2.1.

Setup of the Calibration Field

The calibration field was set on a concrete basketball court. The court was divided into squares of $4\times 4{\mathrm{m}}^2$ in which the rock samples and calibration nets were placed. Figure 1 shows an aerial image of the calibration site and the description of the individual tiles with the final measured coverages of the different materials. The main goal of this investigation was to understand the relation of the distribution of the materials on their spectral signal considering the different pixel sizes as a result of different flight elevations. For this reason, the three different ore materials and reference materials were distributed in 48 tiles. The ore material was laid out in two coverage types—even over the whole tile in different abundances and center-concentrated in the middle of the tile, the final coverage percentage was measured based on the 2.5-cm resolution orthophoto from the airborne mission as the number of pixels covered by the material, shadows, and remaining concrete are visible. In addition to the tiles that contained individual materials, other tiles that contained mixtures of sulphidic and oxidic rock materials were laid out in different concentrations and coverage types as well. Finally, the test field included empty tiles to represent bare concrete to describe the background and also by tiles covered with black net. The nets were black synthetic material similar to the one used by Brook and Ben-Dor¹³ for the supervised vicarious calibration of hyperspectral remote sensing data. Although this correction method was not applied in this research, different coverages of the black nets were used to investigate the statistical set up of the coverage. The test field set up of the tiles was optimized to allow the best separation of the tiles from each other in the hyperspectral data cubes. The spatial accuracy of the setup was verified with a real-time kinematic of Global Navigation Satellite Systems (GNSS).

Fig. 1

(a) Main part of the calibration field—image Hasselblad H5DC 100-mm lens. (b) Short description of the individual tiles.

2.2.

Description of the Geological Target Materials

For the experiment, three typical rock types were provided by the mine operator, including a sulphide ore (Cu 0.54% and Mo 0.0015%) (labeled M1 in this study), an ore of 34% oxidation (M2), and an oxidic waste ore with Cu 0.09% (M3). Additionally, tailing material with a grain size of $45\mu \mathrm{m}$, based on the milling for the processing plant, and sand were laid out as targets. These geological target materials were organized into closed and broken layers with rock samples of 5- to 15-cm diameters.

Following are the description of the three samples following different analysis methods. In detail, the samples can be described as follows and mineral groups are given in Table 1.

Table 1

Mineralogical composition of the sample rock material in mass percent based on TESCAN scanning electron microscope.

Based on energy dispersive spectroscopy, material M1 is the sulphidic sample material containing abundant biotite overprinted by chlorite. The rock comprises 40% to 45% plagioclase, 25% to 30% quartz, 3% biotite, 5% k-spar, 5% sericite, 7% chlorite and clay minerals with k-feldspar, quartz, plagioclase with grain sizes ranging from 0.4 to 1.9 mm and minor tabular biotite. The plagioclase is partially overprinted by sericite. K-spar is rimmed by spotty albite along grain boundaries. Quartz is characterized by isometric grains with irregular boundaries. Biotite is completely overprinted by chlorite and occurs as 0.5- to 0.9-mm tabular grains.

Material M2 is an oxidized ore sample composed of 42% to 45% plagioclase, 25% to 30% quartz, 5% to 10% feldspar, 8% sericite and secondary quartzite, 5% rock material with muscovite overprinted the biotite, and the fractures are filled by secondary quartz. It consists of plagioclase, 0.4- to 2.8-mm quartz, minor interstitial feldspar, and sericite. Quartz grains are 0.2 to 1.4 mm, irregular, and isometric. Accessory minerals are represented by sphene and 0.04- to 0.4-mm apatite. Secondary quartz occurs as fracture fill.

Material M3 is also oxidized, similar to protolith, but is barren of copper and has higher feldspar and overall Al content.

Analysis of the clay minerals and species of the different rock material based on hyperspectral data illustrates the different clay typologies of the samples. The hyperspectral longwave microscopy with $25\text{-}\mu \mathrm{m}$ resolution shows strong signals of saponite, smectite, nontronite, montmorillonite, and weaker signals of illite and kaolinite. In M1, montmorillonite is slightly more dominant than in M2 and M3. M2 has the highest values of nontronite, and in M3 the dominant clay minerals are saponite and smectite.

