Next Article in Journal
SODE-Net: A Slender Rotating Object Detection Network Based on Spatial Orthogonality and Decoupled Encoding
Previous Article in Journal
A Method for the Extraction of Apocynum venetum L. Spatial Distribution in Yuli County, Xinjiang, via an Improved SegFormer Network
Previous Article in Special Issue
A Comprehensive Evaluation of Land Reclamation Effectiveness in Mining Areas: An Integrated Assessment of Soil, Vegetation, and Ecological Conditions
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 17
10.3390/rs17173041
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Monitoring Historical Waste Coal Piles Using Image Classification and Change Detection Algorithms on Satellite Images

by
Sandeep Dhakal
1,
Ajay Shah
2 and
Sami Khanal
1,*
1
Department of Food, Agricultural and Biological Engineering, The Ohio State University, Columbus, OH 43210, USA
2
Department of Food, Agricultural and Biological Engineering, The Ohio State University, Wooster, OH 44691, USA
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(17), 3041; https://doi.org/10.3390/rs17173041
Submission received: 2 July 2025 / Revised: 26 August 2025 / Accepted: 29 August 2025 / Published: 1 September 2025
(This article belongs to the Special Issue Application of Advanced Remote Sensing Techniques in Mining Areas)
Browse Figures
Versions Notes

Abstract

Abandoned coal mine lands, particularly waste coal piles that predate the Surface Mining Control and Reclamation Act (SMCRA) of 1977, pose significant environmental and safety risks. Unlike sites mined after SMCRA—where operators are legally mandated to conduct reclamation—there is no legal obligation for companies or individuals to restore lands disturbed before the law’s enactment. As a result, these historical sites remain largely unmanaged and understudied. This study develops a satellite imagery-based analytical workflow to identify and monitor such historical waste coal piles. Using supervised classification of Sentinel-2 imagery with four machine learning models, we identified waste coal piles in both active mining areas and regions disturbed prior to SMCRA. Among the models tested, Random Forest achieved the highest accuracy for classifying waste coal, with a precision of 86% and a recall of 77%. A subsequent time-series analysis revealed that historical waste coal piles have undergone gradual but consistent vegetation recovery since 1986, indicating a natural reclamation process. These areas showed minimal changes in disturbance magnitude, suggesting the absence of significant disturbing events. In contrast, active mining regions showed substantial disturbance consistent with ongoing operations. The combined classification and change detection approach successfully distinguished historical waste coal piles from those in active mining regions, with a precision of 78% and recall of 100%. These findings highlight the potential of remote sensing and temporal analysis to support the identification and assessment of historical waste coal piles. The proposed approach can help prioritize reclamation efforts and inform policy decisions addressing the long-term environmental impacts of historical coal mining.

1. Introduction

Coal is a crucial energy resource, with the United States (U.S.) holding approximately 22% of global reserves [1]. Since the first commercial extraction of coal in Virginia in 1701, U.S. coal production expanded significantly, reaching nearly a billion tons in 2000 [2]. However, coal extraction generates considerable waste—up to 0.4 tons per ton of coal produced—which includes low-grade coal mixed with rock, shale, and other impurities [3]. Managing this waste is a critical aspect of coal mining operations due to its environmental and logistical challenges.
Before the enactment of the Surface Mining Control and Reclamation Act (SMCRA) in 1977, there was a lack of comprehensive federal legislation regulating surface mining operations [4]. This regulatory gap led to numerous coal mining operations being abandoned without proper reclamation. Many of these sites still contain remnants of coal mining operations in the form of coal spoil, tailings, and refuse piles [5].
Abandoned mine lands often exhibit various forms of environmental degradation and pose risks to surrounding ecosystems and communities, including soil, water, and air pollution, as well as geological hazards for nearby communities [6,7,8]. For instance, the spontaneous combustion of waste coal releases a large amount of CO, SO2, NOx, and other harmful gases, which greatly affect the surrounding residents’ lives [9,10,11]. Additionally, prolonged exposure of waste coal to the surface or groundwater releases toxic elements, further polluting the ecosystem and posing a risk to human health [6,9].
After the enactment of the SMCRA, coal mining operators were mandated to reclaim areas affected by mining operations. However, no company or individual bears responsibility for reclaiming the abandoned mine lands that predate the SMCRA under state or federal laws [12]. Therefore, monitoring these historical sites is crucial to mitigate their negative impact on the ecological environment [13,14].
Conventional methods for monitoring mine lands, such as topographic surveys and photogrammetric studies, are known to be both time-consuming and labor-intensive [15]. Consequently, these techniques are not efficiently scalable for mapping large-scale mine areas. Remotely sensed satellite data has emerged as a cost-effective solution for mapping land cover over large geographic regions. Advancements in sensor and satellite technologies have made it economically feasible to acquire both historical and current spatial information about various land processes. Consequently, remote sensing data has been extensively used by the mining industry for mineral exploration and environmental monitoring purposes in recent decades.
Numerous researchers have explored image classification techniques and change detection approaches to monitor mining areas and disturbed lands [16,17,18,19]. Mao et al. developed a remote sensing model using Landsat imagery to extract coal areas in Huolinhe and Huozhou mining regions in China [20]. Zeng et al. used object-oriented decision trees to extract surface coal mining areas in Inner Mongolia [21]. Their method integrated spectral and spatial characteristics to distinguish mining areas from non-mining areas. More recently, Werner et al. attempted to identify the size and location of various mine features using satellite imagery across different countries and for different commodities like copper, gold, silver, and platinum [22]. More recent studies have advanced mining land monitoring through multi-source datasets. The MineCam dataset combined Sentinel-2 and Sentinel-1 imagery to map over 400 mining sites globally [23]. Similarly, the CUG_MISDataset provided more than 1400 image blocks and 3000 annotated instances in China, supported by a new network architecture that improved segmentation accuracy [24]. These efforts highlight rapid progress in high-precision recognition of mining land occupation and the need to explore diverse methods, as performance remains context-dependent [25].
In recent years, time series analysis of remotely sensed data has gained popularity for monitoring changes within mining areas due to its ability to capture temporal dynamics at a granular (i.e., pixel) level. These methods typically apply algorithms to indices derived from remotely sensed data that reflect vegetation or ecological status. The algorithms often involve fitting, decomposition, and breakpoint detection [26]. Hu et al. assessed the effectiveness of a decomposition algorithm called the Bayesian Estimator of Abrupt change, Seasonality, and Trend (BEAST) in detecting fine-scale human disturbances [27]. Furthermore, a combination of spatiotemporal dynamic weighting, the Best Index Slope Extraction algorithm (BISE), and the moving window methods has been explored to detect mining disturbance and impact [28,29]. Among these, the Landsat-based Detection of Trends in Disturbance and Recovery (LandTrendr) algorithm has proven to be highly effective in identifying disturbances and trends and hence has been widely used in various applications such as land use planning, forest monitoring, and carbon accounting [30]. It has also demonstrated efficiency in detecting changes in surface coal mines [31,32,33], making it the preferred algorithm for change detection.
Several studies have investigated classification and change detection approaches independently to monitor disturbances in active mining areas. However, there is a significant research gap concerning waste coal that existed before the implementation of the SMCRA. Focusing on these pre-SMCRA waste coal piles is crucial because they were created before modern environmental regulations were in place, often resulting in more severe environmental degradation [34]. These historical waste coal piles may pose long-term environmental hazards, including soil and water contamination, increased erosion, and spontaneous combustion. Monitoring these historical sites is essential for improving environmental quality and mitigating historical pollution. This study aims to fill this gap by developing an end-to-end data analytical workflow in Google Earth Engine (GEE) using satellite imagery to map these historical (pre-SMCRA) waste coal piles. The specific objectives of this study are as follows:
  • Identify and map the locations of waste coal piles using machine learning (ML) algorithms trained on satellite imagery.
  • Assess the temporal dynamics of the mapped waste coal piles, focusing on reclamation and disturbance monitoring using the LandTrendr algorithm.
  • Develop an approach to distinguish waste coal piles predating SMCRA from those in active mining areas.

