Deep Learning-Based Multitemporal Spatial Analytics for Assessing Reclamation Compliance of Coal Mining Permits in Kalimantan with Satellite Images

Highlights

What are the main findings?

• A multitemporal U-Net deep learning analysis revealed that changes in mining and reclamation activities across different permit holders closely align with the calculated Reclamation Activity Index (RAI).
• The Reclamation Compliance Ratio (CR) varies among coal mining concessions, showing that while several companies maintain balanced reclamation progress, others still need improvement to achieve proportional restoration.

What is the implication of the main finding?

• The RAI offers a practical and dynamic indicator for monitoring reclamation activities through satellite-based multitemporal analysis, enabling both regulators and industries to track reclamation trends over time.
• The CR findings highlight opportunities to enhance coordination between regulatory bodies and mining companies to promote balanced, sustainable and economically supportive reclamation management.

Abstract

Monitoring reclamation compliance is important to ensure mining activities follow environmental regulations and reduce land degradation. Yet, few studies directly assess compliance by linking multitemporal satellite data with mining permits. This study presents a multitemporal spatial analytics approach to evaluate reclamation compliance in coal mining permit areas in South Kalimantan, Indonesia. Using satellite imagery from 2016 to 2021, a U-Net-based deep learning classification model classified five land surface types (topsoil, subsoil, vegetation, coal bodies and water bodies) with 0.94 accuracy and a Kappa statistic of 0.91. However, this relatively high accuracy was influenced by the dominance of vegetation compared to more challenging classes such as topsoil and subsoil, which remain subject to misclassification. Analysis of temporal transitions revealed patterns of surface disturbance and delayed reclamation, particularly shown by increased subsoil and reduced vegetation. These changes were integrated with coal mining permit boundaries to derived compliance ratios (CR) ranging from 0.32 to 1.44 across nine permit holders, most of which showed moderate to excellent compliance levels. This indicates that reclamation efforts have been generally being implemented, with several permit holders exceeding expectations, while a few others still need to improve. Reclamation Activity Index (RAI) was developed to classify annual performance and showed strong alignment with the U-Net-based deep learning classification model for surface change trends. The proposed approach provides a scalable and practical tool to support evidence-based monitoring and enforcement of mining reclamation policies.

1. Introduction

Reclamation compliance is essential to mitigate environmental damage from coal extraction. Given its economic significance, coal mining operations contributed about 21.3% Gross Domestic Product of South Kalimantan province, Indonesia in 2021 [1]. This makes environmental control even more crucial. A previous study estimated that the conversion of 7618 hectares of forest and shrubland of coal mine area in Kutai Kartanegara could result in emissions of about 271 thousand tonnes of carbon [2]. Reclamation has been well implemented in certain regions in response to these environmental pressures. Nevertheless, the scale and effectiveness of these initiatives are still unknown in comparison to mining operations. As a result, monitoring the actual environmental impact of mining activities depends on long-term observation of land surface changes. With a focus on areas with active or expanding mining activities, it is important [3,4].

While evaluating the consequences of mining, long-term observation is necessary. Knowledge of reclamation workflow enables the accurate interpretation of alterations to the land surface. Usually, mining operations start with the destruction of vegetation and then the removal of layers of topsoil and subsoil. Topsoil restored and planted with vegetation helps restore the ecosystem after mining. However, in some cases, mining activities end without any reclamation efforts. This result in prolonged or even permanent landscape degradation. Regulatory systems have been developed to solve these difficulties and guarantee responsibility in the restoration process, according to Indonesia’s Regulation No. 07 of 2014. Companies must implement approved reclamation policies before they can recover financial guarantees, thereby ensuring the proper execution of reclamation. Reporting that government authorities awarded over 3000 mining permits without ensuring that the necessary post-mining bond obligations were completed, the Indonesian Supreme Audit Agency exposed significant enforcement discrepancies [5]. Reclamation is not only a legal obligation but also an ecological necessity, as it directly contributes to carbon sequestration. Han et al. (2022) demonstrated that reclamation in the Yanzhou Coalfield reduced carbon storage loss by more than 18,000 Mg compared to a no-reclamation scenario, with some sites avoiding over 70% of potential carbon loss [6]. This evidence highlights that enforcing reclamation policies not only restores landscapes but also strengthens regional carbon sink functions [6]. Even with such rules, the pragmatic implementation of compliance guarantees remains difficult. The audit process itself consists of thorough document verification, spatial data cross checking and on site inspections. While it does not replace field audits, remote sensing-based monitoring offers a practical supplement by delivering quick-access data for early stage evaluation of reclamation outcomes.

This study closes this gap by using a multitemporal deep learning strategy to understand surface transitions, including soil stripping, coal extraction, pit flooding and vegetation regrowth. It merges Sentinel-2 data with a U-Net architecture with a ResNet backbone. Sentinel-2’s 10 and 20 m resolution may overlook localized disturbances like small transport highways, even if it is enough for most land-cover change detection combine with deep learning algorithm. Despite this limitation, deep learning has still shown better performance. Artificial neural networks (ANN), for example, attain up to 91.5% accuracy in land cover classification outperforming Support Vector Machines (SVM) and Random Forests [7]. Further improving spatial edge identification, the U-Net plus ResNet model helps distinguish between minor classes such as topsoil and subsoil. These surface change classifications are then combined with official coal mining licenses to derive reclamation compliance ratios (CR), therefore providing a direct assessment of disturbed against recovered areas. To reflect annual variation in restoration efforts, the Reclamation Activity Index (RAI) is introduced as a visual metric that captures annual reclamation intensity relative to vegetation disturbance, supporting the spatial assessment. RAI is intended for straightforward, repeatable monitoring in which past ground surveys are lacking unlike conventional measures. These methods complement each other and provide a scalable and sensible structure for compliance assessment.

