Imaging Spectroscopy for the Mining Life Cycle: A Guide for Drone Applications

1. Introduction

The use of remote sensing, especially imaging spectroscopy, offers large -scale, non-destructive and time-efficient means for mineral, lithology, soil and plant species mapping over spatially extensive areas at the deposit, camp and outcrop scale. Imaging spectroscopy has been an active area of research for almost four decades and is based on selective absorption and reflectance of wavelengths light by different materials. Reviews on the fundamentals are covered by e.g., Clark [1]. Van der Meer et al. [2]. Hunt [3,4,5], Hecker et al. [6], and Manolakis et al. [7]. Geological remote sensing took off with the deployment of the Airborne Visible/ infrared Imaging Spectrometer (AVIRIS) and other airborne and spaceborne instruments such as HyMap, Hyperion, ALI and ASTER [8,9,10,11,12,13,14,15] and has been picked up especially by exploration both for green- and brownfield purposes using medium (5m) to low (30m) spatial resolution satellite and airborne sensors. These instruments collected either multispectral data (up to 10 different wavelength bands offering low spectral resolution) or hyperspectral data (collecting information in hundreds of narrow consecutive bands to resolve small spectral features). Satellite remote sensing, using ultra-high resolution RGB imagery offers some utility for the mining industry, that low spatial resolution data cannot match. This includes the monitoring of small-scale surface changes in tailings, detecting surface expressions of subsidence and mapping mining footprints and mining activities [16,17,18,19,20]. Nonetheless, because it is limited to the visible wavelengths these systems cannot be used to distinguish surface mineralogy. A UAS, equipped with a spectral imaging instrument, is able to bridge this gap and provide high spectral and spatial resolution.

While the term Uncrewed Aerial Vehicle (UAV) describes mainly the aircraft itself, Uncrewed Aerial System (UAS), commonly referred to as “drone”, includes the whole system of aircraft, including the control module, navigation hardware, transmission systems, cameras, software, the ground station and the person(s) controlling the vehicle. The versatility of deploying UASs has led to an increased adoption within the mineral extraction sector. This largely involves applications based on RGB-imagery and photogrammetry, such as stockpile volume surveying, site infrastructure inspections, and some environmental monitoring. However, UASs have the potential to act as interoperable technology with a broad range of applications that include geological pit mapping [21], geotechnical analysis, geophysical survey, rock slope stability assessment, water quality monitoring, erosion and soil loss estimation, acid mine drainage (AMD) mapping, subsidence detection and safety management (e.g., tailings dams, road haulage, etc.), rehabilitation, and post-mining environmental monitoring [22,23]. In complex environments within the mining value chain, the need for Digital Terrain Models (DTM) of high precision and high spatial resolution can be fully addressed with drones equipped with light detection and ranging (LiDAR) technology [24] such as landslide mapping [25,26,27,28,29,30,31,32,33], slope monitoring [34,35,36,37], subsidence modelling [38]) and surface models [31,39,40,41]). Though a few studies have already presented the combined use of UAS and LiDAR with a focus on forestry and agriculture [42,43,44], UASs offer excellent performance in any environment where accurate surface representation is required, especially in inaccessible areas, along steep slopes, or hazardous environments. In such environments, undertaking a field survey with a UAS can ensure the safety of the surveyors. The cost-efficient repeatability of the survey at frequent intervals is also an advantage, that permits the comparison of the 3D surface over time so that the rates and scales of change can be detected.

With advances made to sensor technology, relatively lightweight hyperspectral sensors have recently emerged covering the entire visible near-infrared (VNIR; 400-1000nm) and shortwave infrared (SWIR; 1000-2500nm) regions of the electromagnetic spectrum. Capturing data across these wavelength regions via UASs enables the mapping of a variety of materials over mining sites at high spatial resolution [45,46]. Combining hyperspectral cameras and LiDAR sensors that are attached to UAVs opens new avenues for the use of remote sensing (RS) in the mining industry.

This paper reviews the application of high spatial and spectral resolution spectral imaging data in the mining industry. This review is timely as numerous papers are starting to cover the use of UAS technology in the mining industry, including some limited case studies for hyperspectral UASs for RS studies [47,48,49,50,51,52,53,54,55,56,57,58,59,60]. Although several reviews concentrate on the application of HSI and RS in the broader field of geology [2,4,5,47,48,49,50,51], there is a notable absence of a comprehensive review on the state-of-the-art for HSI and its application to mining both in active and post-mining environments. We focus on the application of emerging UAS-based HSI technology, but we also examine case studies conducted using airborne and spaceborne imaging systems to understand the full potential of UAS in the mining industry. In this paper, we introduce spectral imaging principles in Section 2, we present current general hyperspectral UAV systems in Section 3, and we discuss the design of hyperspectral UAV studies in Section 4, including sampling and validation standards. We then in Section 5 review existing case studies of HSI applied in mining environments, including airborne and spaceborne platforms, to highlight the potential applications for hyperspectral UAV systems. We conclude with Section 6 with a discussion of potential future applications of the technology in the mining value chain.

2. Principles of Spectral Imaging

Imaging spectroscopy is usually referred to when measurements and analysis are taken out with hyperspectral instruments. In remote sensing, we speak more commonly of “hyperspectral imaging” (HSI) or “multispectral imaging” (MSI) when the Earth’s surface is the object of our study. Here, speaking of both methods, we will use the term “spectral imagining” to set it apart from (hyperspectral) imaging spectroscopy. HSI captures the incoming light in many (i.e. hundreds) continuous, narrow bands and can therefore resolve narrow absorption features specific to different minerals and materials. While the definition of “hyperspectral” in terms of number of bands is not rigid, the IEEE Standards Association set up a standard in 2018 for devices that cover the 0.25-2.50um spectral region. For a system to be considered hyperspectral it needs to exceed 32 bands [61]. Anything capturing less than 32 spectral bands is considered multispectral.

Imaging spectroscopy generates three-dimensional data cubes where each image pixel represents four-dimensional information, with three dimensions in the spatial x-, y, and z-coordinates, and spectral information in the fourth dimension. Data cubes are created by using the principles of spectroscopy, where the properties of light are captured across a set of continuous spectral bands to produce a characteristic spectrum on a per-pixel basis. Pushbroom remote sensing employs an array of detectors to capture spectral imaging data of the Earth's surface as the platform moves [62,63]. Unlike whiskbroom systems that scan across the scene with a single detector, pushbroom systems capture entire lines of data simultaneously, offering high spatial and spectral resolution. This method minimizes motion blur and enables efficient data collection by continuously recording data along the platform's path, capturing multiple spectral bands captured simultaneously.

With the use of VNIR and SWIR wavelength ranges, rock material from all stages of the mining value chain can be investigated by interpreting the characteristic absorption features, inflections and signature slopes of each unique spectrum captured by the imaging system. In geologic environments, absorption features detected in the VNIR arise from transitional elements, including iron-bearing minerals and rare-earth elements (REEs), while the SWIR region is commonly used for identifying alteration mineral assemblages related to hydrothermal systems of base and precious metal deposits [64]. The mineral groups that can be detected and mapped working in VNIR-SWIR wavelength regions include carbonates, sulphates, sulfosalts, clays, various iron oxides/hydroxides, and phyllosilicates such as chlorite, talc and muscovite. A detailed account of the cause of those absorption features can be found in [3,4,5].