2. Materials and Methods

2.1. Study Area

The study area is a historical coal mining site located within the Appalachian coal field in Harrison County, Ohio. It spans an area of approximately 100 km2, located between 80°57′–81°4′ west longitude and 40°11′–40°16′ north latitude (Figure 1). Historical imagery reveals mining activities in the area dating back to 1987, with certain sections still undergoing surface coal mining operations. The site also shows evidence of past underground coal mining, as indicated by the scattered presence of several mine-related features and waste coal piles. This site was selected due to its diverse vegetation, varied land cover types, and the presence of both historical waste coal piles and recent mining activities, making it an ideal candidate for addressing the study’s objectives.

2.2. Overview of Methodology

The data analysis workflow utilized in this study is summarized in Figure 2. The initial step involved delineating waste coal piles using supervised image classification. This step identified waste coal piles from both the active mining region and historical sites. Subsequently, the temporal assessment of mining disturbances and reclamation efforts on these piles was conducted using the LandTrendr algorithm. The objective of this step was to develop an approach to isolate historical waste coal piles predating the SMCRA from the classification map obtained from the satellite image classification.
The cloud-based GEE platform was utilized to perform both the image classification and change detection tasks. GEE is a cloud-based platform developed and made publicly available by Google [35] to facilitate efficient preprocessing and analysis of satellite imagery. Specifically, the JavaScript-based Earth Engine Code Editor was utilized to access and process Level-2 surface reflectance (SR) products from Sentinel and Landsat satellites.

2.3. Image Classification

2.3.1. Data Acquisition

This study utilized Level-2A SR products from the Sentinel-2 Multi-Spectral Imager (MSI) for the image classification task. Sentinel-2A and Sentinel-2B, part of the Sentinel-2 constellation, provide a temporal resolution of five days. A total of 20 different Sentinel-2 tiles covering the study area, each with a cloud cover of less than 10%, were used to create a cloud-free annual median composite for 2023 (January to December). These multispectral images encompass 12 bands (Table S1—Supplementary Materials), covering wavelengths from the visible to the shortwave infrared portions of the electromagnetic spectrum [36]. The spatial resolution of these bands varies, with 10 m for visible and near-infrared (NIR) bands, 20 m for red-edge and shortwave infrared (SWIR) bands, and 60 m for atmospheric correction bands.
The Level-2A SR products available in GEE have already been atmospherically corrected using the Sen2Cor algorithm developed by the European Space Agency. This correction includes radiometric calibration, terrain correction, and retrieval of SR. Cloud and shadow masking are also performed in the L2A products using the Scene Classification Layer (SCL), with high-probability clouds, cirrus, and shadows removed. Following these standard preprocessing steps, the images were resampled to 10 m and composited into a cloud-free annual median mosaic for 2023.

2.3.2. Training Samples

While the primary objective of this study is to identify areas of raw and waste coal, four additional dominant land cover classes were included in the classification scheme. This multi-class approach was adopted instead of grouping all non-coal areas into a single category, which allows for a more comprehensive evaluation of model performance across all major land cover types present in the study area. The total number of pixels collected for each class during model development is provided in Table 1. The six classes were selected based on prior knowledge of prevalent land-cover types within the study area. To ensure balanced representation during model training, balanced stratified sampling was applied [37]. This approach involved oversampling minority classes (such as raw and waste coal areas) and undersampling the majority classes to create a more balanced training dataset. This sampling strategy prevents the model from being biased toward dominant land cover types and ensures adequate representation of the coal areas that are the primary focus of this study.
Publicly available thematic maps were used as reference layers to guide the identification of major land-cover classes. The 2019 National Land Cover Data layer (NLCD), obtained from the U.S. Geological Survey (USGS), provided detailed land cover information at a 30 m spatial resolution [38]. This dataset facilitated the extraction of training samples for Forest and Grassland. Additionally, the 2018 Global Human Settlement Layer (GHSL), developed by the European Commission’s Joint Research Centre, aided in identifying built-up regions within the study area [39]. Training samples for water were created using the Normalized Difference Water Index (NDWI) threshold above zero. The Abandoned Mine Land Inventory System (AMLIS) and Resource Conservation and Recovery Act (RCRA) gob pile database, obtained from the Department of Mineral Resource Management (DMRM), Ohio Department of Natural Resources (ODNR), were utilized as reference layers to create training samples for waste coal piles. Additional training samples for waste coal were manually created by cross-referencing high-resolution basemap imagery and the abandoned surface and underground coal mine database. Furthermore, the ground-truth locations of raw coal piles were determined through a site visit to an active mining and coal storage facility near Cadiz, Ohio. During this visit, a WingtraOne (Wingtra AG, Zurich, Switzerland) UAV was used to survey the site and generate a high-resolution orthomosaic. Raw coal piles were then manually delineated from the orthomosaic and incorporated as ground-truth samples in the training dataset. A detailed description of this survey is available in the corresponding Master’s thesis [34].

2.3.3. Image Preprocessing

Raw and waste coals often exhibit spectral signatures similar to built-up areas, particularly roads and highways. To prevent potential misclassification of these land use categories and accurately locate raw and waste coal piles, road and highway features were masked out based on the transportation layer obtained from the U.S. Census Bureau [40]. Specifically, a 10 m buffer was applied on both sides of the line feature of roads and highways to create a mask layer (Figure 1d). To ensure equal weighting of spectral bands in satellite images during image classification, feature scaling was implemented using a unit scale function (0–1).
The spectral profiles of the six land cover classes considered in this study demonstrate distinct surface reflectance (Figure 3), with raw and waste coal showing a similar trend with minor differences in reflectance magnitude. Therefore, the classification algorithms are expected to effectively differentiate raw and waste coal from other land cover classes.

