Fusing LandTrendr BCI and machine learning for spoil dump mapping

Abstract

Monitoring the long-term evolution of open-pit mining spoil dumps is challenging, requiring the decoupling of short-term material deposition from long-term ecological succession. This study presents an automated framework for spoil dump mapping by fusing a novel Bare Coal Index (BCI), the LandTrendr time-series algorithm, and machine learning. Using the 1986–2021 Landsat record, the BCI leverages near-infrared (NIR) and shortwave infrared (SWIR) bands to quantify coal exposure, while LandTrendr tracks vegetation disturbance and recovery from kernel NDVI (kNDVI) trajectories. These spectral and temporal features were integrated with topographic data using a Random Forest machine learning classifier to distinguish internal from external spoil dumps, with boundaries refined by morphological operations. The BCI mapped bare coal with 92% accuracy, significantly outperforming baseline indices. The final fused model achieved 86% accuracy (F1-score = 0.84) in mapping spoil dump typologies, a 22% relative improvement over a disturbance-only approach. This work delivers a robust, scalable framework that fuses multiple data streams for the comprehensive mapping and monitoring of mining landscapes, supporting adaptive environmental management.

Introduction

Mining activities, while crucial for global economic development, generate significant waste materials that are deposited in spoil dump sites. These sites, consisting of overburden, waste rock, and tailings, alter landforms, contribute to soil erosion, and can leach pollutants into local ecosystems, posing substantial risks to environmental health and biodiversity¹. Consequently, the accurate identification, classification, and long-term monitoring of spoil dump sites are critical for effective environmental management, land reclamation, and the promotion of sustainable mining practices.

Traditional spoil dump management, relying on manual surveys and visual inspections, is often time-consuming, labor-intensive, and difficult to scale, particularly for large or remote mining regions². In response, remote sensing has emerged as a powerful tool for monitoring these dynamic landscapes. However, early remote sensing approaches faced their own limitations. For instance, reliance on single-date, low-resolution imagery often failed to capture the complex temporal evolution of mining activities, from initial disturbance to subsequent vegetation regrowth³.

In recent years, more advanced methods have been developed, yet significant research gaps persist. Many studies have relied on spectral indices like the Normalized Difference Coal Mining Index (NDCMI). While useful, these indices often struggle to distinguish spoil dumps from other spectrally similar disturbed surfaces, such as industrial yards or bare soil, and they lack temporal depth for analyzing long-term recovery dynamics⁴. Other approaches have employed machine learning with stacked indices, but these are generally restricted to single snapshots or short-term analyses, failing to reconstruct multi-decadal histories⁵. High-precision methods using hyperspectral or LiDAR data can achieve high accuracy, but their application is severely constrained by high costs, limited spatial coverage, and the lack of historical archives, making them unsuitable for scalable, long-term monitoring⁶.

This study addresses these limitations by proposing a novel, automated framework for spoil dump mapping. Our approach is built on the fusion of a custom Bare Coal Index (BCI) for mapping surface composition, the LandTrendr algorithm for tracking temporal dynamics, and a machine learning classifier for integrating these multi-source data streams. By leveraging the freely available, multi-decadal Landsat archive, this framework reconstructs the 35-year evolution of spoil dumps and, to our knowledge, is the first to explicitly discriminate between internal and external dump typologies based on their distinct disturbance and reclamation trajectories.

The primary objective of this research is therefore to develop and validate this robust machine learning framework for the spatiotemporal mapping of spoil dumps. By fusing spectral, temporal, and topographic information, our hybrid approach provides a comprehensive solution for monitoring mining impacts, guiding sustainable land management, and supporting targeted ecological restoration strategies.