Few satellite-based studies have coupled coal mining categorization with permit-level compliance tracking, despite many having investigated mining dynamics. Ali et al. [8] combined barren areas and coal pits into one class, hiding potential necessary information on the reclamation process. Other studies related to reclamation and post-mining compliance indicate that only 52% of mining license holders have fulfilled these responsibilities. This includes reclamation under supervision, placing reclamation guarantee funds and meeting fund placement requirements for clean mining plan approval certification [9]. This paper proposes a practical solution utilizing deep learning with satellite data to facilitate faster compliance audits, inform regulatory prioritization and promote open environmental governance in resource-rich areas.

2. Materials and Methods

This study’s methodological approach comprises three primary components: Deep learning-based classification of surface changes in coal mining, the development of a RAI and an assessment of compliance with coal mining licenses (Figure 1). The procedure begins with the preprocessing of Sentinel-2 satellite imagery, in which atmospheric correction is implemented to transform Top of Atmosphere (TOA) data into Bottom of Atmosphere (BOA) reflectance utilizing Sen2Cor (V2.12). A set of spectral indicators is generated and the data are combined into annual mosaic composites to capture the change in the coal mining surface. A U-Net-based deep learning model utilizing the U-Net architecture with a ResNet backbone is trained on manually annotated samples that identify essential surface classifications, including topsoil, subsoil, vegetation, coal bodies and water bodies. This model is used to generate yearly land cover maps and detect surface transitions.

Complementing this classification approach, the RAI is introduced to quantify the intensity of reclamation and vegetation disturbance using spectral band comparisons between consecutive years. The RAI is a vegetation-based metric employed to analyze these changes by differentiating between vegetation disturbance, especially mining activity (RAI < 0) and reclamation (RAI > 0), through the comparison of Near-infrared (NIR) and red bands throughout successive years. The categorized surface data and RAI results are subsequently merged with official coal mining permit boundaries to evaluate compliance at the permitted level. This integrated method enables the spatially explicit monitoring of reclamation efforts and supports permit-based compliance assessments in coal mining regions.

2.1. Study Area

Hulu Tapin South Kalimantan Coal Mining region located between Tapin Regency and Sungai Raya in South Kalimantan, Indonesia, is one of the primary coal industry areas in the province. Current estimates indicate that this area possesses roughly 8 billion tons of coal deposits, with a single mining concession having over 3.5 million tons [10].

This study analyzes the mining corridor extending from South Tapin to South Hulu Sungai between about 115°10′ and 115°15′ East Longitude (Figure 2). This region is the core of mining activities in South Kalimantan, resulting in numerous ecological and economic interactions. The mining corridor demonstrates substantial contradictory effects, while mining boosts regional economies and disrupts the environment. Emphasizing the need for ecological balance, this research monitors local mining and environmental sustainability.

Tapin and South Hulu Sungai Regencies share geographic and economic characteristics that benefit agricultural and mining activities. Tapin Regency is optimal for large-scale mining and agriculture because it features a predominantly flat topography, with 67.34% of its area lies 0–7 m above sea level. Most of the land (82.93%) has gradients between 0% and 2%, while only 1.21% exceeds 500 m. The tropical savanna climate affects mining and restoration by influencing soil stability and vegetation regeneration [11].

South Hulu Sungai Regency complements this profile with a few different attributes. It mostly has low-lying terrain suitable for extensive agriculture and mining. Capital Kandangan is the administrative and economic center of agriculture, benefiting from the region’s flat to moderately sloping landscape.

Supporting the physical condition, based on the economic sector’s contributions to the Gross Regional Domestic Product (GRDP) of South Kalimantan from 2019 to 2023 (

Table 1. Economic sector contributions to GRDP of South Kalimantan (2019–2023) in percentages (%). * Preliminary data. ** Provisional data.

Sector	Industry	2019	2020	2021	2022 *	2023 **
Primary	Agriculture, Forestry and Fishing	14.31	14.40	13.58	11.42	11.37
	Mining and Quarrying	19.06	18.28	21.46	32.05	30.82
Secondary	Manufacturing	13.67	13.52	13.55	11.37	10.75
	Electricity and Gas	0.14	0.15	0.15	0.13	0.13
	Water supply and Waste Management	0.42	0.45	0.43	0.36	0.37
	Construction	8.26	8.45	7.96	7.60	7.00
Tertiary	Wholesale and Retail Trade	10.54	10.45	10.07	9.06	8.61
	Transportation and Storage	6.91	6.58	6.13	6.28	7.15
	Accommodation and Food Services	2.12	2.15	2.04	1.79	1.86
	Information and Communication	3.64	3.93	3.83	3.30	3.32
	Financial and Insurance Activities	3.57	3.67	3.50	3.00	3.07
	Real Estate Activities	2.28	2.43	2.37	1.96	1.95
	Business Activities	0.72	0.73	0.70	0.61	0.65
Public	Public Administration and Defence	6.14	6.36	5.93	4.79	4.66
	Education	4.88	5.08	4.78	3.95	3.91
	Health and Social Work	2.01	2.22	2.28	1.94	1.91
Other	Other Services Activities	1.32	1.34	1.25	1.10	1.15
	Total GRDP	100.00	100.00	100.00	100.00	100.00

). Compared to other industries, mining and quarrying are the most significant contributors to the region’s economy. Mining activity has steadily increased from 19.06% in 2019 to an impressive peak of 32.05% in 2022 and has since remained at 30.82% in 2023. This results in mining being the most significant economic contributor compared to other sectors. Mining activity influences various sectors [1] such as manufacturing, trade, accommodation and food services by creating employment opportunities for local people.

Due to the substantial economic impact of mining activities, it is impossible to eliminate or significantly reduce them. A comprehensive and systematic monitoring system is therefore required to ensure that coal mining permit holders comply with post-mining obligations such as reclamation, maintaining balance in mining operations and environmental stability.

