1.1 cm. smec = smectite, ill = illite, musc = muscovite, kaol = kaolinite, chlt = chlorite, unspec = unspecified, sw = short wavelength, lw = long wavelength, lx = low crystallinity, hx = high crystallinity.

Wavelength maps between 2100 and 2400 nm show that the wavelength position of the deepest absorption features near 2205 nm and 2350 nm (green and yellow colors, respectively, in Figure 3c) are the most abundant, and that the 2205 nm feature dominates over the 2350 nm feature because of the more abundant green colors. The wavelength map in Figure 3d shows the result of the calculation of the wavelength image between 2100 and 2400 nm and subsequent mapping over the range between 2185 and 2225 nm to create the wavelength map. These maps show the contrast in the wavelength position of absorption features near 2205 nm in green, red, orange, and yellow colors. The green colors indicate white micas of relatively short wavelength position between 2200 and 2205 nm (rock samples 2–4). The orange–red colors indicate a shift of the Al-OH absorption feature to longer wavelengths up to approximately 2215 nm (rocks 5–9).

The variation in wavelength positions of the three deepest absorption features is highlighted in Figure 4a, which is a color composite of the wavelengths of the first, second, and third deepest features between 2100 and 2400 nm. The different colors indicate different combinations of wavelength positions of absorption features, which are largely the result of varying amounts and types of illite–muscovites, chlorites, kaolinite, and their mixtures. Because each image is stretched according to the statistics of its wavelength position, similar minerals may show in different colors between images. Kaolinite in rock sample 7 of Figure 4a is shown in blue (R: ∼2209 nm; G: ∼2166 nm; B: ∼2320 nm), and white micas surrounding the kaolinite is shown in yellow-green (R: ∼2210 nm; G: ∼2351 nm; B: ∼2115 nm). The scatterplots in Figure 6 show the wavelength positions and the three deepest features for each pixel in the wavelength image of rock sample 7. The deepest features in Figure 6a are largely confined to a narrow cluster between 2209 and 2213 nm and 0.15 and 0.50 reflection. These values are consistent with deepest absorption features of kaolinite and white mica minerals. The wavelength positions of second deepest features in Figure 6b cluster near 2166 and 2350 nm, which is consistent with kaolinite and muscovite, respectively (see mean class spectra of kaolinite and illite–muscovite in Figure 7). The third deepest features cluster at various wavelengths and are generally shallower. The cluster at ∼2320 nm in Figure 6c represents kaolinite. The cluster near ∼2115 nm represents white micas. Blue kaolinite clusters in rock sample 7. Figure 4a exhibits amygdales in the volcanic rock found in the petrographic study of the rocks (Table 3). The magenta-colored pixels in Figure 4c show an 1820 nm feature of kaolinite (see kaolinite spectrum in Figure 7). The blue circular object near the top in rock sample 2 in image Figure 4a corresponds to a xenocryst in the volcanic rock. The maps in Figure 4b result from the classification of the wavelength images between 2100 and 2400 nm using the decision tree $dt\_2100\_2400$ (Table 2). This decision tree classifies pixel spectra by the wavelength position of the deepest and second deepest absorption features which are determined by the type and strength of the molecular bonds in Al-OH, Fe-OH, Mg-OH, and carbonate molecules in mineral lattices. The maps in Figure 4b show two illite–muscovite dominant assemblages, one with relatively short wavelengths in yellow (rock samples 2, 3, 4 and 5) and a second with longer wavelengths in orange (rock samples 6, 7 and 8). The maps also show chlorite dominant assemblages in green (rock samples 10 and 11). Rock sample 9 contains mixtures of illite–muscovite and chlorite. The clusters of yellow pixels in rock sample 11 show white-mica-rich amygdales in the predominantly chlorite-altered rock. In white mica and chlorite mixtures, white micas dominate the spectra, possibly because they are brighter than chlorites.

