Spectral Trajectory-Based Change Detection and Support Vector Machine Algorithm for Assessing Forest Landscape Disturbance Dynamics in Canada’s Athabasca Oil Sands Region

Highlights

What are the main findings?

• We present a long-term remote sensing assessment of the dynamics of forest landscape disturbances in the Athabasca oil sands region of Canada by combining 41 years (1984–2025) of data from the Landsat series.
• Our results demonstrate significant forest cover loss within the study area, an expansion of open-pit mining operations, limited forest reclamation efforts, and, critically, ongoing surface water quality degradation in the Athabasca River.

What are the implications of the main findings?

• The results indicate that mining operations in the Athabasca oil sands region are increasing at a faster rate than previous optimistic projections, intensifying pressure on the surrounding boreal forest ecosystems.

Abstract

Monitoring the long-term cumulative impacts of industrial mining and environmental changes in sensitive ecosystems remains a critical challenge for sustainable resource management. This study investigates 41 years (1984–2025) of land use and land cover (LULC) dynamics in the Athabasca oil sands region (Alberta, Canada) using the historical Landsat archive (TM, ETM+, and OLI). The methodology integrated temporal and spectral features through two complementary phases: automated spectral trajectory-based change detection and a Support Vector Machine (SVM) classification combining original bands with multiple spectral indices, including normalized difference vegetation index (NDVI), soil adjusted vegetation index (SAVI), enhanced vegetation index (EVI), transformed difference vegetation index (TDVI), bare soil index (BSI), and Modified Normalized Difference Water Index (MNDWI). This framework quantified baseline LULC, mining expansion, tailings pond footprints, and reclamation operations across a 12,146-km² study area. The spectral trajectory analysis effectively tracked multiple temporal disturbances, revealing accelerating mining activities alongside forest harvesting, wildfire stress, and insect damage. Additionally, it captured progressive and significant spectral variations in the Athabasca River’s surface water, reflecting cumulative industrial runoff and long-term environmental pressure. Concurrently, the SVM classifications (Overall accuracy ≥ 96.8% and Kappa coefficient ≥ 94%) measured major LULC shifts. Dense coniferous forests declined by 1200 km², while mixed forests decreased from 4500 km² in 1984 to a minimum of 1200 km² in 2010, before recovering to 3040 km² by 2025 due to reclamation efforts. Conversely, active mining infrastructure and tailings ponds expanded from 95 km² (1984) to 900 km² (2025), occupying 7.41% of the landscape. Ultimately, this integrated remote sensing approach captured the complex interactions between accelerating industrial development, climate stressors, and the localized progress of land reclamation, providing a scalable framework for ecosystem monitoring in heavily disturbed landscapes.

1. Introduction

Forest landscape disturbance dynamics are driven by various factors, including deforestation, agriculture, urbanization, climate change, and anthropogenic perturbations such as industrial mining and petroleum operations [1,2,3]. These factors are critical drivers affecting ecological systems and land cover change [4,5]. Accurate inventory of change in forest resources is essential for climate change effects assessment and the estimation of carbon stocks and sequestration in biomass and soils, as the forest environment plays a significant role in the global carbon budget [6,7,8,9]. Furthermore, forest management and land-use land-cover (LULC) changes significantly affect environmental, ecological and health parameters, including biodiversity, soil and air quality, water resources (surface and groundwater), and economic aspects [10,11]. Information on spatiotemporal changes in forestry is used to derive indicators of landscape and ecosystem status to assess impacts at regional and national scales [12]. In Canada, where forest area is a key indicator of environmental health and sustainable development, monitoring changes in forest cover is critical to assessing potential impacts on wildlife habitat, recreational opportunities, and ecosystem services [13,14,15,16]. Moreover, the Canadian boreal forest is an essential natural resource for climate regulation, permafrost preservation, carbon sequestration, and biodiversity, while also supporting local economies through sustainable wood products and providing a home for Indigenous and rural communities [17,18,19].

Situated in the heart of the boreal forest on the banks of the Athabasca River, northern Alberta’s Athabasca region is one of the most significantly modified landscapes in Canada. It is characterized by boreal forests on gently rolling plains, interspersed with vast wetlands [20,21]. Historically, climate change and natural disturbances, such as fires and insect infestation, have been the primary drivers of forest succession in this region [22,23]. Additionally, soil disturbances in these ecosystems directly affect landscape composition, including wetlands, herbaceous vegetation, and forest species. Agricultural practices, deforestation, urbanization, and industrialization further alter hydrological processes, ecosystem productivity, evapotranspiration, soil infiltration, and runoff [10]. Decades of oil sands mining have severely impacted local populations and ecosystems [24,25]. This environmental degradation is driven primarily by artificial tailings ponds located near critical water resources. These permeable basins store sand contaminated with toxic elements, polycyclic aromatic hydrocarbons, and heavy metals [26]. Consequently, these migrating toxins systematically degrade the surrounding boreal forest and wildlife habitats [27]. Moreover, they contaminate groundwater, emit greenhouse gases, degrade air quality, and encroach on Indigenous territories [22,28,29]. The long-term legacy of these environmental effects extends well beyond the lifespan of mining operations and represents some of the most difficult problems facing industry and regulatory bodies [24,30,31]. Assessing these effects is a critical aspect of sustainable development and natural resource management. Recently, the tension between the economic interests of the mining industry and environmental protection has intensified, necessitating reliable information to balance economic benefits against the risks of environmental degradation [32,33,34]. Ultimately, achieving an appropriate balance depends on a thorough assessment of the environmental impacts resulting from industrial mining developments.

Notably, the boreal forests of the Athabasca are under intense pressure from the oil sands industry. Driven by increasing global demand, large-scale industrial development for the extraction, processing, and transportation of petroleum products has accelerated since 1967 [35]. These industrial activities require intensive land use and high-water consumption, polluting the environment through exhaust fumes and leaks from settling ponds. These practices, combined with natural factors, lead to vegetation disruption, deforestation, hydrocarbon spills in the soil and waterways, and drastic topographical changes, complicating restoration efforts [36]. According to Lucas et al. [37], changes in the Athabasca can be gradual, seasonal, or abrupt. Gradual changes reflect vegetation disturbances caused by pollutant inputs or insect infestations [38,39]; seasonal changes are linked to the phenological cycle of vegetation and climate variability [40], while abrupt changes are often caused by forest harvesting, wildfires, mining, or urban expansion [23,41]. Consequently, managing these post-industrial environments presents major challenges, including restoring ecosystem integrity, biodiversity, and functionality. Land restoration is a critical aspect of responsible industrial development, as these activities result in profound landscape transformations that disrupt regional ecosystems, hinder forest development, and risk causing irreversible environmental damage [21]. In these oil sands regions, the assessment of environmental impacts has been the subject of numerous studies using in situ, aerial, and satellite observations. These studies have focused on the input of pollutants and trace elements from operations and their impact on wildlife, aquatic systems, and surrounding vegetation [42,43,44].