Study area and data

Overview of the study area

The Pingshuo open pit mine, located in the Pinglu District of Shuozhou City, Shanxi Province, China, is a crucial component of the Jinbei coal base (Fig. 1)⁷. Covering an area of 380 km², the mine has a coal geological reserve of 12.75 billion tons and a designed production scale of 65 million tons per year. The mine is bounded by coal outcrops on the east, west, and north sides, and by the Danshui Gully Fault on the south side. The study area features complex topography, primarily consisting of mountains and hills, with an elevation ranging from 1270 m to 1505.72 m. The climate is dry and semi-arid, characterized by concentrated precipitation in summer and autumn and frequent sandstorms in winter and spring. The Pingshuo mine has undergone significant land use changes due to mining activities, with extensive land reclamation and ecological restoration efforts since the 1980s.

Fig. 1

Location of the study area. The Pingshuo open-pit mine is situated in Shuozhou City, Shanxi Province, China. Insets show the location of Shanxi Province within China (left) and a true-color satellite image of the mine (right). This map was generated by the authors using ArcGIS 10.8 (Esri, https://www.esri.com). The administrative boundaries are based on the standard map with Review No: GS(2019)1822 from the Standard Map Service System of the Ministry of Natural Resources of the People’s Republic of China.

Data sources and preprocessing

Data sources

This study utilized the Google Earth Engine (GEE) cloud computing platform for data processing and analysis⁸. The primary data consisted of surface reflectance (SR) products from Landsat 5, Landsat 7, and Landsat 8, which are collectively referred to as ‘Landsat’ for brevity. Access to the GEE archive, which contains over three decades of continuous imagery, made these datasets invaluable for the long-term monitoring of land degradation, reclamation, and ecological reconstruction in the mining area⁹.

Data preprocessing

Remote sensing images from Landsat-5, Landsat-7, and Landsat-8 were used in this study, with a total of approximately 728 images from Landsat-5 (1986–2013), 572 images from Landsat-7 (1999–2021), and 234 images from Landsat-8 (2013–2021). The Landsat 5/7/8 remote sensing image data were preprocessed by masking and cropping the images to the study area. To align the data from different Landsat missions, we applied the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) for Landsat-5¹⁰ and the Landsat Surface Reflectance Code (LaSRC) for Landsat-8¹¹ to calibrate sensor radiance and surface reflectance. These calibration methods ensured the consistency of data across different Landsat satellites and time periods.

All image data were in UTM projection with the WGS-84 coordinate system. The Landsat SR (surface reflectance) data includes the “pixel_qa” band, which was used for cloud, snow, and shadow masking. The CFMASK algorithm, a C-language version of the Fmask algorithm adapted by the US Geological Survey, was employed to generate the image quality assessment band¹². This algorithm marks the state of whether a pixel is covered by clouds through a decision tree method and verifies the marked pixels by statistically analyzing the information of the entire image. The quality assessment band was introduced through the select() function, and bitwise conditional judgment functions and shift operations were used to screen out cloud and cloud shadow image pixels. Approximately 20% of the pixels were excluded due to cloud/snow/shadow/quality filtering using the pixel_qa/CFMASK algorithm. This figure represents the average pixel exclusion rate over the entire 1986–2021 study period. Given the 30 m spatial resolution of the Landsat imagery used, the minimum mapping unit (MMU) for this study is a single pixel, representing an area of 900 m².

A total of 10 large spoil dump sites were identified in the study area. The spoil dump sites were categorized into internal dumps (high BCI, disturbed, and reclaimed areas) and external dumps (disturbed and reclaimed, but without high BCI). Each year, 500 samples were collected across the spoil dump sites, with approximately 50 samples allocated to each site. Of these, 70% were used for training and 30% for validation, ensuring representativeness and spatial coverage of the samples.

Methodology

This study uses a three-step approach to identify and monitor dump sites in the Pingshuo mining area. First, BCI was developed using Landsat’s Near-Infrared (NIR), shortwave infrared1 (SWIR1) and shortwave infrared2 (SWIR2) bands. Second, vegetation disturbances were detected using LandTrendr and kNDVI to analyze temporal changes. Third, internal dumps (high BCI, disturbed, and reclaimed) and external dumps (disturbed and reclaimed without high BCI) were classified, with morphological operations refining boundaries. Validation used high-resolution imagery and field surveys, followed by vegetation recovery analysis to assess reclamation success (Fig. 2).

