Perspective Chapter: Optical Remote Sensing for Fluorapatite Content Estimation

Abstract

Remote sensing techniques are an interesting alternative to traditional methods for the rapid abundance prediction and mapping of phosphate mineralization surface states. In this context, a methodological approach based on hyperspectral spectroscopy and X-ray diffraction (XRD) method is proposed for the phosphate surface abundance prediction and exploration in a specific geological context in Tunisia. In this study, partial least square regression (PLSR) method was conducted on hyper-spectral visible-near infrared (VNIR) and short-wave infrared (SWIR) field reflectance spectra of the collected samples and XRD analysis results for phosphate content prediction. Results revealed that carbonate absorption features in SWIR region could be considered for an accurate estimation of phosphate contents. The generated model has shown an interesting performance with an R2 of 0.64, an RMSE of 5.52, and an RPIQ of 2.15, using the training samples set. Moreover, X-ray diffraction (XRD) analysis results were used for the validation purposes. The using validation samples set revealed an R2 of 0.42, an RMSE of 10.29, and an RPIQ of 1.74. All performance coefficients have shown that the generated model can be applied successfully for the content prediction of phosphates. The present study revealed, thus, the contribution of the proposed methodological approach for phosphate exploration in the Chaketma mine site in the Centre West of Tunisia, which can be improved in the future.

1. Introduction

Phosphate is essential to all forms of life, which 90% of its consumption is in agriculture. There is no substitute for phosphate, and it is essential to improve crop yields. Otherwise, phosphate rock is considered as a potential new rare earth element (REE) resource according to their world commercial production that is estimated to be 250 million tons per year [1]. Specifically, apatite is included as one of the known host minerals of rare earth elements by exploration geologists and can provide a feasible resource of REE in future, which demand increases nearly 8% per year [2]. Actually, they are vital to green and emerging technologies [3, 4].

Tunisia is well known for its phosphate reserves, since 1887. Gafsa mining basin is the main region of sedimentary deposits exploited in the south [4]. Indeed, Eocene phosphate basins can be classified into three large sites: the Gafsa basin (Metlaoui, M’Dhilla, Moulares, and Redeyef), the Meknassi basin, and the northern basin (Chaketma and Sra Ouertane), which were recently explored as carbonate phosphates. The extracted raw phosphate was processed in Metlaoui, Moulares, Redeyef, Kef Eddour, and M’Dhilla production centers. Particularly, Chaketma phosphate mine presents a potential large-scale, world-class phosphate development asset in Tunisia. The bulk of the phosphate is located at the base of a massive limestone unit close to the top of a high segmented plateau, which is affected by a series of normal faults.

A first study was conducted in a Tunisian semiarid environment for phosphate mapping using both hyperspectral and multispectral remote sensing technologies as an interesting alternative to standard methods, which are routinely applied in mineralogical and geochemical studies [5]. Indeed, remote sensing tools have been well used to map rocks and minerals because of their detailed information, which can be provided periodically. Hyperspectral remote sensing tools at different scale provide higher spectral resolution data that can be used for mineral exploration and mapping [6, 7, 8, 9, 10, 11]. Several studies detected minerals basing on their spectral absorption features identified from hyperspectral reflectance data, taken by hyperspectral radiometer. Carbonates, clays, were thus studied and analyzed. Recently, near-infrared (NIR) data were used successfully for the prediction of rare earth elements in the largest uranium-phosphate deposit in Brazil. Three partial least squares regressions (PLSRs): full-spectrum partial least squares (PLS), interval partial least squares (iPLS), and successive projections algorithms for interval selection in partial least squares (iSPA-PLS) were conducted to calibrate the measured spectra to predict rare earth elements in topsoil [12]. PLSR is a statistical and a recent technique that generalizes and combines features from principal component analysis and multiple regression. It is particularly useful in the case of predicting a set of dependent variables from a large set of independent variables, called predictors [13].

Studies that pointing out the prediction of fluorapatite phosphate minerals using VNIR-short-wave infrared (SWIR) spectroscopy data and carrying out PLSR methods are missed. The present chapter shows the preliminaries results, derived from a PLSR-based study that was conducted on a particular semiarid Tunisian context aiming to predict fluorapatite phosphate mineral contents using field hyperspectral spectroscopy and X-ray diffraction (XRD) mineralogical analysis.

2. Chaketma mine site

The Chaketma mine, which is located in the Central West of Tunisia, at 200 km far from the southwest of Tunis, is characterized by its sedimentary phosphates. This facies, which is the most common, comes from three main petrographic constituents: granules (pseudo-oolites), nodules, and organic debris (coprolites).