The core scan of the rock material with 0.3-mm pixel resolution in the shortwave infrared (SWIR) bands shows significantly more clay signals (saponite, montmorillonite, and illite) in M2 and M3 than in sample M1. However, M1 was more abundant in nontronite than in the other samples.

Using XRD for bulk samples and size separation, a $<2\text{-}\mu \mathrm{m}$ fraction shows a more detailed picture of the three sample materials. The sulphidic material is rich in quartz and shows in bulk and fine component micas and pyrites. The copper-rich oxidic sample M2 shows high measurements of orthoclase and chlorite. Kaolinite and other clay minerals are only in M2 found in measurable abundances, while other clay minerals were under the set sensitivity of the XRD set up. M3 also shows on the floated samples with typical clay signals, but they are under the response level (Table 2).

Table 2

Results of the XRD on the sample materials M1, M2, and M3 for bulk samples and fractions smaller than 2  μm (in percent).

2.3.

Airborne Mission

The airborne data collection for the calibration field was performed with a Cessna 208B with a dual camera hole equipped with Lidar Riegl Q780, metric camera H5Dc 50 Megapixel with 100-mm fixed focus lens, a hyperspectral scanner Hyspex VNIR1600 (160 bands, 400 to 1000 nm), and SWIR320me (256 bands 1000 to 2500 nm)—both hyperspectral sensors were flown with a field-of-view (FOV) expander. All sensors were flown between 28th and 30th of August 2019. The flight direction was, in each case, south to north. The flight program included flights at six altitudes above ground with numerous repeats in the altitude to assure statistical suitable observations (Table 3). All data were flown within 80 min of solar noon.

Table 3

Summary of the flight set up with altitude above ground, repeat fly overs, and the spatial resolutions of the individual sensors.

The GNSS trajectory of the flight mission was processed against the 5-km far GNSS-reference station of the Erdenet Mining Corporation (EMC). An absolute aircraft position was achieved within 2 cm in $XYZ$. Further orthorectification of the image data for the calibration field was based on 2-cm accurate measured corners of the field and control points on and around the calibration field. Hyperspectral data cubes were rectified with PARGE airborne image rectification from RESE¹⁴ based on a general boresight and further improved with additional control points from the GNSS survey on the calibration field using gridded resampling to optimize between spectral and spatial quality of the orthorectified data. The required spatial pass accuracy to accept the solution was better than a 0.35-pixel size for each individual data cube. Atmospheric correction was performed using atmospheric and topographic correction¹⁵ with support of ground measured spectra of the black nets and the concrete. The atmospheric correction was based on an urban model due to the presence of industrial and city structures as major components of the mission area. The atmospheric correction was based on a fixed water vapor value based on average estimations for each flight level and regression of the water vapor bands.

The wider area of the EMC mine, tailing dam, and Erdenet town in four flight levels (600, 1000, 1500, and 2000 m above ground level) were also flown over with the same sensor set. Data were processed with similar settings as the calibration site and used for checking the results in a practical application.

2.4.

Ground Support Measurements

During the flight campaign, spectral ground measurements of materials of the calibration site were performed using a Spectral Evolution SR-3500 spectroradiometer. The SR-3500 covers a wavelength range 0.35 to $2.5\mu \mathrm{m}$ with a spectral resolution of 3 nm at 700 nm, 8 nm at 1500 nm, and 6 nm at 2100 nm. The measurements were conducted using a bare fiber optic with 25-deg FOV and an approximate measurement distance of 1 m to the ground. Depending on the filling of the tiles, 10 to 20 spectral measurements were conducted. The measured spectra were calibrated to absolute reflectance values using a Zenith Lite™ diffuse reflectance panel of 95% reflectivity. Spectral preprocessing encompassed the removal of wavelength regions heavily affected by noise due to atmospheric water vapor absorptions at 1.4 and $1.9\mu \mathrm{m}$, averaging of all spectra per tile and smoothing the spectra by applying a Savitzky–Golay filter.

2.5.