Although in situ observations and measurements are the most accurate, their spatial coverage is often insufficient to improve diagnoses and predictions in forest environments. Usually, satellite observations combined with temporal in situ measurements allow the extraction of reliable information on the impacts of climate, human footprint, and industrial development. In fact, satellite remote sensing offers remarkable potential for monitoring large areas due to the synoptic nature of the data and readily available archives of imagery. It enables major advances in understanding environmental impacts by quantifying spatiotemporal processes and states of land use, water quality, energy flux, photosynthesis, phenology, leaf area index, and forest disease [45], which cannot be detected by modelling or conventional in situ observations alone. Undoubtedly, remote sensing science and technology can effectively detect spatiotemporal variables related to trends and patterns specifically arising from industrial mining. By enhancing our knowledge of these impacts, these technologies support the development of prevention, adaptation, and mitigation scenarios for the sustainable development of the forest environment [46]. For instance, remote sensing provides industries and government agencies with effective methods to monitor active, abandoned, or rehabilitated mining sites. It enables the evaluation of pre-existing ecosystem variables, such as land cover, soil, water, and vegetation, providing key indicators to assess ecological consistency with previous conditions. Specifically, Gillanders et al. [47] demonstrated the importance of Landsat satellite imagery for monitoring land cover changes to quantify and distinguish different cover classes before land alteration and subsequent reclamation. Using NOAA-AVHRR imagery, Latifovic et al. [48] highlighted the impact of air temperature variations on vegetation cover surrounding mining operations. Furthermore, Thompson et al. [49] employed MSS data to monitor vegetation water stress caused by air pollutants near oil sands sites in Alberta. Ultimately, this technology plays a fundamental role in monitoring restoration activities by detecting environmental impacts caused by the mining industry. Indeed, many operations follow detailed plans to restore sites to their previous state, often incorporating remote sensing into their post-closure monitoring [50].

As a research hotspot in remote sensing, LULC changes detection has seen continuous development of methods and applications across a wide range of practical situations, including post-disaster assessment and land resources management [51,52,53]. Tracking and understanding these changes across vast territories at varying spatiotemporal scales is vital for resource management and decision-making [12]. To achieve this using multi-temporal satellite data, Lunetta et al. [54] emphasize three key factors for change detection: scale, complexity, and temporal frequency. These factors are essential for addressing phenology-induced errors and varying ecological regeneration rates. Grasping these aspects facilitates the selection of appropriate datasets and analytical approaches. With a record now exceeding 50 years and medium spectral, spatial, and radiometric resolutions, Landsat series sensors remain the most suitable space technology for detecting LULC changes at the level of medium-sized plots. Meanwhile, Kennedy et al. [51] provide a framework where change detection methods must meet two fundamental goals. First, methods need to better characterize long-term trends, for example, by leveraging data acquired by sensors onboard the Landsat platforms for more than half a century [39,55,56]. Second, automating change detection is essential for monitoring large areas; it reduces thresholding errors and separates real changes from spectral variability caused by geometric misregistration, variations in illumination and acquisition angles, and seasonal differences [57]. Accordingly, Kennedy et al. [51] proposed a method to automate the monitoring of changes in the forest environment by searching for ideal temporal signatures in a dense stack of Landsat imagery.

This approach relies on the principle that land-use changes alter spectral-temporal trajectories, creating distinct signatures before and after a change event. Specifically, these disturbances manifest as characteristic temporal signatures in the first shortwave infrared (SWIR1) band, corresponding to band 5 for Landsat TM/ETM+ and band 6 for OLI. This method laid the foundations for the “LandTrendrˮ (Landsat-based Detection of Trends in Disturbance and Recovery) algorithm [58], which introduced a flexible temporal segmentation approach utilizing spectral indices instead of the SWIR1 band, and continues to be refined [59]. Rather than detecting “pixel-wise” changes between discrete acquisition dates, the proposed method identifies target signatures across the complete temporal trajectory of spectral values. If an area aligns with the reference profile according to a temporal regression fit, it likely reflects the phenomenon captured by that trajectory. Since the entire trajectory sequence is considered, the method leverages the full depth of image archives to fit curves to each pixel’s history [51]. Furthermore, because detection is based on this curve fitting, thresholding is internally calibrated for each pixel rather than relying on arbitrary user-defined thresholds.

In this article, we examined the impacts of oil sands mining on LULC changes in the Athabasca region over a 41-year period (1984–2025) using Landsat time-series data. We identified and quantified initial land-use classes prior to their conversion and tracked their evolution over time. Our analysis focused on vegetation dynamics, soil and water pollution, mine tailings, and the implementation of restoration and rehabilitation operations. These objectives were achieved by analyzing images acquired by TM, ETM+, and OLI sensors on Landsat satellites. To achieve these objectives, our methodology combined temporal and spectral features through two distinct yet complementary analytical phases. First, we applied the automated spectral trajectory-based change detection method developed by Kennedy et al. [51]. This method leverages image archives based on the observation that LULC changes exhibit distinct temporal evolutions both before and after a change event, generating characteristic trajectories in spectral space. Second, several indices, including NDVI, SAVI, EVI, TDVI, BSI, and MNDWI, were derived and integrated with the original spectral bands in a support vector machine (SVM) supervised classification process to quantify LULC changes in km² across the study site between 1984 and 2025.

2. Materials and Methods

This study evaluates the dynamics of forest landscape disturbances in Canada’s Athabasca oil sands region caused by the mining industry over nearly four decades. To achieve this, it exploits the synergy between the spectral trajectory-based change detection method and the support vector machine algorithm. Figure 1 presents the methodological framework, consisting of five interrelated steps. First, we verified the quality assurance of the downloaded L2SP imagery, including radiometric calibration, atmospheric corrections, and geometric registration. Second, 30 sampling locations representing the major LULC classes in the study area were selected for spectral and trajectory-based change detection analyses. Third, six spectral indices normalized difference vegetation index (NDVI), soil adjusted vegetation index (SAVI), enhanced vegetation index (EVI), transformed difference vegetation index (TDVI), bare soil index (BSI), and modified normalized difference water index (MNDWI) were computed to highlight discrimination among classes, including vegetation covers, wetlands, open water (lakes and rivers), bare soils, and mining wastes. Fourth, these indices were integrated into an SVM classification to analyze land-use change and environmental disturbances associated with mining activities. Finally, we calculated the total area (in km²) of LULC categories affected by expanding mining activities, as well as the potential for rehabilitation and restoration from 1984 to 2025. Note that all our processing steps were carried out using the Canadian PCI-Geomatica (2014) image processing system.

Figure 1. Methodology flowchart.

2.1. Study Area

Considered the third-largest oil reserve in the world after Saudi Arabia and Venezuela, the Athabasca oil sands deposits of Northeastern Alberta cover approximately 140,000 km² (Figure 2). In this region, the three major deposit areas are Athabasca, Cold Lake, and Peace River. Approximately 80% of the available bitumen is located in the Athabasca oil sands area, north of Fort McMurray [60]. These deposits of bitumen consist of a viscous form of petroleum that requires heat or dilution with lighter hydrocarbons to flow [61,62]. The majority of the reserves are contained in the “Lower Cretaceous McMurray Formation”, which consists of a continental sequence of uncemented sands and shales overlying a Devonian limestone surface. According to Sparks et al. [61], about 10% of the bitumen in this deposit is considered economically recoverable using conventional mining methods. The open-pit mining areas are located in the Athabasca River valley, north of Fort McMurray (Municipality of Wood Buffalo), where Cretaceous formations have been eroded so that the oil sands are found less than 50 meters below the surface. The remaining 90% of the deposit is too deep to be exploited via open pits and requires steam injection processes, which makes the oil extracted from Alberta bitumen sands economically viable [63].

Figure 2. Study area: the Athabasca-McMurray oil sands, Alberta, Canada.