Common analytical techniques for imaging spectroscopic data are divided into two broad categories: data-driven and knowledge-based approaches [65]. Data-driven approaches rely only on the data itself and some additional reference data (spectra), commonly called training classes or endmember sets, that are imported to or derived from the image data. Data-driven approaches are categorized into per-pixel (hard classifier) and sub-pixel (soft classifier) approaches with single and multiple labels for each pixel, respectively. Comparison-based per-pixel approaches, such as mineral matching with the use of a reference spectral library, include similarity-based methods such as the Spectral Angle Mapper (SAM) [66,67], least square-based methods such as partial least squares regression (PLSR) and learning-based approaches such as artificial neural networks (ANN) or support vector machines (SVM). Mixture-based, sub-pixel categories include partial and full unmixing methods such as mixture-tuned matched filtering (MTMF) and linear spectral unmixing (LSU). Knowledge-based approaches rely on user knowledge about the spectral behaviour of a target without the use of direct reference data to extract meaningful information from a spectrum. Knowledge-based approaches aim to estimate the quality and/or quantity of either of the main components making up a spectrum including: i) a continuum, ii) absorption bands and iii) residuals or noise. These techniques include absorption and spectral feature modeling. A common feature mapping approach in this category is “minimum wavelength mapping” (MWL). This feature-based approach retrieves the depth, minimum wavelength, area, width, and asymmetry of individual features, known as spectral parameters, for material identification and mapping [6,68,69,70,71]. The output of this technique can be fed into expert systems and decision tree structures (DT) to enable per-pixel classification. Examples of expert systems include the classic USGS Tetracorder and its modern interface called “Material Identification and Classification Algorithm” (MICA) [72,73,74]. Partial absorption modelling, clustering techniques [75], as well as various band arithmetics (e.g., band ratios), also do not require the use of pre-existing reference data. Applying these techniques results in spectral similarity and score images with varying magnitudes of the matched and inferred values and requires specialized visualization and contextualization to read and interpret the resulting output maps. Overview of the spectral processing methods available for geological RS is provided in [2,4,5,47,48,49,50,51]

3. Best-Practice for UAS-Based Spectral Imaging

The collection of field data is complex and highly dynamic and needs to be managed carefully for hyperspectral field campaigns, such as drone surveys. Field data typically includes a variety of instrument data, ancillary data (‘metadata’), as well as geographic coordinates and images, which can lead to interrelated but disjointed datasets.

The terminology here follows the TERN “Effective Field Calibration and Validation Practices” [76]. Data is considered a direct quantitative measurement of the sample in question, either via an instrument or another quantitative method (e.g., instrument readings, raw imagery). Ancillary data is considered as data collected in association with the primary dataset such as geographical coordinates, comments, date and time, descriptive information and imagery and information about the instruments being used. Metadata is considered the information regarding the discovery and the use of the data, such as scale, units, geographical and temporal scales, custodians and licensing. Careful collection of this information is important to ensure that a variety of researchers and stakeholders can reuse the data in the future.

Field calibration and sampling strategies for remotely sensed data have recently been developed for ecological applications and provide some utility for geological surveys [76]. Clearly defined strategies are important as field surveys can be logistically challenging, especially for UAS-based HSI, influencing the quality of the survey data and associated ground sampling and calibration data. Firstly, field data collection can vary depending on the site access conditions and weather conditions. As a result of these factors, sampling strategies often get modified, leading to errors. Secondly, the equipment in larger campaigns can vary, as can the observers and their experience level, resulting in inconsistencies due to observer bias and different objectives. As a result, we suggest the principle that field methods must be discussed prior to any campaign and a consistent sampling strategy (e.g. number/density of samples, sample area dimensions etc.) and protocols (e.g. calibration procedures, repeats, blanks etc.) must be applied. We suggest to follow best practice such as: The use of data entry tools to minimize errors. Provide sufficient data storage via an accessible database with clearly established data licenses during and after a survey. Set up a frequent review period (annual) to identify issues, provide feedback on data completeness, identify unbalanced data collection across sites and potential data bias.

In the following sections, the current state-of-the-art of hyperspectral UAS hardware and survey design are discussed, including sampling guidelines during fieldwork to assist in data interpretation and validation. This includes best practices for UAS-based VNIR and SWIR data collection to acquire meaningful, high-quality data in mining environments. While historically UAS-based VNIR full-frame sensors have been employed for geological mapping applications [59,77,78,79], the inclusion of the SWIR wavelength ranges requires pushbroom-type data acquisition and has only recently become deployable in industry applications due to the miniaturization of cameras, smaller platforms, UAS power requirements and cooling (<-100°C) of the detectors as well as high-performance gimbals.

The inclusion of the SWIR wavelength range offers valuable information for mineral detection, but it also complicates the survey design and flight planning. The SWIR wavelength range introduces dependency on accurate measurements of atmospheric conditions, as well as geometric and radiometric correction and processing of the data. All of which is detailed on in the next sections. Only a limited number of publications using UAS-mounted SWIR cameras are currently available (e.g., [80,81,82,83,84]), which means this research field is only just emerging and our understanding of good methodologies remains nascent. Similarly, the best-practice guidelines for the acquisition and correction of coaligned VNIR-SWIR datasets, and accurate georeferencing and geometrical correction of such combined datasets, remain a topic of current research. In the following sections, different variables influencing VNIR-SWIR drone-based data acquisition and correction are explained. We specifically present considerations around the use of UAS hyperspectral pushbroom scanners in the coaligned VNIR-SWIR wavelength range. For brevity we do not review UAS aviation rules and regulations (which vary considerably from nation to nation), nor general recommendations for UAS pilots on flight planning or safety assessments.

3.1. Hyperspectral Pushbroom UAS

A hyperspectral UAS commonly consists of an airframe platform, a hyperspectral camera, an Inertial Navigation System (INS) and Inertial Measurement Unit (IMU), and a differential Global Navigation Satellite System (GNSS). The airframe provides the platform for all components, i.e., connects the drone arms with the motors and rotors, if multirotor systems are being used, the landing gear/feet, source of energy (e.g., batteries or fuel), an adapter for a gimbal or direct camera attachment, different GNSS and radio antennas for both drone and camera, positioning LEDs and the autopilot system of the drone itself. The platform carries the batteries as well as the payload of the gimbal and the camera. The drone’s maximum allowed take-off mass (MTOM) and certification (C-class label) dictate within which class and in which context and area the UAS can be flown and is dependent on the jurisdiction, registration and license of the drone pilot/piloting company and drone operator. This differs quite substantially between UAV manufacturers and UAS configurations and changes in different countries and jurisdictions. In European countries, European Union Aviation Safety Agency (EASA) rules and regulations [85,86,87] based on recommendations from the International Civil Aviation Organization (ICAO) [88] must be followed, including local regulations in the different countries.