2.3.4. Classification Models

In this study, four ML models, including k-nearest Neighbors (kNN), Gradient Boosting (GB), Random Forest (RF), and Support Vector Machine (SVM), were implemented in GEE to identify the selected land cover classes. The kNN classifier assigns samples based on the majority vote of their k-nearest neighbors [41]. GB sequentially builds decision trees to minimize classification error by correcting residual errors iteratively [42]. RF constructs multiple decision trees using random subsets of training data, enhancing robustness and measuring variable importance through impurity reduction [43,44]. SVM identifies optimal decision boundaries for classification by maximizing margins between support vectors [45].
The hyperparameters of each model (Table S2—Supplementary Materials) were optimized to achieve the best classification performance. Tuning was performed iteratively using a grid search approach within GEE, with candidate values selected from both prior studies and GEE defaults. The final parameter settings (Table 2) reflect the configurations that provided the best balance between classification accuracy and computational efficiency.

2.4. LandTrendr

Landsat time-series data from 1986 to 2023 were utilized to monitor disturbances and reclamation efforts within the study area. The LandTrendr algorithm, configured with parameters listed in Table 3, was employed for this purpose. LandTrendr leverages time series of Landsat imagery to detect and characterize disturbances and subsequent recovery in terrestrial ecosystems. The core functionality of LandTrendr involves fitting a series of temporal segmentation models to Landsat pixel data, which helps in identifying periods of abrupt change and gradual recovery across large landscapes [30]. By doing so, it provides valuable insights into the dynamics of land cover change, allowing for the assessment of environmental impacts, management practices, and conservation efforts.
Coal mining operations destroy vegetation and alter soil structure, making vegetation monitoring essential for assessing ecological changes in mining areas [46]. Vegetation monitoring relies on commonly used vegetation indices, including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Visible Atmospherically Resistant Index (VARI) [47]. Among these indices, NDVI, calculated as the normalized difference in NIR and red bands, has been widely recognized as the most effective index for monitoring vegetation disturbance in mining areas [26]. Consequently, the intra-annual 95th percentile NDVI was selected as the index for the pixel-based time series segmentation algorithm. This metric reduces variability caused by short-term climatic fluctuations, soil moisture dynamics, and phenological cycles, while better capturing peak vegetation conditions and minimizing noise that might obscure long-term recovery trends.
LandTrendr performs continuous segmentation fitting on the time series of each pixel under set parameters to obtain the point with the largest deviation, known as the breakpoint. The process involves two phases: iterative detection of points where notable change occurs and refinement of the selected points using an angle-of-change criterion. Once the maximum number of vertices is reached, straight-line segments are fitted to observed NDVI values from the first to the last year, using simple regression or point-to-point fitting. These breakpoints are then used to determine the timing and location of mining disturbances and reclamation.
This study primarily investigates the temporal dynamics of two land use classes: raw and waste coal. Waste coal piles, remnants of mining activities predating the SMCRA, are expected to show a gradual increase in NDVI values due to natural vegetation growth over time, particularly during the period of highest vegetation growth (20 June to 20 September). In contrast, raw coal within active mining areas is likely to exhibit abrupt or variable changes in NDVI values due to ongoing mining operations, such as excavation and removal of vegetation. This distinct characteristic helps distinguish active mining regions from historical waste coal piles.

2.5. Accuracy Assessment

The accuracy of classification maps, obtained from image classification alone and image classification combined with a change detection approach, was evaluated using a confusion matrix [48]. A total of 10,000 randomly generated pixels was used to assess the accuracy of different models. This matrix provides two significant single-value metrics for understanding classifier performance: overall accuracy (OA), representing the total number of correct predictions across all classes, and the Kappa statistic, which measures the agreement between classification results and true values beyond chance.
For specific classes, performance was assessed using precision, which indicates the probability that a predicted class value is correct, and recall, which represents the probability that a true class value is correctly classified [49]. Additionally, the macro F1-score was used as a comprehensive assessment of a model’s ability to accurately classify across the entire range of classes. It is computed by calculating the F1-score for each class individually and then taking the average of these scores. This approach treats all classes equally, regardless of their prevalence in the dataset, making it particularly useful in scenarios with imbalanced class distributions.

3. Results and Discussion

3.1. Image Classification

Across the study area, all four ML models demonstrated satisfactory performance in classifying six land use classes (Figure 4), with Kappa coefficients and overall accuracies exceeding 80% and 88%, respectively. These metrics indicate the models’ effectiveness in differentiating most land cover classes using the 12-band multispectral imagery. Table 4 provides further insights into model performance across individual classes.
For dominant land cover classes such as Grassland and Forest, the RF and SVM models demonstrated superior performance, achieving precision and recall scores greater than 92%. These scores indicate that RF and SVM models accurately classified these classes with minimal errors, translating to few false positives and false negatives. On the other hand, the kNN and GB models occasionally misclassified ground truth Grassland pixels as Forest, resulting in slightly lower precision scores of 88.76% and 89.50%, respectively. Notably, the classification map produced by the GB model exhibited increased noise and misclassifications (Figure 4 and Figure 5). In contrast, the classification maps generated by kNN, SVM, and RF appeared visually similar, necessitating a metrics-based evaluation to discern the differences in their performance.
The RF model achieved perfect precision and recall scores of 100% for the Raw Coal class, demonstrating the highest classification performance, with no misclassifications. In contrast, kNN, GB, and SVM models misclassified some Raw Coal pixels as Forest or Built-up, resulting in lower recall scores. Notably, the GB model also misclassified Grassland and Forest pixels as Raw Coal, further reducing its accuracy.
For the Waste Coal class, RF maintained strong performance with precision and recall scores of 86.15% and 76.71%, respectively. Other ML models performed significantly worse, with GB and SVM exhibiting precision scores below 50%, and both kNN and GB had recall scores below 50%. These errors were primarily due to confusion between Waste Coal and Forest, Built-up, and Grassland classes, highlighting the models’ limited ability to distinguish this complex class. The superior performance of RF is largely due to its ensemble design, which combines the outputs of many decision trees to reduce variance and improve stability [43]. This structure allows RF to capture subtle, non-linear spectral patterns that help distinguish Waste Coal from spectrally similar classes. In contrast, models such as GB and kNN are more sensitive to overlapping feature distributions, which reduces their ability to separate these challenging classes.
Overall, the performance of ML models varied across land cover types and was notably impacted by class imbalance, with the GB model most impacted. While all models accurately classified dominant classes such as Grassland and Forest, RF consistently outperformed the others in identifying Raw Coal and Waste Coal.
Table 5 lists area estimates alongside performance metrics for each ML model. While kNN, RF, and SVM produced comparable area estimates for dominant classes (e.g., Grassland ~59 km2, Forest ~36 km2, and Water ~2 km2), there were slight variations in the location and extent for minority classes such as Raw and Waste Coal areas (Figure 5). The RF model demonstrated superior performance, achieving an OA of 95.66% and a macro F1-score of 91.20%, and was therefore selected to generate the final classification map used in the change detection analysis.
These findings are consistent with the prior studies focused on mapping mining-related features. For instance, Demirel et al. [50] evaluated the performance of the SVM classifier in monitoring surface coal mines in Goynuk, Turkey, using high-resolution satellite imagery, achieving an OA of 96.9% and 94.4% for the Ikonos (1 m spatial resolution) and Quickbird (0.6 m spatial resolution) images, respectively. In another study, Mao et al. [20] developed a remote sensing model based on the normalized difference in bands 4 and 5 of the Landsat imagery to extract coal areas in China. The authors also validated their approach with an RF model yielding comparable performance. More recently, Zeng et al. [21] developed an object-oriented decision tree to extract surface coal mining features, reporting an OA of 97.07% for a coal mine site in China.