Furthermore, the Athabasca oil sands region is located in the Boreal Plains ecozone, which is composed of three ecoregions: the Wabasca Lowlands, the Slave River Lowlands, and the Boreal Highlands [64]. Characterized by rich and fertile soils, diverse ecosystems, and extensive wetlands, the region’s vegetation cover consists of mixed forests of varying density, ranging from medium to very dense stands. Tree species include white spruce (Picea glauca), black spruce (Picea mariana), balsam fir (Abies balsamea), aspen (Populus tremuloides), balsam poplar (Populus balsamifera), jack pine (Pinus banksiana Lamb.), and tamarack (Larix laricina [Du Roi] K. Koch). Additionally, fens and bogs with tamarack and black spruce are common throughout the region [47,65]. The region is also defined by its sub-humid and subarctic climate, prevalent Brunisolic and Luvisolic soils, and both natural and anthropogenic disturbances [66]. The Athabasca and Clearwater Rivers, as well as McClelland and Kearl lakes, define the main bodies of open water that sustain this ecosystem.

As for the region’s human landscape, according to the 2025 municipal Census, the Athabasca and McMurray regions had a population of 107,740, of whom 10% identified as Indigenous. Industrial operations significantly alter both the land and community. Figure 3 illustrates the ecological and environmental impacts of the oil sands mining industry in Athabasca. In addition to air quality degradation, these impacts include the expansion of oil extraction facilities (Figure 3a), open-pit mining sites (Figure 3b), and the destruction of forest cover (Figure 3c). Moreover, mining activities cause the pollution of surface waters (Figure 3d) and groundwater, which ultimately degrades the living conditions and health of benthic invertebrates, birds, fish, and other wildlife (Figure 3e).

Figure 3. Ecological and environmental impacts of oil sands mining in Athabasca region, showing (a) the mining industry overview, (b) an open-pit mine site, and the resulting impacts on (c) forests, (d) water resources, and (e) wildlife (Public domain images accessed via Google).

2.2. Landsat Data

Managed by NASA and the USGS, the U.S. Landsat program provides the world’s longest continuous historical record of Earth-surface observations from space, offering medium-resolution imagery on a seasonal basis for more than five decades. Since 1972, archived data from Landsat sensors, including the MSS, TM, ETM+, and OLI, have been utilized across disciplines to better understand land surface changes and the environmental impacts of climate and human activity [67,68,69,70]. The evolution of the Landsat satellite series represents a steady, five-decade progression from basic multispectral imaging to highly sensitive, advanced radiometry [71,72,73,74,75,76,77,78].

In the present study, twelve images acquired by three different sensors (TM, ETM+, and OLI) were utilized. Because the spectroradiometric values of the images depend heavily on phenological conditions, particularly in mixed forests, it was necessary to limit the influence of seasonal variations in vegetation cover. Consequently, summer images were selected within a window between 23 July and 24 September, with maximum phenological development of vegetation cover, at approximately four-year intervals to detect, map, and evaluate the environmental impact of oil sands and LULC changes in northern Athabasca between 1984 and 2025. These images are cloud-free or have minimal cloud cover (less than 5%), with any minor clouds restricted to the outer boundaries of the study site. Furthermore, the data are largely unaffected by cirrus and free from topographic shadow effects due to the region’s minimal relief. Table 1 summarizes the basic characteristics of these scenes, which were downloaded from the USGS website as L2SP products. Note that images marked with an asterisk (*) were mosaicked from three individual scenes to cover the entire study area.

Table 1. Landsat Images used in this study. (* Image mosaic generated from three images across two distinct orbits).

Satellite	Sensor	Acquisition Date	Satellite	Sensor	Acquisition Date
Landsat-5	TM	23 July 1984	Landsat-5	TM	25 August 2005 *
Landsat-5	TM	18 September 1987	Landsat-5	TM	14 September 2009
Landsat-5	TM	4 August 1989 *	Landsat-8	OLI	4 August 2014 *
Landsat-5	TM	24 September 1995	Landsat-8	OLI	12 August 2017 *
Landsat-7	ETM+	17 August 1999	Landsat-8	OLI	4 August 2020 *
Landsat-5	TM	7 August 2001	Landsat-8	OLI	25 August 2025

2.3. Data Preprocessing

The images acquired by TM, ETM+ and OLI are preprocessed into Level-2 (L2SP) products by USGS. This means the raw digital numbers were converted to the apparent reflectance at the top of the atmosphere (TOA) using gain, offset, extra-atmospheric irradiance, solar zenith angle and Earth–Sun distance (i.e., radiometrically calibrated). Then corrected for atmospheric effects, including scattering and removal of haze, clouds and water vapor, thus converting the apparent reflectances at the TOA into usable ground surface reflectance products for direct processing, allowing consistent long-term change detection. Moreover, the images were projected in the UTM coordinate system using the WGS84 geodetic reference and were geometrically corrected and orthorectified using ground control points (GCPs) and a digital elevation model (DEM) to correct topographic variations, making them suitable for pixel-level time series analyses. According to Markham et al. [79] and Crawford et al. [80], the Committee on Earth Observation Satellites (CEOS) certified that these L2SP products are Analysis Ready Data (CEOS-ARD).

Theoretically, these corrected images should appear as if they had been acquired with the same sensor under identical atmospheric conditions, illumination, and acquisition geometry [81]. However, although the images were Level-2 surface reflectance products (L2SP) ready for processing, we analyzed their spectroradiometric quality because several factors likely introduced minor spectroradiometric differences. These factors included differing spectral response profiles among the sensors (TM, ETM+, and OLI), nominal spectral differences between homologous bands, varying illumination geometries, and changing phenology (images acquired between 23 July and 24 September). Therefore, the coefficient of variation (CV = $\sigma /\overline{x}$), which measures relative variability and data homogeneity, was used to allow comparison between different uniform regions of our datasets in this analysis step. The results show that regardless of the spectral band analyzed, the recording sensor, or the acquisition period, CV values range between 2% and 5.5%, demonstrating excellent image calibration and data homogeneity. These CV values are significantly lower than the average total errors propagated and associated with Landsat preprocessing steps discussed above, which typically range from 5% to 10% depending on atmospheric conditions [81,82,83]. Moreover, Markham and Helder [84] noted that the 40-year history of Landsat data has shown consistent calibration with total uncertainties of approximately 10% or less across most sensors and bands.

Furthermore, considering that Earth’s natural surfaces do not exhibit a Lambertian spectral behavior, solar and observing zenith angles introduce angular variations in surface reflectances, known as the bidirectional reflectance distribution function (BRDF) effect. The Landsat series sensors (TM, ETM+, and OLI) are strictly nadir-aligned, meaning their optical axes point directly toward the center of the Earth (with a nominal viewing angle near 0°). Because of this narrow swath, these BRDF effects result in relatively minor reflectance variations, typically less than 6% from the image center to the edge [85]. To normalize the BRDF influence on ground surface reflectance images, a semi-empirical approach based on acquisition geometry was proposed by Qi [86]. However, Roy et al. (2016) [85] demonstrated that seasonal shifts in the solar angle led to an artificial, continuous drift in SWIR reflectance that can reach 0.02 to 0.04 in absolute value over an entire time series. Despite this, when evaluating overall error propagation, these angular effects are often overshadowed by other uncertainties. Specifically, atmospheric correction algorithms generally estimate parameters (such as aerosol optical depth) on a band-by-band basis or across coarse spatial grids, rather than resolving the true micro-variations in a heterogeneous atmosphere on a pixel-by-pixel level. Combined with sensor absolute radiometric calibration errors [79] and these atmospheric correction residuals [87], the relative impact of BRDF noise remains comparatively minimal.