Fig. 2

Methodological flowchart. The framework includes three stages: (1) Data preparation using Landsat time-series (1986–2021); (2) Parallel extraction of bare coal (BCI) and vegetation disturbance (LandTrendr-kNDVI); and (3) Integration of results for dump classification, followed by morphological refinement.

BCI development

Spectral band selection

The development of BCI began with an analysis of the spectral characteristics of bare coal and other land cover types. A total of 860 samples, including 660 bare coal (BC) pixels from coalfields, were collected from Landsat imagery¹³. Spectral curves of different coal types (e.g., lignite and bituminous coal) revealed that while reflectance levels varied, the spectral shapes were similar, enabling the use of average spectral values for BCI construction. Key bands for distinguishing bare coal were identified through sensitivity and separability analyses. The SWIR2–SWIR1 band combination showed unique positive values for bare coal fractions ≥ 80%, making it a critical feature for differentiation¹⁴. Additionally, the NIR band demonstrated high sensitivity to bare coal, further supporting its inclusion in the index.

As illustrated in Fig. 3, the separability analysis revealed clear differences among bands and their combinations in distinguishing coal from other land cover types. Among the single bands, NIR demonstrated the highest sensitivity to the bare coal fraction (R² = 0.9723). Furthermore, combinations such as NIR-Red, NIR + SWIR1, and SWIR2-SWIR1 achieved superior separability, with R² values of 0.9409, 0.9801, and 0.7908, respectively. These results confirm that the selected NIR, SWIR1, and SWIR2 bands provide strong discriminatory power for constructing the BCI, thus offering a robust spectral foundation for subsequent classification.

Fig. 3

Sensitivity of Landsat spectral bands to the bare coal (BC) fraction. Panels (A–I) show the linear relationship between spectral values and BC fraction for single bands (A–F) and band combinations (G–I). Each plot includes the linear regression (dashed line), 95% confidence and prediction intervals (shaded areas), regression equation, and R² value.

Index definition and rationale

Based on the spectral analysis, the BCI formula was derived to leverage the distinct reflectance properties of bare coal. The formula combines the NIR, SWIR1, and SWIR2 bands to maximize separability between bare coal and other land cover types (e.g., soil, man-made structures, and vegetation). The BCI is defined as:

$$\:\begin{array}{c}BCI\:=\frac{{R}_{SWIR2}-\:{R}_{SWIR1}}{{R}_{NIR}*\left({R}_{NIR}+\:{R}_{SWIR1}\right)}\end{array}$$

(1)

where $\:{\text{R}}_{\text{N}\text{I}\text{R}}$, $\:{\text{R}}_{\text{S}\text{W}\text{I}\text{R}1}$ and $\:{\text{R}}_{\text{S}\text{W}\text{I}\text{R}2}$ represent surface reflectance in the NIR, SWIR1, and SWIR2 bands, respectively.

In the NIR band (~ 0.85 μm), vegetation generally shows high reflectance, whereas coal and bare soils exhibit much lower values. This makes NIR effective for distinguishing vegetation from bare coal surfaces. SWIR1 (~ 1.6 μm) is highly sensitive to surface moisture, and coal and moist soils respond differently in this region, which enables separation of coal from wet soils or water bodies. SWIR2 (~ 2.2 μm) enhances the spectral contrast between coal and gray infrastructure such as roads and rooftops, reducing confusion with non-vegetated artificial surfaces. Taken together, the combination of NIR, SWIR1, and SWIR2 provides superior separability among coal, vegetation, water, and gray infrastructure, forming a robust basis for constructing BCI¹⁵.