Consisting of a succession of fractured plateaus, the phosphate deposits of the Chaketma site are bounded to the west by the Ghoualguia anticline by a major NNO-SSE normal fault (N160) (Figure 1) controlling the Rouhia hollow. Particularly, in the prospect of Gassaa Kebira, a massive bar of dolomitic limestone, recording a series of Ypresian age and decline of the formation of El Gueria, outcrops thicker than at Gassaa Sghira, in the north of the district, and Kef Oum Ezzine, in the south [14].

Figure 1.

Localisation of the Chaketma mining site as well as the six phosphate perspectives: Map of Tunisia (left) and combination of Aster red-green-blue (RGB) image draped over the digital elevation model (DEM) of the mine site (right).

Kef El Louz, Sidi Ali Ben Oum Ezzine, Gassaa Kebira, Gassat Ezarbat, Kef El Agueb, and Douar Ouled Hamouda are the six individual phosphate prospects, showing thick mineralized zones with more than 15 m at depth [15], which were determined in 2012 by the Celamin Holdings Company. Tunis Mining Service TMS company estimated, however, the mineralization in the six prospects of the Chaketma mine at about 176 Mt. at 19.5% of pentoxide phosphorus P2O5 (Figure 2) and 79 Mt. of phosphate at 30% P2O5 content. Recently, PhosCo Company has revealed that the deposit contains a resource of 148Mt at 20.6% P2O5, confirmed from drilling at only two of the project’s six prospects, precisely, 93 Mt. at 20.3% P2O5 in Gassaa Kebira and 55.5 Mt. at 21.2% P2O5 in Kef El Louz (Figure 3). According to PhosCo, the phosphate resource at Kef El Louz is large, shallow and features simple geology. Drilling results have produced consistent, wide, high-grade phosphate mineralization close to surface, with 50% of the prospect’s known surface mineralization yet to be drilled [16].

Figure 2.

Geological map of the Chaketma mining site (TMS Company, 2012).

Figure 3.

The six phosphate perspectives (PhosCo Company, 2020).

3. Methodological approach

The methodological approach (Figure 4) aiming to fluorapatite content prediction is based on both mineralogical and spectral analysis data. Indeed, the gathered phosphate samples from the top surface, after draying and quartering, were the subject of X-ray diffraction (XRD) analysis, using a PANalytical X’Pert PRO X-ray diffractometer, to identify and estimate the abundance of each mineral. Furthermore, all 25 surface samples in different point locations were measured under natural light with the ASD FieldSpec HiRes spectroradiometer (Analytical Spectral Devices, Boulder, Co.). This spectroradiometer, which is fitted with 10°field-of-view fiber optics, operated in the 350 nm to 2500 nm spectral regions with sampling intervals of 1 nm. The mean of the five surface VNIR-SWIR reflectance spectra scans of each sample was preprocessed before using for the partial least square linear regression (PLSR). The prediction content results were validated basing on their comparison with the measured ones by XRD.

Figure 4.

Flowchart detailing the methodological approach.

4. Results and discussions

4.1 Phosphate mineralogical composition

The interpretation of X-ray diffractograms of all selected samples revealed a high concentration of both calcite and dolomite minerals [5]. The carbonated geological context of the region can explain the presence of these high carbonate mineral abundances. Particularly, dolomite (CaMg (CO₃)₂) and calcite (CaCO₃) showed high concentrations of up to 97% and reached 68%, respectively. The phosphate mineral, fluorapatite (Ca₅ (PO₄)₃F, reached 27.21% in the sector of Sidi Ali Ben Oum Ezzine and 34.05% in the sector of Gassaa Kbira. An overview on result statistics of the fluorapatite is showed in Table 1. Quartz (SiO₂) was also detected. It showed average to low abundance that can reach 30%.

Parameter	Value
Obs	25
Min	0
Max	34.05
Mean	10.79
Median	6.01
Std Dev	10.41
Skewness	0.64
Kurtosis	2.26

Table 1.

Fluorapatite result statistics.

4.2 Phosphate spectral behavior

According to the identified specific vibrational absorption features on the measured reflectance spectra, minerals were detected in convergence with the mineralogical results. Carbonate minerals were highlighted thanks to a specific vibrational absorption feature in the short-wave infrared (SWIR) region at 2336 nm due to CO₃²⁻ ion. Dolomite presents, in particular, a displaced absorption feature at 2326 nm instead of 2336 nm, compared to calcite (calcium carbonate), which presents an absorption minimum at 2270 nm instead of 2298 nm (Figure 5).

