Drone-Based VNIR–SWIR Hyperspectral Imaging for Environmental Monitoring of a Uranium Legacy Mine Site

Abstract

Growing awareness of the environmental cost of mining operations has led to increased research on monitoring and restoring legacy mine sites. Hyperspectral imaging (HSI) has emerged as a valuable tool in the mining life cycle, including post-mining environment. By detecting variations in crystal structure and physicochemical attributes on the surface of materials, HSI provides insights into site environmental and ecological conditions. Here, we explore the capabilities of drone-based HSI for mapping surface patterns related to contamination dispersal in a legacy uranium-rare earth element mine site. Hyperspectral data across the visible to near-infrared (VNIR) and short-wave infrared (SWIR) wavelength ranges (400–2500 nm) were collected over selected areas of the former Mary Kathleen mine site in Queensland, Australia. Analyses were performed using data-driven (Spectral Angle Mapper—SAM) and knowledge-based (Band Ratios—BRs) spectral processing techniques. SAM identifies contamination patterns and differentiates mineral compositions within visually similar areas. However, its accuracy is limited when mapping specific minerals, as most endmembers represent mineral groups or mixtures. BR highlights reactive surfaces and clay mixtures, reinforcing key patterns identified by SAM. The results indicate that drone-based HSI can capture and distinguish complex surface trends, demonstrating the technology’s potential to enhance the assessment and monitoring of environmental conditions at a mine site.

1. Introduction

The mining industry is essential to economic development and technological advancement [1]. Simultaneously, the extensive environmental changes in various regions worldwide highlight the significant footprint of its activities [2]. In recent decades, a heightened awareness of the long-term implications of mining has drawn significant attention to the stewardship of legacy mine sites [3], which stand out as one of the major international environmental concerns stemming from mining practices [4]. Challenged by the scope and complexity of the issues involved [5], presenting data on the magnitude and nature of the problems is central for devising effective solutions for monitoring, remediation, and rehabilitation of these areas [6].

Imaging spectroscopy is a prominent discipline within remote sensing [7], extending the reach of traditional methodologies by incorporating detailed spectral analyses combined with imaging capabilities [8]. By capturing electromagnetic radiation (EMR; see Abbreviations for a complete list of acronyms) across many overlapping, narrow, contiguous spectral bands for each pixel in an image, also referred to as hyperspectral imaging (HSI) [9], variations in features from spectral signatures can be resolved, identified, and mapped [10]. In geoscience, HSI in combination with geochemical and often other mineralogical data is used for characterising materials and semi-quantifying their physicochemical attributes [10,11,12]. Data collected across the visible to near-infrared (VNIR) and short-wave infrared (SWIR) wavelength ranges (400–2500 nm) are used to examine a wide range of geological components, including iron-bearing and other transition element-bearing oxides and hydroxides, clays, carbonates, sulphates, phyllosilicates (e.g., chlorite, talc, and muscovite), and sulphosalts [10]. HSI represents a versatile tool to support multiple stages of the mining life cycle, including the stewardship of post-mining environments [7,13,14].

While satellite-based hyperspectral sensors have evolved in terms of increased spectral resolution, higher signal-to-noise ratio (SNR), and improved global accessibility, constraints related to spatial resolution limit their application where discrete information on smaller scales (i.e., centimetre to decimetre) is required [12,15]. Lower spatial resolution leads to increased mixed-pixel effects, resulting in a gradual decline in predictive accuracy [16]. Uncrewed Aerial Systems (UASs; i.e., drones) have emerged as a flexible and fast alternative for hyperspectral data acquisition with high spatial resolution, delivering data at centimetres ground sampling distance (GSD) and allowing detailed mapping of specific areas for localised studies [7,15]. The potential of UAS-based HSI has been demonstrated in multiple geological and mining applications, including mineral exploration [17,18,19], optimisation of mining operations [20,21,22,23], and environmental monitoring of mining-related areas [24,25,26]. Nonetheless, UAS-based HSI is yet to become fully operational and reach wide-ranging implementation and acceptance [15,27]. The specificity of results to case study sites and the lack of universal applicability in methods [12] require the advantages and drawbacks of different approaches to be considered on a case-by-case basis. This underscores the importance of testing, documenting, and advancing the understanding of UAS HSI capabilities across diverse contexts.