Chlorite-bearing rocks contain more ferrous iron, which can be seen in rock samples 9, 10, and 11 by the brownish colors (higher ferrous drop values in Figure 4d). Illite crystallinity values vary in the illite–muscovite-containing rocks (Figure 5a). For instance, phenocrysts in rock sample 2 (shown as cyan-colored clusters of pixels) have higher crystallinity values than the minerals in the matrix of the rock.

Figure 5b shows the presence of the second feature of kaolinite at 2170 nm by a change of color from blue (illite) to light blue and green (kaolinite). Evidence of kaolinite in rock sample 7 is shown in green.

The Shannon entropy images in Figure 5c show rock textures that result from variations in the spectral mineralogy in the hyperspectral images. For example, the black clusters of pixels in rocks samples 2, 3 and 4 represent white-mica-altered phenocrysts with high crystallinity values. These white micas produce lower Shannon entropy values than minerals in the rock matrix surrounding these phenocrysts. These differences are visible in the images in Figure 5c, and they enhance the microstructure of the rock.

The results of the exploratory analyses were used to create the improved classification scheme for the rock samples and to develop the processing chain to create the mineral maps described in Section 3.2.

3.2. Mineral Maps and Comparison with Petrography

The mineral maps in Figure 5d resulted from the application of the processing chain, which classified wavelength images and summary products using the decision tree $dt\_mineral\_map$ (Figure 2) in Step 5 of the interpretation strategy. The resulting mineral maps show the distribution of illite–muscovite minerals of different composition and crystallinity, and kaolinite and chlorite minerals, and their areal abundance in each map (see

Table 4. Summary of hyperspectral mineral map classes and minerals identified by petrographic analysis.

Sample Number	Mineral Map Classes 1	Percentage Image Pixels 2	Minerals Identified Using Petrography
(1) P2003	ill-musc unspec	45.5	Quartz, hematite, goethite, rutile, sericite
	aspectral	32.2
	ill-musc-lw unspec	6.7
	other	6.1
	ill-musc	5.0
(2) P2004	ill-musc	92.2	Quartz, sericite, goethite
	ill-musc-lx	4.0
	ill-musc-hx	3.3
(3) P2005	ill-musc	92.9	Quartz, sericite, goethite
	ill-musc-hx	3.8
	ill-musc-lx	2.7
(4) P2006	ill-musc-lx	87.2	Quartz, sericite, goethite
	ill-musc	12.1
(5) P2007	ill-musc	92.7	Quartz, sericite, goethite
	ill-musc-lx	3.5
	ill-musc-sw	2.9
(6) P2008	ill-musc-lw	95.7	Quartz, sericite, goethite
	ill-musc-lw-hx	2.1
(7) P2009a	ill-musc-lw-hx	45.1	Quartz, sericite, goethite, accessory chlorite
	ill-musc-lw	44.4
	kaolinite	6.7
(8) P2010	ill-musc-lw	99.1	Quartz, sericite, goethite, hematite, accessory chlorite
(9) P2012	ill-musc-lw-lx	21.8	Quartz, goethite, chlorite, accessory sericite
	ill-musc-lw unspec	21.5
	chlt	14.1
	ill-musc-lx	9.3
	Fe-chlt	8.0
	Fe-chlt unspec	6.3
	ill-musc unspec	6.3
	other	5.0
(10) P2013	Fe-chlt	89.9	Quartz, goethite, chlorite
(11) P2014	Fe-chlt	66.2	Quartz, chlorite, goethite
	ill-musc unspec	9.0
	chlt	8.5
	ill-musc	6.9
	Fe-chlt unspec	5.5
	ill-musc-lx	3.1

¹ Mineral classes as in Figure 2: ill = illite, musc = muscovite, chlt = chlorite, unspec = unspecified, lx = low crystallinity, hx = high crystallinity, lw = long wavelength, sw = short wavelength. ² Classes smaller than 2% image pixels were omitted for clarity.