The system used exemplarily in this manuscript is a BFD SE8 octocopter carrying a Mjolnir VS-620 camera from HySpex and a LiDAR from Velodyne. The system's total MTOM is below 25kg. The Mjolnir VS-620 consists of a VNIR V-1240 and SWIR S-620 hyperspectral camera integrated into one chassis on two optical axes in a co-aligned field of view (for specifications see Table 1), a data acquisition unit operating the two sensors (i7 NUC computer), an internal INS (here, Applanix APX-15), and a radio connection to a ground station for remote access to the Mjolnir (Transmission Control Protocol/Internet Protocol link (TCP/IP)). A LiDAR scanner is mounted underneath the Mjolnir (Figure 1) and triggered in the same software. While the Lidar is run independently of the Mjolnir system it receives time tags in the form of NMEA messages (National Marine Electronics Association) and pulse-per-second signals (PPS)) from the same INS as the HySpex camera. The NMEA 0183 messaging protocol was originally adopted for interfacing marine electronics, but its use has expanded to terrestrial electronics. Figure 1 shows the schematic overview of the AUS. The UAS is driven by two 25Ah lithium polymer high-voltage (LiHV) batteries. The complex interactions and components of a UAS are described in detail in [53] and outside of the scope of this document. To reconstruct the three spatial dimensions of a pushbroom solution, the three axes representing roll, pitch and heading of each mid-exposure of each frame needs to be known with high accuracy. This movement is exaggerated in a UAS due to the lightweight platform and the high spatial resolution [53]. Therefore, this motion is captured by an additional camera-specific set of differential GNSS and IMU to correct for the movement of the camera relative to the airframe [81]. A boresight calibration accounts for differences in the angles between the IMU coordinate system and the camera coordinate system, transforming the INS data to the camera coordinate system. This is achieved via a boresight flight above well-defined Ground Control Points (GCPs) or in reference to a cartographic orthophoto and digital surface model reference. Alternatively, cross-flight pattern-based boresighting is an established standard for this aim [89]. For geometrically stable solutions such as the presented Mjolnir VS, boresighting with GCPs is only necessary once. Gimbals are necessary for multirotor pushbroom UAS primarily to stabilise the imaging payload. This is needed to maintain a consistent orientation relative to the ground. This is especially true in pushbroom systems. Multirotor UAS, due to their inherent instability caused by rotor movements and environmental factors like wind, require gimbals to counteract these movements and keep the imaging payload steady. Proficient gimbal hardware and software have only recently become available for multirotor solutions and is one of the reasons why pushbroom system were previously not widely used on UASs [52].

3.2. Preparation of a Hyperspectral UAS Campaign

The first step of the study design is to determine the area of interest and the available timeframe and necessary airtime. The area that can be covered is influenced by the number of available field days, including the time for transport of equipment. Transportation of all equipment must include the transport of heavy technical components, as well as all lithium batteries, to the survey area.

It is advised to plan all individual data collection and calibration flights in advance of the campaign to gain an understanding of the area and the total necessary airtime under ideal flight conditions (wind speed, sun angle, atmospheric conditions). However, based on practical experience almost all flight plans will require adjustment in the field based on on-site conditions. Flight planning must consider the sun angle, available illumination, change in surface level and assumed related battery drainage, terrain-following capability and limits, and local legislation and limitations around uncrewed aviation. The variability in the size of the flight lines after data correction and any possible resulting data gaps in the collected data also must be considered. The authors therefore advise an overlap of 25% between individual flight lines to avoid data loss during mosaicking. The survey objective will determine the required ground pixel size or spatial resolution. This in turn dictates the optimal flight altitude and the required sampling point spread function compatible with the pixel size requirements. The speed of flight is adjusted to achieve maximum SNR. A test flight prior to the first data acquisition flight can help the operator determine the optimal integration times of the VNIR and SWIR camera to avoid over- or undersaturation in light albedo pixels. This will give information on the relationship between speed, altitude and SNR and will likely lead compromises with the survey objective in mind. The speed of different systems will depend on the systems light sensitivity, i.e. lenses’ F-number, pixel size on RAD and spectral sampling and will differ from system to system. To give an example, for the UAS described in Table 1, a good pay-off between a high spatial resolution and a high coverage per flight for flights taking place in Central Europe during summer is achieved by flying at the maximum permitted height of 120m height above ground level (AGL), resulting in a corresponding pixel size of 6cm/pixel with a ground speed between 2-4 m/s. Maximum height AGL for flight planning will differ based on drone certification, pilot competency and country regulations.

The changes of sun angle and illumination strength during the day determine the optimal data acquisition timeframe. The best illumination is available around midday (sun close to solar noon), and it is recommended that surveys are planned closely around this timeframe, with solar zenith angle lower than 70° (optimal) or lower than 80° (acceptable). However, when dealing with steep terrain, areas of interest might only be illuminated at certain times of the day (especially important for vertical drone or tripod measurements). The weather is a key factor that must be considered at the planning stage, as well as in the field. The questions to be asked are: What is the weather expected to be like at the study location? How quickly does it change? What are the operating boundaries of the used drone? Rain poses a significant threat to the success of the campaign and cloud cover detrimentally affects the quality of the SWIR data. Depending on the drone manufacturer and the confidence of the pilots, the wind speed and likelihood of gusts must also be considered. [53] suggest a windspeed of <5m/s as best practice for multi-rotor systems. Optimal boundary conditions for high-quality data are a solar zenith angle lower than 70°, a cloud-free sky and dry rock surfaces. The described here system achieves approximately 12 minutes of total flight time per flight, including approximately 2 minutes for the take-off and the first INS alignment flight, 8 minutes for data acquisition, and 2 minutes for the second INS alignment flight and landing. These 12 minutes are influenced further by air temperature, wind speed, changes in terrain surface level and the gradient of these changes, speed/ampere during battery charging, the take-off position and its distance to the area of interest, fail-safe procedures, wildlife interference, and pilot confidence. Including the preparation time before and after landing, one flight is 30 – 45 minutes including a total of 12 minutes of airtime. Depending on the payload, e.g. if an additional LiDAR is mounted, flight time can increase up to 25min. Figure 2 shows the drone platform carrying the hyperspectral camera and lidar payload.

3.3. Execution of a Hyperspectral UAS Campaign

We advise using the first flight of the day as a test flight for the drone to collect data and calibrate the INS. Calibrating the internal camera magnetometer of the APX to the new current location is advised, and to follow UAV manufacturer recommendations for the instruments necessary to enable autopiloting and navigation. If accurate Digital Surface Models (DSMs) for flight planning are not available, a LiDAR flight can be carried out for an accurate surface model of the study area. This could also be achieved with a smaller drone that offers a 3D surface reconstruction through photogrammetry workflows. A detailed DSM of the study area allows for precise flight planning and adjustment of the preferred flight altitude (e.g., to ensure that 120m AGL are kept), a constant pixel size and precise fail-safe procedures. Ideally, this is available before the survey. In addition to the acquisition of hyperspectral data, data from different sensors is also usually collected, to improve the correction of the HSI and aid the interpretation of the collected data. Auxiliary data collected alongside the flight campaign often includes the collection of ground-truthing data for interpretation (including spectral and/or physical samples for physiochemical testing). This is discussed in more detail in the following Sections. The collection of data to calibrate the drone-based data for reflectance can be additional and is not necessarily imperative. Examples include collecting downwelling irradiance, using sky-ward videography to record visual changes in atmospheric conditions, and by placing large homogeneous (often calibrated) reference panels in the FOV of the UAS survey [53].

3.4. Data Correction and Post-Processing

The geocoding of the pushbroom hyperspectral data is done by forward ray tracing from the sensor position to a readily available digital surface model. This can either be done in a raster representation or a projection onto a surface mesh. The result of such a correction is a set of x/y/z coordinates for each hyperspectral image pixel, i.e., a ‘hyperspectral point cloud’, which is a standard output of (e.g.) the PARGE software [89]. Starting from this representation, the geometry for each pixel is well described and can be used directly for visualization, rectification to a standard grid or for terrain-based simulations. The latter is important to estimate the irradiance at the spectral image data position.