3.2. Variable Importance

Figure 6 illustrates the relative importance of Sentinel-2 bands in differentiating Raw and Waste Coal from other land cover classes using the RF model on the training dataset. The key variables that played a significant role in the model performance include B1 (Coastal Aerosol), B9 (Water Vapor), B2 (Blue), and B12 (SWIR 2). Conversely, B6 (Red Edge 2), B7 (Red Edge 3), B3 (Green), and B8 (NIR) showed lower influence.
The higher importance of B1 and B9 may seem unexpected, as these bands are commonly associated with atmospheric effects rather than surface properties. However, because this study used the Sentinel-2 Level-2A SR product, which incorporates atmospheric correction through Sen2Cor, the influence of atmospheric interference is substantially reduced. As a result, the importance of B1 and B9 is more plausibly attributed to their ability to capture genuine spectral contrast between coal-related materials and other surfaces. Specifically, raw and waste coal piles often contain fine particulate matter and elevated moisture content, properties that these bands are effective at detecting. Additionally, B2 (Blue) and B12 (SWIR 2) bands are effective in identifying specific mineral compositions and moisture levels [51], which are characteristic of waste coal piles. In contrast, bands such as B6 (Red Edge 2), B7 (Red Edge 3), B3 (Green), and B8 (NIR) showed lower influence, as they are more suited for vegetation monitoring [51] and less sensitive to the unique spectral properties of coal materials.
These findings also confirm the effectiveness of integrating multispectral satellite data with ML for detecting and delineating mining-related features. By specifically focusing on waste coal piles, this study advances previous research and demonstrates that ML models can discern subtle spectral variations associated with these features.

3.3. Temporal Statistics of NDVI

We also examined the temporal dynamics of the raw and waste coal areas identified through image classification. Figure 7 illustrates the average NDVI time series of these two land cover classes for the period of 1985 to 2022. The NDVI time series graph for raw coal shows a sharp decline in NDVI between 2007 and 2009, indicating significant disturbances, likely due to increased coal stockpiling activities (Figure 7a). The areas containing raw coal were likely used for stockpiling only after 2008. Following that period, NDVI values recovered slowly, as indicated by the shallow slope of the black line. This slow recovery reflected minimal reclamation efforts, with the vegetation health remaining relatively poor. Therefore, the area with raw coal was characterized by distinct disturbances with minimal reclamation.
The fitted line for waste coal (solid orange line) exhibited a slight upward trend, indicating a slow but steady improvement in vegetation health or coverage over the years (Figure 7b). This gradual increase in NDVI values suggested natural vegetation recovery in the absence of active reclamation efforts. The waste coal pile, remnants from historical coal mining activities predating 1986, has not experienced significant recent disturbances. Consequently, pixels with no significant disturbance represented the areas with waste coal. These areas showed a slow but consistent improvement in vegetation cover, reflecting a natural recovery process over several decades.

3.4. Greatest Disturbance and Reclamation Mapping

Using LandTrendr, we computed the magnitude of disturbance or reclamation by measuring the difference in NDVI values before and after the occurrence of the event. Figure 8 highlights the significant disturbance events and their corresponding years of occurrence. In Region 1, the disturbance magnitude was notably high (>470), indicating ongoing surface coal mining operations post-2000. The threshold value of 470 was determined from zonal statistics at known ground-truth locations within the study area, making it specific to the spectral and ecological conditions of this mining landscape. This threshold was effective in distinguishing active mining disturbances from historical waste coal piles in the current study area. If applied to other regions with different environmental conditions, the workflow could be adapted by calibrating thresholds using local training data or percentile-based breakpoints to ensure consistent performance.
Conversely, Region 2, which contains waste coal piles predating the SMCRA, showed little to no change in disturbance magnitude, suggesting the absence of significant disturbance since 1986. Similarly, Figure 9 illustrates the significant reclamation events and their corresponding years of occurrence. The SMCRA of 1977 mandates that coal mining operators reclaim affected areas to their original state after completing mining operations [34]. Consequently, the region with the greatest disturbance (Figure 8—Region 1) also exhibited a high reclamation magnitude (Figure 9—Region 1), indicating effective reclamation efforts in areas recently disturbed by coal mining. Additionally, Figure 9b revealed that reclamation efforts in active mining regions were more recent. On the other hand, Region 2 displayed no significant change in reclamation magnitude, suggesting minimal vegetation recovery in areas with historical waste coal piles.
In summary, areas with active mining operations exhibited high disturbance magnitudes, indicating ongoing mining activities. In contrast, historical waste coal piles did not show such a change, reflecting the absence of recent disturbances or reclamation efforts. The lack of significant NDVI fluctuations in these areas suggests that they have remained largely unchanged since the cessation of historical mining activities [34].
By combining the classification map obtained from image classification with the greatest disturbance map and applying a disturbance magnitude threshold of <470, it is possible to effectively distinguish between waste coal in active mining areas and that from the pre-SMCRA period. Figure 10 illustrates this differentiation, highlighting the clear contrast between actively mined regions and those containing historical waste coal piles. The resulting classification map was post-processed by removing waste coal piles smaller than one acre to reduce noise from spectrally mixed or uncertain pixels.
The evaluation using a confusion matrix indicates that the proposed approach performs exceptionally well, achieving an OA of 99.94%, a kappa coefficient of 87.87%, and a macro F1-score of 93.98% (Table 6). Specifically, it attains a precision of 78.57% for historical waste coal, with some instances being misclassified as other land cover types. The method also exhibits a perfect recall score of 100% for the historical waste coal, indicating that all actual instances were correctly identified. These results highlight the robustness and effectiveness of the combined image classification and change detection approach in accurately distinguishing historical waste coal piles predating the SMCRA from those in active mining areas.
The LandTrendr-based temporal analysis revealed significant changes in waste coal pile extent over the study period (1986–2023). This approach aligns with Kennedy et al. [30], who developed the LandTrendr algorithm for detecting forest disturbance and recovery using time series Landsat data. Our results also indicate significant variability in the temporal dynamics of waste coal piles pre- and post-SMCRA. While areas with active mining operations exhibited disturbance magnitudes, historical waste coal lacked recent disturbances and reclamation efforts. The absence of significant NDVI fluctuations in these areas suggests that they have remained largely unchanged since the cessation of historical mining activities. Similar temporal dynamics have been documented in studies by Pflugmacher et al. [52] and Cohen et al. [53], who applied LandTrendr to assess forest disturbances and subsequent recovery processes.