2.2. Sentinel 2 Preprocessing

The primary data source for this study was Sentinel-2 satellite imagery derived from the European Space Agency’s (ESA) open-access Sentinel Scientific Data Hub between 2016 and 2021 [12]. The acquisition dates were not uniform across months or seasons. The study area in Kalimantan is near the equator, where frequent cloud cover limits data availability and makes it difficult to obtain images from consistent time periods. Level-1C images with minimal cloud cover were chosen to ensure the dataset’s quality, emphasizing clarity. The imagery was further processed with ESA’s Sen2Cor tool to provide Level-2A orthoimages, which deliver BOA or surface reflectance products adjusted for atmospheric distortions [13].

$$\mathrm{NDVI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{Red}}}$$

(1)

$$\mathrm{NDCI}={\displaystyle \frac{\mathrm{SWIR}1-\mathrm{NIR}}{\mathrm{SWIR}1+\mathrm{NIR}}}$$

(2)

$$\mathrm{BAI}={\displaystyle \frac{\mathrm{Blue}-\mathrm{NIR}}{\mathrm{Blue}+\mathrm{NIR}}}$$

(3)

$$\mathrm{SAVI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{Red}}{\mathrm{NIR}+\mathrm{RED}+\mathrm{L}}}\ast (1+\mathrm{L})$$

(4)

$$\mathrm{NDWI}={\displaystyle \frac{\mathrm{Green}-\mathrm{NIR}}{\mathrm{Green}+\mathrm{NIR}}}$$

(5)

$$\mathrm{BSI}={\displaystyle \frac{(\mathrm{SWIR}1-\mathrm{Red})-(\mathrm{NIR}+\mathrm{Blue})}{(\mathrm{SWIR}1+\mathrm{Red})-(\mathrm{NIR}+\mathrm{Blue})}}$$

(6)

$$\mathrm{NDMI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{SWIR}1}{\mathrm{NIR}+\mathrm{SWIR}1}}$$

(7)

The Sentinel-2 dataset consists of 13 spectral bands with various resolutions: 4 bands at 10 m (Red, Green, Blue, NIR), six bands at 20 m (Vegetation Red Edge (VRE) 1, VRE 2, VRE 3, Short-Wave Infrared (SWIR) 1 and SWIR 2) and three at 60 m (Water Vapour and SWIR-Cirrus). This research primarily utilized the 10-m and 20-m resolution bands, as the more effective spatial resolution enhances analytical precision.

Preprocessing involved atmospheric correction using Sen2Cor to reduce the impacts of aerosols and water vapor, guaranteeing that spectral readings closely indicate surface conditions. The change is vital to improving the reliability of the classification. The following steps included image compositing and feature stack preparation via ArcGIS Pro 3.5, enabling the systematic input data arrangement for U-Net-based deep learning model training.

This study employed the original bands from Sentinel-2 data and bands derived from spectral indices on coal mining mapping for the training model inputs. Spectral indices are mathematical formulas employed to evaluate remote sensing image bands, enhancing the differentiation of detected objects. We calculated a total of seven spectral indices and used nine spectral bands from 10 and 20 m resolutions of Sentinel 2, resulting in 16 composite bands in total, which include as the Normalized Difference Vegetation Index (NDVI), Normalized Difference Coal Index (NDCI), Built-up Area Index (BAI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Water Index (NDWI), Bare Soil Index (BSI) and Normalized Difference Moisture Index (NDMI), among others as shown on Formulas (1)–(7).

These indexes aim to improve coal mining surface change and reclamation area differentiation. The NDVI is frequently utilized for assessing diverse vegetation components, and its calculation for Sentinel-2 imagery aids in identifying regions with less vegetation cover, perhaps signaling prospective coal mining locations [14,15]. The NDCI enhances the differentiation between mined regions and other land cover types by normalizing the difference between the SWIR-1 and NIR bands, which are closely linked to mining attributes in satellite imagery [16,17]. The BAI can enhance water bodies area classification by employing Normalized discrepancies between the blue and NIR bands, consequently improving the identification of coal mining areas [18].

The SAVI was employed to reduce the impact of soil brightness on vegetation detection, particularly in regions with sparse vegetation. Where the soil adjustment factor L (commonly set to 0.5) helps minimize background soil noise [19,20]. The NDWI facilitates the identification of water bodies by amplifying the contrast between aquatic and terrestrial surfaces, which is crucial for detecting waterlogged mining regions [21]. The BSI is especially effective for identifying bare soil, which is common to mining activities [22,23]. The NDMI is employed to assess moisture content in vegetation and soil, strengthening the identification of environmental changes caused by mining [24,25]. These indicators collectively improve the classification of land cover and provide a thorough examination of surface shifting in coal mining regions.

2.3. Coal Mining Surface Change Classification U-Net Based Deep Learning Model

The application of deep learning in remote sensing has been considerably enhanced by using ANN, particularly models such as U-Net and ResNet [26,27,28]. Deep learning algorithms have resulted in substantial improvements in the classification of land use and land cover (LULC), particularly for detecting coal mines [29,30,31].

The utilization of deep learning techniques has expanded due to the increasing availability of remote sensing data and the advancements in computing capacity [32,33,34]. The U-Net-based deep learning model was trained using Sentinel-2 images co rrected for atmospheric effects using Sen2Cor in this investigation. The dataset was obtained using an adjusted topographic map of the coal mining area in Kalimantan for training data partitioned into training and validation sets, with 10% of the dataset being used for validation (

Table 2. Land cover classes distribution of training dataset.

Land Cover Classes	Distribution (%)
Topsoil	17
Subsoil	16
Vegetation	40
Coal Body	10
Water Body	17

). During the training process, U-Net was optimized to identify spatial and spectral patterns associated with topsoil, subsurface, coal bodies, vegetation and water bodies using ResNet-34 as a feature extractor. The validation phase evaluated the model’s performance on new data and modified parameters to enhance accuracy. The trained model was employed to classify these surface components, thereby facilitating the analysis of the variations in the coal mining surface over time.

To monitor coal mining and reclamation progress, topsoil, subsoil, coal bodies, vegetation and water bodies on the coal mining surface must be monitored. By identifying surface components, the model creates a realistic map of transitions, including reclamation, soil stripping and plant removal. These changes can be studied periodically to assess the extent to which mining alters the land surface and the effectiveness of reclamation in restoring it. The model identifies specific reclamation phases, including soil recovery and vegetation regeneration. These are crucial for monitoring the success of environmental restoration efforts in mining regions.

