4.1.

Hyperspectral Image Preprocessing

Different levels of preprocessing are available to correct the original hyperspectral images including geometric and atmospheric correction (ATCOR). Bad characteristics of an image that have been caused by the sensor can be removed by geometric or sensor error correction techniques. These errors involve stripes and other noise type distortions.^24,25 For example, several vertical lines with no information can be produced by detectors in Hyperion sensor. The information from these areas can be covered by replacing their pixel values with the average pixel values of their neighboring pixels.²⁵

Due to scattering and absorption of solar radiation and spectra by atmospheric gases and aerosols, the hyperspectral images have atmospheric effects. These atmospheric effects must be removed in order to use hyperspectral data for quantitative remote sensing.²⁶ ATCOR methods have evolved through the years so that they can be categorized as scene-based empirical approaches and radiative transfer model (RTM)-based approaches.^25–28 Several scene-based empirical approaches were developed to correct the atmospheric effects on the hyperspectral images. Kruse²³ computed the average spectrum of a scene by internal average spectrum of a scene. Then the corresponding spectrum of any pixel is divided by the average spectrum to get relative reflectance spectrum for each pixel. This method was mostly applied for areas without any vegetation. Roberts et al.²⁹ introduced flat field correction approach. They normalized the input image to an assumed area with neutral spectrally reflectance (with flat topographic and spectral feature). The empirical line approach³⁰ uses field reflectance spectra for bright and dark targets to linearly correlate the raw input imaging spectrometer data. Reinersman et al.³¹ developed empirical “cloud shadow” method for ATCOR over the darker water surfaces. The quick ATCOR approach³² is developed to remove atmospheric effects from multispectral and hyperspectral images in VNIR and SWIR bands. This approach drives atmospheric compensation factor directly from imaging spectrometer data.

RTMs are physically based codes that try to simulate the transfer process of an electromagnetic wave in the atmosphere.³³ They need field-based measurement of the atmosphere conditions at the time of image attainment and produce apparent reflectance or scaled surface reflectance.^25,34 There are several RTM-based algorithms to model the ATCOR. Atmospheric removal algorithm was developed by Gao et al.³⁵ This model performed to retrieve ground reflectance spectra from hyperspectral data by a theoretical modeling technique. It simulates the absorption and scattering effects of gases and aerosols existing in the atmosphere. ATCOR developed by Richter³⁶ includes a large database of ATCOR functions. It covers a wide range of atmospheric conditions. Kruse²⁷ developed ATCOR now algorithm, which is model-based ATCOR software using MODTRAN-4 code to reduce atmospheric and topographic effects in the data.

Fast line-of-sight atmospheric analysis of spectral hypercubes was developed by Adler-Golden et al.³⁷ It retrieves the land surface reflectance without using ground measurements. Staenz et al.³⁸ developed imaging spectrometer data analysis system to process hyperspectral data by removing sensor and calibration artifacts. It converts at sensor radiance to surface reflectance and analyses the hyperspectral data. Qu et al.³⁹ developed high-accuracy atmospheric correction for hyperspectral data. This model corrects the water vapor and other gases (such as ${\mathrm{CO}}_2$ and methane).

4.2.

Dimensionality Processing

Once corrected, the hyperspectral data still have redundant spectral information, which need to be processed. To reduce the computational costs and processing time without losing the useful information, the data should be dimensionally diminished. Dimension reduction techniques can be categorized into different groups such as unsupervised and supervised, linear and non-linear approaches. Principal component analysis (PCA), minimum noise fraction (MNF), and independent component analysis (ICA) are the most popular unsupervised dimension reduction approaches. PCA is a linear widely used technique that searches to magnify the variance within the new and lower subspace.⁴⁰ MNF is a well-known denoising technique that transforms noisy original hyperspectral data cube to channel images with the increasing noise level.⁴¹ ICA is a statistical mechanism that transforms the data into maximally independent and non-Gaussian subparts.^42,43 This method uses virtual dimensionality to determine the number of independent components, which should be retained.²⁵ Supervised methods are performed to extract more useful information using the prior knowledge of training samples. Commonly used supervised dimension reduction algorithms include Fisher’s linear discriminant analysis⁴⁴ and discriminative locality alignment (DLA).⁴⁵ Local Fisher’s discriminant analysis⁴⁶ and semisupervised DLA⁴⁵ are some variants of those techniques.