In this paper, we present preliminary results from hyperspectral UAS-based monitoring for environmental impact, in the context of ongoing research in the M4Mining project (https://m4mining.eu, accessed on 15 April 2025). This initiative aims to develop drone- and satellite-based integrated remote sensing approaches for mapping and monitoring active and legacy mining sites. This research assesses the current capabilities of UAS-based HSI in monitoring the environmental conditions of the Mary Kathleen Uranium legacy mine site, in Queensland, Australia. Previous hyperspectral investigations at the site were limited to mineral exploration around the existing open pit using airborne (HyMap, 4.5 m pixel resolution, 126 spectral bands) [28,29,30] and none assessed environmental aspects or utilised hyperspectral drone-based systems for higher spatial resolution mapping at centimetre scale. Our study targets two key areas of the site related to the site’s tailings storage facility (TSF) and evaporation pond (EP). In total, seven UAS flight missions were executed over two days over the areas of interest (AOIs). The seven flights cover a total of 20 ha. Two standard spectral processing methods (i.e., Band Ratios—BRs; and Spectral Angle Mapper—SAM) were utilise for surface mineral mapping to address site-specific aspects. The results demonstrate the ability of UAS-based HSI to highlight differences in mineral surface patterns within critical areas and establish a foundation for future research efforts and regular mapping on site. We believe current capabilities provide a robust basis for narrowing down monitoring efforts to areas of potential environmental concern, showcasing the potential for HSI to enhance the effectiveness of environmental impact assessments and remediation efforts.

2. Study Site Background and Context

The Mary Kathleen mine site (Figure 1) comprises an area of approximately 12 km², which includes the open pit, waste rock dumps, TSF, and EP. The region is characterised by high evaporation rates (ca. 2700 mm annually) and receives an average rainfall of around 500 mm per year, mostly between December and February. In these months, the area experiences intense and localised storms that cause quick runoff, temporary watercourses, and flooding [31,32]. Although sporadic, these events are recognised as one of the primary drivers of sediment transport in the area [33,34]. Sulphates are deposited in stream beds and banks along the flow path of surface waters, as well as downstream from the EP, where impermeable rock bars force groundwater to the surface [34].

Mining operations at Mary Kathleen were conducted in two phases between 1956 and 1982. Processing operations included a sequence of crushing and grinding before leaching the oxidised sulfidic ore with sulfuric acid (H₂SO₄). The metallurgical processing tailings were disposed of in a 1.3 km² large TSF, approximately 2 km north of the open pit [33], and are estimated to comprise between 5.5–7.5 million tonnes of material [31]. It hosts a variety of primary minerals, including quartz, andradite, grossular, albite, allanite, epidote, diopside, actinolite, scapolite, and chlorite. Secondary mineral phases comprise mainly gypsum and jarosite [34], along with magnetite, muscovite, anatase, rutile, hematite, and siderite [36].

Rehabilitation efforts (1982–1985) aimed at ensuring the mine site and the surroundings were left in safe and satisfactory conditions, aligned with the future land use plans for the area (i.e., cattle grazing), with no anticipated need for maintenance and minimum monitoring [37,38]. Measures considered the climate (semi-arid with short and intense rainfall periods and high evaporation rates) and soil conditions (high alkalinity capable of neutralising acid leachate and binding heavy metals) [38]. However, the neutralising potential and adsorption capacity of the tailings were found to be overestimated, which led to limited sulphide oxidation and associated acid mine drainage (AMD) in the upper part of the TSF [34]. In addition, higher-than-anticipated seepage rates have mobilised these acid, metal-rich, radioactive waters into the adjacent EP and local catchment system [32,34].