The radiometric data correction and post-processing described below are based on the “Drone atmospheric correction method” (DROACOR) created by ReSe Applications [90,91,92] and are considered industry standard when it comes to HSI drone-based data correction. The previously mentioned gathered auxiliary data can be included in the data correction and post-processing. The necessary information for processing the data in the DROACOR software are the time of day, location (LAT/LON), solar and observation angles, terrain height, platform altitude, and the sensor internal geometry. DROACOR uses a physical inversion process from at-sensor radiance to ground reflectance based on pre-calculated Look-up-Tables (LUT), based on the Libradtran radiative transfer code [93]. Thus, the HSI data must be calibrated traceably. For uncalibrated data, an inflight calibration based on calibrated reflectance targets can be performed. Some more information is gathered directly from the acquired images by DROACOR prior to the reflectance retrieval. This parametrization consists of four steps for HSI: i) an estimation of the aerosol optical thickness at 550 nm, ii) an (optional) inflight radiometric calibration using the aforementioned calibrated reflectance target, iii) a spectral shift detection and correction based on atmospheric absorption features and adaption of look-up-tables (LUTs), iv) an estimation of the total column of atmospheric water vapour. The estimation of the water vapor and aerosol optical thickness are retrieved from the average image spectrum using spectral fitting. The sensor-specific atmospheric LUT can be created based on the spectral recalibration as a subset of a more generic LUT [70]. The main processing step of DROACOR is the reflectance retrieval. The reflectance retrieval is relying on the calibrated at-sensor radiance data. The reflectance retrieval takes the relative distance between the earth and the sun, the path scattered radiance, the direct solar ground flux, the diffuse flux, and the off-nadir and diffuse ground-to-sensor transmittance into consideration. The adjacency effect from nearby objects created by the atmospheric scattering is relevant for low altitude data acquisitions specifically for the irradiance term, which is governed not only by the direct illumination but also by aerosol scattering and indirect scattering affected by the neighbouring pixels. Wavelength within regions known for high atmospheric feature absorption can be removed or interpolated to keep a continuous spectrum. After the polishing of the reflectance data (optional), the variable illumination in terrain and the bidirectional reflectance distribution function (BRDF) effects can be corrected. The goal of the topographic illumination correction is to transform the first-order bottom of atmospheric reflectance ("scaled or apparent reflectance”) to spectral albedo values (“absolute reflectance”). DROACOR handles this by a modified Minnaert approach which ensures that the effect of BRDF is not overcorrected. If a system with a large FOV is used, the BREFCOR method can be applied which corrects the BRDF effect considering the observation angle [91].

3.5. Sampling and Validation in Geological Remote Sensing Studies

The validation of data products based on RS data shows a wide range of approaches and is based on either, on-site physical sampling for subsequent laboratory-based chemistry or in situ spectral measurements. Sampling is usually approached from a geochemical perspective, though the ethical aspect of sampling is not to be discarded. While geochemical sampling for exploration or environmental studies is discussed in various textbooks (e.g., [94]) and focuses largely on randomization of sampling, avoiding sampling bias, the available sample medium (rock or soil), sampling interval, grain size fraction, sampling depth, background variation, etc. The ethics of sampling geological sites is discussed less frequently, see [95]. These practices include: the “leave no trace” approach, acquiring proper sampling permits from land managers (governmental, tribal, private), minimizing oversampling and implementing an adequate collection of metadata that leads to an organized archive of samples and associated data to surpass a project’s lifetime. Sampling for the purpose of spectral geological studies is seldom taught or considered and usually follows the standard sampling approaches applied in the field of geology. Similarly, the standard practices of sampling and their ethics are rarely discussed in RS publications even though sampling plays a frequent role in spectral field studies. Sampling for RS studies is often dictated by the requirement to validate and interpret observed surface spectral patterns. In contrast to standard geological field sampling, spectral sampling campaigns for an HSI survey serve different purposes, which include the spectral validation of pixels in the sampling location for drone- and/or satellite-based data, the provision of spectral surface sampling spots with handheld spectrometers at georeferenced ground control points (GCPs), the physical sampling for analysis in the laboratory to aid geological, geochemical, mineralogical or environmental interpretation of the HSI data, as well as the establishment of the GCPs for spectral validation of the atmospheric compensation of the collected data.

In the following, a guideline for sampling during hyperspectral UAS campaigns is described. Sampling point identification: The location of relevant sampling points must follow a pragmatic approach. Sampling points must be within the FOV of the UAS survey and represent areas large enough to be assumed spectrally homogeneous within a sufficient number of collected image pixels. A reference target should at least cover a diameter of 5x5 times the spatial resolution of a system to ensure that its centre observation is not affected by surrounding object reflectance. The margin is required to cover the optical point spread function (PSF) which reaches well into neighbour pixels for optical reasons and also to avoid influences of potential geometric blur due to sensor motion during the data acquisition, which mostly affects direct neighbour pixels. Those sampling points can be marked clearly before the start of the UAS survey to locate the sampling locations in the visual field of view (FOV) of the drone data. The sampling point must be left undisturbed between the marking of the point, during the survey, and until physical or spectral sampling is completed. The sampling point location relative to the extent of the survey should be chosen based on known geological features in the area (based on historic or recent geological mapping, known faults, etc.), known anthropogenic activity (e.g., location of waste piles in a legacy mining area), the expected spatial and spectral variation in the area and other factors influencing the mineralogical or geochemical composition of the surface that is of interest in the mapped area. Furthermore, satellite or existing airborne surveys from the area can be analysed to identify spectrally homogeneous areas for sampling before the survey. The number of sampling points and the frequency of sampling in an area will depend on the scope of the study, the accessibility of the area and associated safety precautions for the area, the timeframe in which sampling needs to commence, the access time to the target area and the availability of landowners, the budget for subsequent relevant geochemistry or mineralogy performed on the samples, and lastly the expected frequency of the spectral variation in the surface. Samples should only be taken for analysis to help with the interpretation of the spectral data and the area. Sample point markers: Samples should be marked in a manner that makes them identifiable in the resolution of the drone-borne data. This means marking the sampling spot in a relevant size with a VNIR- and SWIR-active, ideally non-permanent paint or marker. For example, if an image pixel is around 10cmx10cm in size, the sampling area must exceed one pixel, ideally more than 8 pixels, e.g., a 50x50cm square area in this scenario (25 pixels) (see Figure 3). Some studies discuss the use of chalk-based easily dissolvable paint to mark sampling points (“leave no trace approach”). Also discussed is the use of orange plastic cones, such as for infrastructure projects, as they show both a distinct feature in the VNIR and SWIR imagery and are large enough to be identified. Measuring the sample position using a GNSS system must be carefully considered, as the GNSS signal of both the UAS and the ground measuring device will likely show a couple of cm-m of deviation/error in the position, making a direct match of the sample location in the drone data non-trivial. Having visual markers that can be identified in the drone data and allowing the labelling of those using the GPS position of the sampling points is advised. To label the points accurately, they must be chosen in meaningful areas and spaced out sufficiently to tell them apart.