Threshold optimization

Thresholds for BCI classification were determined using the Otsu method, which automatically identifies an optimal cut-off based on bimodal distributions¹⁶. This approach avoids subjective threshold selection and ensures robust applicability across multi-year, multi-scene datasets. To determine the optimal BCI thresholds for bare coal detection, a pixel-based simulation program was employed. The simulation modeled mixed pixels using a linear spectral mixing approach, combining soil, man-made surfaces, and bare coal in varying proportions¹⁷. Results indicated that the BCI value increased sharply for bare coal fractions ≥ 80%, with a corresponding threshold range of (0, 3.776). For practical application, a threshold of 0.8 was selected to identify pixels with ≥ 90% bare coal fraction, ensuring high accuracy while minimizing false positives from gray roofs and Gobi Desert areas. Water bodies, which could interfere with BCI detection, were masked using the Normalized Difference Water Index (NDWI) with a threshold of − 0.05¹⁸. This optimization process ensured robust and precise identification of bare coal areas.

Disturbance detection via LandTrendr-kNDVI

LandTrendr algorithm

The LandTrendr algorithm was used to detect disturbances in the mining area by analyzing time-series Landsat data¹⁹. The algorithm identifies changes in vegetation cover over time, making it suitable for detecting mining disturbances^{20, 21}. The LandTrendr algorithm was applied to the time-series data to identify the temporal trajectories of vegetation changes, providing a comprehensive understanding of the disturbance patterns in the mining area.

We specified the annual compositing window as the median method within the DOY 60 to DOY 330 range. Outliers were removed using Z-score and model-based residual analysis. Missing values were filled using linear and spline interpolation, with regression models applied for longer gaps. The LandTrendr parameters were set as follows: $\:maxSegments\:=\:5$, $\:spikeThreshold\:=\:0.1$, $\:vertexCountOvershoot\:=\:3$, $\:recoveryThreshold\:=\:0.2$, and $\:pvalThreshold\:=\:0.05$. kNDVI was computed using a Gaussian kernel ($\:\sigma\:\:=\:0.5$) for smoothing and trend preservation.

Kernel normalized difference vegetation index (kNDVI)

The kernel Normalized Difference Vegetation Index (kNDVI) was used in combination with the LandTrendr algorithm to enhance the detection of vegetation changes²². The kNDVI is a modified version of the traditional NDVI that uses a kernel function to account for the non-linear relationships between spectral bands. This makes the kNDVI more sensitive to changes in vegetation cover, providing a more accurate assessment of the disturbed areas.

Disturbance detection

The results of the LandTrendr-kNDVI analysis were refined using post-processing techniques to improve the accuracy of disturbance detection. These techniques included filtering and smoothing algorithms to reduce noise and false positives, ensuring that the detected disturbances were accurate and reliable. The post-processing techniques also helped to identify the boundaries of the disturbed areas, providing a clear map of the mining disturbances.

The integrated LandTrendr-kNDVI framework generated a disturbance map by identifying pixels with significant vegetation loss. This map served as the foundation for subsequent spoil dump classification, distinguishing active mining zones from reclaimed or stable areas.

Reclamation was defined as the condition where kNDVI stabilizes above 0.45, with a recovery slope of at least 0.02 per year, sustained for a minimum of 3 years after disturbance. Disturbance events were identified as declines with a slope $\:\le\:-0.05$ per year. These criteria were aligned with the LandTrendr parameters ($\:recoveryThreshold\:=\:0.2$, $\:pvalThreshold\:=\:0.05$). Representative trajectories are provided (Fig. 4) to illustrate disturbance versus reclamation.

Fig. 4

Example LandTrendr-kNDVI temporal trajectories. The plots show time-series for pixels experiencing (a) disturbance and reclamation, (b) disturbance only, and (c) stable vegetation. Gray dots are annual kNDVI values; colored lines represent the LandTrendr fitted segments (red = disturbance, green = reclamation, blue = stable).

Spoil dump classification framework

Spoil dump classification

Internal dumps were defined as areas that are disturbed, have a BCI value above the threshold, and are reclaimed. The BCI was used to identify areas with bare coal surfaces, and the reclamation status was determined based on the presence of vegetation cover. This step involved classifying the disturbed areas into internal dumps based on the BCI and reclamation status.