The architecture specifically designed for semantic segmentation incorporates a contracting path and an expansive path by the standard convolutional network architecture. This study integrates U-Net and ResNet-34 providing further insight into the performance improvements this combination yields. The Kaiming team at Microsoft Research Institute introduced ResNet to resolve degradation issues associated with training error rates in deeper neural networks. This approach balances the time constraints frequently encountered when training more extensive models that are limited by their number of layers [35]. Furthermore, it is crucial to acknowledge that this combination exhibits potential for classification tasks in remote sensing applications. It is accurate and firmly in the face of potential variations that may arise during real-world utilization scenarios. This method has been demonstrated to be beneficial in monitoring general coal mining surface change and tracking reclamation efforts. Consequently, it is an valuable tool for the long-term monitoring and management of coal mining region.

2.4. Confusion Matrix Accuracy Assessments

The accuracy and robustness of a U-Net-based deep learning model can be achieved by using a confusion matrix. The confusion matrix provides a detailed view of the model’s accuracy and identifies errors made by the model. A confusion matrix is managed in a tabular format, with rows representing the predicted class and columns representing the actual classes [36]. A confusion matrix was applied to evaluate the accuracy of the classification results based on true and false values [37]. In this study, the ground truth data were obtained from the adjusted topographic map provided by the Ministry of Environment and Forestry. Unlike common approaches that rely on limited sample points, the confusion matrix was constructed using the entire AOI within the reference map. This strategy reduces potential bias from selective sampling and ensures that the reported accuracy and Kappa statistics reflect the full spatial variability of the study area. For model training, representative subsets were also selected from the same adjusted topographic map to cover all land surface classes (topsoil, subsoil, vegetation, coal bodies and water bodies). In addition to overall accuracy and Kappa, the evaluation also reports Producer’s Accuracy (PA) and User’s Accuracy (UA) for each class, providing insights into omission and commission errors across land cover categories. The model’s overall accuracy and the Kappa value serve as critical metrics for assessing the performance of U-Net-based deep learning models in detecting surface changes in coal mining and monitoring reclamation activities [38]. The metrics assess the model’s reliability and effectiveness, confirming its capability to identify changes and activities within coal mining regions accurately.

In coal mining surface change detection, a model may demonstrate high overall accuracy. However, this may result from the model mostly predicting the majority class (vegetation) while improperly recognizing the minority class (coal bodies). The confusion matrix helps clarify this issue and reveals potential biases in the model. It also supports model improvement by enhancing its ability to identify minority classes.

General accuracy and the Kappa value are critical measures in assessing the performance of U-Net-based deep learning models for coal mining activity detection. Overall accuracy measures the proportion of correctly classified samples, while the Kappa coefficient adjusts for the agreement expected by chance, making it more reliable in cases of class imbalance [38]. Offering a more reliable assessment than simply accuracy, the Kappa value gauges the agreement between the expected and actual classes while controlling for the possibility of agreement. Conversely, overall accuracy is the percentage of accurately categorized samples out of all the samples.

Moreover, the confusion matrix is rather useful especially in cases of considerable class imbalance, which is typical in real-world data [39,40]. This approach helps academics recognize biases, improve the model and determine whether further data is required. Finally, in the evaluation and enhancement of U-Net-based deep learning models in the detection of coal mining surface changes with confusion matrix is an effective tool.

2.5. Coal Mining Surface Change and Reclamation Workflow

Almost all coal mining areas in Kalimantan employ open-cast mining techniques. The open-cast mining technique has several advantages, including lower operational costs, easier distribution, higher productivity and easier monitoring. However, the open-cast mining technique has some disadvantages, including higher environmental impact, wider land cover, dust pollution, water contamination and social conflict; therefore, reclamation is a must. Reclamation activity is in the post-mining rehabilitation stage, as shown in Figure 3. Mining activity has three phases: pre-mining, mine excavation and post-mining rehabilitation. The pre-mining phase involves clearing vegetation and soil stripping, which involves removing topsoil from another area. This soil is reinstated during post-mining rehabilitation to create multilayered coal bodies with subsoil. After that, the mine excavation phase focuses on obtaining coal and transporting it to the destination buyer. The coal mining area enters the post-mining rehabilitation phase after no coal bodies remain on the site. This phase focuses on restoring topsoil and revegetation using plant species commonly used in Kalimantan, such as Bamboo, Rain Tree, Albizia, Cajuput Tree, Solomon Albizia and Pula Tree. These fast-growing and adaptive species help accelerate ecosystem recovery. The reclamation process is completed when the area becomes vegetated again, but monitoring is still required even after the reclamation is finished.

Monitoring each stage of the coal mining process is vital, from pre-mining to post-mining rehabilitation. The alterations in mining operations may be observed by remote sensing visual interpretation (Figure 4), which demonstrates the visual interpretation of several mining cover classes (topsoil, subsoil, coal deposits and water bodies). The visual analysis of remote sensing pictures can be improved by applying a false-color composite rather than the actual color. A false-color composite employing the SWIR band can differentiate the soil properties, indicating that the topsoil retains more moisture than the subsoil. Differentiating between topsoil and subsoil is crucial for comprehending its relevance to mining or reclamation processes. The texture observed in the satellite imagery can identify the coal bodies. Water bodies need to be dried out when mining activities are necessary. As determined by remote sensing analysis, mining operations ceased if water levels were up.

Remote sensing monitoring can utilise a false-colour composite that focuses on the surface reflectance response to each class of land cover, such as SWIR bands, to distinguish between topsoil and subsoil (Figure 5). Vegetation shows stronger reflectance in the NIR band than in other bands. Coal bodies also have higher responses in the SWIR bands. In contrast, water bodies appear weak in SWIR, which helps distinguish moisture levels between topsoil and subsoil in coal mining areas.

Utilizing the capacity of remote sensing can facilitate detailed monitoring of coal mining workflow. Utilizing a U-Net-based deep learning model for satellite imagery classification is a faster approach to understanding the coal mining surface change and reclamation workflow. Coal mining permits can be approached using multitemporal datasets based on classification, and the transition can be monitored to ensure compliance with regulations in coal mining.