Additionally, sensitivity to geometric misregistration among homologous pixels is a weakness of trajectory-based change detection methods; however, rigorous geometric registration obviously improves overall accuracy. Fortunately, misregistration can largely be overcome with sophisticated geometric correction methods [88]. For instance, according to Storey et al. [89] and Yan and Roy [90], the geometric quality and geolocation accuracy of the L2SP product are less than 4.1 m, while the achieved geometric RMSEs across all images (as indicated in metadata files) are less than or equal to 4.5 m. Furthermore, our geometric analyses across all considered images and homologous band pairs demonstrated very slight spatial variability, caused by geometric and topographic distortions, ranging from 2.5 to 4.1 m (less than 0.13 pixels). This is a highly satisfactory level of accuracy for trajectory change detection [91] and multi-sensor data analysis over time [88].

2.4. Spectral Indices

Since the emergence of remote sensing as a new discipline in the early 1970s, vegetation indices (VIs) have been developed as spectro-radiometric measurements of the spatial and temporal distribution of photosynthetically active vegetation [92]. Using the red and NIR bands, the NDVI was proposed by Rouse et al. [93] at the dawn of remote sensing. To minimize the effects caused by soil background (color and brightness) on the NDVI, the SAVI was suggested by Huete [94]. Moreover, to overcome the limitations of linearity and saturation, reduce the noise of atmospheric effects, and remove the artefacts of soil optical properties, the EVI was also developed by Huete et al. [95]. Furthermore, to describe the vegetation cover fraction independently of the soil background, to reduce the saturation problem, and to enhance the vegetation dynamic range linearly, Bannari et al. [96] established the TDVI.

Furthermore, several soil-specific indices have been proposed in the literature to distinguish bare soils from other land-use classes [97]. Rikimaru et al. [98] proposed the BSI to enhance the detection of exposed bare soil, particularly within the context of forest canopy mapping. This index was frequently used to identify uncultivated built-up areas and to map, inventory, and classify bare soil properties [99,100,101]. Moreover, it was used for excavation detection and monitoring mining activities [102], distinguishing open-pit areas from other land-use types, while also enabling the detection of waste dumping sites and disturbed land [103].

Additionally, Gaos [104] formulated the NDWI for total liquid water content within the soil-vegetation system and drought monitoring, exploiting the NIR and SWIR1 bands. While, McFeeters [105] NDWI was proposed to delimit areas of open water bodies and improve their visibility. This version was improved by Xu [106] to the Modified Normalized Difference Water Index (MNDWI), replacing the NIR band used in the NDWI with a SWIR1 band. According to Cheng et al. [103], MNDWI successfully highlights water bodies and tailings ponds within mining areas. The MNDWI and NDVI were combined for monitoring and analyzing temporal changes in mining sites in Alberta (Canada) and the tailings dam site in Brumadinho (Brazil). This integrated approach supports effective monitoring and decision-making in the mining industry and provides quantitative guidance for the safe operation and closure of mine sites [107]. These six indices (NDVI, SAVI, EVI, TDVI, BSI, and MNDWI) were selected for consideration in the present study.

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\left({\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\rho }_{Red}\right) / ({\rho }_{NIR}+{\rho }_{Red})$$

$$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=1.5\times \left({\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\rho }_{Red}\right) / ({\rho }_{NIR}+{\rho }_{Red}+0.5)$$

$$\mathrm{T}\mathrm{D}\mathrm{V}\mathrm{I}=1.5\times \left({\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\rho }_{Red}\right)/\left(\sqrt{{(\rho }_{NIR}^2+{\rho }_{Red}+0.5) }\right)$$

$$\mathrm{E}\mathrm{V}\mathrm{I}=2.5\times \left({\mathsf{\rho }}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\rho }_{Red}\right)/\left({\rho }_{NIR}+{6\times \rho }_{Red}-7.5\times {\rho }_{Blue}+1\right)$$

BSI = [(ρ_SWIR1 + ρ_Red) − ( ρ_NIR + ρ_Blue)]/[(ρ_SWIR1 + ρ_Red) + ( ρ_NIR + ρ_Blue)]

$$\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}=\left({\mathsf{\rho }}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}-{\rho }_{SWIR1}\right) / ({\rho }_{Green}+{\rho }_{SWIR1})$$

2.5. Support Vector Machine

In tandem with innovations in remote sensing technology, numerous image processing methods and algorithms have been developed to advance thematic mapping. Although many supervised classification methods exist, the Support Vector Machine (SVM) has gained significant prominence. Originally rooted in the Generalized Portrait Method proposed by Vladimir N. Vapnik in 1964 at the Institute of Control Sciences, this early framework served as the precursor to modern SVM architecture [108]. Today, SVM is widely utilized across machine learning applications for its robust capacity to handle both linear and nonlinear classification. When datasets are not linearly separable, radial basis functions (RBFs) transform the data into a higher-dimensional space to enable linear separation. This adaptability yields highly balanced predictive performance across diverse applications [109,110]. Based on statistical learning theory, SVM makes no assumptions about data distribution; instead, it employs an iterative search for an optimal decision boundary (hyperplane) to separate training data in an N-dimensional space [111].

In a complex mining environment, SVM has repeatedly outperformed alternative classification methods. For instance, the algorithm has demonstrated superior efficacy in mapping land use within open-pit quarries [112] and has accurately identified opencast iron mines and solid waste using medium spatial resolution imagery [113]. Additionally, Karan and Samadder [114] found that SVM provided the highest accuracy when monitoring the temporal evolution of open-pit coalmines. When integrated with multi-temporal change detection, the algorithm enables precise monitoring of both active surface mining and subsequent reclamation efforts [115]. To address the computational complexities of massive remote sensing datasets, Li and Narayanan [116] successfully developed an integrative SVM approach tailored specifically for mapping mining waste. More broadly, the literature underscores SVM’s high classification accuracy when processing medium-resolution data characterized by high-dimensional feature spaces [117,118,119,120]. Given these proven capabilities, SVM was selected as the primary classification method for this study. The classification framework integrated six derived indices with the six original spectral bands. Training sites for the major classes (Section 2.4) were rigorously selected, and classification was implemented following a standard procedure that assumed equal a priori probabilities. To ensure strict pixel classification, a 90% probability threshold was applied; only pixels meeting or exceeding this threshold were assigned to a class, while all others were rejected.

In the present study, to discriminate forest species from other land cover classes, including water, bare soil, and mining operations, the RBF kernel SVM classifier was trained. To neutralize the effect of extreme spectral profiles from mines and water on forest dynamics, the variables were normalized using a RobustScaler. Class skewness was corrected by applying class weights inversely proportional to their frequency. Optimal hyperparameters were determined through grid search coupled with cross-validation, identifying the combination (C = 100 and γ = 0.01) as offering the best compromise. Finally, hyperplane distances were converted into calibrated probabilities using isotonic regression.

2.6. Trajectory-Based Change Detection

In this study, we adapted an approach from Kennedy et al. [51] to analyze land-use dynamics related to mine development and rehabilitation. Spectral trajectories were derived for each sampling location from the SWIR1 band reflectance of geometrically and radiometrically normalized image stacks. The efficacy of this spectral region is well documented for soil and forest cover characterization [121,122,123], land-use change detection [47,51], and fire severity assessment [121,124]. It reliably discriminates between mature forests and exposed bare soil [125,126], maps coniferous forest structure and condition [55,127], and is sensitive to canopy moisture, foliage greenness, and photosynthetic activity [128]. Furthermore, it is highly sensitive to variations in chlorophyll and woody biomass properties [129], and effectively distinguishes between herbaceous and shrub vegetation [130].

