rad.Data Spectral Analytics presents: How site-specific material libraries can improve the overall mapping of hyperspectral imagery using ReSens+ mapping algorithm “EnGeoMap”

ReSens+Mine Ops meets ReSens+Space

Hyperspectral imagery in mining operations opens the possibility of mineral & ore-classification, and (semi-)quantitative maps in a diverse and changing mining environment. Tripod or drone-based (UAV) measurements can generate updated and pixel-precise ore-classification maps of a mine within 1-2 hours based on site-specific and georeferenced hyperspectral library data.

A new map can be added to the geological model of the mine each time a new block has been produced or a new active mine face has been opened. The pixel size can vary between 1.5 – 15 cm depending on the distance of the scanner to the scanned object and should be chosen based on the structurally relevant pixel-size of the geological features of the scanned area.

Based on the integral scanning of entire surfaces with a known range of spectral features at a resolution equal or below the structurally relevant spatial resolution of geological features, ReSens+MineOps can deliver precise (semi-) quantitative and reproducible remote chemical analysis to mine operators and geologists. 

Mineral deposits that were only determined by point-like data (e.g. drill core) and subsequent simulation-modelling between the point sources can be measured and visualized over a large area and quantitative predictions can be adjusted quickly.

Mine face scanning will complement drill cores and their use for resource modelling process. It will provide fast, extensive and statistically relevant results over large open areas and will add a new level of (semi-)quantitative precision to the mineral extraction process. This will make mineral processing more balanced and robust and will help to improve ore recovery rates.

In the following, a site-specific library of the mine Skouriotissa in the Republic of Cyprus is presented and a mapping result based on publicly available, multispectral satellite Sentinel-2 is shown and explained. The mapping was done by rad. Data’s EnGeoMap algorithm.

Skouriotissa – copper-gold-pyrite mine

Skouriotissa is an operating copper-gold-pyrite mine (VHMS-type Cu deposit). Visiting the mine in March 2018, we took several samples from two different positions and hyperspectral imaging scans in the VNIR and SWIR range (400 – 2500 nm) of a smaller open pit in the area. This work was conducted during a joint field work of the German Research Centre of Geosciences GFZ, Potsdam and the Geological Survey Department of the Republic of Cyprus, GSD. The scans were taken with the HySpex VNIR and SWIR system of Norsk Elektro Optikk.

The principal component analysis based on processed reflectance data can be seen in Figure 2.


Figure 2: Principal component analysis (PCA) of one HySpex scan of Skouriotissa pit “Three Hills”. The PCA shows spectrally similar areas in the mine in the same color coding.

Lower into the pit, we acquired a mine face scan and took three samples each from five different positions. Figure 3 shows the small mine face and the position of sample collection.

After measuring the samples with the hyperspectral VNIR and SWIR sensors “HySpex” we built a customized spectral library for this specific mine site . Each sample is assigned a unique spectrum, a fingerprint (Figure 4). These fingerprints can be used for subsequent mapping with rad.Data’s EnGeoMap algorithm.

Site Specific Library


Figure 4: Samples from Skouriotissa and the corresponding spectra in the library. The spectra are not smoothed in any way.

The full spectral library plot of the HySpex sensor in Figure 5 shows, how well the samples can be distinguished from each other by their spectra. Each sample position was sampled three times, distinguishing between weathered and fresh surfaces. The spectral library used in this example focuses on the weathered surface, when possible. One spectral fingerprint for each sample material is defined. These spectral fingerprints and their distinction allows us to find and map these materials in spectral imagery e.g. satellite data, overflights or mine face scans. A selection from the full geochemistry is presented in table 1 and shows geochemical similarities.

 
Table 1: A selection of samples from Skouriotissa and a selection of their main geochemistry

Sample

SiO2 (%)

Al2O3 (%)

Fe2O3 (%)

MgO (%)

CaO (%)

Sko1_B_1b

51.38

12.91

18.26

7.09

1.26

Sko1_B_1c

34.13

17.66

23.77

5.40

5.30

Sko1_B_2a

59.03

13.95

13.11

4.11

0.47

Sko1_B_2b

63.71

11.87

11.96

3.60

0.51

Sko1_B_2c

52.59

12.78

17.69

5.28

0.64

Sko1_B_3a

59.87

13.35

12.10

4.78

0.27

Sko1_B_3c

63.08

12.51

10.87

3.89

0.26

Sko1_B_4a

62.37

12.24

8.86

7.20

0.25

Sko1_B_4c

59.51

12.34

13.34

6.54

0.22

Sko1_B_5b

56.39

12.69

14.00

7.28

0.16

Sko1_B_5c

62.27

12.61

9.24

6.61

0.27

 


Figure 5: Spectral library plot of HySpex (408 bands) of the samples.