2.6. RAI for Assessing Disturbance and Recovery in Mining Landscapes

This study tries to understand the reclamation process in coal mining with two approaches. First, a deep learning algorithm is used for classification. Then, based on the change in multi-temporal data classification, it can be determined which areas are involved in reclamation activities and which areas are involved in mining activities. Second, more straightforward approach only uses the formula for two different time satellite imagery in the exact location. This approach offers the benefit of a quicker methodology compared to deep learning, which requires extensive training with data to ensure the model’s accuracy. The index approach is also more straightforward, as evidenced by a Geographic Information System (GIS) dashboard, where only inputting the formula yields reclamation activity.

$$\mathrm{RAI}={\displaystyle \frac{({\mathrm{NIR}}_{\mathrm{post}}-{\mathrm{RED}}_{\mathrm{post}})-({\mathrm{NIR}}_{\mathrm{pre}}-{\mathrm{RED}}_{\mathrm{pre}})}{({\mathrm{NIR}}_{\mathrm{post}}+{\mathrm{RED}}_{\mathrm{post}})+({\mathrm{NIR}}_{\mathrm{pre}}+{\mathrm{RED}}_{\mathrm{pre}})}}$$

(8)

The RAI is inspired by NDVI, which distinguishes vegetation characteristics using the NIR and red bands. The NIR band shows high reflectance from leaf cells, while the red band has low reflectance due to chlorophyll absorption. However, reclamation means not only the establishment of vegetation but also the conversion of other land cover types, especially subsoil, to vegetation (Figure 6). Typically, the index formula utilizes a single remote-sensing image. However, this study aims to develop a formula that can monitor changes between two images. Therefore, according to Formula (10), the formula focuses on the different times of NDVI.

The values of the RAI classification are shown in

Table 3. Classification of RAI values.

RAI Range	Class	Interpretation
>0.3	High Reclamation	Strong NDVI increase; large-scale restoration.
0 to 0.3	Moderate Reclamation	Partial recovery or early vegetation regrowth.
0	Stable/Neutral	Minimal change; stable cover or minor transitions.
−0.3 to 0	Moderate Disturbance	Vegetation loss; early mining or land clearance.
<−0.3	High Disturbance	Severe degradation; active or expanding mining.

. A high RAI value does not mean only that the area has high vegetation. Still, it has also undergone drastic changes from pre-imagery, such as the transformation of water bodies into vegetation. Because NDVI values are low in water bodies, lower RAI values may not only indicate sparse vegetation but also disturbances such as mining activity. This is especially true near water bodies where NDVI values can be extremely low. The RAI approach can be an alternative solution for monitoring reclamation activity and become a trend for indexing using a two-time dataset, allowing for quick monitoring of any changes.

2.7. Reclamation CR in Coal Mining Permit

Ensuring reclamation compliance is a fundamental responsibility of coal mining permit holders, forming the basis of environmental accountability in coal extraction activities. Public perception often opposes mining because of its ecological impacts. However, regulated mining with proper post-mining reclamation can provide both economic benefits and landscape recovery.

Post-mining reclamation is regulated under National Indonesian Law No. 4/2009 on Mineral and Coal Mining and Ministerial Regulation No. 7/2014. These regulations clearly assign reclamation and post-mining recovery responsibilities to coal mining permit holders. While the government acts as a supervisor, the full responsibility of planning, funding and executing the reclamation effort lies with the permit holder themselves. On the other hand, the Government has a role in monitoring reclamation compliance to protect environmental interests. Reclamation does not mean returning the land to its original state. Restoring to pre-mining conditions, especially in terms of elevation, may impose an environmental burden by extracting large quantities of fill material from hills or rivers in other areas.

$$\mathrm{CR}={\displaystyle \frac{\mathrm{Total}\mathrm{Area}\mathrm{of}\mathrm{Reclamation}\mathrm{Activity}}{\mathrm{Total}\mathrm{Area}\mathrm{of}\mathrm{Mining}\mathrm{Activity}}}$$

(9)

Monitoring reclamation compliance activity can be achieved by using multitemporal data, which involves collecting data at least two different times in the coal mining area. Eventually, understanding the transition of land cover is crucial. Reclamation activity encompasses not only vegetation but also the transition from other land cover, such as topsoil, to vegetation. There is a change in land cover. Additionally, mining activity cannot be limited to the subsoil alone. Current mining areas may have originally been forested landscapes prior to coal extraction activities. As shown in

Table 4. Transition categories representing mining and reclamation activities based on land cover changes for calculation CR.

Mining Activity	Reclamation Activity
Topsoil → Subsoil	Topsoil → Vegetation
Topsoil → Coal Bodies	Subsoil → Topsoil
Topsoil → Water Bodies	Subsoil → Vegetation
Subsoil → Coal Bodies	Coal Bodies → Topsoil
Subsoil → Water Bodies	Coal Bodies → Vegetation
Vegetation → Topsoil	Water Bodies → Topsoil
Vegetation → Subsoil	Water Bodies → Vegetation
Vegetation → Coal Bodies
Vegetation → Water Bodies
Coal Bodies → Subsoil
Coal Bodies → Water Bodies
Water Bodies → Subsoil
Water Bodies → Coal Bodies

, there are class changes in mining activity and reclamation activity for the CR formula. The goal is to preserve vegetation for environmental safety while understanding mining history to identify activity patterns.

The CR is calculated from the combined multitemporal mining activity and reclamation activity within the coal mining permit area (Formula (9)). CR value classification is shown in

Table 5. CR classification for post-mining reclamation performance.

CR	Classification	Description
CR ≥ 0.70	Excellent	Most of the disturbed area has been reclaimed
0.40 ≤ CR < 0.70	Moderate	Reclamation effort is present but insufficient
CR < 0.40	Poor	Reclamation lags far behind mining

, which has three levels of difference. An excellent CR indicates that compliance is generally accomplished. Moderate means that the coal mining permit required further monitoring. Poor CR indicates potential regulatory concerns regarding coal mining permits. The CR approach can be an alternative and fast way to understand how well the coal mining permit holder responsibility of their reclamation activity.