Setup of Supporting Laboratory Experiments

To support the modeling of the linear quantitative relationship between coverage and spectral features of clay minerals, a laboratory experiment similar to the calibration field was set up. This included four Petri dishes with different watery mixtures of concrete, hardened and dried in the muffler oven, and then set up with different sizes from 5 to 15 mm of circular set tailing material from the Erdenet mine with a suitable covering thickness (i.e., optically thick, $>100\mu \mathrm{m}$). The tailing material was used. It can be expected to have similar clay features as the ore material and has a uniform grain size of $45\mu \mathrm{m}$ due to processing of the ore. All samples were dried. The set up of the four dishes was completed with 5% and 40% spectralon reflectance panels. These panels were used for the correction of the measured data to reflectance values using an empirical line correction. The realization of additional coverage measurements was done via spectral binning of different sized circular areas around the circular tailing material. In total, 110 setups were measured, distributed over the different concrete settings. Absorption features were measured on continuum removed spectra in which absorption features present in the spectra are normalized against a convex hull covering the total spectral range.

2.6.

Measurements of Spectral Parameters

Spectral data from the data cubes were analyzed using ENVI image analysis functionalities under visual controlled conditions, meaning unclear pixel locations over the calibration field were excluded from the measurements. For the imagery acquired at a flight altitude of 480 m, spectra were collected and averaged within a $3\times 3\text{pixel}$ window, and for the 650-m altitude data, this average window was defined as a $2\times 2\text{pixel}$. Faulty bands were spectrally averaged with neighboring bands. Furthermore, all spectral measurements were attributed with material, coverage type, and flight altitude. For the analysis of individual spectral bands, spectra were normalized using the continuum removal procedure.¹ In total, 351 valid spectra were measured over the different materials, altitudes, and repeats. For the description of the 2200-nm absorption band, the average of the continuum removed the value of the bands at 2100 and 2240 nm, which was subtracted by the value at the band on 2192 nm, while for the description of the 1400- and 1900-nm bands only one reference band on the lower site (1370 and 1850 nm) was used.

To separate the individual clay minerals for all samples, a spectral linear unmixing procedure was applied using the spectra from the United States Geological Survey (USGS) library and the concrete spectra from the field measurement. For the linear unmixing, water bands were excluded and unit sum constrains were not used under the assumption that in addition to the concrete and the clay minerals, further minerals were also part of the rock material laid out over the calibration tiles. For the analysis, only evenly laid out calibration tiles were used. Tiles with unclear visual allocation were excluded.

3.1.

Linearity Checks on Calibration Targets in the Calibration Field

As a first-order validation, 47 scenes with black net fields were used to verify the viability of the calibration field concept and the methodology of the measurement techniques for the spectra. The black net folded four times provides a reflectance below 5% over the full spectral bandwidth of the hyperspectral sensor used.¹³ An average of the reflectance of each tile with different coverages of black net over all bands of the SWIR range was calculated and the damping of the signal of the concrete based on the coverage percentage with black net was observed. Figure 2 shows the correlation between the coverage percentage of the pixel with black net and the averaged reflectance signal over all bands of the SWIR sensor. There is a good linearity between coverage and reflectance. The coverages below 50% and at 100% show well-clumped points. However, the variability for the three measurements at 71% coverage is distinctively above the expected values. This can be explained with the spatial rectification of the data better than a 0.35-pixel size. It shows that the spatial rectification error for a pixel with a partial coverage of more than 50% of the pixel size by a central target hampers the derivation of correct quantitative information, and therefore, is excluded from further analysis. Removing the measurements at 71% cover (red dots in Fig. 2) provides an $R^2$ of 0.98. This supports the feasibility of the methodological approach of the calibration field.

Fig. 2

Correlation between coverage percentage of the pixel with black net and the average reflectance value of the pixel in the SWIR bands—all flight altitudes included.

3.2.

Usage of Clay Mineral Absorption Bands 1400, 1900, and 2200 nm

Clay minerals show strong absorption features in the SWIR at 1400 nm (caused by OH), 1900 nm (caused by molecular water), and 2200 nm (caused by combined overtones of Al-OH).² Measurements of the depth of the absorption band and the area of the absorption were taken from continuum removed spectra. Figure 3 shows a sample spectrum of M2 with the dominant clay features.

Fig. 3

Spectrum of material M2 based on grab sample scan with Hyspex SWIR with 0.3-mm resolution and corrected with spectrolons.