In order to overcome some problems that arose from traditional dimension reduction methods, sparsity-based techniques such as sparse-graph-based discriminant analysis⁴⁷ and multiple-features-combining⁴⁸ method were developed to enhance the dimension reduction results. The incorporates the spectral, texture apart from traditional methods; new methods based on machine learning communities have been developed for processing hyperspectral images. A number of metric learning algorithms such as relevant component analysis,⁴⁹ neighborhood component analysis,⁵⁰ information-theoretic metric learning,⁵¹ and ensemble discriminative local metric learning⁵² solve the problems in hyperspectral analysis. In addition, different manifold learning-based dimension reduction methods are proposed to display the non-linear structure in hyperspectral images including locally linear embedding,⁵³ Laplacian eigenmap,⁵⁴ isometric mapping,⁵⁵ neighborhood preserving embedding (NPE).⁵⁶ Hypergraph learning methods have also been introduced to explore the multiple adjacency relationship in hyperspectral data and discover the complex geometric structure between hyperspectral images.⁵⁷ Discriminant hyper-Laplacian projection,⁵⁷ semisupervise hypergraph embedding,⁵⁸ local pixel NPE,⁵⁹ and spatial–spectral regularized sparse hypergraph embedding⁵⁷ are some types of graph embedding methods for dimensionality reduction of hyperspectral images.

4.3.

Endmember Extraction Methods

Due to enhanced spectral resolution of HSI, extraction of hyperspectral endmember is a fundamental step in the hyperspectral data processing. An endmember is a pure and diagnostic signature that can be used to specify a spectral class. It should be noticed that each endmember does not need to be a pure pixel.⁶⁰ There are several endmember extraction algorithms that are broadly grouped into two main classes including convex geometric-based and statistical-based categories.¹² In widely used linear unmixing model, the spectra of each pixel vector are regarded to be a linear combination of endmembers weighted by their corresponding abundances.^61,62 There are two main groups of endmember extraction methods within the linear spectral unmixing algorithms as pure-pixel and non-pure-pixel methods. In these methods, the hyperspectral data are considered as simplex, and its endmembers are constituting its vertices. The pure-pixel approach assumes that there is at least one endmember in the image scene and applies simplex growing algorithms or maximum volume transform to find the endmembers within the hyperspectral data cloud.^63,64 The popular algorithms involve pixel purity index (PPI), and N-finder algorithm (N-FINDR), vertex component analysis, automatic target generation process, and convex cone analysis are typical pure-pixel or simplex growing methods.^47,48 In order to overcome the issues in pure-pixel endmember extraction, non-pure methods have been developed that use the shrinking simplex algorithms or minimum volume transform techniques.⁶⁴ Optical real-time adaptive spectral identification system (ORASIS), minimum volume constrained non-negative matrix factorization (MVC-NMF), minimum volume enclosing simplex (MVES), minimum volume simplex analysis (MVSA), and abundance separation and smoothness constrained non-negative matrix factorization are developed as non-pure-pixel extraction algorithms.⁶⁴ The ORASIS as the linear mixture model, developed by the Naval Research Laboratory, consists of series of stepwise algorithms including prescreener, basis selection, and endmember selection to unmix each pixel of dataset to different endmembers.^65,66 This method considers all the data pixel inside a simplex with endmember vertices. According to Chang,⁶⁶ the functionality of this method declined in hyperspectral data with low signal-to-noise ratio (SNR). MVC-NMF has been developed by Miao and Qi⁶⁷ as a non-pure pixel-based algorithm, which is convex geometry-based algorithm. This method tries to fit simplex while its volume is minimal and encompass the data cloud.⁶⁷ Nouri et al.⁶⁸ used particle swarm optimization to improve MVC-NMF method for mineralogical unmixing of hyperspectral data with high SNR.⁶⁸ MVSA is another linear mixture model for endmember extraction that was developed by Li and Bioucas-Dias.⁶⁹ This algorithm tries to fit a minimum volume simplex to the hyperspectral data, which contains the abundance fractions to belong to the probability simplex.⁷⁰ In this method, it is assumed that pure pixel may not exist in the hyperspectral data, which can address a common situation in hyperspectral data with highly mixed pixels.⁷⁰ Motivated by Craig’s belief, Chan et al.⁷¹ developed linear-based model called MVES algorithm in which vertices of the simplex enclose all the observed pixels. In this method, the simplex should estimate all the endmember signatures with high fidelity.⁷² Non-linear geometrical unmixing methods have been used less than linear techniques. Broadwater⁷³ introduced non-linear kernel methods to solve the unmixing problem in high-dimensional hyperspectral data. Heylen⁷⁴ introduced geodesic simplex volume maximation methods under non-linear mixing assumptions to extract hyperspectral endmembers. The statistical-based endmember extraction methods use statistical representations.⁶³ Based on parametric, non-parametric, and spatial statistic representations, different statistical-based unmixing methods have been developed. Stochastic mixing model, which assumes Gaussian distribution for each endmember, is an example of statistic-based algorithms that used parametric representations.⁷⁵ ICA is proposed by non-parametric statistical unmixing approaches. ICA-based abundance quantification algorithm and combination of ICA and independent factor analysis are some kinds of non-parametric statistical endmember extraction approach.⁶³ Automated morphological endmember extraction, iterative error analysis, and spatial–spectral endmember extraction (SSEE) are statistical unmixing methods that used spatial statistics to improve the endmember selection.^12,63