3. Results

3.1. Deep Learning Classification

Prior to the preparation of training data, several spectral indices were calculated to highlight different surface conditions of coal mining areas as shown in Figure 7. Training data from coal mining areas in Kalimantan were visually interpreted and classified into five land cover classes. Following the Export Training Data for Deep Learning and Train Deep Learning Model workflows, these training samples were then used to develop a U-Net deep learning classification model in ArcGIS Pro. The training and validation results of the U-Net deep learning network are shown in Figure 8. The model was trained using 205 Sentinel-2 images, which were cropped into 256 × 256-pixel tiles, producing a total of 6931 training patches (features).

Representative samples were generated by creating square training tiles around coal mining areas across Kalimantan using the adjusted topographic map as a guide. Mining sites are typically surrounded by vegetation. As a result, a considerable portion of vegetation was also included in the sampled patches. This reflects the actual landscape context but also contributes to class imbalance that was later addressed by balancing strategies during training. To balance each land cover type, 256 × 256-pixel tiles with a stride of 128 × 128 pixels were used to increase the training samples.

This improved model generalization by strengthening training. The information includes imagery of topsoil, subsoil, vegetation, coal bodies and water bodies. The model was trained for 200 epochs with a batch size of 8 and a learning rate 0.001. The images used for training were selected from coal mining areas with minimal cloud cover to ensure data suitability for analysis. The extended training process allowed for better adaptation to variations in the dataset and helped mitigate overfitting, though regularization methods like Dropout and L2 were not explicitly used. ArcGIS Pro’s built-in optimization managed the weight adjustments to improve model stability and reduce overfitting.

A validation dataset consisting of imagery with 16 bands and five classes was used for model evaluation. Validation loss was tracked during training to monitor model performance. Validation loss was primarily used to assess the model’s performance on unseen data and to track overfitting.

The model categorizes coal mining areas using U-Net and ResNet backbones. The graph shows its training and validation loss. Training and validation loss are initially high but decrease as training progresses. This indicates that the model is learning and improving its predictive accuracy. Losses stabilize and converge with time, demonstrating that training does not appreciably reduce error. Validation loss may occasionally increase due to challenging data, but it does not affect learning. The close agreement between the training and validation loss curves indicates that the model generalizes effectively to new data without overfitting or underfitting.

3.2. Deep Learning Confusion Matrix

The U-Net deep learning classification model was applied to the dataset. The results were then compared pixel by pixel with the reference data from the adjusted topographic map provided by the Ministry of Environment and Forestry to generate the confusion matrix. Figure 9 presents four comparative maps showing the reference data and classification results from the U-Net, SVM and Maximum Likelihood methods. The “Reference” map represents the ground truth dataset, while the other three illustrate their respective classification outputs. Visually, the U-Net model provides clearer boundaries and better separation between topsoil, subsoil and vegetation classes compared to SVM and Maximum Likelihood, which tend to blur or overgeneralize transitional areas.

As summarized in

Table 6. Comparison of overall accuracy and kappa coefficient between classification methods.

Method	Overall Accuracy	Kappa
U-Net (Deep Learning)	0.94	0.91
SVM	0.89	0.84
Maximum Likelihood (MaxL)	0.87	0.81

, the U-Net model achieved the highest overall accuracy (0.94) and Kappa coefficient (0.91), outperforming SVM and Maximum Likelihood. These results show that the deep learning approach can better capture spatial complexity and subtle spectral variations in mining regions. It is particularly effective in distinguishing between similar surface types such as topsoil and subsoil. The improvement in accuracy further demonstrates the robustness of U-Net in handling heterogeneous land-cover conditions typical of coal mining environments.

One likely reason for the persistent misclassification between topsoil and subsoil is the heterogeneous and dynamic nature of surface mining sites. In these areas, topsoil is often stripped and mixed with underlying subsoil, creating transitional zones that are spectrally ambiguous as shown in Figure 10. This challenge is compounded by the spectral similarity of soil layers in Sentinel-2 imagery and its moderate spatial resolution, which can produce mixed pixels at class boundaries. Future research could use higher-resolution UAV imagery or hyperspectral data. These datasets provide finer spatial detail and stronger spectral discrimination to better separate topsoil and subsoil classes.

Table 7. Confusion matrix (pixel counts) with User’s Accuracy (UA) and Producer’s Accuracy (PA).

Confusion	Topsoil	Subsoil	Vegetation	Coal Body	Water Body	Total	UA
Topsoil	93,551	1328	1376	13	380	96,648	0.97
Subsoil	656	192,363	1630	450	480	195,579	0.98
Vegetation	1300	19,380	421,826	610	610	455,426	0.93
Coal Body	31	1560	980	24,560	105	27,236	0.90
Water Body	220	47	710	50	21,204	22,231	0.95
Total	107,458	214,678	426,522	25,683	22,779	797,120
PA	0.87	0.90	0.99	0.96	0.93

summarizes User Accuracy (UA), Producer Accuracy (PA), Overall Accuracy and Kappa. Looking at the table, Subsoil had the highest UA at 0.98. However, it has a low Producer Accuracy because it is mixed with vegetation and topsoil. Vegetation, with the highest PA of 0.99 and good UA of 0.93 because the AOI is dominated by vegetation with more than 50% of the total area. The PA for Topsoil is 0.87 because this class is usually mixed with Subsoil in some areas. The Kappa value of 0.91 and overall accuracy of 0.94 demonstrate that the predicted and actual classes agree well.

The high overall accuracy was partly influenced by the dominance of vegetation in the study area. Vegetation is easier to classify than more challenging classes such as subsoil and topsoil. The model achieved high accuracy (OA 0.94, Kappa 0.91). However, errors remain, mainly in reclaimed areas where young vegetation or bare soil can be misclassified.

The confusion matrix results show that the model is effective in distinguishing coal mining areas. However, this method is more practical and scalable than manual field surveys. It is therefore useful for large-scale environmental monitoring. It may not be as precise as direct field measurements. However, it serves as a useful tool for tracking mining and reclamation activities while supporting environmental research and land management.