As Figure 6 shows, the sample spectra are spectrally dominated by iron and additionally influenced by the clay and water content and show these features in their spectra. The principle indicator for clay mineralization is the spectral clay feature at 2200 nm wavelength range. The strong spectral features for water in the samples can be attributed to water bound by clay-minerals. The intensity of the water signal (feature depths) apparently correlates with the clay content and may be exploited as an additional source of information. The carbonate feature around 2300 nm is also visible in  some of the spectra, especially 1b and 1c as validated by the geochemistry in table 1.


Figure 6: Hyperspectral Library of the samples, highlighting the areas of the spectra which are associated with showing iron- clay- and carbonate content and the features associated with water.

Site-specific spectral libraries, such as the one we built from the presented samples, can be deployed to identify individual materials with great precision and deliver semi-quantitative information through remote sensing. Thanks to the EnGeoMap algorithm of rad. Data, composite spectral information from satellite scenes can also be analyzed semi-quantitatively. Due to the rather large area each satellite scene pixel covers (20*20m) the resultant spectra contains signals from various of the above measured samples. With the EnGeoMAP algorithm we are capable to determine the individual contribution of the different materials present at the surface of the mine and rank them in their relative (not absolute) abundance. 

In the following, we tested the site-specific spectral library on a Sentinel-2 multispectral satellite scene dated July 2019 of the northern part of the Republic of Cyprus. The Sentinel-2 raw data is publicly available and free of charge, but has to be corrected for atmospheric conditions and checked for cloud coverage. Sentinel-2 is a satellite sensor built for atmospheric monitoring and climate & environmental protection purposes. It is focussed on these purposes, which is why it has 9 specific bands at wavelength positions important to map atmospheric conditions and vegetation. It is not ideal to map minerals but having that in mind, we can already salvage the data at hand and map some seriously interesting patterns.

Sentinel-2 Analysis

To use the spectral library built from the sample it has to be resampled to the spectral properties of Sentinel-2. This means resampling 408 bands of the HySpex, covering 400 – 2500 nm nearly continuously to the 9 sparsely distributed bands of Sentinel-2. As seen in Figure 7, a lot of the spectral information gets lost during this process, but the spectra are still distinguishable. (Future images provided by the German EnMAP hyperspectral imaging satellite will deliver full hyperspectral data).


Figure 7: Spectral library in Sentinel-2 band set.

Deploying the spectral library and using it to map the Sentinel-2 scene taken over the mine, results in a mineral map as seen in Figure 7. Sentinel-2 has a pixel size of roughly 20m*20m. The different materials that occur in this 20m*20m pixel are mixed spectrally, resulting in an average pixel spectrum from all the material spectra that can possibly be present on the surface of said pixel and that are represented in the site-specific library. Unmixing these mixed pixels based on the library and the “known” present materials is possible and gives an idea about the content of each pixel. This is presented based on the six pixel surrounding sample points from the sampling in March 2018, see Figure 8.


Figure 8: Left and Center: Sentinel-2 scene and a zoom in into the RGB satellite image and the highest abundance mapping by rad.Data’s EnGeoMap, Right: Six pixel in the RGB scene with their corresponding spectral mixtures calculated by rad.Data. “Contains Modified Copernicus Sentinel-2 Data (2019)”

Figure 8 shows that each of the six pixel on the right highlighted in the black frame are spectrally dominated by different spectral library entries. The material spectrum with the highest influence in the mixed spectrum of the 20m x 20m satellite scene is assigned to each pixel in the color coding. On one hand, this example shows how well the site-specific materials can already be distinguished even under less than ideal conditions (multispectral 9 band sensor and 20 * 20m pixel). On the other hand, major improvements can be achieved when using hyperspectral sensors with the full range of spectral information in  UAV-based or tripod-based mine(-face) scanning from a close distance or by applying these data to readily available commercial super-spectral WorldView-3 scenes with 16 bands within the VNIR & SWIR range. The evidently high variation of materials in a small area (mine-face or even mine-pit) can be resolved by the smaller pixel sizes of near-field scanning. The material variation can be mapped confidently and can play an important role in the geological model of the deposit and for mine planning.