External dumps were defined as areas that are disturbed and reclaimed but do not meet the BCI threshold. These areas were classified as external dumps based on the absence of bare coal surfaces and the presence of vegetation cover. This step involved differentiating between internal and external dumps based on the BCI and reclamation status.

Morphological refinement

Morphological operations, such as dilation and erosion cycles, were applied to refine the boundaries of the identified spoil dumps²³. These operations helped to smooth the boundaries and remove small noise artifacts, improving the accuracy of the classification. The morphological refinement step ensured that the classified spoil dumps were clearly defined and accurate.

Validation and accuracy assessment

The Random Forest (RF) classifier, a widely used machine learning algorithm, was employed in this study for its robustness in handling high-dimensional data and multicollinearity, as well as its resistance to overfitting^{24, 25}. Its applicability in remote sensing has been widely recognized²⁶. The final classification result was determined by a majority vote mechanism from multiple decision trees. The dataset was split using both spatial and temporal blocking to avoid leakage, with 70% training, 15% validation, and 15% testing. Per-class precision and recall were reported with 95% confidence intervals (block bootstrap, n = 1000). Kappa and MCC were also computed to complement OA and F1^{27, 28}. To avoid data leakage, a spatial blocking strategy was applied. The following parameters were used for training: 500 trees, no depth limit, a minimum of 2 samples per leaf, and the square root of the total number of features for each split.

In addition to spectral indices such as BCI and kNDVI, terrain-derived features were also included as inputs to the Random Forest classifier. These features, extracted from a 30 m DEM and resampled to Landsat resolution, included elevation, slope, curvature, topographic position index (TPI), and roughness²⁹.

The classified dumps were validated using high-resolution imagery and field surveys to ensure the accuracy of the classification. The validation process involved comparing the classified dumps with the ground truth data to assess the accuracy of the classification. The accuracy assessment was performed using metrics such as the overall accuracy and the Kappa coefficient, providing a quantitative measure of the classification performance.

Results and discussion

Bare coal extraction results

Bare coal distribution and accuracy assessment

The spatial and temporal distribution of bare coal areas from 1986 to 2021 was mapped using BCI and morphological post-processing techniques. No significant bare coal areas were detected in 1986–1987 likely due to limited mining activities during this period. From 1988 onward bare coal areas began to emerge and expand reflecting the intensification of mining operations.

Figure 5 illustrates the gradual expansion of bare coal areas from 1986 to 2021. Between 1986 and 1987, no significant bare coal areas were detected, likely due to the limited mining activities during this period. However, from 1988 onward, as mining activities increased, bare coal areas began to emerge and expand, especially from 1988 to 1992, with small patches appearing in the southwestern part of the study area. This early phase of expansion is likely linked to the initial mining operations and the establishment of mining pits.

Fig. 5

Spatiotemporal expansion of bare coal (1986–2021). The map shows the year of first detection for each bare coal pixel, identified via BCI. The color ramp from blue (early years) to red (recent years) illustrates the expansion of mining activities.

From 1992 to 1998, the bare coal areas expanded northward, which corresponds with the development of new mining pits and an increase in coal extraction. During this period, mining activities spread from the original mining zones into surrounding areas, driven by the growing demand for coal and the improvement of mining infrastructure. From 1999 to 2021, there was a significant eastward expansion of bare coal zones, suggesting that large-scale mining operations and infrastructure development were major drivers of this spread. This could be attributed to the growing scale of mining activities and the increasing industrialization of the region.

A noteworthy observation is the emergence of bare coal areas in the northeastern region between 2015 and 2021, which then gradually extended westward. This expansion is likely linked to the intensification of mining activities and the opening of new extraction sites in this region. The recent opening of new mines could be a response to the increasing demand for coal, driven by both domestic and global market dynamics. It is important to note that this expansion could also be influenced by factors such as policy changes, technological advancements, and shifts in coal demand. For instance, as mining companies scale up production to meet market needs, new mining areas are developed, resulting in the further spread of bare coal zones into previously untouched regions.