3.3. Multitemporal Coal Mining Surface Change

With the good accuracy of the U-Net Deep learning classification model, this model can be used with a multitemporal dataset to obtain the multitemporal coal mining surface change. Mining activities in the AOI between 2016 and 2021 resulted in notable land cover transitions, as illustrated in Figure 11. The vegetation trend decrease shows the increase of mining activity, as the expanding subsoil on the AOI shows the decrease in vegetation. On the other hand, high coal body values indicate that the ore is still present on the site. When the coal values are lower, it means the material has already been transported or exported to the buyer country. Water bodies fluctuate over time. In 2019, the water level reached its lowest point while the subsoil reached its highest extent.

Furthermore,

Table 8. Land cover composition as percentages (%) of total area (8469.30 Ha) for 2016–2021.

Year	Topsoil	Subsoil	Vegetation	Coal Body	Water Body
2016	1.8%	18.0%	71.2%	7.0%	2.0%
2017	3.7%	18.6%	69.7%	5.0%	3.0%
2018	7.1%	26.9%	61.8%	2.5%	1.6%
2019	3.4%	40.8%	51.8%	2.7%	1.3%
2020	11.5%	23.1%	59.4%	3.3%	2.7%
2021	7.0%	28.2%	58.1%	3.8%	3.0%

shows 2016–2021 land cover statistics for topsoil, subsoil, vegetation, coal and water bodies. The extent of each land cover type over the years is indicated in hectares. From 2016 to 2021, vegetation was the dominant land cover category. Its coverage decreased from 71.2% in 2016 to 51.8% in 2019. Vegetation dominated the study region for years, suggesting its broad coverage. However, coal and water bodies cover less territory, with coal bodies covering the least. Water Bodies ranged from 3% in 2017 to 1.3% in 2019.

Changes in subsoil and topsoil occurred over the years, with subsoil showing substantial growth, particularly in 2019, reaching 51.8%. Topsoil varied more significantly, peaking at 11.5% in 2020 before declining to 7% in 2021 (Figure 12). The fluctuations in topsoil and subsoil may be related to different land management practices or environmental changes in the study area. These variations could have important implications for soil conservation and land reclamation.

3.4. Multitemporal RAI

With the RAI formula applied to the multitemporal dataset, annual RAI values were calculated for each mining permit and then averaged. These results were compared with the multitemporal coal mining changes detected by the U-Net deep learning classification model.This comparison was used to examine whether the RAI trends are consistent with the classification results. BE’s RAI declined from 2016–2019 due to mining expansion without reclamation, with recovery starting in 2019–2020 (Figure 13). KB had relatively small mining activity. However, it showed a negative RAI value in the early period due to delays in reclamation. It improved significantly when reclamation began to be carried out. AS showed more active reclamation dynamics than mining activities. This resulted in a steadily improving RAI trend, indicating good compliance with land restoration.

EL showed a sharp increase in mining from 2018 to 2019. However, reclamation did not keep pace, causing the RAI to drop sharply during that period. Although, large-scale reclamation in 2019–2020 managed to increase RAI again. B2 stands out as a concession with two consecutive years of significant mining expansion. Even though it was followed by fairly extensive reclamation, the RAI values remained unstable and indicated high ecological pressure. Meanwhile, MU operated on a small scale but was able to maintain balance. Its RAI trend increased gradually due to a relatively stable reclamation pattern. AM had large mining activities and irregular reclamation patterns. As a result, its RAI fluctuated and did not show consistent land recovery.

Figure 14 shows that positive RAI means reclamation exceeded mining, while negative values mean mining dominated. In general, RAI fluctuations are quite significant between years and between companies. Several companies, such as B1 and EL, showed negative RAI in early years. However, these later shifted to positive, indicating progress in reclamation efforts. Meanwhile, companies such as KB (average RAI –0.028), AM (–0.020) and BE (–0.017) showed a predominantly negative annual RAI trend. This indicates that their reclamation activities lagged behind mining activities in almost every year. The average RAI helps explain how reclamation progresses over time. An average value close to zero or positive indicates a balance or consistency between mining and annual reclamation. In contrast, a negative average value shows a lag in reclamation implementation, which may be substantial overall but fails to keep pace with yearly mining activities.

3.5. Reclamation Compliance Assessment in Coal Mining Permits

Multitemporal coal mining changes were detected using the U-Net deep learning classification. Land cover transitions were then calculated and categorized into three groups: areas under mining activity, areas under reclamation activity and areas that remained unchanged. These transitions from 2016 to 2021 of each year were then used to calculate each coal mining permit’s CR total. Figure 15 illustrates the KB concession’s mining pattern. The activity began in the central area in 2017–2018 and later expanded rapidly toward the south and east. Reclamation only began in 2019–2020. However, most ex-mining land remained disturbed, as reflected in the low compliance rate of 0.32. This reflects a delayed and irregular reclamation pattern within the operating area.

In contrast to KB, the EL concession shows a balanced pattern of mining land development. Figure 16 illustrates its active and sustainable reclamation activities throughout the five-year period. In early 2016–2017, reclamation activities were even wider than mining. This indicates early efforts to restore or rehabilitate previously affected land. In subsequent years, mining area expansion occurred progressively.However, this expansion was almost always accompanied by the development of new reclamation areas. This pattern indicates that each mining expansion was followed by reclamation, resulting in a high compliance rate (0.73).

Between these two contrasting cases, the B2 concession area presents an intermediate pattern of mining and reclamation dynamics. Figure 17 shows the B2 concession area’s massive and widely distributed mining expansion from west to east. The activity began in 2017–2018 and reached its peak in 2018–2019. Reclamation activities were also carried out on a large scale, especially in 2019–2020. This period marked the highest level of reclamation, indicating strong efforts in land restoration. This phenomenon shows that the trend quantity of the reclamation process is small.It remains limited compared with the land disturbance caused by coal mining activities. Overall, B2 reached a moderate compliance rate (0.66). However, reclamation efforts were unable to match the scale of mining disturbance.