6.1.

Lithological Mapping

High-spectral-resolution properties of the hyperspectral remote sensing make it as an important potential to accurately recognize and map the earth surface materials.^29,33 Hundreds of contiguous spectral bands of hyperspectral images can be used to extract pixel spectrum for mineral and rock discrimination.³³ Performing mineral and lithological mapping from hyperspectral images is generally due to comparison of unknown spectrum to a reference spectrum.⁸² Spectral absorption characteristics of different minerals are mostly related to vibrations of ${\mathrm{Fe}}^{2+}$, ${\mathrm{Fe}}^{3+}$, Al-OH, Mg-OH, OH-, and ${\mathrm{CO}}_3$ in the VNIR and SWIR. Minerals with ${\mathrm{Fe}}^{2+}$ and ${\mathrm{Fe}}^{3+}$ have an absorption peak at 1.03 and $0.64\mu \mathrm{m}$, respectively. Al-OH-bearing minerals have a significant absorption peak at 2.15 to $2.22\mu \mathrm{m}$, whereas the most important absorption peak position of Mg-OH minerals occur at 2.30 to $2.39\mu \mathrm{m}$. ${\mathrm{CO}}_3$ group minerals show a distinct absorption peak at $2.35\mu \mathrm{m}$.^{14,16,18–20} VNIR-SWIR region is mostly used for alteration minerals mapping including Fe-OH, Mg-OH, Al-OH, and ${\mathrm{CO}}_3$-bearing minerals. However, these spectral features can be applicable for the Mg, Fe, and Al silicate minerals such as olivine, pyroxene, mica, and amphibole. Mineral spectral features of silicate minerals mostly attributed to vibration of Si–O bonds in TIR region.^14,16,19

Kumar et al.⁸⁵ proposed an automated lithological mapping approach on AVIRIS-NG hyperspectral data from gold-bearing granite-greenstone belt of the Hutti area (India). In that approach, they employed spectral enhancement techniques such as PCA and ICA and different machine learning algorithms (MLAs) to get an accurate lithologic map (Fig. 4). However, they utilized conventional geological map and spectral enhancement products derived from ASTER data to generate a high-resolution reference lithology map. Among different MLAs, they found SVM using joint mutual information maximization (JMIM)-based optimum bands to have a better performance with higher accuracy.

Fig. 4

(a) Reference lithology map derived from spectral enhancement products using ASTER and (b) lithological classification map generated from SVM using JMIM-based optimum bands of AVIRIS-NG hyperspectral data from gold-bearing granite-greenstone belt of the Hutti area (India).⁸⁵

Salehi et al.⁸⁶ investigated the performance of HyMAP hyperspectral images and Sentinel-2, ASTER and Landsat-8 OLI spaceborne multispectral data for lithological mapping of mafic–ultramafic rocks in lichen-covered area in south west of Greenland. They performed EnMAP geological mapper (EnGeoMAP) and iterative spectral mixture analysis (ISMA) algorithms. They used structural similarity index measure to compare the output of classification results of airborne data with available geological map and spaceborne data with reference HyMAP data. According to their results, the HyMAP and spaceborne multispectral data can provide geological maps comparable to geologic maps over the less accessible arctic regions. Three different ultramafic rocks (dunite, peridotite, and pyroxenite) and one mafic rock (gabbro) has been mapped based on VNIR and SWIR spectral analyzing.⁸⁶ They found EnGeoMAP algorithm to have a better performance for dunitic rocks and ISMA method for peridotite and pyroxenite units.

