Lithological and Hydrothermal Alteration Mapping Using Terra ASTER and Landsat-8 OLI Multispectral Data in the North-Eastern Border of Kerdous Inlier, Western Anti-Atlasic Belt, Morocco

The principal components analysis (PCA) is an effective statistical method used to analyze the correlation in multispectral and hyperspectral datasets (e.g., Tripathi & Govil, 2019). Among its functionalities, the reduction of the dimensionality of input data and the separation of the noise components by suppressing irradiance effects dominate all spectral bands. Thus, PCA is an image enhancement technique that transforms correlated spectral bands into a set of uncorrelated variables called principal components or PC bands (Rajan Girija & Mayappan, 2019). In this study, we applied the PCA transform to both of ASTER and OLI VNIR/SWIR bands. Tables 3 and 4 display the eigenvalues percentages and the eigenvector loadings matrix derived from ASTER and OLI data, respectively.

The Minimum Noise Fraction (MNF) transformation is a statistical method similar to the PCA method. It is used for linearly transforming the reflectance input image to reduce spectrally its dimensionality and noise (Green et al, 1988). It determines the input data noise statistics in terms of Eigen-values, and these values were used to select the lower MNF bands that contain the most spectral information. The first MNF new bands are usually used as FCC to enhance lithology due to their high eigenvalues loadings.

The maximum likelihood (ML) classification algorithm is a parametric classification applied for supervised classification. It makes the assumption that the data for each class and band will be distributed normally. It determines the probability that a pixel belongs to a particular class (Fal et al, 2019). All pixels are classified, and each pixel is assigned to the class with the highest probability if the probability threshold is not selected.

The Support Vector Machine (SVM) classification algorithm is one of the broadly applied supervised classification algorithms based on statistical learning theory (De Boissieu et al, 2018). The SVM machine learning logic is based on a maximized hyperplane to classify the image according to the given regions of interest. The selection of the hyperplane is the one that represents the largest separation between two classes to minimize the spectral mixture (Ge et al, 2018). Commonly, the use of the SVM method in the geological and mineralogical studies provides good mapping accuracies (Bachri et al, 2019; Ge et al, 2018; Soulaimani et al, 2014a).

The spectral indices are computed by dividing the multiplied or added reflectance value of pixel bands by the corresponding values of pixels in another band (or another physical quantity, e.g., spectral radiance) displaying the result as a grayscale image (Corrie et al, 2011; Ninomiya, 2010). The band ratios are computed based on distinctive spectral band characteristics, resulting in an output image that identifies the targeted mineralogical composition in the image. The VNIR and SWIR spectral characteristics of various minerals associated to hydrothermal processes were analyzed. Ninomiya (2002) used several spectral bands for mapping quartz-rich, carbonate, and silicate rocks using TIR data from the ASTER sensor. The applied spectral indices were developed on the basis of SWIR data by Ninomiya (2004), including hydroxyl-bearing minerals, calcite, and alunite.

The CEM classification method is a target signature approach, which limits the desired target signature with a specific gain while minimizing the effects caused by other unknown signatures (Chang et al, 2000; Farrand & Harsanyi, 1997; Liu et al, 1999). This classification method has been applied to the ASTER image to enhance the target signatures. The ENVI 5.3 software was used to apply this method in the study area. The image generated by the CEM method is in gray levels, where each pixel gray level value in the image reflects the amount of the targeted mineral fraction abundance that was identified. The mineral detection is then performed based on the resulting grayscale image, and thematic classification is performed by detecting the desired mineral in a separate image (Pour et al, 2019b).

The PCA transformation is applied to the nine bands of ASTER and the seven bands of Landsat-8 OLI (Tables 3 and 4). The first three principal components (PC1, PC2, and PC3) contain more than 96.05% and 99.49% of all the inner information in ASTER and OLI images, respectively. Figure 3 shows the eigenvector loading trend of the first five ASTER and OLI PC images for better visualization of the data in Tables 3 and 4.

Figure 3.

Eigenvector loading trend graphical representation extracted from the first five (a) ASTER and (b) OLI PC bands from data in Tables 3 and 4

The PC1 image usually enhances the main information in the image, such as the general geological contrast, topography, and structural features. The ASTER PC1 eigenvector loadings show negative contributions for all bands (Figure 3a) with a little decrease from the first band (-0.20736) and the ninth band (-0.37795). As demonstrated in the study area on the extracted PC1 image (Figure 4a), the Cambrian formations (Sa, Ds, and Di) are highlighted by dark pixels, while Neoproterozoic units (B, Vc, Re, Gb, Qz, Rs, and Ra) are highlighted by relatively bright pixels. PC2 highlights the Ds unit with dark pixels following the high negative contribution in the bands 8 and 9 of ASTER (-0.26804) and (-0.34255), respectively (Figure 3). This is due to CO3 absorption feature in 2.33 μm corresponding to ASTER band 8 (Hunt & Ashley, 1979). Iron-bearing minerals exhibit distinctive absorption features at ASTER bands 1, 2, and 3 and reflectance in band 4 (Pour & Hashim, 2012). Gabbro intrusions are discriminated by bright pixels in these two PCs (Figure 4c and d) due to their richness in basic ferrous minerals.