Figure 5.

Phosphate samples: (a) photo of the phosphate surface and (b) ASD VNIR-SWIR field reflectance spectra, taken on the surface of the different sampling points.

Fluorapatite is, however, featureless within the spectral range from 2000 to 2500 nm [17]. Some spectral features could, however, be considered for a judicious diagnostic and an efficient characterization of phosphate-enriched rock. Indeed, the flatness of rock phosphate spectra as compared to the spectrum of dolomite around the spectral range from 2230 to 2400 nm as well as the absorption kink at 2209 nm have been used for this purpose. The defined absorption features were confirmed basing on the spectrum of a fluorapatite sample that was collected from a Tunisian carbonate geological context. Considering the both carbonate minerals in this case region, the phosphate-enriched dolomitic rock showed a decrease in the value of the absorption depth at 2326 nm as well as for the absorption minimum at 2270 nm. The phosphate-enriched limestone rock also showed a similar behavior but at different wavelengths, at 2336 nm and 2298 nm, respectively. The presence of fluorapatite was still detected according to 2209 nm absorption feature [5] within both phosphate-enriched dolomitic rock and phosphate-enriched limestone rock spectra.

4.3 Fluorapatite content prediction

4.3.1 PLSR regression method

The PLSR regression method was conducted using the resulting XRD analysis data and the measured VNIR-SWIR hyperspectral data in an attempt to predict fluorapatite contents within phosphate samples. No transformations have been applied on the used reflectance spectra.

Aiming to verify the prediction capability of each generated PLSR model for the training dataset, an optimum number of latent variables (LVs) were determined through the leave-one-out cross-validation procedure [13, 18], which was repeated for all samples to predict fluorapatite contents. The resulting optimal PLSR model was then applied to the validation dataset, which represents 30% of selected samples. Seventy percent of samples were, however, selected for training purpose (Figure 6).

Figure 6.

The selected samples for training and validation purpose.

According to the previous studies [19], fluorapatite concentrations were sorted in increasing order, interpreted as the optimal method. The performance of the generated PLSR model was assessed through the comparison of predicted content values of fluorapatite dedicated for validation with the measured ones. Thus, the coefficient of determination R² in the training and the validation sets, the root mean square errors (RMSE) and the ratio of the performance to interquartile (RPIQ) were calculated. The most significant wavelengths n in the PLSR were justified [20] and used for PLSR model generation.

4.3.2 PLSR linear regression

The generated PLSR model for the fluorapatite content prediction was generated using hyperspectral reflectance spectra and its respective concentrations, estimated by X-ray diffraction analysis. The most important VNIR and SWIR spectral regions were used by the PLSR for the fluorapatite content prediction of 400 to 450 nm, 500 to 570 nm, 2000 to 2100 nm, 2180 to 2200 nm, and 2300 to 2350 nm (Figure 7). The spectral bands corresponding to clay (Al-OH) and carbonate (CO₃) occurrences were particularly used. The generated model has shown an R² equal to 0.64, an RMSE of 5.52, and an RPIQ at 2.15. An R² equal to 0.42, an RMSE of 10.29, and an RPIQ at 1.74 were, however, determined using the validation dataset (Figure 8). It has been shown that these interesting prediction accuracies, which were revealed using these preliminary results, are promoted for the prediction of fluorapatite contents.

Figure 7.

The most spectral regions used in the PLSR regression for the prediction of fluorapatite concentrations.

Figure 8.

Accuracies of the PLSR model, generated for the fluorapatite content prediction.

The present chapter highlighted the usefulness of the hyperspectral reflectance spectra for the content prediction of the fluorapatite around the Chaketma phosphate mining site.

The detailed methodological approach based on PLSR regression method, which was conducted using VNIR-SWIR reflectance spectra, taken on the surface of the different point sampling, and the corresponding X-ray diffraction (XRD) results, was presented. Results have shown that fluorapatite, the main phosphate mineral in the Chaketma mine, was predicted according to the interesting prediction accuracies, showing an R² = 0.64, an RMSE = 5.52, and an RPIQ = 2.15.

A thorough study will be conducted in the future with the aim of further improving the prediction performances of the fluorapatite content modeling, through the application of the data preprocessing methods. Moreover, multispectral image data, such as SENTINEL-2 and Aster data, as well as hyperspectral image data, such as Hyperion, will be used for the mapping of this phosphate mineral. Such results will be of great interest for locating and estimating the phosphate reserve.