Harris et al.⁸⁷ performed several methods on airborne PROBE hyperspectral data from the test site in southern Baffin Island, Canada to produce lithological–compositional map. After masking out and eliminating water bodies, ice, snow, and vegetation areas from analysis, they applied MNF transformation to reduce the large dataset to a fewer number of components that have the main information. They performed matched filtering (MF) to extract the user defined endmembers due a partial unmixing method. The end members were selected by color discrimination in MNF composite images. They discriminated one lithological group (metatonalites) and three compositional units (psammite, quartzite, and monzogranites).

Zhang and Li⁸⁸ implemented an evolved SAM approach using EO-1 Hyperion hyperspectral images in lithological mapping in two different arid areas in China. They employed two different methods to improve the lithological mapping performance. In the first method, they used the mean of spectral derivatives and mean of original spectral data in SAM to improve class separability. In the second approach, multiple reference spectra were utilized to accommodate the spectral variability. As the result, they mapped Carboniferous-Permian volcanic-sedimentary rocks, Jurassic sedimentary rocks, granite, and quaternary sediments in Junggar area and Cambrian, Ordovician, Silurian, Devonian sedimentary rocks including dolomite, shale, marl, siltstone, quartzose sandstone, and sandstone, and quaternary sediments in Kalpin area.

Dadon et al.⁸⁹ used EO-1 Hyperion hyperspectral data for stratigraphic and lithologic mapping. They performed SAM supervised classification after some preprocessing and endmember extraction processes. They separated different plutonic (monzogranite, granite, and granodiorite), volcanic (green tuffs and basaltic lava), and sedimentary (limestone, sandstone, conglomerate, and dolomite shale) rocks form the Dana Geology National Park (south-west Jordan) after stratigraphic and lithologic mapping process.

Pal et al.⁷⁷ considered the VNIR and SWIR bands of Hyperion spaceborne hyperspectral data and ASTER and Landsat 8 multispectral images. They implemented four-step technique for lithological mapping of Udaipur area, Rajasthan, western India. They employed a hybrid classification method involving minimum distance, SAM, spectral information divergence, and SVM to optimize lithological mapping. Pal et al.⁷⁷ used this method to discriminate different sedimentary and metamorphic lithoclasses including quartzite, phyllite, graywacke, dolomite, mafic-metavolcanics, migmatite, graphitic metapelites, and quartzite-arkose-conglomerate in the Udaipur area.

6.2.

Mineral Mapping and Mineral Exploration

Mineral mapping and determination of surface composition for mineral exploration purposes is one of the most important applications of hyperspectral remote sensing. There are different ore deposits classifications based on different fundamental genetic processes: magmatic, hydrothermal (magmatic), sedimentary, metamorphic, and mechanical process. Different mineralization systems are dominantly associated with different alteration associations and mapping surface mineralogy can be considered as vectors to ore deposits.⁹⁰ A schematic overview of intrusion-related mineral deposits and associated alteration patterns have been shown in Fig. 5.⁹¹ Figure 6 shows some diagnostic different alteration assemblages along with some of the main ore deposits. Most of the alteration minerals have distinct spectral features in VNIR-SWIR and TIR region.⁹²

Fig. 5

Generalized alteration-mineralization zoning pattern for PCDs.⁹¹

Fig. 6

Common alteration minerals in different ore deposits.⁹²

Magmatic hydrothermal deposits include several important ore deposit types such as porphyry, epithermal, skarns, intrusion-related, volcanogenic massive sulfide (VMS), iron oxide copper gold (IOCG) uranium rare earth element (REE), alkaline complexes, and spreading centers sea-floor smokers.⁹³ Hydrothermal intrusion-related ore deposits directly formed within or with a distance with igneous intrusions. There are spatial alteration patterns and mineral assemblages in these deposits. However, they have some genetic relationship with the fractures and faults in the host rocks. They include alteration patterns such as feldspathic, sericitic, silicic, greisen, calc-silicate, and/or advanced argillic assemblages.