These patterns highlight the growing scale and industrialization of mining activities, which increasingly impact the regional environment. The analysis of the temporal and spatial changes in bare coal areas not only illustrates the direct consequences of mining but also sheds light on the broader socioeconomic and political factors influencing this expansion. Further research should explore the deeper drivers behind these trends, such as policy, technological developments, and market forces, and examine how to balance mining activities with environmental protection to achieve sustainable resource development.

The accuracy of the BCI-based bare coal extraction was validated using high-resolution imagery and field survey data. Key metrics including overall accuracy (OA) omission error (OE) and commission error (CE) were calculated to evaluate performance. The BCI achieved an OA of 92% with OE and CE values of 8% and 7% respectively. Error sources and limitations include spectral confusion where bare coal shares similar spectral characteristics with certain land cover types such as gray infrastructure (e.g. concrete asphalt) and sediment-laden water leading to occasional misclassification. Mixed pixels are another issue where the moderate spatial resolution of Landsat imagery (30 m) results in mixed pixels particularly at the edges of bare coal areas where coal fractions may be below the detection threshold. Atmospheric interference also affects data quality as cloud cover and atmospheric haze in some images necessitate the exclusion of certain time points from the analysis. Threshold sensitivity is a concern as the BCI threshold optimized using Otsu’s algorithm may not fully capture low-coal-fraction pixels leading to underestimation in some regions³⁰. Mitigation strategies include the integration of hyperspectral data to improve material discrimination the application of machine learning classifiers to handle spectral ambiguities and the use of higher-resolution imagery (e.g. Sentinel-2) for edge refinement and mixed-pixel analysis³¹.

Precision comparison of BCI with NDVI and NDCMI

The performance of BCI was compared with the NDVI and the NDCMI. BCI significantly outperformed NDVI and NDCMI in detecting bare coal. BCI achieved markedly higher accuracy, with an F1-score of 0.92 compared with 0.74 for NDVI and 0.68 for NDCMI. Omission error was reduced to 8% (versus 22% for NDVI and 26% for NDCMI), while commission error was limited to 7% (compared with 18% and 21% for NDVI and NDCMI, respectively).

The superior performance of BCI arises from its spectral robustness and temporal stability. By exploiting shortwave infrared (SWIR) and thermal infrared (TIR) bands from Landsat imagery—spectral regions sensitive to the distinctive reflectance characteristics of bare coal—BCI effectively discriminates coal from vegetation, soil, and other substrates³². Moreover, its capacity to capture subtle temporal variations in coal exposure makes it particularly well suited for long-term monitoring of mining activities.

By contrast, NDVI and NDCMI show inherent limitations in this context. NDVI, designed primarily to characterize vegetation, performs poorly in separating coal from other non-vegetated surfaces such as soil or anthropogenic structures³³. NDCMI, although tailored for coal detection, relies predominantly on mid-infrared bands, which limits sensitivity to low-coal-fraction pixels and contributes to higher omission errors³⁴.

Overall, BCI offers a robust and scalable framework for detecting bare coal, outperforming established indices in both accuracy and consistency. Its integration with time-series analysis and morphological post-processing further enhances its utility for monitoring mining dynamics and supporting land reclamation. Nonetheless, challenges remain, particularly spectral confusion and mixed-pixel effects. Future work should explore the integration of hyperspectral datasets and advanced machine learning algorithms to further refine coal detection. Applying BCI across diverse mining regions and environmental settings will be critical to validate its generalizability and operational value.

Vegetation disturbance extraction results

The LandTrendr algorithm was applied to extract the spatiotemporal patterns of mining disturbances and reclamation areas from 1986 to 2021, with the results presented in Fig. 6. The changes in disturbance and reclamation are shown at four-year intervals, clearly illustrating the spatial and temporal distribution of these activities. The legend indicates the time series, with different colors representing the disturbance and reclamation spatial ranges detected in different periods.