The airborne imagery analyzed for evenly laid out material coverages of 8%, 23%, and 31% of the tiles with materials M1 to M3, summarized over all flight altitudes recorded (total $n=54$), showed strong differences in the strength of the absorption bands at 1400, 1900, and 2200 nm and their variability as shown in Fig. 4. The absorption at 2200 nm showed the strongest relationship to the coverage of the pixel with the material M2. It also shows the smallest standard deviation and a clear increase of absorption signal with coverage. For the materials M1 and M3, similar behaviors could be found. For this reason, the further detailed analysis of the clay absorption bands is focused on the 2200 nm. The poor correlations in 1400 and 1900 nm are due to overlays with strong water vapor bands and the difficulties in atmospheric correction (Fig. 4).

Fig. 4

Average and standard deviation for the depth of the three absorptions bands of clay minerals for rock type M2. $N$ per case and coverage is 18% of the tile.

In a next step, even and centered laid out coverage tiles below 50% were analyzed for the dependencies of the depth and area of the absorption band at 2200 nm to different parameters.

A relative absorption depth normalized to the coverage in percent was calculated to analyze the influence of the flying height on the depth of the absorption band. This relative depth was based on the assumption of linearity as shown above. The average relative absorption band depth shows no significant difference between the different flying heights. The error interval is increasing with the flying altitude, which might partly be due to the usage of the pixel-based averaging of the spectral measurements in the lower altitudes (480 m: $3\times 3\text{pixels}$, 650 m: $2\times 2\text{pixels}$). Figure 4 shows that, for the designed calibration field, the flying altitude is not of major importance and that using lower altitudes allows better statistics with reduced efforts for the set up of the calibration field, which is also supported by an analysis of variance. Comparing the three lower altitudes shows that up to 950 m, similar behavior can be found for all three target materials, but there is an increase on the confidence interval of the width for the average with increasing altitude.

Figure 5 shows the comparison of the relative absorption depths of the 2200-nm feature for the three materials between only center-filled and evenly filled tiles. While the averages are similar, the centered laid out samples have a higher error based on a smaller number of samples and the above-described complications of the spectral data on the pixels. This means again that evenly laid out calibration tiles with target rock samples are preferable. For estimating linearity of the coverage/absorption band solution, the centered tile design can be used to get supporting data points, especially in the lower coverage percentages, but the values measured have to be weighted lower.

Fig. 5

Average and 95% confidence intervals of the relative depth of the 2200-nm absorption band plotted by the flying heights, an example for rock type M3 ($n=75$).

Figure 6 shows the linearity of the tile coverage in relation to the absorption depth and the area for the three materials, respectively. The absorption depth shows similar tendencies for the oxidized materials M2 and M3, while material M1 shows a weaker increase with the coverage due to the lower clay contents. The linearity of the absorption band depth in evenly filled tiles is $R=0.70$ for M2 and $R=0.82$ for M3 higher than for M1 with $R=0.65$. For the target materials solely laid out in the center of the tiles, the correlation between coverage and absorption strength of the 2200 nm band was found to be weaker, which is most likely due to the aforementioned higher errors in this setting. The correlations observed for the individual scenarios per flying altitude and coverage type show better fits to the quadratic correlation function between coverage and absorption band depth. As shown in Fig. 7, this could be related to the fact that the lowest measurements were measured near the sensitivity threshold of the sensor.

Fig. 6

Comparison of averages of the relative absorption band depths for the three materials evenly distributed as well as centered and confidence interval of the averages.

Fig. 7

(a) Depth and (b) area of the 2200-nm absorption bands summarized over all heights with a trend line.

3.3.

Sensitivity of the Detection in Low Coverage and Linearity in the Supporting Laboratory Experiment

The results of the laboratory experiment were used to obtain more information on the minimum sensitivity and the linearity to expect from a calibration field. They show a high linearity between the cover of the measurement area with tailing material over the concrete and the absorption depth at 2200 nm. The parameters of the linearity are dependent on the concrete background, but all show similar increases and an $R^2$ of 0.95. Figure 8 shows the data for the four different concrete types in the Petri dishes together—the linearity for the different concrete types is easy to see in separate point clouds. The figure also shows one example for concrete No. 3, showing the strong linearity per concrete background.

Fig. 8

Absorption band depth in normalized reflectance at 2200 nm of tailing material versus the coverage percent of the tailing material over the concrete in the Petri dish. (a) All concrete types together and (b) only for sample concrete type No. 3.