Metadata: The collection of metadata for each of the sampling points is crucial. This includes the accurate position of each point, a photograph of the sampling area, and a description of the sampling area. It can include further handheld devices to preliminary describe the sample, such as handheld point spectrometers, to collect spectral information, or handheld XRF or LIBS to collect relative elemental information. The physical sampling of the surface within the marked sampling area commences after the successful UAV data collection. Ideally, only the surface that contributes to the spectral signal is sampled. This means sampling only the uppermost surface in a relevant volume to allow the splitting of the sample for XRD/XRF or other analyses and hyperspectral data acquisition under laboratory conditions. For this, the surface of samples with larger grain sizes (e.g., cobble- or boulder-sized grains via the Udden-Wentworth scale) visible in the drone data (pointed upwards) should be marked for later identification. Higher-resolution photographs can also assist in this task. It also means that weathered surfaces are most likely the main contributor to the spectral signal that is being collected by the sensors but are not the main contributor in the chemical or mineralogical composition of a bulk sample. This is why a good on-site description of the samples and photographs are required to assist in the following interpretation steps. Due to the factors described above (bulk vs. surface geochemistry), the dominance of minerals in spectral signal does not always equal a mineralogical dominance in the sample. As a result, we step away from calling the discussion of these samples a “validation” sample and refer to it as another tool to interpret both the hyperspectral data and the derived products.

3.6. Examples from Different Published Studies

A review of different case studies where spectral imaging is utilized for mining applications returns a wide range of sampling practices and analytical approaches to validate and interpret the HSI data. Few publications detail best practices and sampling strategies to aid in the interpretation of the data and to connect spectral surface measurements to quantified surface geochemistry and mineralogy collected using other methods. A selection of laboratory-, ground- and airborne-based studies that deployed HSI in the raw materials industry (from exploration through to reclamation stages) was evaluated to understand the range of sampling protocols currently used in published research, and specifically focusing on the number of samples per square kilometer, both for physical and spectral samples (Appendix A, Table A1).

Exemplarily, six recent studies are described regarding their sampling and validation procedures. Spectrally active mineralogy across the near-infrared region (NIR) and the relationship to copper grade were analysed by [96] for ore sorting applications. More than 150 samples in total were scanned using a NIR active point spectrometer. The samples originate from the oversized output of a semi-autogenous grinding (SAG) mill at a South American mining operation. Sample validation was done using XRF (on 65 out of 150 samples), XRD (30 out of 150 samples) and petrographic analysis (30 out of 150 samples). XRD and XRF are based on the bulk of half a sample, and thin sections were prepared using the surface that was sampled by NIR as much as possible. Ultimately, the authors of the study were able to find and validate the presence of hydrothermal alteration minerals in a porphyry copper system. Self-identified limitations of this study include the surface-nature of NIR measurements and the relation to bulk composition using surface measurements. [58] presents a radiometric correction method for long-range targets for two study sites. At the first scanned site in Greenland, the presence of different carbonates in field samples is analyzed by portable XRF (average of 4-6 measurement spots per sample) and thin section analysis. A second study site is in the Corta Atalaya pit (Riotinto, Spain) in the Iberian Pyrite Belt (IBP). Here, samples of the main lithologies were taken for validation (number of samples not disclosed), the sample locations were recorded using handheld GPS and a portable spectroradiometer with 8mm FOV was used to acquire spectra of the representative fresh and altered rock surfaces. A validation and quality assessment of the corrected spectrally imaged data pixel is provided via visual comparison with field spectra from the measured GPS positions (±5 m). A direct validation with methods other than spectral sampling is not stated. In [21] enhanced digital outcrop models are presented for the Corta Atalaya open pit located in the IBP as well. Fifty-seven field samples were collected, representing the main exposed lithologies, and spectrally imaged under laboratory conditions to create a spectral library and define the spectral lithotypes. This spectral interpretation was validated using XRD on selected samples including a Rietveld quantification of the mineral phases [21]. From the suite of 57 field samples, a total of 15 samples were validated by XRD, and the remaining 42 samples were labelled based on visual interpretation of the samples and sample spectra. The first drone-based, fully corrected, oblique, hyperspectral SWIR survey of an outcrop was published for limestone lithology by [84]. Labels of micro- and macro-facies for the site were provided by the geologist on site. In total, 10 spectral measurements at each sampling site were collected with a handheld spectrometer. Spectra collected during the hyperspectral drone campaign were within a ±1 m area of the marked handheld spectrometer sampling sites and assumed to be within the relevant microfacies. Maps of CO₃ and AlOH features and the extent of different lithologies were not validated by any standard method (e.g., XRD), and no physical sampling was performed. [9] detail best practice procedures for multispectral satellite ASTER and hyperspectral airborne HyMap surveys from 2006-2008 carried out by teams at the CSIRO Exploration and Mining, the Geological Survey of Queensland, Geosciences Australia, and James Cook University. The large number and extent of mineral mapping products (e.g., mapping Mount Isa from around 140 ASTER scenes (500k km²) or 88 HyMap flight lines (8250 km²)) were validated by visiting over 200 field sites and collecting over 500 samples for physiochemical testing. The samples were analyzed using a variety of techniques, including XRD, whole rock chemistry, SEM, EMPA, and stable carbon isotopes. [97] investigates the area around the inactive and rehabilitated U-REE mine site (Mary Kathleen) in northern Queensland using airborne HSI HyMap data. The spatial distribution of Ca-bearing silicates and Fe-rich minerals (oxy-hydroxides) was mapped based on the USGS spectral archive endmembers. However, no on-site sampling, neither physical nor spectral, are discussed. All six studies are included in Appendix A, comparing an estimated number of physical sampling (per sqkm) and number of spectral handheld point spectra taken on site (per sqkm) for 14 studies in total Table A1.

4. Spectral Imaging Applied to the Resources Sector

Currently, drone-based HSI data has been featured in only a few published case studies; however, there is a wealth of studies based on airborne, ground-based, and satellite-borne imaging spectroscopy in the mining sector, of which their review can provide important lessons for the application and interpretation of drone-based HSI data. In the following section, relevant RS studies in the mining environment are reviewed and categorized based on different stages in the mining life cycle from exploration to the mining extraction phase, and finally closure and rehabilitation. Satellite-borne RS data, together with field-based (close-range) sensing systems (handheld hyperspectral point spectrometers) have played a crucial role during green- and brownfield exploration programs. From an exploration perspective, spectral imaging is a relatively mature technology to map spectrally active minerals and alteration footprints of hydrothermal mineral systems. However, once a deposit is discovered or enters the operational phase, spectral RS technology is not widely employed for mineral identification and mapping. In this phase, the common RS nadir perspective (bird’s eye view) of airborne and satellite sensors needs to be complemented by drone-borne and ground-based scanning systems to cover steep vertical walls that are omitted by orbital sensors.

Here, we need to emphasize that the current review primarily focuses on the applications of spectral RS data in the mineral industry and does not cover the pictorial characteristics of high-resolution satellite data commonly utilized in the mining sector for logistics applications, land-change detection, and mapping (including topographic mapping). Appendix B lists the reviewed studies and puts them in the context of the area of application, used sensor and platform and primary mapping objective (Table A2).