Figure 4.

PC images of ASTER: (a) PC1; (b) PC2; (c) PC3; (d) PC4; (e) PC7

The PC1 eigenvectors demonstrated a negative contribution in all the seven bands with a maximum value of (-0.09192) in band 1 and a minimum value of (-0.60072) in band 6 (Figure 3b). PC2 shows vegetation in white pixels due to the high positive contribution in the band 5 (0.842454), and the units of Ds, Sa, and Di are enhanced by dark pixels due to the high negative contribution in band 7 (-0.45639) (Figure 3b).

PC3 highlights the units of Vc and Ra with white pixels due to the high positive contribution in band 4 (0.690656), and Gb lithological unit was identified with a dark tone due to the negative contribution (-0.46946) in band 6 (Figure 3b) (Fe-bearing minerals reflectance band). Thereafter, PC4 (Figure 5d) shows the Qs and Ds lithological units with bright pixels due to the high positive contribution in band 7 (0.717502) and the Gb lithological unit with black pixels due to the high negative contribution in band 6 (-58625) (Figure 3b).

Figure 5.

PCA and MNF extracted images of: (a) OLI PC1; (b) OLI PC2; (c) OLI PC3; (d) OLI PC4; (e) OLI PC5; (f) OLI PC6; (g) ASTER MNF 1; (h) ASTER MNF 2; (i) OLI MNF1

Besides, PC5 displays the Ra, B, and Sa in bright pixels due to the high positive contribution in band 2 (0.523961) (Figure 3b). PC6 (Figure 5f) shows a good capability to differentiate Sa formation in white tone due to the high positive contribution in band 3 (0.815732).

The FCC of the selected bands is used for discrimination of the different lithologies. The resulting MNF images of ASTER MNF1 and MNF2 as well as MNF1 of OLI are presented in Figure 5g, h, and i. Furthermore, the FCC is used for both PCA and MNF images (Figure 6). The best combination for discriminating lithological units of the study area using ASTER data is MNF1, PC4, and PC2 as RGB images (Figure 6a). On the other hand, the OLI (MNF1, PC5, and PC3) as RGB shows a high capability to enhance lithology visualization in the study area (Figure 6b).

Figure 6.

(a) FCC ASTER MNF1, PC4, and PC2; (b) FCC OLI MNF1, PC5 and PC3

The analysis of the results from the OLI image shows that both ML and SVM classification methods demonstrated good discrimination of the study area lithology (Figure 7). According to the resulting lithologic maps, the five lithological units, brown dolomite with stromatolite (Ds), brown dolomite with marl intercalations (Di), rhyolitic-andesitic volcanic and ignimbrite with epiclastic deposits (Ra), siltstone and argillite (Sa), and quartzite sandstone (Qs), are well mapped in both classification methods. These units show high values of overall accuracy and Kappa coefficient.

Figure 7.

Lithological classification maps: (a) MLC classification using ASTER image; (b) SVM classification using ASTER image; (c) MLC classification using OLI image; (d) SVM classification using OLI image

Figure 8.

User and producer accuracies: (a) ML ASTER; (b) SVM ASTER; (c) ML OLI; (d) SVM OLI

Based on the confusion matrix results, the ML classification method has an overall accuracy (of 91.74%) and a Kappa coefficient (of 0.90) higher than the SVM classification method with an overall accuracy of 88.82% and a Kappa coefficient of about 0.86 (Figure 9).

Figure 9.

Overall accuracy and Kappa coefficient of MLC and SVM classifications using ASTER and OLI: (a) overall accuracy; (b) Kappa coefficient

In addition, several lithological units were well enhanced by both classification methods, and the results are presented in Figure 7. The ML classification method has well enhanced two lithological units: brown dolomite with stromatolite (Ds) and brown dolomite with marl intercalations (Di) (Figure 7a and 7c), as observed in the ML classification results compared to the SVM classification. The results obtained from both ML and SVM classification methods using the ASTER data did not show satisfactory lithological discrimination compared to that from the OLI data.