For each hypothetical time trajectory, initial shape parameters were estimated at each sampling location and fed into a fitting function to optimize the alignment between hypothetical and observed trajectories. These trajectories represent temporal transitions corresponding to major disturbances and active rehabilitation. By comparing each location’s trajectory against seven predefined disturbance curves (Figure 4) using statistical fitting routines, we predicted the most closely resembling scenario to define the local disturbance regime. These scenarios successfully capture a diverse range of landscape conditions and transitions. For instance, stable vegetation maintains consistent spectral trajectory over time (Figure 4a), whereas mine-induced disturbance appears as an abrupt, step-like increase in SWIR reflectance due to the removal of low-reflectance vegetation and the exposure of bare soil, gravel, and rock (Figure 4b). Conversely, active revegetation is represented by long rehabilitation trajectories (Figure 4c), and mining disturbances followed by progressive recovery show initial sparse growth developing into dense, stable vegetation over time as waste materials decrease (Figure 4d). In addition to direct clearing, vegetation near mining operations frequently experiences stress and stunted growth driven by soil, air, or water contamination (Figure 4e,f). Finally, adjacent water bodies reflect pollution and tailings disposal impacts resulting from mining and industrial activities (Figure 4g).

Figure 4. Seven spectral trajectories representing different land-change scenarios related to the development and rehabilitation of mining activities in the Athabasca Oil Sands region (Adapted from Kennedy et al. [51]).

2.7. Sampling Locations Selection

For sampling locations where trajectory change detection was implemented, a reference map of land-use classes was established using the earliest Landsat-TM image (1984) and SVM classification (Figure 5). Existing ancillary data, historical vegetation maps [131], and the work of Gillanders et al. [47] supported this process. In addition to the six spectral bands, six indices (NDVI, SAVI, EVI, TDVI, BSI, and MNDWI) were integrated into the classification to improve spectral discrimination among vegetation cover, wetlands, open water, bare soil, and mining waste.

Figure 5. Thematic map of our study area based on 1984 TM-image classification, showing the extent of main land cover classes and sampling site positions used for spectral trajectory change detection.

The 30 locations selected across the study area are displayed in Figure 5. These represent various LULC classes, including dense coniferous, dense broadleaf, and both open and dense mixed forests, as well as wetland shrubs, bare soil, exposed mine land, water (rivers and lakes), and reclaimed areas. Additionally, areas of mining activity, tailings, claims, and revegetated sites were selected based on known historical activity in the imagery time series and the presence of distinct spectral signals from relatively unmixed pixels. For water sampling, eight locations were chosen along the main river passing through the middle of the study area. One was located mid-river near mining activities, while three others were situated upstream, downstream, and at the southern junction of two watercourses (the Athabasca and Clearwater rivers). Additionally, four other locations were sampled: one from a secondary watercourse southeast of the site, one from the center of a lake, and two representing clear and polluted water conditions. Regarding vegetation cover, samples were picked on both sides of the river at varying distances to assess the impact of mining on the forest and rehabilitation processes. The collection included three samples of dense mixed forest, five of dense broadleaf, one of open broadleaf, one of sparse conifer, and five of dense conifer. Finally, two other samples were added, one from a shrubby wetland and another from a tailings pond. All these sampling locations were compiled into a shapefile within a GIS environment to accurately define their spatial reference for subsequent trajectory analyses.

3. Results and Analysis

Landsat time-series imagery reveals the spatiotemporal expansion of surface mining operations within the northeastern Alberta oil sands region over a 41-year period from 1984 to 2025, with substantial development starting from 2001 onward (Figure 6). The distinct contrast between active mining and the surrounding boreal forest environment facilitates easy detection, highlighting the efficacy of Landsat data for long-term monitoring. Open-pit mining requires the wholesale removal of boreal forest canopy and organic overburden to expose underlying geologic strata, manifesting in the imagery as highly reflective, irregularly shaped, light-grey terraced excavations. Following excavation, bitumen is extracted by mixing the oil sands with hot water to form a slurry; gravitational separation then isolates the heavy sand at the bottom, leaves fine clay and water suspended in the middle, and allows aerated bitumen to float to the surface. Once the bitumen is skimmed, the remaining wastewater and fine solids are discharged into tailings ponds. In the satellite imagery, these ponds appear as large, smooth expanses ranging from dark-brown to black colors, bounded by highly contrasting, light-beige borders of artificial containment dikes constructed from coarse silica sand and gravel (Figure 6).

Figure 6. Time series of twelve Landsat images (RGB: NIR, SWIR1, SWIR2) used in the present study, acquired between 1984 and 2025 over the Athabasca-McMurray oil sands region.

Figure 5 presents the thematic LULC map of the study area, derived from baseline imagery acquired in 1984 by the Landsat-TM sensor. SVM classification was performed using historical auxiliary data, including the regional vegetation cover map established by Reid and Sherstabetoff [131], and supported by the ecological descriptions outlined by Gillanders et al. [47]. The map delineates distinct classes: dense conifer and broadleaf (trembling aspen) forests, open and dense mixed-wood forests (balsam poplar and white spruce), wetlands, shrubs, and a dominant mixed vegetation class. East of the river system and the primary mining footprint, dense conifer stands, predominantly composed of jack pine (Pinus banksiana), white spruce (Picea glauca), and black spruce (Picea mariana), constitute the primary landscape cover. The open water class delineates the Athabasca and Clearwater Rivers, as well as McClelland and Kearl lakes. Finally, two classes, representing early-stage mining operations and the initial tailings ponds, capture anthropogenic landscape modifications. This classification, and in parallel all other SVM classifications, was completed with an overall accuracy of ≥96.8% and a Kappa coefficient of ≥94% by measuring the main land cover changes. Errors of commission and omission occurred during transitions between forest classes because mixed forests combine regeneration layers (young shoots on recolonized former mining sites) with mature, dense canopies. Consequently, the SVM occasionally struggles to define a clear hyperplane between these subclasses of varying leaf density. Additionally, minor commission errors in the water class stem from spectral confusion with peatlands and wetlands containing standing water. Despite these localized variations, overall user and producer accuracies remain satisfactory (>93%).

The 30 sampling locations used to extract the spectral trajectories are displayed in Figure 5, representing distinct land cover classes. These sites were selected based on their continuous presence throughout the image time series and the various mining development and rehabilitation scenarios illustrated in Figure 4. Spectral trajectories were then generated from homogeneous and nearly pure pixels. Based on these profiles, the trajectories were classified into seven categories that match ground-truth observations across the 41-year study period: stable, stressed, modified, and rehabilitated forest sites, alongside stable, polluted, and completely transformed water sites.

3.1. Trajectories of Forest Sites (Stable and Stressed)

As expected, the method successfully captured abrupt disturbances, areas of relative stability, and ongoing rehabilitation processes. Temporal positions and spectral values of the spectral trajectory’s starting and ending vertices provide essential information regarding LULC changes in the 30 sampling locations. Figure 7 illustrates representative spectral trajectories of stable and stressed forest sites obtained from the analyzed image stack. The mixed dense forest site (25) and wetland-shrubs site (7) are located on either side of the Athabasca River, far from mining activities. Both sites are characterized by relative temporal stability over the 41 years of analysis (Figure 7a). The slight variations observed could be attributed to phenological differences, as images were acquired in mid-September in 1987, 1995, and 2009, whereas the remaining images were recorded in July and August.

Figure 7. Spectral trajectories of stable (a) and stressed forest (b) sites, (a), and spectral signatures of stressed forest sites (c,d).