Fig. 6

Spatiotemporal patterns of mining disturbance and reclamation (1986–2021). Maps show the year of (a) initial vegetation disturbance and (b) onset of reclamation, as detected by the LandTrendr-kNDVI algorithm. The colors represent different time periods, showing the progression of mining and restoration.

Figure 6 indicate that mining activities caused significant disturbances in the study area, exhibiting a certain spatial regularity. From 1986 to 2021, the disturbance area in the west advanced from the southwest to the northeast, while in the east, it moved from east to west, consistent with the direction of mining operations. Reclamation activities gradually expanded from the southwest to the northeast, accompanying the mining activities. This spatial pattern suggests that mining and reclamation activities were closely related and influenced each other’s spatial distribution.

The temporal distribution of mining disturbances and reclamation activities also showed distinct characteristics. Mining disturbances were more intense in the early stages, with a larger area affected, while reclamation activities gradually increased over time. This indicates that the pace of reclamation activities was relatively slower than that of mining, but the reclamation efforts were continuously strengthened.

The vegetation disturbance extraction results further revealed the impact of mining activities on vegetation. Mining activities led to significant vegetation loss and degradation in the study area, with the vegetation cover in the disturbed areas being much lower than that in the non-disturbed areas. However, with the implementation of reclamation activities, the vegetation cover in the reclaimed areas gradually increased, indicating that reclamation activities had a positive effect on vegetation restoration.

Dump classification results

Extraction of potential areas for internal and external dumps

The extraction of potential areas for internal and external dumps is crucial for understanding the spatial distribution of mining - related land disturbances. Areas where the surface was first disturbed, bare coal appeared, and then recovered were identified as potential internal dump areas. In contrast, areas where the surface was disturbed and recovered without the appearance of bare coal were considered potential external dump areas. The identification of these areas helps in assessing the extent and pattern of mining activities and their impact on the environment³⁵. The potential areas for internal and external dumps are shown in Fig. 7. However, the extraction results still contained some non - dump interference information, discontinuous edge areas, and holes, which could affect the accuracy of the analysis and needed further processing.

Fig. 7

Potential internal and external dump areas before morphological refinement. This map shows the initial classification based on the presence (internal) or absence (external) of bare coal during the disturbance-reclamation cycle.

Extraction of accurate areas for internal and external dumps

To improve the accuracy of the extracted areas, mathematical morphology processing was applied. This technique is widely used in image processing for tasks such as image segmentation and edge smoothing. It involves selecting certain geometric structural elements, such as square, rectangular, and rhombus structures, and utilizing their basic properties to segment and recognize images. Common morphological operation methods include dilation and erosion, opening and closing operations, morphological gradient, Top-hat transform, skeleton extraction, and watershed analysis. For binary images, the four commonly used morphological processing methods are dilation, erosion, opening, and closing operations (Table 1)³⁶.

Table 1 Morphological processing rules for the boundary of dump site.

In this study, the potential areas for internal and external dumps were first converted into binary images. Then, image processing rules were established to perform morphological processing on the extracted potential areas. The main goal was to remove the non-dump interference information, smooth the discontinuous edge areas, and fill the holes. By applying these morphological operations, the extracted areas were refined, and the final accurate areas for internal and external dumps were obtained. The results are shown in Fig. 8.

Fig. 8

Final distribution of internal and external dumps after morphological refinement. The map shows the refined boundaries of both dump types after morphological operations were applied to the areas in Fig. 7 to remove noise and smooth edges.

The accurate extraction of internal and external dump areas provides a more reliable basis for further analysis of the spatiotemporal patterns of mining-related land disturbances and the development of effective reclamation strategies. Understanding the dynamics of the mining activities and the reclamation efforts over time can help in optimizing the reclamation strategies and improving the overall environmental impact of mining operations. The morphological processing techniques used in this study can effectively enhance the accuracy of the extracted areas, making it easier to identify and analyze the potential and accurate areas for internal and external dumps.