On the other hand, siltstone and argillite (Sa) and volcanic conglomerates (Vc) are well mapped by the SVM classification (Figure 7b and 7d). According to the resulting accuracies (Figures 8 and 9) of the ASTER image, the lithological units of rhyolitic volcanic and ignimbrite with epiclastic deposits (Re) are better classified using the ML method than the SVM method. The ML method showed a spectral mixture between the lithologic units of rhyolitic volcanic and ignimbrite with sandstones and conglomerates (Rs) and quartzite sandstones (Qs).

The abundance images of hydrothermal alteration mineral occurrences were calculated. For the mineralogy of Ninomiya (2004), the pseudocolor ramp image of the calcite mineral discriminates Ca-bearing alteration minerals as bluish pixels (Figure 10a). In addition, the kaolinite-rich zones characterize the argillic and phyllic alteration zones which are highlighted by OH(a) and OH(b) that moreover enhance the pyrophyllite and montmorillonite (Figure 10b and c). Then, the index images were thresholded to extract all pixels with reflectance values well above the threshold value to be combined as shown in Figure 11.

Figure 10.

Resulted images of the Ninomiya spectral indices: (a) Calcite; (b) OH(a); (c) OH(b); (d) Alunite

Figure 11.

Images of the Ninomiya spectral indices superimposed on the true color composite of ASTER image

The CEM classification allows the enhancement of alteration minerals separately to identify each specific mineral (Pour et al, 2019b). The Precambrian basement-hosted cupriferous mineralization can be presented in the vein structure, where the porphyry alteration systems are common (Azmi et al, 2014). Azmi et al. (2014) proposed a genetic model for the Hercynian-remobilized porphyry-type mineralizations. Following this model, nine hydrothermal alteration minerals have been selected for classification to highlight argillic, phyllic, and propylitic alteration features (Figure 12). The spectral signatures of the hydrothermal alteration minerals were extracted from the USGS spectral library. This spectral library includes spectral signatures of minerals with a spectral resolution of 3 nm (Kokaly et al, 2017). The spectral signatures are resampled to the ASTER bands (Figure 12). The resulting pseudocolor ramp images of the CEM classification are shown in Figure 13. The results show a good correlation with those of the Ninomiya spectral indices presented in Figures 10 and 11. This confirms that the CEM classification is an efficient method of classification for mineralogical aspects (Moradpour et al, 2020; Pour et al, 2019b). The CEM images of the argillic hydrothermal alteration-related minerals (e.g., alunite, kaolinite, montmorillonite, and pyrophyllite) in addition to the images of the CEM classification of phyllic hydrothermal alteration-related minerals (e.g., illite, muscovite) show a very good correlation with the images of the OH(a) and OH(b) spectral indices, respectively (Figure 13). The Di (dolomite with marl) lithological unit shows an abundance of calcite, illite, kaolinite, montmorillonite, muscovite, and pyrophyllite. Sa (siltstone and argillite) unit is marked by chlorite, calcite, muscovite, and montmorillonite. On the other hand, the dolerites intercutting the quartzite sandstone unit are associated essentially with alunite, illite, kaolinite, muscovite, and montmorillonite (argillic/phyllic alteration).

Figure 12.

Laboratory reflectance spectra of hydrothermal alteration minerals from USGS spectral library (left) and spectra resampled to ASTER bands (right)

Figure 13.

ASTER resulting images of a high abundance of hydrothermal alteration minerals using CEM: (a) alunite; (b) chlorite; (c) calcite; (d) epidote; (e), illite; (f) kaolinite; (g), montmorillonite; (h) muscovite; (i) pyrophyllite

This region highlights a high density of tectonic lineaments extracted from HH and HV polarization images of ALOSPALSAR (Jellouli et al, 2021), which suggest a structural control of the hydrothermal alteration process. Furthermore, the eastern Kerdous (Amlen Valley shear zone) presents several hydrothermally altered regions that show high similarity with our findings. Hyperspectral data with high-spatial resolution (5 m) allow the detection of several zones of Al-OH, Fe3+-Fe2+, and Mg-Fe-OH/CO3 hydrothermal alteration minerals. Some detected zones were depicted as a potential sites for copper and manganese mineralizations after applying fuzzy logic modeling for 20 layers derived from hyperspectral dataset (Hajaj et al, 2023b).

In addition, kaolinite developed in argillic alteration zones correlates well with the CEM images and spectral indices (Figure 13). To verify the accuracy of the CEM classification, the USGS library spectra were used for validation. The analysis using spectral angle mapper (SAM) allowed us to identify the matching scores. Table 5 (Hitlist) shows the matching score of the selected image spectra of the highly altered regions. The computed matching scores are ranging from 0.73 to 0.91.