4. Future Directions

This study applied ML and change-detection algorithms to medium-resolution satellite imagery to identify historical waste coal piles. The small size of these piles and the limited availability of training and testing datasets may introduce uncertainties in the performance of the ML models when applying this approach to other regions. Integrating data from larger geographic regions and leveraging advanced ML techniques may help address these limitations. As the method has been tested in a single study area, its effectiveness under different geographic and environmental conditions remains unverified. Future work should therefore prioritize integrating data from larger and more diverse geographic regions to evaluate model transferability and strengthen its generalizability. Additionally, integrating high-resolution drone or commercial imagery with medium-resolution satellite data could further improve mapping accuracy, enable more detailed characterization of surface properties, and enhance detection of smaller or partially obscured piles [54].
We analyzed trends in NDVI within the mining region to assess reclamation efforts and monitor ongoing coal mining operations. To minimize short-term climatic and seasonal variability, we used the intra-annual 95th percentile NDVI, which helped reduce the influence of geographic variation in vegetation growth driven by climate, soil, and topography. Moving forward, future work could be strengthened by integrating ancillary datasets. In particular, DEM-derived terrain indices such as slope or aspect may serve as stable, time-invariant covariates in relatively undisturbed landscapes. However, for actively mined or reclaimed areas where topography is continually altered, the use of terrain indices should be limited to stable zones. Incorporating such terrain indices could improve the quantification of site-specific changes and enable a more comprehensive assessment of reclamation success.

5. Conclusions

This study developed a comprehensive data analytical workflow using remotely sensed data to classify and monitor waste coal piles predating the SMCRA regulations. Waste coal piles were first delineated through supervised image classification of Sentinel-2 imagery using four ML models—kNN, RF, GB, and SVM. Among these, the RF model achieved the highest performance, with a precision of 86%, a recall of 77%, an OA of 96%, and an F1-score of 91%. The LandTrendr algorithm was then applied to a multi-decadal time series to assess disturbance and reclamation. This analysis revealed that historical waste coal piles have experienced a gradual and consistent increase in vegetation cover since 1986, reflecting natural reclamation processes. These areas also showed little to no change in disturbance magnitude, indicating the absence of major new disturbances. In contrast, active mining areas exhibited consistently high disturbance magnitudes due to ongoing surface coal mining operations. These contrasting patterns, captured in the disturbance mapping, enabled clear differentiation between historical waste coal piles and active mining sites.
The proposed workflow achieved a precision of 78.6% and a recall of 100% in identifying historical waste coal piles, demonstrating its reliability for monitoring legacy mining impacts. By integrating image classification with long-term change detection, this approach advances understanding of historical mining disturbances and provides a practical framework to support reclamation planning and policy development. Moreover, the workflow is adaptable to other coal-producing regions, offering a scalable solution for data-driven environmental management and sustainable land rehabilitation.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/rs17173041/s1, Figure S1. Zonal NDVI time-series for the reclaimed areas; Table S1: Characteristics of Sentinel-2 bands; Table S2. Hyperparameters and search spaces for kNN, RF, and SVM; Table S3. Confusion matrix for training data—kNN; Table S4. Confusion matrix for training data—GB; Table S5. Confusion matrix for training data—RF; Table S6. Confusion matrix for training data—SVM; Table S7. ML model performance metrics for training data; Table S8. Confusion matrix for test data—kNN; Table S9. Confusion matrix for test data—GB; Table S10. Confusion matrix for test data—RF; Table S11. Confusion matrix for test data—SVM.

Author Contributions

Conceptualization, S.D., A.S. and S.K.; methodology, S.D. and S.K.; software, S.D.; validation, S.D. and S.K.; data curation, S.D.; writing—original draft preparation, S.D. and S.K.; writing—review and editing, S.D., A.S. and S.K.; funding acquisition, A.S. and S.K. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the U.S. Department of Energy, National Energy Technology Laboratory under Award Number DE-FE0032204.

Data Availability Statement

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Acknowledgments

The authors would like to thank the Division of Mineral Resources Management, Ohio Department of Natural Resources, for providing GIS data layers related to Resource Conservation and Recovery Act (RCRA) gob piles and the Abandoned Mine Land Inventory System (AMLIS).

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

AMLISAbandoned Mine Land Inventory System
BEASTBayesian Estimator of Abrupt change, Seasonality, and Trend
BISEBest Index Slope Extraction algorithm
DMRMDepartment of Mineral Resource Management
EVIEnhanced Vegetation Index
GHSLGlobal Human Settlement Layer
GEEGoogle Earth Engine
GBGradient Boosting
kNNk-nearest Neighbors
LandTrendrLandsat-based Detection of Trends in Disturbance and Recovery
MLMachine Learning
MSIMulti-Spectral Imager
NLCDNational Land Cover Data layer
NIRNear-infrared
NDVINormalized Difference Vegetation Index
NDWINormalized Difference Water Index
ODNROhio Department of Natural Resources
OAOverall Accuracy
RFRandom Forest
RCRAResource Conservation and Recovery Act
SCLScene Classification Layer
SWIRShortwave Infrared
SVMSupport Vector Machine
SMCRASurface Mining Control and Reclamation Act
SRSurface Reflectance
USGSU.S. Geological Survey
U.S.United States
VARIVisible Atmospherically Resistant Index