Additionally, the forest area-surrounding site 25 was harvested in 1989, making it more exposed to dust from forest harvest operations and pollution from nearby mining activities. As evidenced in the 1989 imagery (Figure 6), this clear-cutting operation severely altered dense conifer sites (12 and 24) and the dense mixed wood site (19). This disturbance induced a strong spectral trajectory deviation that lasted until 1995 (Figure 7b), followed by a period of slow natural regeneration between 1995 and 2000. Between 2000 and 2010, these same sites experienced progressive stress from insect damage (Figure 7b), as documented by Pouliot and Latifovic [39]. Fieldwork by these authors revealed that insect-driven tree mortality primarily affects coniferous forests, occurring mostly in the southeast near rivers or open areas, characteristics that directly match our study site.

Within the dense forest, several locations (sites 8, 12, 18, 19, and 24) exhibit clear signs of stress (Figure 7b); specifically, coniferous site 18 and mixed-forest sites 8 and 19 are experiencing significant environmental disturbances. Site 8 (similar to site 12), located south of the McClelland Lake wetland complex, and site 19, positioned in the southeast of the study area, have both become progressively encircled by mining activities. Their spectral trajectories reflect cumulative stress from mining pollution, shifting climate regimes, and secondary compounding effects. These include drought-induced hot summer winds, secondary fires, and insect infestations, all of which exert increasing pressure on northern natural ecosystems. Situated east of Fort McMurray near the Clearwater River, site 18 initially exhibited canopy densification between 1984 and 2008 before undergoing an abrupt successional shift in 2009. Subsequent disturbances in forest areas surrounding this site have included a massive infestation due to mountain pine beetle and a series of localized wildfires [132,133]. As a result, the spectral trajectory of site 18 captures the cumulative impact of these disturbances between 2008 and 2020, as well as a significant trend of natural post-fire regeneration extending through 2025 (Figure 7b).

Wildfires are an integral natural disturbance in the boreal forest biome; however, exacerbated by severe drought and driven by strong winds, the exceptionally intense May 2016 wildfire devastated the southwestern, southeast, and southern portions of the study area. Burn scars of varying dimensions are clearly visible in the southern portion of the 2017 imagery (Figure 6). In this false-colour composite, the scars are distinguished by pronounced blue and cyan tones, which contrast sharply with the surrounding healthy vegetation represented in shades of bright and dark brown. Notably, the expansion of this wildfire effectively surrounded the city of Fort McMurray, prompting a complete regional evacuation [23]. Located on the edge of the burned areas, sites 18 and 19 experienced convection currents driven by air superheated by the fire. Influenced by these combined anthropogenic and natural factors, the spatiotemporal spectral signatures of both sites reveal distinct, band-specific behaviors (Figure 7c,d). Specifically, pronounced fluctuations were observed in chlorophyll absorption within the red band, alongside remarked variations in NIR-derived biomass density (particularly for site 19), and SWIR-indicated canopy structure and moisture content. Moreover, the 1995 spectral profile reflects a historical environmental disturbance caused by a lightning-induced forest fire near site 18 [134].

3.2. Trajectories of Polluted Water Sites

Analysis of the eight selected water samples, whether extracted from rivers or lakes across the study area, revealed that none exhibited the spectral characteristics of uncontaminated water (Figure 8). Although located away from active infrastructure, a small lake north of the study area (site 28) exhibited a consistently high SWIR-1 signal (~12%). The relative stability of this signal between 1984 and 2020, prior to the lake’s eventual excavation for an open-pit mine, indicates that the water body was chronically impacted by surrounding mining waste throughout this 36-year period. Moreover, after the water has passed through mining areas for approximately 45 km, the three sites located furthest downstream (sites 27, 29, and 30) display unstable trajectories over time (Figure 8a), indicating persistent water quality degradation.

Figure 8. Spectral trajectories (a) and spectral signatures (b,c) of polluted water sites.

Likewise, the four samples collected from the middle of the Athabasca River exhibited varying, yet significant, pollution levels (Figure 8a). These include sites 6, 13, and 14, situated in the center of the study area and flanked by mining activities on both banks, as well as site 10, located approximately 70 km downstream from the furthest extent of mining exploitation. Fluctuations in spectral response were most pronounced at sites 14 and 30. Site 14 exhibited sharp variations, particularly between 2005 and 2009, followed by a gradual decline from 2009 to 2017. Similarly, site 30 demonstrated high variability, most notably during 1985–1996 and 2005–2010 intervals. To provide an in-depth analysis of these sites’ dynamics over the 41-year study period, multi-temporal spectral signatures were generated using six spectral bands (Figure 8b,c). The resulting spectral signature curves do not reflect the typical signatures of clear water, algal absorption, or aquatic vegetation. Instead, reflectance increases toward the VNIR and SWIR regions, driven primarily by high turbidity. Typically, the primary distinction between the spectral signatures of pure and polluted water lies in how suspended contaminants alter light absorption, scattering, and reflection across different wavelengths.

Pure water typically absorbs nearly all electromagnetic radiation, exhibiting very low reflectance in the VNIR (~0.05) and dropping to near-zero in the SWIR. However, mining-related pollutants, including acid mine drainage, toxic elements from bitumen processing, and emulsified hydrocarbons, disrupt this characteristic signature. Hydrocarbons form surface films or emulsions that drastically alter the water’s optical properties. When coupled with high concentrations of suspended sediments, these pollutants induce intense light scattering across the VNIR spectrum, creating the irregular, arc-shaped profiles seen in Figure 8b,c. While individual diagnostic absorption features cannot be resolved due to the broad spectral resolution of Landsat bands, the physical presence of petroleum induces a significant macro-scale increase in the SWIR baseline reflectance. This results in unusual band values of around 0.14 in 2009 (Figure 8b), and around 0.12 for the other years. Indeed, hydrocarbons and toxic organic compounds can elevate broadband SWIR reflectance to this level depending on their physical state, thickness, and concentration. Theoretically, thick crude oils and heavy bitumen exhibit a baseline reflectance ranging between 0.10 and 0.20 in these spectral regions, whereas thin hydrocarbon films remain closer to 0.05 [135]. While these signatures are often dominated by sediment load, they also coincide with areas heavily impacted by industrial activities and potential hydrocarbon contamination. Ultimately, the contaminants highlighted in this analysis are highly consistent with the rising polycyclic aromatic hydrocarbon (PAH) levels observed in the region’s lake sediments [136], an issue historically exacerbated by the lack of an integrated regional environmental monitoring program.

3.3. Trajectories of Rehabilitated and Un-Rehabilitated Sites

As part of government rehabilitation mandates, exploited mining areas must be revegetated following the conclusion of extractive operations. This process typically involves capping disturbed areas with layers of uncontaminated sand and sediment before replanting with native species [137,138]. Spectral trajectory curves of six sites (1, 2, 4, 17, 23, and 26) indicate that rehabilitation occurred at various times between 1995 and 2025 (Figure 9a). Sites 2 and 26, originally characterized by very sparse mixed forest cover, have undergone natural densification over several decades. In contrast, sites 17 and 23 initially featured dense forest cover that was cleared for open-pit mining and subsequently revegetated; their rehabilitation sequences began in 2020 and 2005, respectively, as shown in the imagery of Figure 10 and the trajectories in Figure 9a. Sites 1 and 4, which both featured dense mixed forest in 1984, were converted into containment dikes (Figure 10) protecting a mine tailings spillway (small-scale for site 1; large-scale for site 4) before undergoing revegetation in 1995 and 2020, respectively. These results suggest that the current restored canopy density and spectral profiles are comparable to the baseline native state existing prior to industrial disturbance. Ultimately, these spectral trajectories successfully validate the a priori scenarios established in Figure 4d.

Figure 9. Spectral trajectories of rehabilitated (a), and un-rehabilitated (b) sampling sites.