). Note that the sericite indicated in the petrographic descriptions refers to fine-grained white mica minerals, which includes the group of illite–muscovite minerals. The mineral maps in Figure 5d also show microstructural features that result from the variation in mineralogy. The spectra of rock sample 1 predominantly contain shallow absorption features. The pixels in the aspectral class have absorption features between 2100 and 2400 nm and shallower than 0.05 reflectance (Figure 7, p2003 aspectral). The spectra with deeper features indicate white micas (yellow colors) and minerals with features between 2340 and 2400 nm (green), which are not diagnostic but could include chalcedony. Sedimentary structures are shown by the alternation of yellow, black, and green layers. Rock samples 2–5 contain predominantly Al-rich white micas with wavelength positions between 2200 and 2205 nm and of varying illite–muscovite crystallinity (low crystallinity values in grey, intermediate values in yellow and high values in cyan; see Figure 7 for representative mean spectra of rock 2). The variation in crystallinity values shows the presence of xenocrysts (grey pixels in rock sample 2) and phenocrysts (cyan pixels in rock samples 2 and 3, yellow pixels in rock sample 4). Rock samples 6–8 contain predominantly Al-poor white micas with wavelength positions between 2210 and 2215 nm and varying illite–muscovite crystallinity (i.e., low crystallinity values in dark blue, intermediate values in orange and high values in lavender; see Figure 7 for the representative mineral spectra of rock samples 2, 6, and 7). Rock sample 7 contains kaolinite in amygdales. Rock samples 9–11 contain white micas and chlorites (in green, see Figure 7 for the endmembers of chlorite endmembers) and their mixtures. The amygdales in rock 11 show as clusters of yellow pixels.

The qualitative comparison of the mineral maps and the results of the petrography and microprobe analyses (see summaries in Table 3 and Table 4) confirmed the presence of the illite–muscovite minerals of different compositions. They also confirmed the presence of chlorites in the respective rocks as well as various microstructural characteristics in the maps that were highlighted by mineralogical difference in the maps. The difference between the samples dominated by illite–muscovite and chlorite was also observed in micro-photographs (see Appendix D), the latter being more ferruginous. The difference in illite–muscovite crystallinity values could not be analyzed using the applied methods. However, they might be related to different microstructural features of the rocks. Quartz, goethite, and hematite (see Table 4) minerals could not be observed in the hyperspectral maps because of the lack of absorption features between 100 and 2500 nm. Amphiboles were not be found in the mineral maps, possibly from the alteration of amphibole to chlorite.

The visual interpretation of representative mean spectra (in Figure 7) and the classification of selected reference spectra using the same decision trees in

Table 5. Classification of reference spectra from the USGS spectral library version 7 [41] using $dt\_mineral\_map$ (Table 2, Figure 2). chlt = chlorite; epid = epidote; ill = illite; musc = muscovite; sw = short wavelength; lw = long wavelength; lx = low crystallinity; hx = high crystallinity.

Class	Count	Reference Spectrum
ill-musc-sw	2	Muscovite_GDS113_Ruby; Muscovite_GDS113a_Ruby
phengite	2	Illite_GDS4.2_Marblehead; Illite_GDS4_Marblehead
epid/chlt	5	Chlorite_HS179.1B; Chlorite_HS179.2B; Chlorite_HS179.3B; Chlorite_HS179.4B; Chlorite_HS179.6
ill-musc-hx	3	Muscovite_HS146.1B; Muscovite_HS146.3B; Muscovite_HS146.4B
ill-musc-lx	2	Illite_IMt-1.a; Illite_IMt-1.b_lt2um
ill-musc-lw-hx	2	Muscovite_GDS116_Tanzania; Muscovite_GDS116a_Tanzania
kaolinite	3	Kaolinite_CM9; Kaolinite_KGa-1_(wxl); Kaolinite_KGa-2_(pxl)

assisted in the interpretation of the mean spectra. It confirmed the decision tree classified reference spectra in the right classes, according to the wavelength positions of absorption features and crystallinity values.