Upon discharge, the oxygenation of seepage waters leads to the hydrolysis of dissolved Fe²⁺ and consequent precipitation of Fe³⁺ phases [31]. Co-precipitation and adsorption cause the enrichment of metals (U, Y), metalloids (As), rare earth elements (REEs) (Ce, La), and radionuclides (U-235, U-238) in Fe-rich sediments, which occur immediately below the dam, in the drainage channel, and to a lesser extent in the EP [31,34]. Alkalis and alkaline-earth elements (Ca, K, Mg, Na, Sr), along with Mn, sulphate, REEs (Ce, La), and radionuclides (U-235, U-238, Ra-226, Ra-228) remain in solution until pH neutralisation and evaporation cause their precipitation in sulphate-rich evaporative sediments (Figure 2) [31,34]. Latest assessments indicate the need to monitor the long-term patterns of contamination dispersion in the area [31,32,33,34].

3. Materials and Methods

This research is based on hyperspectral data collected across the VNIR–SWIR wavelength ranges from distinct sensors and at multiple scales. These include imagery captured by a UAS with a hyperspectral camera, a LiDAR (Light Detection and Ranging) sensor, and in situ ground spectral measurements from a portable Analytical Spectral Device (ASD). The specifications of hyperspectral instruments employed for data acquisition are listed in

Table 1. Specifications of hyperspectral instruments employed for data acquisition.

Data Source/Instrument	Spectral Range	Nº. Bands	Spectral Sampling	Spatial Resolution
In situ measurements: ASD TerraSpec Halo 1	350–2500 nm	358	6 nm	Single measurements, spot size of 1 × 1 cm
UAS: HySpex Mjolnir VS-620 2	400–2500 nm	410	3 nm (VNIR); 5.1 nm (SWIR)	6–10 cm

¹ Malvern Panalytical Ltd., Chipping Norton, NSW, Australia. ² HySpex Division–Norsk Elektro Optikk AS, Oslo, Norway.

.

The UAS (Figure 3) consists of a BFD SE-8 aircraft carrying a HySpex Mjolnir VS-620 hyperspectral camera and a Velodyne (San Jose, CA, USA) LiDAR on a Gimbal H16 from Gremsy (Ho Chi Minh City, Vietnam), totalling a 25 kg payload. The Mjolnir comprises VNIR-1240 and SWIR-620 cameras built into a single chassis, operating on two optical axes co-aligned in the same field of view (FOV). Collectively, they cover the 400–2500 nm wavelength range acquiring data across 490 channels with spectral sampling between 3 nm (VNIR) and 5.1 nm (SWIR). The system also encompasses a differential Global Navigation Satellite System (GNSS), an Inertial Measurement Unit (IMU), and an Inertial Navigation System (INS). More detailed information on the system specifications is published by Koerting et al. [15]. The LiDAR is hard-mounted underneath the camera and is triggered alongside the hyperspectral imagers, allowing both the LiDAR and the hyperspectral data to receive paired time tags. Its output is used to produce a digital surface model (DSM) that contributes to precise raytracing of the HSI data to the DSM and enables both a 3D representation of the data and a precise co-registration of VNIR and SWIR data.

During the fieldwork, 13 ground control points (GCPs) were identified visually for their geological significance and visibility in the UAS data, marked visibly for the UAS FOV (Figure 4), and sampled spectrally with the ASD. The size of targets was defined to represent a diameter of at least 5 × 5 times the spatial ground sampling resolution of the system. Considering a GSD of 6 cm flying at a relative altitude of 120 m, each GCP covered an area of approximately 50 cm × 50 cm. Depending on the visible heterogeneity of GCPs, three to five spectral measurements (i.e., the reflectance spectrum from a spot with a size of ca. 1 cm × 1 cm of the surface) were acquired with the ASD to serve as a reference for the in situ conditions of selected targets.

Raw data of the sites were acquired with the hyperspectral UAS in a total of seven flights, each covering between 2–4 ha. Optimal integration times of the cameras, as well as the relationship between speed, altitude, and SNR were determined a priori in a test flight to avoid over- or undersaturation of image pixels. All missions were executed under good illumination and weather conditions, between 11:00–16:00 to ensure solar zenith angle lower than 70°. Missions were executed at a speed of 3 m/s, at 120 m height above ground level (AGL), and with a 25% overlap between flight lines, resulting in a spatial resolution of 6 cm/pixel. The flight lines are planned to overlap with neighbouring flight lines to avoid gaps in data and ensure full hyperspectral coverage of the AOIs. Considering a maximum roll of less than 5°, an overlap of 30% between flight lines was previously suggested [27], while the system provider (Norsk Elektro Optikk–Oslo, Norway) suggests a minimum overlap between flightlines of 25% (advised best practice, verbal communication by pilot in command). Figure 5 illustrates the two sites surveyed with the UAS and the respective GCP locations. In addition, a second DSM was created via photogrammetry from a DJI drone by the Geological Survey of Queensland of Site 2 and was used to assess the link between topography and visible surface patterns under investigation (Appendix A, Figure A1). Following the acquisition of hyperspectral UAS data, the surfaces of GCPs were physically sampled for geochemical and mineralogical analysis.

The spectral data of GCPs were analysed to identify mineral or mineral group endmembers in the spectra and assign a geological label to each GCP. Preliminary mineral classification labels are based on the internal ASD standard library [39]. Ultimately, considering field observations and the geological context of the area, a site-specific spectral endmember set was compiled (

Table 2. Mineral or mineral group endmembers interpreted from ground control points, with respective sources, diagnostic features, and surface patterns to which they relate.

Endmember	GCP ID 1	Spectral Features 2	Surface Pattern 3
Gypsum	MK7	1400 nm; 1580 nm; 1750 nm	Evaporitic sediments
Gypsum–Chlorite	MK18	1750 nm; 2260 nm	Evaporitic sediments
Clay–Gypsum	MK12	1750 nm; 2200 nm	Early phase of evaporites formation
Clay–White Mica–Ferric Oxide	MK15	520 nm; 860 nm; 2200 nm; 2350 nm	Clay and soil surficial layers
Clay–Carbonate	MK16	2200 nm; 2340 nm	Waste rock capping (rehabilitation)
Carbonate–White Mica	MK10	2155 nm; 2250 nm; 2340 nm	Waste rock capping (rehabilitation)
Amphibole–Epidote–Chlorite	MK19	2260 nm; 2320 nm; 2390 nm	Stable surfaces

¹ One GCP was selected to represent each endmember and compose the spectral library in case the surface mineral or mineral group was identified in more than one GCP. Duplicate endmembers not included in the table correspond to Amphibole–Epidote–Chlorite (MK8), Carbonate–White Mica (MK11), and Gypsum (MK9, MK20, MK21, MK22). ² Diagnostic absorption features identified based on [10,11,41,42]. ³ Assumed surface patterns based on the literature [31,32,33,34,36,37,38], field observations, and spectral sample analyses.

). These endmembers were used to assemble a reference spectral library for further comparison-based analyses. The ASD spectra representing the in situ spectral condition of GCPs, and the respective endmembers are displayed in Figure 6. The previously mentioned geochemical and mineralogical analysis on ground-truth field samples from the GCPs is discussed in [40].