4.1. Exploration Sector

Hyperspectral remote sensing has traditionally been employed in mineral exploration, and the exploration sector is responsible for the majority of published case studies over the past decades [8,12,98,99,100]. This has been accomplished mainly by using airborne hyperspectral and/or satellite-borne multispectral datasets at regional scale, with a limited number of studies focused on mineral mapping at a deposit scale prior to drilling. A well-documented example in this context is the study of the Tertiary channel iron ore deposits in Rocklea Dome in the Hamersley region, Australia, covered by weathered/transported materials. The airborne hyperspectral data together with surface and subsurface drill core spectral dataset (VNIR-SWIR) were then used for 3D mineral mapping for exploration purposes [101,102]. By combining the surface and subsurface data, it was demonstrated that about 30 % of exploration drill holes were sunk into the barren ground and could have been potentially avoided if airborne HSI had been consulted for drill hole planning. Figure 4 shows one of the resulting maps suggesting areas where drill cores were sunk into barren ground. In a more recent study, [103] used high-resolution airborne data to map alteration mineralogy over a porphyry copper deposit in Iran. They integrated the results with geochemical, geophysical and geological evidence maps using the fuzzy logic to generate a drilling favourability map. The new generation of spaceborne hyperspectral imaging such as PRISMA, EnMAP, and GaoFen-5 (GF-5) are also predominantly used for mineral exploration and alteration/lithologic mapping purposes [98,104,105,106,107]. Yet, the high spectral resolution (<10nm) and high SNR (>400:1) of these imaging systems can find numerous applications within the mining sector, as outlined in this paper. A notable advantage of spaceborne imaging spectroscopy is its extensive coverage on a global scale. For instance, a single EnMAP scene encompasses approximately 1000 km2. This coverage will further expand with the arrival of next-generation satellite systems, including CHIME (ESA) and SBG (NASA), which offer wide swath passes (120 - 180km) and short revisit times of >10 days for continuous global monitoring.

In contrast to ground-based hyperspectral imaging, which is still paving its way in the exploration sector, HSI is firmly established in core scanning technology during green and brownfield exploration. Hyperspectral core scanning is a non-destructive and time-efficient solution to characterize the alteration footprints of ore deposits in 3D. Operational, industrial scanning systems are often a combination of high-resolution digital (RGB) cameras, a spectrometer capturing VNIR and SWIR spectral ranges, and a laser profilometer. The systems include either discontinuous step-and-measure data acquisition (e.g., GeoLogr by Hyperspectral imaging or the HyLogger by CSIRO), allowing for variable integration times, or a continuous, imaging data acquisition per interval (e.g., Corescan, Terraocore, GeologicAi, Plotlogic, DMT ANCORELOG) (examples here include [108,109,110,111,112,113,114].Examples of where hyperspectral core scanning has been deployed and results are published include, but not limited to, analyses of samples and downhole intervals from Cu porphyry deposits to facilitate ore sorting [115,116], porphyry deposits alteration mineral characterization [114,117,118], coal quality studies [111], Au-Cu-Zn VMS mineralization associated with volcanogenic massive sulfide (VMS) deposits [116,119,120,121], unconformity related Uranium deposits [122], analysis of basement rocks [123], characterization of REE-bearing minerals [124], and lithology discrimination in iron ore deposits [125].

4.2. Operational Mining and Extraction Sector

HSI data was used in various research studies during the extractive, active phase of mining. A drone-borne HSI survey was demonstrated to find potential failure zones related to non-structural geological factors data [80,130]. Also, the ability of HSI to readily map clay minerals and differentiate between their species is important for safety measures [131,132,133]. Imaging spectroscopy has been used to detect fugitive gas emissions (including, but not limited to CO2 and CH4 from industrial processes [134], a capability that can be naturally extended to mining facilities. Few studies using HSI have been applied to active mining environments and with the aim to integrate spectral data into the mining process in a production basis. Most of these studies rely on hyperspectral outcrop scanning as an established method to map surface mineralogy and lithology in natural rock outcrops and open pit faces. While not embraced by the mining industry at large, it enables a data-guided, close-to-face identification and sorting by mapping quality (ore, waste, contamination) before or after blasting, and before loading and transport of the rock mass. Commonly, a tripod-mounted hyperspectral camera enables the acquisition of an image via the rotation of the camera head. This is often referred to as “ground-based” HSI. While a correction to absolute or relative reflectance is still a topic of ongoing research, it often requires the placement of one or more calibration panels near the mine face or outcrop. This conventional requirement introduces different challenges, including health and safety issues associated with access to the mine face and bringing people into heavy machinery-dominated environments. This might and the lack of robust equipment to place into an active mine site might explain the reluctance to adopt HSI in mining. Standard methods of reflectance retrieval and correction approaches for ground-based HSI include empirical line correction based on ground-truth spectra, terrain and geometric correction and incident light correction using photogrammetry or LiDAR-based 3D models for close- and long-range applications [50,58,135,136,137,138,139,140,141]. Sophisticated drone-based atmospheric and geometric correction is currently limited to nadir-based data acquisition. However, in the future, this will hopefully be available for the oblique scanning of mine faces via UAS and eliminate the necessity of a person accessing the mine face with a tripod. An overview of currently published case studies on ground-based mine face scanning gives great insight for future drone-based HSI studies [51]. The capability of VNIR-SWIR data for lithological mapping in outcrops was explored for REE mineral-, crude oil and shale mapping in outcrops and samples via airborne HSI [142,143,144,145]. REE detection in outcrops in the FEN complex in Norway was performed, by identifying narrow REE absorption features using noise reduction methods [146] and can prove applicable to mining operations planned in the area. Long-range outcrop scanning (<1km distance) was performed in the eastern Alaska Range to map outcropping porphyry deposits [147,148]. Automatic mapping of mine face geology has been published in combination with LiDAR-based 3D modelling of open pit surfaces [132,149,150,151]. Lithology and iron ore mapping in an unspecified open pit mine in the Pilbara region in the VNIR spectral range was tested [149,152] as well as mapping mine faces in iron ore and gold mining using ground-based HSI [153]. Outcrops in a Lithium-pegmatite mine in Northern Portugal were scanned using ground-based and drone-based SWIR [154] and underground HSI outcrop mapping has been performed in several instances, e.g., to map clay materials to assess sealing properties in the Swiss Mont Terri rock laboratory [135] and in a Li-underground mine in Zinnwald/Cìnovec [155]. The first drone-based, fully corrected, oblique, hyperspectral SWIR survey of an outcrop was published for limestone lithology [84]. A review of close-range, ground-based hyperspectral studies from 2019 [51] covers many of these studies in great detail. White mica characterization by HSI means was reviewed in [156] in relation to different deposit types such as base metal sulfide, epithermal, porphyry, orogenic gold, sedimentary rock-hosted gold, iron oxide copper gold and unconformity-related uranium deposits. White mica chemistry and the white mica 2200nm combination feature wavelength position are meaningfully connected. The wavelength position of the white mica in Cresson Pit at the Cripple Creek & Victor mine in cripple creek, Colorado, USA is mapped in Figure 5. A general trend towards longer wavelength positions due to lower Al content with increasing depth in the pit corresponds to increased proximity to the high-temperature causative intrusion at depth.

Hyperspectral point spectrometry has shown preliminary success detecting hyperaccumulators of Nickel both for phytomining and monitoring the extent of the metal crop, detecting native hyperaccumulators and for remediation of Nickel-enriched soils [157]. A first drone-borne and tripod-based study using hyperspectral imaging was performed in the context of geometallurgy in a small active copper mine and leach operation. It concludes that HSI can become a useful tool to detect the distribution of clays and carbonates [80]. Clays are well distinguishable using HSI and are of interest in active mining environments as they affect recovery in floatation circuits, and are responsible for lower permeability in leaching, excess fines in communition, increased dust on tailings heaps and weakened rock mass. Carbonates are of interest as major short-term acid consumers in most leach operations. However, several criteria are identified as unmet at the time of the publication in 2021 to meet large-scale geometallurgical needs: rugged equipment and easy data acquisition, adequate turn-around times achieving results within days of the data acquisition, consistent data quality with differing weather conditions, robust illumination compensation methods and robust methods to reduce effects of differing water content and shadows on the data.

4.3. Closure and Rehabilitation

Mining activity is commonly associated with environmental damage due to the disposition of large volumes of rocks that are potentially harmful to the environment. Numerous abandoned mines have left an environmental legacy of AMD and toxic metals in downstream watersheds with adverse effects on flora and fauna [158,159,160]. Many of the here presented studies show success and the limits of using lower spatial resolution spaceborne or airborne data . The knowledge generated in these studies is very promising when applied to higher spatial resolution drone-based data, as it enables a more precise pattern detection for rehabilitation efforts. UAS-borne HSI sensors are not yet being used to routinely monitor tailings material and peer-reviewed research documenting UAVs being deployed over tailings sites has only recently been published. The use of UAVs over TSFs has been focused to monitoring potential subsidence [23], rather than for material identification and/or routine monitoring of environmental contamination and its impact. AMD commonly originates from an intensified process of sulfide hydro-oxidation occurring due to an increased effective surface of the crushed/milled rocks that is deposited in tailings storage facilities (TSFs), leaching pads or mine waste dams following mining and mineral processing. Spectral imaging can effectively identify contaminations and determine their sources and impacts on the water cycle and vegetation health [73,161]. Furthermore, HSI can be used to predict the potential sources of AMD discharge as well as acid neutralization capacity of naturally occurring geologic materials, including carbonate minerals and propylitic alteration assemblage minerals. In particular, it can be used to map key pH-sensitive secondary iron oxides/hydroxides/sulphates minerals (e.g., jarosite, copiapite, schwertmannite, ferrihydrite, and goethite) enabling RS pH mapping of the areas affected by AMD [158,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176].

[165,172] demonstrate the ability to map pH in hyperspectral airborne data (AVIRIS and HyMap) based on PLS regression modelling field pH and spectral data. Multitemporal pH prediction was performed for the Brukunga Pyrite Mine in South Australia and in the Leadville, Colorado mining district [177,178]. Mineral mapping in relation to soil pH was also explored for open pit mining lakes of the Sokolov region in the Czech Republic [78,179,180,181]. Though there has been some success in mapping iron features using spaceborne multispectral data [182], the majority of studies highlight the necessity for hyperspectral data to effectively differentiate pH-sensitive mineral species. This is, however, highly dependent on the spatial resolution of the data [183]. The use of same-day satellite and airborne imagery acquired over the same site was explored to understand the accuracy of map AMD proxies based on spatial and spectral resolution [59]. The study includes imagery from Sentinel-2 (S2, 443-2190nm 4 bands, 10m2 pixels, simulated) and PlanetScope (PS, 464-888nm, 4 bands, 3m2 pixels), drone-based Nano-Hyperspec (Nano, 400-1000nm, 270 bands, <50mm2 pixels) imagery in combination with the ground-based handheld VNIR-SWIR point data. Fe(III) oxyhydroxides are regarded as a proxy for AMD. A ferric (Fe(III) iron) band ratio (665/560nm) is explored as an AMD proxy for a comparison of RS Fe(III) reflectance values with ground-based values. The relationship between the two decreases with decreasing sensor spatial resolution (correlation coefficient between RS and ground-truth ferric iron: Nano < PS < S2), while mixed pixels between water and land surface increases with decreasing spatial resolution (mixed pixel occurrence: Nano > PS > S2). This demonstrates varying degrees of success in the use of non-intrusive AMD surveying tools. Figure 6 shows the Fe3+ band ratio maps for the different datasets from [59].

The effect of spatial resolution in detecting AMD waste was addressed in a simulated study indicating that 30 m spatial resolution of current (and future) hyperspectral satellite systems will be sufficient to identify hazardous zones around mining sites. The optimal resolution to achieve this aim has been determined to be 15 m [184]. An example of the application of EnMAP hyperspectral data in mapping ferric iron absorption features arising from iron oxides (hematite-goethite) and sulphates (mainly jarosite) around the Chuquicamata copper mine is shown in Figure 7.

High-resolution drone-based VNIR data was acquired alongside physicochemical field data and laboratory analyses from water and sediments along the Tintillo [163]. The acidic Tintillo River collects drainage from the Rio Tinto massive sulfide deposit with acidic water and dissolved metals (Fe, Al, Cu, Zn, amongst others) and transports them into the Odiel River with known neutral pH. Drone-based HSI VNIR data was collected and used to map the pH variation in the river, as well as any mineralogical trends in the riverbank Figure 8.

An active Cu-Au-Pyrite Skouriotissa mine and the surrounding area in the republic of Cyprus were studied via surface samples and satellite imagery. Within the REMON project (Remote Monitoring of Tailings Using Satellites and Drones) [185], the area was analysed using super- and multispectral satellite data (WorldView-2 and 3 and Sentinel-2) and ground-based scans. Representative minerals each associated with varying pH environments are mapped, representing very low (Jarosite), low (Goethite), and neutral (Hematite) pH values. Figure 9 shows the difference in mapping those minerals using data acquired by WorldView-2 (4x4 m pixels, 16-band superspectral) and Sentinel-2 (20x20m pixels, 9-band multispectral) [186]. Similarly, the mine face and surrounding area of the abandoned Cu-Au-Py mine Apliki, South of the Skouriotissa mine, was analysed using ground-based HSI and WorldView-2 data to map surface alteration patterns showing stockwork alteration [187,188,189].

Hydrochemical parameters of mining lakes can be monitored using spectral imaging [190]. The minerals of environmental concern such as asbestos and talc, which occur in ultramafic rock complexes and/or arise from anthropogenic sources such as mining activities, may be detectable using spaceborne imaging spectroscopic data [162,191]. In addition, spectral imaging can predict heavy metal content, and identify and map chemical/geochemical contents of wastes and residues [162,167,192,193,194,195,196,197,198,199]. Spectral imaging has proved to be useful for quantitative measurements of dust emissions from mining activities in nearby areas [200,201,202]. Likewise, it has been utilized to monitor the soil moisture content of tailings surfaces and predict/prevent undesirable dust emissions from the mining environment [203]. Downwind movement of acid-generating minerals could also be monitored with HSI data [170]. From an economic perspective, mine tailings can contain large quantities of high-value metals, which could be assessed (in combination with in-situ and laboratory data) for re-mining [164,204,205]. Although HSI data has provided encouraging results in mapping the mineralogy of leach pads, the spatial resolution of current (and likely near-future) spaceborne HSI sensors might not be sufficient for leach pad-scale mapping and most likely, it requires UASs with high spatial resolution [206]. Spaceborne HSI has also been used to provide classification maps of land cover around mining sites, aimed at understanding the effect of mining activities on landscape and geo-environments at regional scale [207]. During land reclamation and remediation, the technology could be used to monitor/evaluate reclamation and ecological restoration [46,169,208,209,210]. It also could be used for monitoring landscape structure, vegetation change, and soil pollution during mining activities and closure [211]. By linking the pioneer vegetation fraction derived from airborne hyperspectral data to pH, [212] devised a monitoring tool for spatial assessments of post-mining landforms. Although not making use of spectral imaging, hyperspectral point spectrometers have been successfully used to identify the presence of salts and crusts forming over tailings and the spread of AMD, as well as satellite hyperspectral data validated during field surveys. An example is a study over a tailings site in Mexico [213]. In addition, mapping of soils and tailings material from legacy mines in Nova Scotia, Canada was shown using satellite HSI ground-truthed by hyperspectral point spectrometers [214], successfully mapping the extent of tailings dispersal, depicting clays, chlorite and hydrous amorphous material (quartz).