References

  1. Energy Information Administration. Annual Coal Report 2022. 2023. Available online: https://www.eia.gov/coal/annual/pdf/tableES1.pdf (accessed on 26 June 2024).
  2. Committee on the Study of the Control of Respirable Coal Mine Dust Exposure in Underground Mines; Board on Earth Sciences and Resources; Board on Environmental Studies and Toxicology; Board on Health Sciences Policy; Division on Earth and Life Studies; Health and Medicine Division; National Academies of Sciences, Engineering, and Medicine. Monitoring and Sampling Approaches to Assess Underground Coal Mine Dust Exposures; National Academies Press: Washington, DC, USA, 2018; p. 25111. ISBN 978-0-309-47601-0. [Google Scholar]
  3. Hanak, D.P. Environmental Life-Cycle Assessment of Waste-Coal Pellets Production. Clean Energy 2022, 6, 1–14. [Google Scholar] [CrossRef]
  4. Kite, M.S. The Surface Mining Control and Reclamation Act of 1977: An Overview of Reclamation Requirements and Implementation. Land Water L. Rev. 1977, 13, 703. [Google Scholar]
  5. Bureau of Land Management. Abandoned Mine Land Program Policy Handbook; United States Department of the Interior: Washington, DC, USA, 2007.
  6. Adiansyah, J.S.; Rosano, M.; Biswas, W.; Haque, N. Life Cycle Cost Estimation and Environmental Valuation of Coal Mine Tailings Management. J. Sustain. Min. 2017, 16, 114–125. [Google Scholar] [CrossRef]
  7. Hu, Z.; Xia, Q. An Integrated Methodology for Monitoring Spontaneous Combustion of Coal Waste Dumps Based on Surface Temperature Detection. Appl. Therm. Eng. 2017, 122, 27–38. [Google Scholar] [CrossRef]
  8. Jendruś, R. Environmental Protection in Industrial Areas and Applying Thermal Analysisto Coal Dumps. Pol. J. Environ. Stud. 2017, 26, 137–146. [Google Scholar] [CrossRef]
  9. Bell, F.G.; Stacey, T.R.; Genske, D.D. Mining Subsidence and Its Effect on the Environment: Some Differing Examples. Environ. Geol. 2000, 40, 135–152. [Google Scholar] [CrossRef]
  10. Finkelman, R.B. Potential Health Impacts of Burning Coal Beds and Waste Banks. Int. J. Coal Geol. 2004, 59, 19–24. [Google Scholar] [CrossRef]
  11. Bian, Z.; Dong, J.; Lei, S.; Leng, H.; Mu, S.; Wang, H. The Impact of Disposal and Treatment of Coal Mining Wastes on Environment and Farmland. Environ. Geol. 2009, 58, 625–634. [Google Scholar] [CrossRef]
  12. U.S. Environmental Protection Agency. Abandoned Mine Lands Team: Reference Notebook; 2004. Available online: https://semspub.epa.gov/work/11/176040.pdf (accessed on 26 June 2024).
  13. Agboola, O.; Babatunde, D.E.; Isaac Fayomi, O.S.; Sadiku, E.R.; Popoola, P.; Moropeng, L.; Yahaya, A.; Mamudu, O.A. A Review on the Impact of Mining Operation: Monitoring, Assessment and Management. Results Eng. 2020, 8, 100181. [Google Scholar] [CrossRef]
  14. Gunathunga, S.U.; Gagen, E.J.; Evans, P.N.; Erskine, P.D.; Southam, G. Anthropedogenesis in Coal Mine Overburden; the Need for a Comprehensive, Fundamental Biogeochemical Approach. Sci. Total Environ. 2023, 892, 164515. [Google Scholar] [CrossRef] [PubMed]
  15. Anderson, A.; Schultz, D.; Buchman, N.; Nock, H. Landsat Imagery for Surface-Mine Inventory. Photogramm. Eng. Remote Sens. 1977, 43, 1027–1036. [Google Scholar]
  16. Maxwell, A.E.; Warner, T.A.; Strager, M.P.; Conley, J.F.; Sharp, A.L. Assessing Machine-Learning Algorithms and Image- and Lidar-Derived Variables for GEOBIA Classification of Mining and Mine Reclamation. Int. J. Remote Sens. 2015, 36, 954–978. [Google Scholar] [CrossRef]
  17. Chen, W.; Li, X.; He, H.; Wang, L. A Review of Fine-Scale Land Use and Land Cover Classification in Open-Pit Mining Areas by Remote Sensing Techniques. Remote Sens. 2017, 10, 15. [Google Scholar] [CrossRef]
  18. He, D.; Le, B.T.; Xiao, D.; Mao, Y.; Shan, F.; Ha, T.T.L. Coal Mine Area Monitoring Method by Machine Learning and Multispectral Remote Sensing Images. Infrared Phys. Technol. 2019, 103, 103070. [Google Scholar] [CrossRef]
  19. Nugroho, G.; Sofan, P.; Nugroho, U.C.; Suwarsono; Pambudi, A.I.; Yulianto, F.; Ichsan, N. The Use of Texture Analysis and Band Transformation on Multispectral Imagery to Map Open-Pit Mines Using Machine Learning. IOP Conf. Ser. Earth Environ. Sci. 2022, 1109, 012071. [Google Scholar] [CrossRef]
  20. Mao, Y.; Ma, B.; Liu, S.; Wu, L.; Zhang, X.; Yu, M. Study and Validation of a Remote Sensing Model for Coal Extraction Based on Reflectance Spectrum Features. Can. J. Remote Sens. 2014, 40, 327–335. [Google Scholar] [CrossRef]
  21. Zeng, X.; Liu, Z.; He, C.; Ma, Q.; Wu, J. Detecting Surface Coal Mining Areas from Remote Sensing Imagery: An Approach Based on Object-Oriented Decision Trees. J. Appl. Remote Sens. 2017, 11, 015025. [Google Scholar] [CrossRef]
  22. Werner, T.T.; Mudd, G.M.; Schipper, A.M.; Huijbregts, M.A.J.; Taneja, L.; Northey, S.A. Global-Scale Remote Sensing of Mine Areas and Analysis of Factors Explaining Their Extent. Glob. Environ. Change 2020, 60, 102007. [Google Scholar] [CrossRef]
  23. Jabłońska, K.; Maksymowicz, M.; Tanajewski, D.; Kaczan, W.; Zięba, M.; Wilgucki, M. MineCam: Application of Combined Remote Sensing and Machine Learning for Segmentation and Change Detection of Mining Areas Enabling Multi-Purpose Monitoring. Remote Sens. 2024, 16, 955. [Google Scholar] [CrossRef]
  24. Zhu, Y.; Chen, W.; He, W.; Wang, R.; Li, X.; Wang, L. CUG_MISDataset: A Remote Sensing Instance Segmentation Dataset for Improved Wide-Area High-Precision Mining Land Occupation Recognition. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2024, 17, 16476–16486. [Google Scholar] [CrossRef]
  25. Çığşar, B.; Ünal, D. Comparison of Data Mining Classification Algorithms Determining the Default Risk. Sci. Program. 2019, 2019, 8706505. [Google Scholar] [CrossRef]
  26. Xiao, W.; Deng, X.; He, T.; Guo, J. Using POI and Time Series Landsat Data to Identify and Rebuilt Surface Mining, Vegetation Disturbance and Land Reclamation Process Based on Google Earth Engine. J. Environ. Manage. 2023, 327, 116920. [Google Scholar] [CrossRef]
  27. Hu, T.; Myers Toman, E.; Chen, G.; Shao, G.; Zhou, Y.; Li, Y.; Zhao, K.; Feng, Y. Mapping Fine-Scale Human Disturbances in a Working Landscape with Landsat Time Series on Google Earth Engine. ISPRS J. Photogramm. Remote Sens. 2021, 176, 250–261. [Google Scholar] [CrossRef]
  28. Yang, Z.; Li, J.; Shen, Y.; Miao, H.; Yan, X. A Denoising Method for Inter-Annual NDVI Time Series Derived from Landsat Images. Int. J. Remote Sens. 2018, 39, 3816–3827. [Google Scholar] [CrossRef]
  29. Zhu, D.; Chen, T.; Zhen, N.; Niu, R. Monitoring the Effects of Open-Pit Mining on the Eco-Environment Using a Moving Window-Based Remote Sensing Ecological Index. Environ. Sci. Pollut. Res. 2020, 27, 15716–15728. [Google Scholar] [CrossRef]