The UAS data were pre-processed, including georeferencing, orthorectification, and atmospheric corrections to at-surface reflectance. Georeferencing and geometric rectification was performed using the PARGE software (Version 3.5, Build 341, August 2023), tracing the data from the sensor pixel positions to the available DSM, resulting in a set of three-dimensional—x, y, and z—coordinates for each hyperspectral image pixel (i.e., hyperspectral point cloud) [43]. Radiometric correction and post-processing were carried out based on the industry standard Drone and Atmospheric Correction framework (DROACOR, Version 2.0.2., Build 138, November 2023) [44,45,46]. The processing routines consider time, location, solar angles, terrain height, platform relative altitudes, sensor geometry, and pre-calculated look-up tables (LUTs). DROACOR uses a physical inversion process from at-sensor radiance to ground reflectance based on the LUTs and the libRadtran radiative transfer code [47]. Thus, the HSI data must be calibrated traceably. Some more information is gathered directly from the acquired images before the reflectance retrieval. This parametrisation consists of four steps for HSI: (i) an estimation of the aerosol optical thickness at 550 nm; (ii) an (optional) inflight radiometric calibration using calibrated reflectance targets; (iii) a spectral shift detection and correction based on atmospheric absorption features and the adaption of LUTs; and (iv) an estimation of the total column of atmospheric water vapour. The estimations of the water vapour and aerosol optical thickness are retrieved from the average image spectrum using spectral fitting to simulated at-sensor radiance spectra. The main processing step of DROACOR is the reflectance retrieval relying on the calibrated at-sensor radiance data. The wavelength regions that are known for high atmospheric feature absorption can be removed or interpolated to keep a continuous spectrum. The main log file factors of reflectance processing with DROACOR are presented for each flight and flight line in Appendix B (Figure A2, Figure A3, Figure A4, Figure A5, Figure A6, Figure A7 and Figure A8).

Spectral processing involves extracting qualitative and/or quantitative information from remotely sensed data based on albedo- and wavelength-dependent characteristics of materials [48]. In geological remote sensing, methods can be characterised into knowledge-based and data-driven approaches [49]. Knowledge-based methods incorporate user knowledge about the characteristics of absorption features (e.g., position, depth, asymmetry, width) to extract information from spectra without directly relying on reference data [49]. In contrast, data-driven approaches are mathematically based methods that derive information by treating image pixels as n-dimensional vectors (i.e., n number of spectral bands) and modelling them based on reference spectra (endmembers) of labelled or unlabelled data [49,50]. A method from each category was applied in this study to analyse the hyperspectral data. The choice of methods considered potential for resolving spectral surface patterns in the case study area, time efficiency, applicability for real-time processing, scalability between distinct spatial resolutions, complementarity of product map types, and reproducibility for multi-temporal approaches. Band Ratio (BR) and Spectral Angle Mapper (SAM) were found applicable under these considerations.

BR is a knowledge-based absorption modelling technique that produces feature intensity maps. The calculations are sensitive to the shape or gradient of absorption features within specific spectral ranges and can be used to derive relative intensity maps that highlight surface patterns, changes, and features that are indicative of mineral compositions [51]. As it relies on basic mathematical operations and spectral bands, BR is a rapid method that demands minimal computational resources. Nonetheless, it represents a simplified measure without considering spectral feature complexities or nuances, and outcomes may be ambiguous for interpretation [52]. In this study, preceding the spectral processing routines, BR was employed to map and mask vegetation deploying the Normalised Difference Vegetation Index (NDVI). Following that, other BR were selected based on the interpretation of reference targets and applied according to the published indices in the Index DataBase [53] to map the targeted surface patterns (

Table 3. Selected BRs applied in the data, their formulae with respective wavelengths used in calculations, index ranges, and targeted surface patterns.

Band Ratio 1	Formula	Index Range	Targeted Surface Pattern
NDVI	(800 nm − 678 nm) ÷ (800 nm + 678 nm)	−1.00–1.00	Vegetation covers
Reactivity	(2210 nm + 2395 nm) ÷ (2285 nm + 2330 nm)	0.66–2.11	Reactive areas
Clay Mixtures	(2168 nm × 2224 nm) ÷ (2198 nm)	1.15–4.34	Clay mixtures

¹ Band ratios selected according to [53].

).

SAM is a data-driven method that generates per-pixel classification maps using site-specific spectral endmembers as a reference. The method treats pixels and the reference spectra as n-dimensional vectors (where n corresponds to the number of spectral bands) and assesses their spectral similarity by computing the angle between them [54,55]. The smaller the angle, the more closely an unknown spectrum matches the reference spectrum. The endmember with the closest angle to the pixel spectrum is assigned as a match for a particular image pixel. The outputs are classification images where each pixel is colour-coded according to the closest-matching reference spectrum [55]. A user-defined angle threshold can be applied to reduce false positives, and image pixels that fall over this threshold for all endmembers remain unclassified [22,55]. SAM also produces a rule image per endmember, indicating the fit value (angle) of each pixel for given endmembers. This allows the identification of the “goodness” of fit of each endmember spectrum across the image. In this study, SAM was applied using ENVI software (Version 5.7, 2023) to classify the data according to the reference spectral library compiled from the UAS image pixels within the GCPs (Figure 7), labelled after the interpreted endmembers (Table 2).

4. Results

4.1. Mapping and Masking of Vegetation Covers

The surveyed areas exhibit widespread vegetation cover, affecting surface information characterisation by either covering bare earth signal or contributing to spectral mixtures as expected in rehabilitated legacy mines. The NDVI quantifies the relative reflectance of red (678 nm) and near-infrared (800 nm) light to assess the presence and condition of vegetation [56]. The index was used to map pixels influenced by vegetation across the scene, masking the affected areas. Different thresholds were tested to visually validate outputs, and NDVI-masked images were assessed by cross-referencing with true-colour hyperspectral images. An optimal threshold was established at 0.15 NDVI (i.e., >0.15 identifying vegetation), with approximately 50% of the pixels remaining available for further data processing and interpretation (Figure 8).

4.2. Data-Driven Mapping of Endmembers

SAM classified the UAS data by the seven endmembers comprising the reference spectral library. Figure 9 depicts the endmembers that represent the best match of image pixels to the reference spectra, thresholding a match at the maximum angle of 0.15 radians. Image pixels with no assigned colour (greyed out) include those that could not be matched to any spectrum within the angle threshold (i.e., unclassified) and those masked due to vegetation (NDVI > 0.15). In both sites, SAM classifies the central area as Gypsum, Gypsum–Chlorite, and Clay–Gypsum mixtures, represented by red, pink, and dark blue pixels, respectively. In addition, these endmembers are also observed beyond the EP, north of Site 2. The mapping of gypsum-related endmembers highlights the elongated pattern along the flow path of surface waters, which match the terrain’s depression zones for both sites. Meanwhile, the surrounding areas are generally classified as the endmembers representing mixtures of clay minerals and white mica (i.e., yellow, green, and light blue pixels). In contrast, SAM notably classifies part of the western portion of the EP’s barrier (Site 2) as a mixture of Clay–Gypsum.

4.3. Knowledge-Based Feature Modelling

4.3.1. Mapping of Reactive Areas

The distribution of reactive surfaces was mapped through the features present in stable, non-reactive surfaces (i.e., the “Reactivity” BR) (Figure 10). This BR compares the sum of reflectance values at two pairs of wavelengths in the SWIR range (i.e., reflectance at 2210 nm and 2395 nm are summed and divided by the sum of 2285 nm and 2330 nm) [53].

The spectral analyses indicate that endmembers associated with gypsum (i.e., suggesting reactivity) exhibit a general drop in reflectance towards the longer wavelengths within this range. Therefore, lower index values were associated with higher reactivity, as these areas exhibited weaker absorption features at the selected wavelengths. As a result, the BR captures the distinction between endmembers associated with gypsum and those that are not, thereby highlighting areas interpreted as having increased surface reactivity (Figure 11).

4.3.2. Mapping of Clay Mixtures

Diagnostic absorption features of clay minerals (i.e., at 2200 nm) identified in several instances suggest their presence in various mixtures, potentially associated with distinct patterns or processes occurring at the sites. The BR employed examines the diagnostic absorption feature of clay by multiplying the reflectance values at the shoulders (2168 nm and 2224 nm) and normalising with the reflectance at the minimum absorption point (2198 nm) [53]. In Figure 12, high index values are indicated in the west of Site 1, in the EP’s barrier and surroundings on Site 2, as well as in patches within the flow path of surface water run-off.

5. Discussion

5.1. Vegetation Cover

The study aimed to analyse surface patterns by focusing exclusively on bare-earth pixels. Therefore, quickly mapping and masking vegetation cover represents a fundamental step in the workflow. The high spatial resolution of UAS-based hyperspectral data (ca. 6 cm/pixel) allows for a detailed differentiation of surfaces, facilitating the classification of image pixels using straightforward BR calculations and enabling an effective masking of the pixels affected by vegetation. The results suggest that real-time processing of the UAS imagery should not be hindered by vegetation. Nonetheless, considering that revegetation plays a fundamental role in reversing the damage caused by mining activities [57], plans for investigating UAS HSI data focused on vegetation cover, health, and possible patterns to complement assessments of environmental conditions are underway. This will require different survey approaches, such as in situ spectral sampling of representative vegetation and the inclusion of GCPs following best practices in vegetation remote sensing. The mapping of vegetation variations and detection of stressed areas may be possible with the high spatial resolution of UAS data, and a link to the corresponding distribution of contaminants of bare earth pixels is anticipated to create a more comprehensive understanding of the sites.

5.2. Data-Driven Mapping of Endmembers

The map product of SAM was validated by comparing the classification of the GCPs in the image with their expected label, and the algorithm classifies the GCPs as expected in all seven instances. In both sites, the algorithm successfully delineates and differentiates areas visibly abundant in evaporitic sediments, presumed to represent the reactive surfaces of the sites. These areas are classified as gypsum as well as endmembers that represent mixtures of gypsum with chlorite or clay. Figure 13 emphasises a visually uniform area within Site 1 where SAM can distinguish between different endmembers, whose spectra are also displayed in the figure.

Considering the surroundings of the reactive surfaces, SAM indicates a distinction between the sites. Site 1 is dominated by the Clay–White Mica–Ferric Oxides endmember, associated with Carbonate–White Mica, whereas Site 2 contains Clay–Carbonate (see Figure 9). The endmembers that include white mica may likely represent surfaces that are less affected or transported across the sites, suggesting greater stability and non-reactiveness, as also observed in the unscathed parts of the pond’s barrier. Alternatively, Clay–Carbonate appears to concentrate at Site 2 around the evaporites within the zone of terrain depression. The occurrence of this endmember at Site 1 is mainly near Gypsum–Clay, which could represent a transition stage in the evaporite formation. This could suggest that evaporite formation may be occurring in the western part of the barrier, where the elevation is lower compared to the rest of the structure. Furthermore, SAM effectively highlights key areas for environmental monitoring, including the path and accumulation of gypsum extending beyond the EP, as observed north of Site 2. A drawback of the method was noted at Site 2, where notably large patches within the reactive areas remain unclassified despite their visual association with the surrounding sediments. This suggests that the reference spectra lack relevant signatures for the area, highlighting the importance of a high-quality spectral library for generating data-driven outputs. The choice of GCPs and spectral sampling points likely contributed to an inconclusive spectral endmember library. While with the SAM technique a portion of unclassified pixels remains, the M4Mining project is returning to the site in May 2025 to collect more UAS-based hyperspectral data to enable a multi-temporal analysis, a comparison of mapping results after several flooding events since 2023 and the collection of more spectral mineral endmembers.

5.3. Knowledge-Based Feature Modelling

The application of the BR focused on the features expected in stable, non-reactive surfaces (i.e., the “Reactivity” BR) has generated satisfactory outputs that allows mapping the distribution of reactive surfaces. Besides effectively highlighting the areas visibly dominated by evaporitic sediments, this BR captures the unstable minerals suggested also by the SAM classification to be present at the edge of the confined area as well as in the EP’s barrier. A contrasting observation between the SAM and BR outputs is noted in the west of Site 1, where mixtures of clay, white mica, and carbonate materials are expected, based on the data-driven results. Future surveys of the topography of this site are expected to provide complementary information regarding the potential impact of surface water in the area and allow for more conclusive interpretations.

With regards to clay minerals, the BR indicates high index values to areas with an expressive classification of pixels as clay-related endmembers by the SAM algorithm. These include the west of Site 1, the barrier and the EP at Site 2, as well as the patches classified as Gypsum–Clay within the reactive surfaces. Additionally, medium index values throughout both sites match the overall classification of Clay–White Mica–Ferric Oxides on Site 1 and Clay–Carbonate on Site 2. Figure 14 presents a ternary diagram as an RGB image overlaying the clay index map (red) with the inverse rule images for the “Clay–White Mica–Ferric Oxides” (green) and “Clay–Gypsum” (blue) endmembers. These SAM endmembers represent two distinct types of clay occurrences as interpreted in the spectra from GCPs. The RGB map shows that pixels mapped by the BR are also represented by one or both SAM endmembers, indicated by the general absence of the primary colour red. This suggests that the “Clay Mixtures” BR may be used as a general approach to map key locations associated with clay minerals.

6. Conclusions

The results indicate that UAS-based HSI can effectively capture and distinguish complex trends at the site, highlighting critical areas and the distinct patterns to which they are associated. The study showcases its potential to enhance the assessment and monitoring of potential environmental impacts, providing valuable HSI data to enable the identification of specific mineral compositions and surface patterns that may visually appear homogeneous. Figure 15 presents two examples in which the data allowed for the identification of otherwise indiscernible patterns, including the changes in the EP’s barrier suggesting degradation (Figure 15A) and the complex composition of a central portion in the reactive area of Site 2 (Figure 15B). This ability to highlight variations in surface properties can offer valuable insights into reactivity and contamination dispersal, representing a powerful tool to enhance monitoring and remediation efforts at the Mary Kathleen site. Overall, UAS-based HSI data add a layer of information that enhances the analysis and interpretation of surface conditions. While different analytical methods were deployed—each potentially highlighting distinct patterns observed on-site—we believe that these approaches are complementary and collectively strengthen the overall interpretation of the site. Integrating the outputs of different analyses reveals areas of overlap, and the use of multiple methods introduces a level of substitutability. For instance, if the quality of subsequent data collection is compromised, one method may perform better than another and still provide reliable results.

At the current stage of our research, we gained a better understanding of the progress of the original remediation at the Mary Kathleen mine site and where future mitigation efforts should be focused, besides identifying additional areas for in-depth and multi-temporal analysis. While the EP was supposed to confine sediments and discharge waters, the barrier’s integrity is compromised. The DSM (Appendix A, Figure A1) highlights the disruption point allowing the seepage, and the hyperspectral data products confirm the accumulation of evaporites beyond the structure, suggesting reactivity toward the catchment area. Moreover, mapping results from the SAM and BR calculations indicate an ongoing degradation on the western part of the TSF’s barrier, where a relatively lower elevation is noted through the DSM. The disruption of the structure and consequent seepage is perceived as a matter for immediate remediation, while the risks associated with the structure’s integrity justify a swift investigation. This study is part of active research within the M4Mining project, and future efforts aim to increase both the understanding of the Mary Kathleen site, as well as of the methodologies discussed in this paper for real-time deployment. The next steps will expand the available data and advance experiments with real-time approaches for UAS-based HSI data, as the site will be revisited in May 2025.