Jintanzi gold province (China) is one of the intrusion-related gold deposits. This deposit is mostly related to quartz veins within the granites and associated with different order faults. Using spaceborne Tiangong-1 HSI and airborne short-wave infrared airborne spectrographic imager (SASI) data, Liu et al.⁹⁴ tried to produce alteration mineral maps via SAM algorithm. They considered JPL and endmembers extracted by SSEE as the reference spectra. Using either set of the reference spectra, they detected alteration muscovite, kaolinite, chlorite, epidote, calcite, and dolomite as the hydrothermal minerals. Figure 7 shows the mineral maps derived from image endmembers from SASI and Tiangong-1 hyperspectral data. According to their results and previous studies, the distribution of the alteration minerals has been distinctly controlled by the gold-bearing quartz veins and faults.

Fig. 7

Mineral maps derived from the use of image endmembers applied to (a) SASI data and (b) Tiangong-1 HSI data from Jintanzi gold province (China).⁹⁴

Epithermal Au–Ag deposits have been considered as important gold and silver sources in the world, which commonly distributes along the convergent plate margins.⁹⁵ Extensive geological studies have been done on different epithermal gold-silver deposits. However, several researchers have considered the application of remote sensing data in order to investigate and model the hydrothermal alteration mineral mapping in these deposits.^96–99 Silisic, potassic, argillic, propylitic, and zeolitic are the most common hydrothermal alteration minerals in epithermal deposits. Rodalquitar deposit (southern Spain) is one of the most popular epithermal gold and silver deposits that have been extensively studied in geochemical aspect to understand the mineralization process. Also numerous geological remote sensing studies have been conducted to generate mineralogic map of the area. van der Meer et al.⁹⁷ performed wavelength mapper and QuanTools to derive absorption feature from HyMAP hyperspectral images of Rodalquiar epithermal deposit. They used absorption feature position as a key to find the mineral chemistry variations such as Al–Mg versus OH, which can be interpreted in terms of fluid compositions and temperature. Figure 8 displays the alteration mineral maps of Rodalquiar epithermal system constructed via Quantools and wavelength mapper.⁹⁷

Fig. 8

From top, mineral map constructed with Quantools and mineral map constructed with the wavelength mapper using HyMAP data of Rodalquilar epithermal system (southern Spanish Gabo de Gata volcanic area).⁹⁷

PCDs constitute the largest source of copper and a major source of molybdenum, gold, and silver in the world.¹⁰⁰ These deposits commonly have diagnostic broad alteration patterns with distinctive minerals. Typical alteration types in PCDs include potassic alteration in the core, which is surrounded by sericitic, argillic, and propylitic zones.⁹¹ Subtle variations in the alteration minerals could be considered as the high economic potential in PCDs. For example, zonation of chemical variation of white mica in phyllite alteration is a proxy of chemistry of ore forming fluid.^16,101 Figure 9 shows the variation of white mica chemistry, which has different SWIR wavelength characteristics in typical cross section of PCD.¹⁶ Although several spaceborne and airborne hyperspectral data have been used for porphyry Cu exploration, there are a few literatures that investigated them.^102–104 Sarcheshmeh PCD (southern Iran) located in the Orumieh-Dokhtar magmatic arc is one of the best sites for studying alteration related to PCD. Using Hyperion data, Zadeh et al.¹⁰⁴ (2014) identified characteristic alteration minerals including biotite, muscovite, illite, kaolinite, goethite, hematite, jarosite, pyrophyllite, and chlorite via subpixel mixture tuned matched filtering (MTMF) method (Fig. 10). Discriminating minerals such as biotite and iron oxide is one of the most important evidence for PCD exploration. Bishop et al.¹⁰² used SAM and MTMF techniques to discriminate and map the alteration mineral assemblages from Hyperion data in the Pulang, PCD (Yunnan, China). They determined argillic alteration, iron-oxide-, and sulphate-bearing minerals in the target deposit.

Fig. 9

Typical cross section of PCD, showing generalized Al(OH) $2.200\mu \mathrm{m}$ absorption of different white mica in phyllic alteration.^16,101

Fig. 10

Final classification image map of alteration minerals at Sarcheshmeh deposit (southern Iran) derived from MTMF algorithm. Bio, Mu, Il, Kao, Goe, Hem, Ja, Pyr, and Ch indicate biotite, muscovite, illite, kaolinite, goethite, hematite, jarosite, pyrophyllite, and chlorite, respectively.¹⁰⁴

IOCG (±U-REE) deposits are a group of magmatic-hydrothermal deposits, which are structurally controlled and commonly associated with different alteration zones. The alteration zones of host rocks begin with sodic alteration, which changed to sodic-calcic and potassic toward the main Fe-oxide mineralization.¹⁰⁵ Deposits with more developed potassic alteration commonly have intensive magnetite mineralization. Surficial rocks often show sericitic and silisic alteration. Investigation of Corriveau et al.¹⁰⁶ show that each alteration type in this deposit type has a unique spectral signature. Potassic and sodic alterations are mostly distinguished by the location of hydroxyl feature (OH-) near $2.200\mu \mathrm{m}$. According to their results, airborne or spaceborne hyperspectral data have a potential capability to map exposed alteration maps around the IOCG deposits. Tappert et al.¹⁰⁷ used the infrared reflectance spectra to discriminate the high-Al and low-Al phengite as a potassic mineral identified at the Olympic Dam IOCG deposit (south Australia). The phengite minerals in the heavily sericitized ore-bearing rocks have lower Si content, higher Al content, and lower Mg content that the phengites formed in the weakly sericitized altered host rocks. High-Al phengite has a spectral absorption feature at $2.206\mu \mathrm{m}$, whereas low-Al type is recognized by the absorption feature at $2.213\mu \mathrm{m}$. Therefore, spectral absorption feature can be representative of the sericitic alteration degree due hydrothermal alteration related ore forming process. Laukamp et al.¹⁰⁸ used HyMap airborne hyperspectral data to derive high-resolution alteration mineral map of Mount Isa Inlier IOCG deposit (Australia). They employed hyperspectral data as a tool for hydrothermal alteration detection and fluid pathway identification related to Cu–Au IOCG deposits.

VMS deposits are kind of metal sulfide deposit, mostly copper and zinc, which formed near the sea floor. Similar to other magmatic hydrothermal ore deposits, they have alteration patterns, which have been formed via circulation of ore forming fluid through the host rocks.¹⁰⁹ The hydrothermal alteration patterns include potassic, advanced argillic, argillic, sericitic, chloritic, and carbonate propylitic from the inner mineralized core to the peripheral area.¹¹⁰

There are different hyperspectral remote sensing studies on the several important worldwide VMS deposits to characterize the hydrothermal systems. van Ruitenbeek et al.¹¹¹ used HyMAP hyperspectral data to evaluate white mica distribution pattern around Panorama VMS deposit (western Australia). They used the advantages of different absorption features at 2.185 to $2.235\mu \mathrm{m}$ of Al-rich and Al-depleted white mica to map the alteration patterns and alteration related fluid. Using AVIRIS-NG hyperspectral data, Samani et al.¹¹² discriminated alteration minerals including calcite, muscovite, and chlorite in the Ambaji-Deri area (northwestern India). They performed different image processing approaches such as MNF, PPI, $N$-dimensional visualization, and SAM classification. They matched the endmember spectra with the USGS spectral library. Figure 11 shows the observed spatial distribution of different hydrothermal minerals in the Ambaji-Deri region. Laakso et al.¹¹³ used different scale hyperspectral data such airborne, laboratory, and field methods for VMS mineral exploration in Canadian Arctic. They took the advantages of the spectral Al-OH and Fe-OH absorption features in the SWIR wavelength region of biotite and chlorite, which reflect the chemical compositional changes with the distance to the mineral deposit. Fe-OH has an absorption feature at $2.254\mu \mathrm{m}$ in proximal areas to the ore deposit while it changes to $2.251\mu \mathrm{m}$ in the distal areas. In addition, proximal areas have the Al-OH absorption feature at $2.203\mu \mathrm{m}$ in contrast with the absorption feature at $2.201\mu \mathrm{m}$ in distal areas.

Fig. 11

Spatial distribution of the carbonate minerals at northeast of Khokhar Bili and alteration minerals around Jharivav and Amblimal.¹¹²

Skarn deposit is another type of magmatic hydrothermal deposits, which have been formed due to metamorphism and hydrothermal alteration of carbonate-bearing rocks. They are considered as the potential source of different ore minerals such as copper, tungsten, iron, molybdenum, zinc, and gold. There is a zonation pattern in skarn deposits from garnet in the proximal to pyroxene in distal areas and pyroxenoid such as wollastonite in the skarn and marble contact.¹¹⁴ However, individual skarns may show systematic compositional changes in these zonation pattern. Thus detail mineralogy of skarns and their zonation can be useful in ore exploration processes.