Furthermore, the multitemporal mining activity and reclamation activity of each coal mining permit holder with U-Net-based deep learning model can also be seen in Figure 18. There are gaps in mining activity compared with reclamation activity in the early period 2016–2017 until 2018–2019. Most permit holders dominated mining activity compared with reclamation. BE and AM especially have high mining and small reclamation activity. The mining activity peaked in 2017–2018 and 2018–2019. B1 and B2 also showed massive mining increases but relatively higher reclamation compared with AM. The years 2019–2020 show a significant increase in the reclamation activity. For almost all coal mining permit holders, reclamation activity is higher than mining activity. This trend shows there is a turning point in environmental restoration. Mining activity continued in 2020–2021. However, most coal mining permit holders showed lower reclamation activity compared to 2019–2020.

At the permit level, each coal mining permit holder operates at a different scale. This variation occurs because the study area includes both small and large coal mining permits. B1 and B2 show mining expansion followed by increased reclamation activity. Although AM conducts reclamation every year, its volume remains too low compared with its mining operations. Though operating on a small scale, MU remains active in reclamation activities. BE, EL and BR also show positive progress in their reclamation efforts. This trend is consistent with that of other coal mining permit holders in 2019–2020. On the contrary, KB shows that minimum reclamation activity requires evaluation. AS shows minimum activity on mining activity, and reclamation is higher. Even with different sizes of coal mining permits, the resulting trend is relatively the same from one coal mining permit to another. Because of this issue, it is important to know the compliance rate.

To further evaluate their performance, the compliance rate of each coal mining permit holder can be seen in Figure 19. The graph compares the extent of mining activity with reclamation activity. Permit holders such as EL (CR 0.73) and BR (0.75) are included in the Excellent category. They have large reclamation areas (EL: 530.85 Ha and BR: 37.42 Ha) and comparable levels of mining activity. AS also falls into the Excellent category with a CR of 1.44. However, its small mining area makes this value less representative, even though reclamation was carried out consistently.

Meanwhile, other permits show Moderate to Poor compliance. The Moderate category includes permit holders such as B2 (CR 0.66), MU (0.67), B1 (0.60) and BE (0.51). All four have large mining and reclamation volumes, such as B2, which mines up to 1155.87 Ha and reclaims 768.90 Ha. However, their average RAI is generally negative (B2: −0.005, BE: −0.017), indicating that reclamation does not parallel mining activities yearly. MU has an RAI close to zero (−0.002), reflecting a relatively balanced distribution of reclamation throughout the year. B1 has also started to show improvements in RAI in recent years. On the other hand, KB (CR 0.32) and AM (CR 0.41) fall into the Poor category. They have low RAI values (KB: −0.028, AM: −0.020) and show a significant imbalance between mining and reclamation. AM, for example, mined 1170.85 Ha but only reclaimed 485.36 Ha. This condition shows that reclamation is lagging in total and annually, becoming a significant concern in environmental regulatory supervision. The combination of CR, RAI and activity volume indicators offers a clearer assessment of each company’s performance. Together, these indicators provide a more complete understanding of reclamation compliance and effectiveness.

4. Discussion

Reclamation trends identified in this study align with the sustainability report from PT Kaltim Prima Coal (KPC), a subsidiary of Bumi Resources. The comparison is shown in Figure 20 [41]. According to official data, KPC’s reclamation dropped to 1085.4 hectares in 2019 before rising to 1697.48 and 1683.79 hectares in 2020 and 2021. This significant increase indicates a strengthened dedication to land restoration after previous years of decreased reclamation results.

A similar pattern is observed in Arutmin’s case, as shown in

Table 9. Cumulative reclamation performance of arutmin (bumi resources) from 2011 to 2017.

Year	Land Opening	Land Arrangement	Planting	Number of Trees
	(ha)	(ha)	(ha)	(Thousand)
2011	12,500	4700	4500	3600
2012	13,200	4900	4700	3700
2013	13,700	5100	4900	4000
2014	13,800	5200	5000	4100
2015	14,100	5400	5200	4200
2016	14,500	5700	5500	4400
2017	15,000	6100	5900	4700

[41]. The data from this Bumi Resources subsidiary show that land preparation and planting cover only about 50% of the cleared land. This ratio matches the CR trends in this research, demonstrating that reclamation compliance is still poor but improving. This indicates an improvement in coal mine reclamation management. Corporations are now showing stronger commitment to land restoration. Relevant authorities must consistently monitor reclamation activities. This ensures that they meet regulatory standards and produce meaningful ecological outcomes.

It should be noted that the Sentinel-2 imagery used in this study has a spatial resolution of 10–20 m. This means that a small area in the image can represent a much larger surface on the ground. Consequently, minor patches visible in the imagery may correspond to substantial land areas in the field. This highlights the importance of careful interpretation when assessing reclamation progress. In addition, the U-Net model may face reduced performance under complex surface conditions. This is especially true in areas where topsoil and subsoil have similar spectral characteristics. Such conditions can introduce uncertainties in detailed class-level assessments. According to this study shows that satellite imagery-based monitoring is a realistic approach to assess mine reclamation. Deep learning also provides consistent results in evaluating reclamation responsibilities.

5. Conclusions

These findings not only reveal spatial patterns but also reflect a broader narrative in how land is used, altered and restored. Reclamation is a story and history that cannot be captured solely by a single moment. This study presents an approach to assessing reclamation compliance using deep learning applied to multitemporal satellite imagery. Spatial analytics were used to track mining and reclamation activities across coal mining permit areas. The RAI provided additional insight into annual reclamation performance. Our findings showed that the U-Net-based deep learning model can classify mining land cover with 0.95 accuracy. The CR showed that compliance with reclamation varies among concessions. Most mines conduct reclamation, but the reclaimed areas often differ from the disturbed ones. This emphasizes spatial-temporal assessment in mine reclamation monitoring. Integrating spatial and temporal approaches is essential for assessing mine reclamation. These approaches help evaluate both environmental impacts and governance sustainability. Furthermore, future research should explore how reclamation stories differ across country. This is especially important for nations that depend heavily on coal mining as a key part of their economy. This cross-national understanding is crucial for examining how reclamation policies and compliance operate in different contexts. It helps reveal variations in social, economic and environmental conditions that influence implementation. Such insights provide a foundation for global learning toward more sustainable mining management.