Feature importance based on the Random Forest Gini index showed that elevation (0.32) and slope (0.27) were the most influential variables, followed by TPI (0.18), curvature (0.13), and roughness (0.10). This ranking demonstrates that terrain features substantially enhanced classification performance by complementing spectral indices in distinguishing internal and external dumps³⁷.

As shown in Table 2, model performance improved substantially with the addition of BCI, topography, and morphological refinement. LandTrendr alone produced an F1 of 0.69. Adding BCI raised the F1 to 0.81, which corresponds to an ~ 18% relative improvement. Incorporating topography and morphology further enhanced the performance, with the full model reaching an F1 of 0.84. This corresponds to an overall relative improvement of ~ 22% compared with LandTrendr only, underscoring the critical role of BCI and morphological refinement in improving dump classification.

Table 2 Ablation study results for different model configurations.

Error analysis revealed occasional confusion between bare coal and gray infrastructure such as roads, industrial yards, and rooftops. Minor overlaps with water bodies were also observed, although this effect was limited by the sparse water distribution in the study area.

Future work

Our framework, which fuses a novel BCI with LandTrendr temporal segmentation, achieved a notable 22% relative accuracy gain over baseline methods. By leveraging the 35-year Landsat archive, this scalable approach provides a new benchmark for monitoring mining landscapes, particularly in its unique ability to differentiate internal and external dump trajectories. Building on this foundation, we identify several key directions for future research.

While the BCI performed well, its robustness must be tested across diverse geological and environmental contexts to establish its generalizability. A primary limitation of our current validation is the lack of independent site-level testing, so future work will implement more rigorous schemes, such as leave-one-dump-out and seasonal-split cross-validation, to better assess model performance and prevent spatial autocorrelation³⁸. To further enhance classification precision, especially along dump boundaries, our framework could be integrated with higher-resolution data sources. Fusing the Landsat-derived trajectories with imagery from Sentinel-2 or unmanned aerial vehicles (UAVs) would offer a powerful tool for refining object edges, improving mixed-pixel analysis, and enabling near-real-time monitoring³⁹.

Beyond these technical refinements, a critical avenue for future research is to move toward a comprehensive ecological assessment. Our finding that vegetation recovery in internal dumps was 19% lower than in external dumps warrants deeper investigation into the long-term impacts on soil properties, biodiversity, and ecosystem services. Such studies are essential for developing targeted, evidence-based reclamation strategies that promote sustainable ecosystem recovery⁴⁰. Finally, while topographic variables demonstrably improved model accuracy, their specific contributions were not systematically evaluated. Future analysis will therefore focus on optimizing the selection and extraction of terrain-derived features to reveal the underlying geomorphological drivers of spoil dump distribution and further enhance classification performance⁴¹.

Conclusion

This study successfully developed and validated a novel framework for the long-term mapping of open-pit mining landscapes. By fusing a custom spectral index (BCI), temporal segmentation (LandTrendr), and a machine learning classifier, our approach provides a robust and scalable solution for tracking the 35-year evolution of mining activities and reclamation efforts in the Pingshuo mining area.

The key findings demonstrate the power of this fused methodology. First, the BCI achieved high accuracy (92% OA) in identifying bare coal, significantly outperforming traditional indices. Second, the LandTrendr algorithm accurately captured the spatiotemporal trajectories of land disturbance and subsequent vegetation recovery. Finally, our machine learning model, which integrated these spectral and temporal data streams, successfully classified internal and external spoil dumps. The subsequent application of morphological operations further refined these mapping results, offering a nuanced understanding of mine waste distribution critical for land management.

The implications of this research extend beyond the study area. The developed machine learning framework provides a replicable and scalable tool for stakeholders to monitor mining operations and assess reclamation effectiveness. By offering a cost-effective alternative to expensive, data-limited surveys, this approach helps bridge the gap between scientific monitoring and operational management. Ultimately, integrating advanced remote sensing analytics, such as the data fusion techniques demonstrated here, into policy and planning is crucial for balancing resource extraction with long-term ecological stewardship.