5. Discussion

The Future of Drone-Based hyperspectral Imaging

Hyperspectral RS technology holds significant appeal for the mineral industry because unlike multispectral RS data, which only maps mineral groups collectively, hyperspectral data possesses the capability to identify individual minerals and, above that, can characterize the variations in the chemistry of specific (vector) minerals, thereby highlighting mineral zonations and lithological boundaries. This publication aimed to present the current state-of-the-art of HSI research covering various points in the mining value chain and to understand the remaining challenges and opportunities for high-resolution HSI within mining environments.

Currently, the mining industry is predominantly sample-based and relies on the interpolation of typically lab-measured parameters to acquire 2D/3D information about the mineralogy, lithology, and geochemistry of a deposit or tailings. Hyperspectral RS can close this gap by providing continuous data coverage complementing existing mapping capabilities. This makes hyperspectral RS data indispensable for a wide range of applications including monitoring land cover changes, detection and quantification of soil erosion, water quality monitoring, management of waste materials and TSFs, monitoring reclamation and restoration efforts, assessment of vegetation regrowth and ecosystem recovery, identification of areas prone to AMD and contamination, and detecting contaminant plumes and fugitive gas emissions from the mining facilities. Areas in the resource sector yet to benefit from the routine use of HSI include stockpile mapping, geometallurgy and ore processing, re-mining/ re-valorisation of TSFs and waste, close-to-face sorting to aid resource allocation and transport, surface mineralogy mapping to aid block modelling and deposit modelling, open pit lake monitoring, and active TSF monitoring. Undoubtedly, multi-scale RS data can create synergetic effects providing a complete picture of the mined commodity facilitating our understanding of mineral variability from microscopic scale (crucial for mineral processing) to regional scale (significant for development, operations, and eventually closure of mines). In this sense, spaceborne hyperspectral data is expected to complement high-spatial and/or high-temporal resolution multispectral data as well as high-spectral resolution drone data for mapping and monitoring aims.

While new hyperspectral spaceborne missions provide higher spectral resolution than ever before, a major impediment is limitations in the spatial resolution of the data that is not exceeding 30m [215]. For instance, tailings and AMD-prone surfaces have been analysed mostly via airborne or drone-borne HSI data of high spatial resolution, due to the limited spatial extent of these surfaces and associated surface patterns that need to be resolved. Although the target mineralogy is relatively well known, applying the methodology to spaceborne data is still a challenge with concerns to spatial resolution and the repeatability of the method for monitoring purposes.

Some other factors contributing to the past limited utilisation of HSI data within the mining industry include:

i.

The sparse availability of commercial turn-key solutions, outside of core scanning systems covering both the data acquisition and analysis/interpretation.

ii.

Difficulties in sensing the vertical faces of a mine (which is currently being filled by ground-based and drone-based systems). Though ground-based solutions (tripod-based) have provided data on vertical faces, their deployment in an open pit environment was at -best prototypical. Some truck-mounted systems have been deployed, suggesting safer practices at open pit sites.

iii.

The scalability of the results from regional to close-range sensing and vice versa is an ongoing topic of research and has only recently produced comparative studies and the effects of scale on data interpretation.

iv.

The inability of 3D modelling software systems (e.g., Datamine, MinePlan, Leapfrog, Vulcan) (at least until recently) to take (semi-) quantitative mineralogical data into account, deal with complex colour-coding and display legends for 4D spectral data cubes.

v.

Concerns about the repeatability of data over the highly dynamic mining sites and seasonally variable AMD. The consistency of the data over time is a challenge that has yet to be addressed fully.

vi.

Methodological limitations for time-relevant data acquisition, visualizations, and processing. Current techniques of data acquisition and processing are still labour-intensive, costly and time-consuming (especially for airborne HSI data) and heavily rely on the expertise of the interpreter.

vii.

A lack of service providers in the space to offer e.g. UAS-based HSI data collecton and interpretation to non-expert users in the mining industry.

viii.

And a shortage of well-documented and publicly available case studies with quantified, validated results and clear value propositions.

We see UAS-based HSI to kickstart a renewed potential for the mining industry. UASs provide high-spatial resolution imaging data at lower costs than hyperspectral airborne imagery. Drone data can be collected at a temporal frequency enabling mining engineers and service providers to re-collect data on an on-demand basis. This enables higher turn-around times and less planning effort than current HSI airborne campaigns. Drone-based HSI however poses its own challenges:

i.

Currently, the turnaround time from flight to data products takes >8h which is not practical within a typical shift-system at a mine site.

ii.

Commercially available SWIR UASs only operate in nadir mode and are not able to adjust the viewing angle to scan steep terrain or sloping surfaces.

iii.

Likewise, the reflectance retrieval for oblique scanning angles (i.e., mine faces or steep terrain) is an active topic of research as is the correction for atmosphere, geometry and illumination effects within near-real-time (within one shift, ca. 4h). A real-time data correction, analysis and visualisation of hyperspectral drone data is currently not possible.

iv.

Current airtimes of SWIR UASs do not meet mining demands, especially in large-scale mining operations.

v.

The setup, preparation and operation of a hyperspectral UAS, while research-ready, does not yet meet easy application standards for non-expert users.

vi.

An open issue in geological RS is the scaling effect and how the signal evolves from a microscopic to an outcrop scale, and eventually to regional scale as captured by satellite data with moderate spatial resolution. While there have been sporadic studies in the literature about the subject [109,148,216,217], the scaling effect on mineral mapping is not fully understood.

vii.

And lasty, in today’s operational hyperspectral UAS community, there are few interactions between hardware suppliers and the people in charge of processing the data. An often-under-communicated fact is that systems can show high amounts of spatial and spectral misregistration, resulting in the observation of spectral and spatial mixtures and in data analysts trying to solve the issue of non-linear spectral mixtures from the wrong end (software solutions) instead of the hardware optimization. Countless articles are trying to solve non-linear mixing of the data, typically concluding that ground physical properties are the reason for spectral mixing while hardware plays a similarly strong role [218,219,220,221,222,223,224]. Large spatial misregistration is an enormous contributor to non-linear spectral signature mixing and needs to be taken into account in all processing steps. It is proposed that only <10% pixel spatial misregistration of the spectral fidelty of each pixel is upheld [225,226]. With the advent of hyperspectral UAS systems flooding the market, including mining, spectral hardware providers are therefore encouraged to provide test reports, calibration reports and necessary guidance for their system so that both the potentials and limitations of each collected dataset can be gauged effectively and taken into account for the accuracy and robustness of derived results and maps.

viii.

Lastly, HSI data analysis is complex and data products are not easy to produce, interpret, or reproduce.

Luckily, many of these points are addressed in ongoing research projects funded by governmental or industry incentives. For example, EU- and national-funded projects are underway deploying UASs in the context of the critical mineral strategy and the European Raw Materials Initiative. One of these projects is the M4Mining project (Multi-scale, Multi-sensor Mapping and dynamic Monitoring for sustainable extraction and safe closure in Mining environments, www.m4mining.eu) funded by the European Union. This project aims to develop solutions for HSI UAS systems and identify data analysis and map standards for the mining industry in particular to enable an end-user-ready hardware and software solution. We strongly believe, that identifying the state-of-the-art in deploying spectral imaging within mining environments is the first step for these projects to identify the remaining challenges and opportunities and find solutions to fully integrate hyperspectral imaging in the mining industry as a standard tool.