  30. Kennedy, R.E.; Yang, Z.; Cohen, W.B. Detecting Trends in Forest Disturbance and Recovery Using Yearly Landsat Time Series: 1. LandTrendr—Temporal Segmentation Algorithms. Remote Sens. Environ. 2010, 114, 2897–2910. [Google Scholar] [CrossRef]
  31. Dlamini, L.Z.D.; Xulu, S. Monitoring Mining Disturbance and Restoration over RBM Site in South Africa Using LandTrendr Algorithm and Landsat Data. Sustainability 2019, 11, 6916. [Google Scholar] [CrossRef]
  32. Han, Y.; Ke, Y.; Zhu, L.; Feng, H.; Zhang, Q.; Sun, Z.; Zhu, L. Tracking Vegetation Degradation and Recovery in Multiple Mining Areas in Beijing, China, Based on Time-Series Landsat Imagery. GIScience Remote Sens. 2021, 58, 1477–1496. [Google Scholar] [CrossRef]
  33. Guo, J.; He, T.; Xiao, W.; Lei, K. Time Series Procession for Monitoring Land Disturbance Caused by Surface Coal Mining in China. J. Clean. Prod. 2024, 448, 141585. [Google Scholar] [CrossRef]
  34. Dhakal, S. Mapping and Volume Estimation of Waste Coal in Abandoned Mine Lands Using Remote Sensing and Geospatial Techniques. Master’s Thesis, The Ohio State University, Columbus, OH, USA, 2024. [Google Scholar]
  35. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
  36. Copernicus Sentinel-2 (processed by ESA). MSI Level-2A BOA Reflectance Product. Collection 1. European Space Agency. Sentinel Online. [CrossRef]
  37. Dong, S.; Guo, H.; Chen, Z.; Pan, Y.; Gao, B. Spatial Stratification Method for the Sampling Design of LULC Classification Accuracy Assessment: A Case Study in Beijing, China. Remote Sens. 2022, 14, 865. [Google Scholar] [CrossRef]
  38. Jon Dewitz National Land Cover Database (NLCD) 2019 Products (Ver. 3.0, February 2024) 2024. Available online: https://www.sciencebase.gov/catalog/item/5f21cef582cef313ed940043 (accessed on 26 June 2024).
  39. Pesaresi, M.; Politis, P. GHS-BUILT-C R2023A—GHS Settlement Characteristics, Derived from Sentinel2 Composite (2018) and Other GHS R2023A Data 2023. Available online: https://data.jrc.ec.europa.eu/dataset/3c60ddf6-0586-4190-854b-f6aa0edc2a30 (accessed on 26 June 2024).
  40. Esri Federal Data Transportation 2024. Available online: https://hub.arcgis.com/maps/f42ecc08a3634182b8678514af35fac3/about (accessed on 26 June 2024).
  41. Altman, N.S. An Introduction to Kernel and Nearest-Neighbor Nonparametric Regression. Am. Stat. 1992, 46, 175–185. [Google Scholar] [CrossRef]
  42. Freund, Y.; Schapire, R.E. A Short Introduction to Boosting. J. Jpn. Soc. Artif. Intell. 1999, 14, 771–780. [Google Scholar]
  43. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  44. Pal, M. Random Forest Classifier for Remote Sensing Classification. Int. J. Remote Sens. 2005, 26, 217–222. [Google Scholar] [CrossRef]
  45. Cortes, C.; Vapnik, V. Support-Vector Networks. Mach. Learn. 1995, 20, 273–297. [Google Scholar] [CrossRef]
  46. Feng, Y.; Wang, J.; Bai, Z.; Reading, L. Effects of Surface Coal Mining and Land Reclamation on Soil Properties: A Review. Earth-Sci. Rev. 2019, 191, 12–25. [Google Scholar] [CrossRef]
  47. McKenna, P.B.; Lechner, A.M.; Phinn, S.; Erskine, P.D. Remote Sensing of Mine Site Rehabilitation for Ecological Outcomes: A Global Systematic Review. Remote Sens. 2020, 12, 3535. [Google Scholar] [CrossRef]
  48. Townsend, J.T. Theoretical Analysis of an Alphabetic Confusion Matrix. Percept. Psychophys. 1971, 9, 40–50. [Google Scholar] [CrossRef]
  49. Ting, K.M. Confusion Matrix. In Encyclopedia of Machine Learning; Sammut, C., Webb, G.I., Eds.; Springer: Boston, MA, USA, 2011; p. 209. ISBN 978-0-387-30768-8. [Google Scholar]
  50. Demirel, N.; Emil, M.K.; Duzgun, H.S. Surface Coal Mine Area Monitoring Using Multi-Temporal High-Resolution Satellite Imagery. Int. J. Coal Geol. 2011, 86, 3–11. [Google Scholar] [CrossRef]
  51. Loyd, C. Landsat 8 Bands. Available online: https://landsat.gsfc.nasa.gov/satellites/landsat-8/landsat-8-bands/ (accessed on 5 February 2025).
  52. Pflugmacher, D.; Cohen, W.B.; Kennedy, R.E. Using Landsat-Derived Disturbance History (1972–2010) to Predict Current Forest Structure. Remote Sens. Environ. 2012, 122, 146–165. [Google Scholar] [CrossRef]
  53. Cohen, W.B.; Yang, Z.; Healey, S.P.; Kennedy, R.E.; Gorelick, N. A LandTrendr Multispectral Ensemble for Forest Disturbance Detection. Remote Sens. Environ. 2018, 205, 131–140. [Google Scholar] [CrossRef]
  54. Ren, H.; Zhao, Y.; Xiao, W.; Zhang, L. Unmanned Aerial Vehicle (UAV)-Based Vegetation Restoration Monitoring in Coal Waste Dumps after Reclamation. Remote Sens. 2024, 16, 881. [Google Scholar] [CrossRef]
Remotesensing 17 03041 g001aRemotesensing 17 03041 g001b
Figure 1. Location of the study area (a) the Ohio state boundary (b) Annual median true color composite (R, G, B) image from 2023 (c) Annual median false color composite (NIR, R, G) image from 2023 (d) Transportation layer mask. In (c), red hues indicate vegetation, whereas blue hues indicate built-up areas including manmade structures, roads, and barren lands.
Figure 1. Location of the study area (a) the Ohio state boundary (b) Annual median true color composite (R, G, B) image from 2023 (c) Annual median false color composite (NIR, R, G) image from 2023 (d) Transportation layer mask. In (c), red hues indicate vegetation, whereas blue hues indicate built-up areas including manmade structures, roads, and barren lands.
Remotesensing 17 03041 g001aRemotesensing 17 03041 g001b
Remotesensing 17 03041 g002
Figure 2. Overview of methodology.
Figure 2. Overview of methodology.
Remotesensing 17 03041 g002
Remotesensing 17 03041 g003
Figure 3. Spectral signatures of different land cover classes. Surface reflectance values were obtained after masking the transportation layer. The built-up layer specifically excludes roads and highways.
Figure 3. Spectral signatures of different land cover classes. Surface reflectance values were obtained after masking the transportation layer. The built-up layer specifically excludes roads and highways.
Remotesensing 17 03041 g003
Remotesensing 17 03041 g004
Figure 4. Classification map (a) kNN (b) GB (c) RF (d) SVM.
Figure 4. Classification map (a) kNN (b) GB (c) RF (d) SVM.
Remotesensing 17 03041 g004
Remotesensing 17 03041 g005
Figure 5. Location and extent of Raw Coal and Waste Coal (a) kNN, (b) GB, (c) RF, and (d) SVM.
Figure 5. Location and extent of Raw Coal and Waste Coal (a) kNN, (b) GB, (c) RF, and (d) SVM.
Remotesensing 17 03041 g005
Remotesensing 17 03041 g006
Figure 6. Relative importance of input variables based on the RF model.
Figure 6. Relative importance of input variables based on the RF model.
Remotesensing 17 03041 g006
Remotesensing 17 03041 g007
Figure 7. Zonal NDVI time-series for areas with (a) raw coal and (b) waste coal. The dotted line represents the original NDVI value, while the dark line represents fitted NDVI values.
Figure 7. Zonal NDVI time-series for areas with (a) raw coal and (b) waste coal. The dotted line represents the original NDVI value, while the dark line represents fitted NDVI values.