4. Discussion

4.1. Strengths

The interpretation strategy presented here has the following strengths: (i) the focus on the wavelength positions of absorption features, which are physical mineral properties rather than statistical measures, (ii) the possibility of using spatial patterns in different wavelength maps and summary products, (iii) its versatility, (iv) the possibility of the built-in expert knowledge, and (v) the option to automate processing and classification steps.

Focus on wavelength positions of absorption features. The wavelength positions of absorption features in the SWIR are directly related to the energy level of atomic and molecular bonds in the crystal lattices and are therefore physics based [42]. This makes them less susceptible to changing the measurement conditions and matrix effects in rocks (caused by other minerals and their geological structures); they can be considered robust indicators of the presence of minerals that contain these bonds. In addition, minor changes in bond strength caused by the substitution of Al by Fe and Mg in octahedral sites in white micas manifest as small nanometer-scale changes in absorption wavelengths of the ∼2200 nm feature [43]. These minor wavelength shifts are readily observable in the wavelength images.

Evaluation of spatial patterns in wavelength maps and summary products. The possibility of evaluating spatial patterns in wavelength images and summary products (for example in Figure 3, Figure 4 and Figure 5) allows for direct analyses of the relationships between the spectral features and geological textures and micro structures. Wavelength images (and some of the summary products) represent physical mineral properties and therefore show the distribution of physical mineral parameters, which is not the case in rule images created by measuring statistical similarity between reference and pixel spectra. Furthermore, wavelength images help to identify image noise, i.e., typical spatial patterns created by noise, including incoherent shapes and/or striping.

Versatility. The strategy is versatile because it allows for the modification of the wavelength ranges of wavelength maps, their stretching intervals, the type of summary products that enhance specific mineralogical variation, as well as the addition of newly developed decision trees or the modification of existing trees. Processing chains can readily be adapted to create mineral maps that are specific for certain mineral assemblages and geological settings.

Incorporation of expert knowledge. The use of expert knowledge reduces uncertainty in the analysis and interpretation of hyperspectral images. Expert knowledge was incorporated in several steps of the interpretation strategy, including (1) in the creation of wavelength images and maps, summary products, and decision trees for the exploratory analysis in Step 1, (2) the exploratory analysis and interpretation of maps and spectra in Step 2, (3) the comparison with reference spectra and other mineral analyses in Step 3, and (4) the development of the improved classification scheme and decision tree to create mineral maps in step 4. However, once the processing chain for the creation of mineral maps is developed, the processing can be performed by non-experts, and the mineral maps can be readily understood.

Automation of processing and classification steps. All processing and classification steps in the interpretation strategy were created using HypPy software command line statements. This made it possible to automate the steps in which images, maps, and spectra were created and processed, but the steps that required analysis and interpretation by a human operator could not be automated. Once the processing chain for the classification of the rock samples is created (in Step 4 in Figure 1), the execution of the procedure is automated and applied to all hyperspectral images. Automation has the advantage of the reproducibility of results, the image processing and classification processes become time efficient, and the risk of human-induced errors is reduced. The processing time for the execution of the processing chain largely depends on the time required to create wavelength images. However, a test carried out on a PC with a 12th generation i5-12600 (3300 Mhz) Intel processor with 6 cores showed that the calculation of a wavelength image from a 1.12 Gb hyperspectral image of 271 × 1415 pixels took 2 m and 30 s, while the creation of a wavelength map from the wavelength image only took 3 s.

4.2. Weaknesses

Weaknesses of the interpretation strategy may lie in (i) the dependence on expert knowledge to develop the various classification procedures, and (ii) the focus on deepest features that cause bias toward spectrally dominant minerals.