Figure 10. Sampling locations for tracking significant historical LULC changes across the Athabasca oil sands region between 1984 and 2025 (i.e., the color of the small circles is solely to enhance visibility of sampling locations).

Furthermore, spectral trajectories (Figure 9b) and multi-temporal imagery (Figure 8) illustrate sites that underwent significant historical changes in LULC without subsequent rehabilitation. Initially dominated by lichens on rocky soil, site 11 transitioned to a mining area, then a tailings dike, and ultimately became a tailings pond around 2015. As previously noted, site 28 remained in its natural state from 1984 until 2020, after which it was excavated for a large open-pit mining operation (Figure 9b and Figure 10). Meanwhile, a majority of locations (sites 5, 9, 15, 16, 20, 21, and 22), initially covered by dense broadleaf, coniferous, and mixed forests, were progressively cleared for mining activities between 1984 and 2025. Site 5 transitioned to active mining in 1990, whereas sites 20 and 22 were transformed into tailings ponds in 2006. Site 16 became a tailings storage facility in 2010, and site 21 served as a tailings dike from 2010 to 2020 before transitioning into a tailings pond. Finally, sites 9 and 15 exhibited complex trajectories, initially operating as open-pit mines (1995–2020 and 2005–2016, respectively) before their eventual conversion into tailings ponds (Figure 9b and Figure 10).

3.4. LULC Areas in km²

The LULC dynamics and associated mining expansion within the study area from 1984 to 2025 are illustrated in Figure 11. A substantial expansion of mining sites and associated infrastructure occurred, including the development of tailings ponds and containment dikes (Figure 11a). The cumulative footprint of these industrial activities grew nearly tenfold, increasing from 95 Kkm² in 1984 to 900 km² in 2025, ultimately accounting for approximately 7.41% of the total study area (12,146 km²). Forest cover was significantly impacted by mining expansion, forestry operations, and natural disturbances. Specifically, the study area underwent several distinct phases of degradation, driven by intensive forest harvest (1989–1995), widespread forest loss due to insect infestations (2000–2010), and severe wildfires in May 2016. These transitions in forest composition and soil conditions underscore an unstable ecological history. The bare soil class expanded from approximately 550 km² in 1984 to 4760 km² by 2009 due to intensive forest harvesting, insect infestations, and severe wildfires. However, between 2017 and 2025, this trend reversed due to natural regeneration, with bare soil stabilizing at approximately 13.5% of the total study area (~1700 km²) by 2025 (Figure 11b).

Figure 11. Dynamic of LULC area changes (km²) over time in the Athabasca oil sands region from 1984 to 2025: (a) mining operation sites, (b) soil, wetland and vegetation cover sites.

Surrounding vegetation classes exhibited varied responses to these cumulative stressors. The dense coniferous forest class experienced a sharp decline of 1200 km², dropping from 1800 km² in 1984 to 600 km² by 2005. Meanwhile, the mixed forest class (comprising dense and open broadleaf and mixed-wood) exhibited the most severe degradation. Following cumulative disturbances, it decreased from 4500 km² in 1984 to a minimum of 1200 km² in 2010, before recovering to 3040 km² by 2025 (Figure 11b). Despite the widespread degradation of tree canopy cover across the study area, wetland-shrubs, which occupy nearly half of the study area, remained relatively stable over time. However, this class experienced significant disturbances between 2000 and 2010, corroborating the work of Montgomery et al. [139], before showing a gradual increase, particularly between 2015 and 2025 (Figure 11b). Following the wildfires and logging events, marshes spontaneously emerged in topographic depressions, progressively recovering their ecological wetland functions. Moreover, collaborative initiatives between industry stakeholders and researchers have successfully designed and developed functional, experimental wetlands [140].

4. Discussion

The Athabasca oil sands in Alberta, Canada, represent a region experiencing profound environmental disturbances and recovery processes driven by both anthropogenic activities and natural factors. These drivers include mining exploration, forest resource exploitation, urban expansion, wildfires, climate change, insect infestations, forest regeneration, and restoration efforts. Collectively, they induce major spatiotemporal shifts in LULC and forest ecosystems, with cascading consequences for water and air quality, disease risks, food security, human health, and biodiversity. Mitigating these effects effectively requires accurate spatiotemporal data to connect LULC dynamics with ecosystem properties and assess cumulative impacts. Based on Landsat series images acquired over a 41-year (1984–2025), the method of spectral trajectory-based change detection accurately tracks and highlights distinct abrupt trends of development in the study area. Through the spatiotemporal analysis of 30 selected locations, we observed clear site-specific trends: some sites remained relatively stable, while others experienced environmental stress, waterway pollution, and various anthropogenic disturbances including mining activities, forest harvesting, forest fires, insect damage, water pollution, and site rehabilitation and revegetation.

4.1. Forest Sites

Stable forest sites are located at significant distances on either side of the Athabasca River. In contrast, the observed variations at stressed sites could be attributed to phenological differences, since some images were acquired in mid-September and others in July and August. This stress is also due to dust from forest harvesting operations (sites 12, 19, and 24), mining pollution (site 25), or cumulative impacts and progressive stress (sites 18 and 19) due to climate change (secondary wildfires) and insect infestations [39,132]. This severe insect invasion caused widespread ecological stress, damaging vast tracts of forest across the Athabasca and Fort McMurray regions [141,142]. These biotic pressures were compounded by climatic factors, including mild winters and premature spring thaws. Combined with increased fire frequency, fuel management challenges, and industrial “edge-effectsˮ from oil sands operations, the ecosystem became exceptionally vulnerable to ignition. This heightened sensitivity ultimately culminated in the catastrophic wildfire of May 2016 [23,143].

4.2. Water Sites

Understanding the impacts of oil sands exploitation on water resources is vital for ensuring sustainable resource development while protecting local communities and aquatic ecosystems. Spectral trajectories and multi-temporal spectral signatures (1984–2025) of water sites extracted from the center of the Athabasca River, adjacent to open-pit mines, reveal distinct anomalies corrupted by industrial activity. Rather than reflecting the typical patterns of clear water in the VNIR and SWIR spectra, these signatures align with highly turbid and polluted waters (Figure 8b,c). In the SWIR region specifically, hydrocarbons and toxic organic compounds elevate reflectance in proportion to their thickness and concentration [135]. Consequently, these spectral signatures are dominated by sediment load, heavy metals, and potential hydrocarbon contamination resulting from mining operations. This spectral evidence of contamination aligns with the elevated levels of polycyclic aromatic hydrocarbons (PAHs) documented in regional lake sediments [136], an issue historically exacerbated by inadequate environmental monitoring.

Furthermore, these remote sensing observations are heavily corroborated by broader hydrological literature. Zhang et al. [144] note that oil sands mining and bitumen extraction influence the Athabasca River via three primary pathways: atmospheric deposition, groundwater infiltration, and localized surface runoff from disturbed landscapes. Extensive terrain alteration and disrupted hydrological connectivity facilitate both surface runoff and the subsurface seepage of oil sands process-affected water from tailings pond management facilities into nearby streams. Indeed, the environmental impacts of these effluents are evident, stored in containment ponds located dangerously close to the Athabasca River. This process-affected water contains a toxic cocktail of naphthenic acids, benzene, phthalates, polycyclic aromatic hydrocarbons (PAHs), and heavy metals such as mercury, arsenic, nickel, thallium, vanadium, chromium, and selenium [26,145,146,147,148]. Because these tailings ponds are permeable, contaminants routinely leach into the groundwater and evaporate into the atmosphere, severely damaging the surrounding boreal forest ecosystems and wildlife habitats [27].