There are several hyperspectral remote sensing studies to detect alteration zones around skarn deposits as a proxy to skarn type ore deposit exploration. Xu et al.¹¹⁵ used spaceborne Hyperion and close-range field hyperspectral data from Dapingliang skarn copper deposit (China) to identify the mineral zones around the skarn deposit. They distinguished pixels related to skarn using SAM from spectra overall shape and absorption bands spectral shape. Field data were applied to directly identify alteration mineral instead of just surface materials, which probably do not have any direct relationship with the ore mineralization. Tian et al.¹¹⁶ used the SWIR spectral analyses to detect alteration minerals including sericite group minerals (montmorillonite, illite, and muscovite), kaolinite, and carbonate minerals (calcite, ankerite, and dolomite), with minor chlorite, halloysite, and dickite around Jiguanzui Cu-Au (Eastern China) deposits. According to their results, Fe-OH absorption feature of chlorite (2.241 to $2.263\mu \mathrm{m}$) becomes shorter to the distal areas. Also minimum Al-OH absorption feature of $2.209\mu \mathrm{m}$ can be a useful vectoring tool to the ore deposit.

6.3.

Environmental Geology

Environmental geology is one of the geological applications to solve environmental problems derived from interaction of human with geologic environments like mineral resources. For instance, the mining projects and abandoned mines are one of the main agents contributing to cause serious ecological pollution. Vegetation, soil, water resources, and human can be threatened by wastes that were left over including lead, zinc, cadmium, and some toxic minerals.^117,118 These wastes with a widespread distribution have the high potential of pollution. As mentioned in the foregoing sections, hyperspectral remote sensing has a powerful potential to detect earth surface materials such as different minerals and elements. Therefore, it can be used to characterize and monitor the mineralogy of mine waste surfaces to predict the potential source of pollution including metal leaching and acidity. Since the application of hyperspectral remote sensing in geology, different studies have been done to monitor pollution from various mining areas.^{117,119–125}

One of the processes that have been caused pollution in the mining area is pyrite oxidation.^123,125 This process may produce acidic water, which has gradually crystallized different Fe-bearing secondary minerals in the way from the mining waste. These minerals include copiapite,⁶¹ jarosite $[(\mathrm{K},{\mathrm{H}}_3\mathrm{O},\mathrm{Na}){\mathrm{Fe}}_3{({\mathrm{SO}}_4)}_2{(\mathrm{OH})}_6]$, schwertmannite $[{\mathrm{Fe}}_8{\mathrm{O}}_8{(\mathrm{OH})}_6{\mathrm{SO}}_4]$, ferrihydrite $[{\mathrm{Fe}}_5{\mathrm{HO}}_8,{4\mathrm{H}}_2\mathrm{O}]$, goethite $[\alpha \text{-}\mathrm{FeO}(\mathrm{OH})]$, and hematite $[\alpha \text{-}{\mathrm{Fe}}_2{\mathrm{O}}_3]$.¹²³ These secondary minerals are spectrally identifiable due to their diagnostic spectral reflectance signatures. Swayze et al.¹²³ used AVIRIS spectral data to evaluate mine waste in the California Gulch Superfund Site. They used 0.4 to $2.5\mu \mathrm{m}$ spectral range of hyperspectral data to detect secondary Fe-bearing minerals and create a mineral map to highlight the areas, which have the potential for acidic drainage and predict the leaching and acid generation sites at Leadville mining district (Fig. 12). They performed Tetracorder algorithm to produce the Fe-bearing minerals map. Davies and Calvin¹²⁶ have used airborne hyperspectral data to characterize acidic mine waste at the Leviathan mine Superfund site in the eastern Sierra Nevada. They used SAM and MF algorithms to map minerals related to acid mine drainage.

Fig. 12

Spectral traverse and AVIRIS mineral maps overlaid on a high-spatial-resolution aerial photograph of the Venir mine-waste pile. Mineral maps show the spectrally dominant Fe-bearing secondary minerals.¹²³

Mielke et al.¹²⁵ applied spaceborne hyperspectral and multispectral data to monitor mine waste mineralogy in south Africa. They derived iron feature depth (IFD) as a new index to detect the extend of waste minerals distribution. This index is justified with primary and secondary iron-bearing minerals. They performed material identification and characterization algorithm on the EO-1 Hyperion data for mineral identification. According to their results, integration of hyperspectral (EnMAP) and multispectral (Sentinel-2) for mineral mapping and IFD monitoring could be a promising proxy of mine waste.