To get information on the sensitivity of the detection and the influence of the background, we analyzed the low coverage percentages of the tailings materials (low coverage exposes more background). Figure 8 shows the results for the measurements below 10% coverage. The graph shows that the linearity can be identified from coverage values between 2% and 6% coverage. In summary, the concrete types with the lowest water content showed the highest reflectance values and the best linearity between low coverage values and the absorption depth at 2200 nm of the tailing material. In contrast, the concrete types with higher water content showed higher variability (the extreme example is concrete type 4, Fig. 9). Concrete type No. 2 shows a good sensitivity below 4% coverage with the tailing material. Using the 1400- and 1900-nm absorption band depths as measure of the free water in the concrete indicates that concrete types 1 and 4 have the highest free water content. This affirms the importance of in-depth knowledge of the utilized background material in the field and in laboratory experimental set ups for quantitative spectral analyses

Fig. 9

(a) Measurements of relative absorption band depth 2200 nm on continuum-removed data in tailing material for lowest coverages in percent on the $x$ axis for the 12 sets of data and (b) summary for the four different concrete types for the same parameter.

3.4.

Separation of the Clay Minerals Using Unmixing Methods

To separate the individual clay minerals, a linear unmixing approach was carried out using the USGS spectral library¹⁶-based spectra and the field spectra for the concrete. Then, abundance data for evenly filled calibration tiles for the materials M1 to M3 were collected. Figure 10 shows the results for the repeats for the flight altitude at 650 m. Strong correlations between the tile material coverage and the abundance values were found for the materials M2 and M3 for the clay species nontronite ($R^2=0.93$) and kaolinite ($R^2=0.87$) as well as chlorite ($R^2=0.76$) for material M2. For material M3, nontronite shows again the strongest correlation, while chlorite has an $R^2$ of 0.88, stronger than kaolinite ($R^2=0.8$). Material M1 shows a weaker correlation in the estimation of the abundance with linear unmixing, with illite with $R^2$ of 0.53, and kaolinite with $R^2$ of 0.32 showing the strongest dependencies. As shown in Fig. 9, the increase of the abundance signal with the coverage in M1 is very small compared to the strong increase in the oxidized M2 and M3. The concrete abundance value is decreasing as expected with the increasing coverage of target material strongly independent of the cover material ($R^2\mathrm{s}$ between 0.59 for M1 and 0.95 for M3).

Fig. 10

Abundance of (a) kaolinite and (b) nontronite separately measured over the calibration tiles based on linear unmixing for the three target minerals.

Comparing the results of unmixing, it shows clearly the differences in the quantity measurements using various methodologies. For instance, nontronite has strong spectral features of the SWIR and the microscopic mid-IR samples, but cannot be quantified with standard sample preparation for XRD. This points to the importance of special sample preparation of ground samples and the selection of different approaches of quantification measurements for correlation with spectral measurements. It also enforces the point that small abundances of clay minerals are important to exploration geologists and small abundances can be detected with IR. It also shows the opportunities to understand, with the help of such experiments and calibration sites, locally specific sensitivity and the relationship of individual clay minerals and apply this data to mineral exploration.

3.5.

Adaption of the Data into the Larger Area Airborne Mineral Exploration

The calibration field was analyzed in the context of a larger survey of $160{\mathrm{km}}^2$, where the calibration site was utilized as a check for the whole map area. The coverage based on the calibration of the depth of the 2200-nm absorption band shows reasonable results over the mine where the rock material comes from, but also for known exploration targets. These results are based on the continuum removed data for the whole area. Applying linear unmixing to the area mapping data with similar library data as for the calibration site and using again the calibration site as a check, it shows similar behaviors for kaolinite, chlorite, and nontronite (Fig. 11). The example map shows the calibrated appearance based on nontronite-containing material for an area of known oxidization similar to material M2/M3. While the values using the correlation found on the calibration site are in a reasonable range, it is expected that the highest value are especially over represented and a better calibration with more different materials would improve the results.

Fig. 11

(a) Results of the airborne mapping linear unmixing for some clay minerals measured over the calibration site and (b) sample map of calibrated nontronite values for a known exploration area north of Erdenet city.