As a result, contaminant concentrations exhibit distinct spatial patterns across the watershed. Pollutant levels are markedly higher in tributaries whose catchments have been physically disturbed by mining activities [146,147,148,149]. Within the main stem of the Athabasca River, pollutant concentrations peak near active development zones and tailings treatment facilities compared to upstream baselines (specifically at monitoring sites 6, 13, and 14), and remain elevated far downstream within the Athabasca Delta (site 10). This spatial degradation is further reflected in chronic water quality violations. In a foundational study, Kelly et al. [145] revealed that Albertan water quality guidelines for the protection of aquatic life were exceeded for cadmium, copper, lead, mercury, nickel, silver, and zinc near or downstream of industrial sites. Decades of subsequent research indicate that this trend persists; Alexander and Chambers [150] identified long-term elevations of selenium, dissolved arsenic, and total vanadium relative to pre-development baselines, while Mahoney et al. [151] recently confirmed that aluminum and methylmercury concentrations consistently continue to exceed regulatory thresholds. Eventually, decades of intensive oil sands exploitation have left a substantial legacy of environmental damage that continues to threaten both local aquatic ecosystems and downstream human populations [24,25].

4.3. Rehabilitation Sites

Characterizing and restoring forest cover in the Athabasca oil sands region requires long-term monitoring, a task uniquely suited to remote sensing methodologies. In Alberta, regulatory frameworks mandate strict rehabilitation processes for disturbed lands [138], dictating that restoration efforts must integrate site-specific landscape characteristics to return ecosystems to their original states using native species. Furthermore, re-establishing viable wildlife habitats depends heavily on critical structural and environmental factors, including stand structure, species composition, biodiversity, and water availability. However, despite successful localized rehabilitation at sites 1, 2, 4, 17, 23, and 26, mining activities inflict persistent, long-term ecological losses. To offset these impacts, Gillanders et al. [47] proposed the establishment of dedicated protected areas. To be effective, these reserves must comprehensively represent Alberta’s boreal ecosystems and be sufficiently large to support viable native populations, maintain core ecological processes, and withstand wildfire, a natural and vital driver of boreal forest dynamics.

Compounding these landscape-scale conservation challenges are severe subterranean limitations that complicate surface-level restoration. Even when surface revegetation appears successful, restored mining soils frequently exhibit profound physical, chemical, and biological discrepancies compared to their natural counterparts. In particular, Wu [152] highlights critical deficiencies in organic matter content, baseline nitrogen levels, and nitrogen mineralization rates, underscoring the long-term functional gap between reclaimed and undisturbed boreal soils. Ultimately, detecting these complex dynamics, ranging from subterranean failures to macro-scale deforestation, requires the robust multi-decadal satellite tracking previously mentioned. As illustrated by the spectral trajectories in Figure 9b and the imagery in Figure 10, sites 5, 9, 15, 16, 20, 21, and 22 experienced extensive LULC changes between 1984 and 2025 with no subsequent rehabilitation. Originally characterized by dense broadleaf, coniferous, and mixed forests, these locations were progressively cleared and converted into open-pit mines, oil extraction facilities, tailings ponds, and tailings dikes.

4.4. Landscape Disturbances

This study demonstrates that the large-scale development of Athabasca oil mining has a direct and profound impact on the surrounding boreal forest cover. This development caused extensive landscape disturbances, including the physical clearing of pristine land, severe fragmentation of contiguous habitats, and structural modification of the overall ecosystem. Over the study period, the total footprint of mining activities grew nearly tenfold, expanding from 95 km² in 1984 to 900 km² in 2025, ultimately accounting for 7.41% of the total 12,146 km² study area by the end of the period. This rapid expansion of mining infrastructure resulted in distinct, class-specific shifts across the landscape between 1984 and 2025. Coniferous forests experienced a steady net decline of 1200K km². Mixed forests (including dense and open broadleaf and mixed-wood) exhibited the most severe degradation; this class decreased sharply from 4500 km² to 1200 km² between 1984 and 2010, before undergoing a partial recovery to 3040 km² by 2025. Meanwhile, wetland-shrubs remained relatively stable over time, experiencing only minor disturbances between 2000 and 2010, followed by a gradual increase from 2015 to 2025 [139].

The structural fragmentation observed in this study has divided once-contiguous forest canopies into smaller, isolated patches. This introduces severe edge effects that degrade the health of the remaining forest margins, ultimately diminishing overall tree growth and biomass carbon intake [153]. Consequently, significant disruptions to regional carbon sequestration are certain to persist through ongoing land-use change and accelerated deforestation [154]. Beyond the direct physical removal of forest canopy cover, decades of intensive oil sands production raise ongoing concerns regarding the indirect impacts of atmospheric emissions, specifically dust, sulfur, and nitrogen, on adjacent, undisturbed forests. Interestingly, long-term monitoring highlights that these emissions have not triggered widespread structural degradation. Instead, data show a neutral to slightly positive fertilization effect, which has enhanced tree growth and understory vegetation productivity directly beneath the timber stands [44].

5. Conclusions

This study demonstrates the efficacy of synergizing long-term Landsat spectral trajectory analysis (1984–2025) with SVM classification for robust spatiotemporal monitoring in the Athabasca boreal region. Rigorous preprocessing, including radiometric calibration and topographic rectification, enabled the extraction of environmentally meaningful spectral trends. These trends successfully decoupled natural disturbances, such as wildfires and insect outbreaks, from intensive anthropogenic activities, while tracking the complete life cycle of open-pit mining, forest harvesting, and subsequent revegetation. Furthermore, extending trajectory analyses to adjacent aquatic systems like the Athabasca River highlighted the broader hydrological footprints of oil sands exploitation, capturing fluctuations in optically active water quality parameters linked to sediment loads and industrial contaminants. However, field validation is necessary to confirm these remote sensing findings regarding aquatic targets.

Over the four-decade study period, the long-term LULC analysis revealed a profound transformation of the boreal landscape. Pristine wetlands and dense forest stand experienced severe fragmentation, particularly near major development hotspots such as Fort McKay, Fort McMurray, and Kearl and McClelland lakes. Dense coniferous forests declined significantly, and mixed forests dropped sharply from 4500 km² in 1984 to a low of 1200 km² in 2010 before partial recovery to 3040 km² by 2025 due to reclamation. Conversely, active mining infrastructure and tailings ponds expanded dramatically from 95 km² to 900 km², ultimately occupying 7.41% of the landscape. These findings underscore the cumulative pressure that accelerating industrial footprints place on regional ecosystems.

Finally, the Landsat series provides invaluable spatiotemporal data for monitoring the oil sands region, particularly for landscape-scale analyses. However, while Landsat archives can effectively characterize the evolution of an ecosystem over time, many critical ecological properties cannot be resolved remotely with sufficient reliability and accuracy, and establishing a true baseline of historical ecosystem functioning remains challenging. To overcome these limitations, future research should combine Landsat-derived trajectories with ground-based measurements and high-frequency data acquired by modern very-high spatial resolution satellites (e.g., WorldView, Pléiades, and PlanetScope). Such integration will enhance the accuracy of ecological monitoring and better support evidence-based reclamation strategies in heavily impacted boreal landscapes. Moreover, with the rapid development of artificial intelligence, it is possible to transform the assessment and temporal monitoring of forest canopies. This shifts the field from manual, reactive mapping to automated, real-time digital solutions, predictive risk modeling, and multi-sensory data fusion, enabling precise tracking of canopy health, dynamics, and structural changes over several decades.