Dependence on expert knowledge to develop processing chain. The interpretation of hyperspectral image maps and spectra and the development of a processing chain to create mineral maps requires spectral geological expert-knowledge. This might be seen as a limitation because it requires investment in understanding mineral spectroscopy and the analysis and interpretation of hyperspectral images. In addition, care must be taken when multiple experts create classifications of a similar set of samples from a specific geological setting. Agreement on the use of the same processing chain with the same classification procedure is then required to obtain reproducible results.

Bias toward dominant spectral minerals. The focus on the deepest absorption features may result in bias toward the identification of minerals that produce deep absorption features and dominate the spectral mineralogy. It is important that the interpreter of the various classifications is aware of this potential bias so that the interpretation strategy is adapted to include the detection of minerals with shallow absorption features. One example of such strategies is the adjustment of the depth stretch of absorption features in the process of calculating wavelength maps so that shallow features are also visible. Another example is the use of multiple absorption features (including the shallower features of less dominant minerals) within specific wavelength ranges.

Pixel counts of the classes in the mineral maps provided semi-quantitative estimates of the spectrally observed minerals. These estimates do not represent true mineral concentrations because they are based on the dominant spectral minerals in each pixel. To determine true mineral concentrations, the following must be considered: (i) the content of non-spectrally active minerals (such as quartz in the applied wavelength range between 1000 and 2500 nm), (ii) the presence of mixtures of spectrally active minerals within pixels, and (iii) the response of mineral absorption feature depths to increasing concentration in all pixel spectra. These are topics for further research.

4.3. Application

We demonstrated the interpretation strategy on laboratory-acquired SWIR hyperspectral images of hydrothermally altered volcanic rock samples at high spatial and spectral resolution between 1000 and 2500 nm. The rock samples were coherent and intact; and therefore, the spatial patterns in the distribution of spectral minerals could be evaluated and allowed the identification of sedimentary and volcanic microstructures including the presence of phenocrysts, xenocrysts, and amygdales, cf. [44]. The strategy can also be applied to hyperspectral images of rock samples that are fragmented, such as rock chips resulting from drilling [45]. Focus is then less on the analysis of spatial patterns in the imagery and more on the classification of the minerals in the rock fragments.

The study has shown that the classification procedures used in the interpretation strategy can be applied automatically to multiple images from rocks acquired from the same type of geological setting with similar mineralogical compositions using the same processing chain. Therefore, the method is potentially suitable for the automated classifications of large sets of hyperspectral images. The analyses and interpretation can be applied over large depth intervals with the same classification procedure and with reproducible results. Care should be taken when working in different geological settings. The classification procedures may have to be adjusted by comparison with complementary mineralogical analyses to include the changed mineralogical characteristics in that setting. This may result in the use of wavelength maps over different wavelength ranges, other summary products and decision trees, and a modified processing chain.

The interpretation strategy presented herein was developed for academic geological research but may also be suitable where mineralogical information is required for informed decision making, for example in mineral exploration [46] and the exploration for hydrocarbons and geothermal energy [45].

The interpretation strategy has not been tested on hyperspectral images acquired with other cameras or covering other wavelength ranges (e.g., visible, near and long-wavelength infrared), on multi-spectral images and on airborne and spaceborne hyperspectral images at lower spatial resolution. Investigating the suitability of the strategy on those data sets are topics for future research.

5. Conclusions

The novel interpretation strategy presented here involves the application of a series of processing, analysis, and classification steps for the extraction of geological and mineralogical information from the hyperspectral images of rocks and the creation of mineral maps. This strategy differs from previous strategies because of the extensive exploratory analysis of spatial patterns in many different wavelength images, maps, summary products, and classifications that were created without the use of statistical matching algorithms. The results of the exploratory analysis were subsequently used for the development of a customized and automated processing chain to create mineral maps. We conclude that the interpretation strategy provides an effective means to embed expert knowledge in the interpretation of hyperspectral images of rocks and into a processing chain to create reproducible mineral maps without relying on statistical matching criteria.