Remotesensing 17 03041 g007
Remotesensing 17 03041 g008
Figure 8. Greatest disturbance mapping (a) Magnitude of change and (b) Year of detection. In (a), the values in the legend represent the difference in scaled NDVI values multiplied by 10,000 (Landsat SR scaling factor in GEE). Values range from 200 to 1588, with lower values (green) indicating minor disturbances and higher values (red) representing more severe disturbances. In (b), the values in the legend indicate the timing of the greatest NDVI-based disturbance event between 1986 and 2023.
Figure 8. Greatest disturbance mapping (a) Magnitude of change and (b) Year of detection. In (a), the values in the legend represent the difference in scaled NDVI values multiplied by 10,000 (Landsat SR scaling factor in GEE). Values range from 200 to 1588, with lower values (green) indicating minor disturbances and higher values (red) representing more severe disturbances. In (b), the values in the legend indicate the timing of the greatest NDVI-based disturbance event between 1986 and 2023.
Remotesensing 17 03041 g008
Remotesensing 17 03041 g009
Figure 9. Greatest reclamation mapping (a) Magnitude of change and (b) Year of detection. In (a), the values in the legend represent the difference in scaled NDVI values multiplied by 10,000 (Landsat SR scaling factor in GEE). Values range from 300 to 1663, with lower values (green) indicating minor reclamations and higher values (red) representing major reclamation events. In (b), the values in the legend indicate the timing of the greatest NDVI-based reclamation event between 1986 and 2023.
Figure 9. Greatest reclamation mapping (a) Magnitude of change and (b) Year of detection. In (a), the values in the legend represent the difference in scaled NDVI values multiplied by 10,000 (Landsat SR scaling factor in GEE). Values range from 300 to 1663, with lower values (green) indicating minor reclamations and higher values (red) representing major reclamation events. In (b), the values in the legend indicate the timing of the greatest NDVI-based reclamation event between 1986 and 2023.
Remotesensing 17 03041 g009
Remotesensing 17 03041 g010
Figure 10. (a) The location of waste coal piles within the study area (b) A waste coal pile predating the SMCRA (c) Waste coal at an active mining site. The purple areas indicate the extent of waste coal identified through image classification only, while the black areas show the extent of waste coal determined through image classification followed by change detection.
Figure 10. (a) The location of waste coal piles within the study area (b) A waste coal pile predating the SMCRA (c) Waste coal at an active mining site. The purple areas indicate the extent of waste coal identified through image classification only, while the black areas show the extent of waste coal determined through image classification followed by change detection.
Remotesensing 17 03041 g010
Table 1. Dataset for image classification with land-cover class definition.
Table 1. Dataset for image classification with land-cover class definition.
ClassesNumber of PixelsDescription
Raw Coal31Coal extracted or exposed for removal from a mine, yet to be processed
Waste Coal398Remnant of coal mining, including coal fragments, shale, rock, and other impurities
Forest7201Land dominated by mature, woody vegetation
Water2621Water, including retention ponds, streams, and standing water
Built-up2384Areas affected by human activities, including manmade structures and barren areas
Grassland18,243Land dominated by herbaceous, non-woody vegetation
Note: Each pixel represents an area of 10 m × 10 m on the ground.
Table 2. Final parameter settings for ML models.
Table 2. Final parameter settings for ML models.
ModelsParametersValues
kNNk5
searchMethodAUTO
metricEuclidean
GBnumberOfTrees10
shrinkage0.005
samplingRate0.7
lossLeastAbsoluteDeviation
seed0
RFnumberOfTrees50
variablesPerSplit3
minLeafPopulation1
bagFraction0.5
seed0
SVMdecisionProcedureVoting
svmTypeC_SVC
kernelTypeRBF
shrinkingTRUE
gamma1
cost10
Table 3. LandTrendr parameters.
Table 3. LandTrendr parameters.
ParametersValues
maxSegments6
spikeThreshold0.9
vertexCountOvershoot3
preventOneYearRecoveryTRUE
recoveryThreshold0.25
pvalThreshold0.05
bestModelProportion0.75
minObservationsNeeded6
mag>300
dur<4
preval>300
mmu11
Table 4. Summary of model performance across individual classes.
Table 4. Summary of model performance across individual classes.
ClasseskNNGBRFSVM
PrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecall
Raw Coal100%83.33%2.84%66.67%100%100%80.00%66.67%
Waste Coal70.59%49.32%45.61%35.62%86.15%76.71%46.79%69.86%
Forest88.76%96.02%89.50%92.59%94.15%96.94%92.77%95.39%
Water97.87%71.04%72.46%77.22%96.46%73.75%97.41%72.59%
Built-up96.90%68.35%68.19%87.71%90.60%89.23%97.05%77.61%
Grassland95.22%94.96%95.25%87.97%97.25%96.81%95.90%96.79%
Table 5. Land cover area estimates and ML model performance metrics for test data.
Table 5. Land cover area estimates and ML model performance metrics for test data.
ModelArea (km2)Model Metrics
Raw CoalWaste CoalForestWaterBuilt-UpGrasslandOAKappaF1-Score
kNN0.0430.59138.9751.9304.97059.56892.79%86.99%83.13%
GB1.1850.86537.1202.5559.62154.73188.89%80.88%63.24%
RF0.0450.83636.3412.0407.43659.37995.66%92.26%91.20%
SVM0.0481.40736.6551.9705.81960.17894.32%89.80%81.43%
Table 6. Accuracy assessment of the classification map obtained after image classification and change detection.
Table 6. Accuracy assessment of the classification map obtained after image classification and change detection.
ClassesOtherPre-SMCRA Waste CoalTotalPrecision
Other997209972100%
Pre-SMCRA Waste Coal6222878.57%
Total997822
Recall99.94%100% 87.97%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Dhakal, S.; Shah, A.; Khanal, S. Monitoring Historical Waste Coal Piles Using Image Classification and Change Detection Algorithms on Satellite Images. Remote Sens. 2025, 17, 3041. https://doi.org/10.3390/rs17173041

AMA Style

Dhakal S, Shah A, Khanal S. Monitoring Historical Waste Coal Piles Using Image Classification and Change Detection Algorithms on Satellite Images. Remote Sensing. 2025; 17(17):3041. https://doi.org/10.3390/rs17173041

Chicago/Turabian Style

Dhakal, Sandeep, Ajay Shah, and Sami Khanal. 2025. "Monitoring Historical Waste Coal Piles Using Image Classification and Change Detection Algorithms on Satellite Images" Remote Sensing 17, no. 17: 3041. https://doi.org/10.3390/rs17173041

APA Style

Dhakal, S., Shah, A., & Khanal, S. (2025). Monitoring Historical Waste Coal Piles Using Image Classification and Change Detection Algorithms on Satellite Images. Remote Sensing, 17(17), 3041. https://doi.org/10.3390/rs17173041

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop