Impact of Coal Mining on Growth and Distribution of Sabina vulgaris Shrublands in Mu Us Sandy Land: Evidence from Multi-Temporal Gaofen-1 Remote Sensing Data

Abstract

Sabina vulgaris is a keystone shrub species endemic to arid northwestern China, renowned for its exceptional drought tolerance, sand fixation capabilities, and critical role in desert ecosystem stability. This study investigates the impact of coal mining activities on the spatiotemporal dynamics of S. vulgaris shrublands in the ecologically fragile Mu Us Sandy Land, focusing on the Longde Coal Mine adjacent to the Shenmu S. vulgaris Nature Reserve. Utilizing seven periods (2013–2025) of 2 m resolution Gaofen-1 (GF-1) satellite imagery spanning 12 years of mining operations, we implemented a deep learning approach combining UAV-derived hyperspectral ground truth data and the SegU-Net semantic segmentation model to map shrub distribution via GF-1 data with high precision. Classification accuracy was rigorously validated through confusion matrix analysis (incorporating the Kappa coefficient and overall accuracy metrics). Results reveal contrasting trends: while the S. vulgaris Protection Area exhibited substantial expansion (e.g., Southern Section coverage grew from 2.6 km² in 2013 to 7.88 km² in 2025), mining panels experienced significant degradation. Within Panel 202, coverage declined by 15.4% (58.4 km² to 49.5 km²), and Panel 203 showed a 18.5% decrease (3.16 km² to 2.57 km²) over the study period. These losses correlate spatially and temporally with mining-induced groundwater depletion and land subsidence, disrupting the shrub’s shallow-root water access strategy. The study demonstrates that coal mining drives fragmentation and coverage reduction in S. vulgaris communities through mechanisms including (1) direct vegetation destruction, (2) aquifer disruption impairing drought adaptation, and (3) habitat fragmentation. These findings underscore the necessity for targeted ecological restoration strategies integrating groundwater management and progressive reclamation in mining-affected arid regions.

1. Introduction

Sabina vulgaris (commonly known as creeping juniper or sand juniper) is an evergreen procumbent shrub belonging to the Juniperus genus of the Cupressaceae family. As a drought-adapted species, it exhibits robust root systems, dense foliage, rapid growth rates, and extensive ground coverage [1]. This plant demonstrates exceptional resistance to drought, cold, and windblown sand, enabling it to form dense stands even in extreme arid sandy environments [2]. As a keystone ecological species in China’s northwestern arid regions and a protected plant in Shaanxi Province, S. vulgaris provides critical ecosystem services, including sand fixation, soil structure improvement, water conservation, and climate regulation [3]. Natural populations are primarily distributed across Inner Mongolia, Shaanxi, Ningxia, Qinghai, Gansu, and Xinjiang at elevations of 1100–2800 m, inhabiting rocky slopes and fixed/semi-fixed sand dunes in major wind-swept areas such as Hunshandake, Mu Us, Helan Mountains, Yin Mountains, Qinghai Lake Basin, Qilian Mountains, Tianshan, and Altai Mountains [4]. The species’ extensive root network, multi-branched stems, and low, dense canopy confer remarkable windbreak and erosion control capabilities. Its unique adaptive traits include self-propagation via adventitious roots and stem elongation under sand burial, allowing shrub mounds to reach heights up to 10 m [5]. These characteristics make S. vulgaris vital for stabilizing fragile desert ecosystems and combating desertification.

S. vulgaris demonstrates exceptional water acquisition and retention capabilities in arid environments, primarily attributed to its complex root system and physiological–ecological adaptation mechanisms [6]. From the surface vegetation layer, the species’ dense canopy foliage effectively reduces soil evaporation while sustaining growth through photosynthesis [7]. The root system initiates functionality in the topsoil (0–10 cm), where lateral and adventitious roots form an extensive network in the subsoil layer (10–60 cm)—the primary water absorption zone. This mid-soil root architecture enables efficient water uptake and redistribution, particularly critical during drought periods [8]. In deeper soil strata (>60 cm), the taproot penetrates toward groundwater tables, directly extracting moisture to ensure stable water supply. This groundwater access constitutes a survival keystone for S. vulgaris under extreme aridity [9].

Beyond these root-based adaptations, S. vulgaris employs foliar water uptake through canopy condensation. Its leaves and branches absorb atmospheric dew during nighttime, supplementing water resources for growth and survival [10]. To minimize water loss, the species reduces stomatal conductance under drought conditions, partially closing stomata to limit transpiration [11]. Thickened cuticles on foliage surfaces further suppress epidermal evaporation, enhancing drought resistance [11]. Physiologically, the plant exhibits osmoregulatory adaptations by lowering cellular osmotic potential, thereby improving water absorption capacity and maintaining turgor pressure during drought stress [12]. While both photosynthetic and transpiration rates decline under water stress, the disproportionate suppression of transpiration leads to increased water-use efficiency [13]. These integrated mechanisms—including root architecture optimization, stomatal regulation, cuticular modification, and cellular osmoregulation—collectively enable S. vulgaris to thrive in extreme arid conditions, demonstrating remarkable ecological plasticity (Figure 1).

The Mu Us Sandy Land in the Ordos Basin represents a typical continental semi-arid region, serving as China’s frontline against desertification and a critical area for combating land degradation. During the 20th century, the region experienced severe soil degradation, shifting sand dunes, and pronounced water erosion [14]. Research indicates that these changes were primarily driven by climatic rather than anthropogenic factors [15]. To mitigate desertification and restore ecosystems, Chinese forestry authorities implemented comprehensive vegetation restoration measures, including sand control afforestation, enclosure restoration, conversion of cropland to forest/grassland, grazing bans, aerial seeding, and artificial afforestation. These efforts have transformed most of the sandy land into wind-sheltered grassland areas [14,15,16]. As vegetation plays a pivotal role in halting desertification progression, reducing wind erosion, and facilitating local habitat recovery [17], since the 1950s, Shenmu County in Shaanxi Province has established S. vulgaris nature reserves by extensively planting this pioneer species [18]. The shrub’s exceptional drought tolerance, cold resistance, and salt tolerance make it particularly well-suited for arid desert environments. Its extensive root system with high branching density achieves up to 90% ground coverage, effectively stabilizing sand dunes and preventing their migration [19].

However, since the 1980s, local coal mining has caused multifaceted negative impacts on surface vegetation and ecological environments [20]. First, it directly destroys surface vegetation, leading to soil acidification, heavy metal pollution, and fertility decline. These deteriorations in soil conditions are particularly detrimental to shrub growth, especially as their shallow root systems are more prone to absorbing heavy metals, thereby inhibiting growth or even causing death [21]. Second, coal mining disrupts water resources, causing groundwater depletion and surface water drying up. S. vulgaris, which relies on shallow water sources, faces moisture shortages, resulting in stunted growth or mortality. Simultaneously, vegetation destruction exacerbates soil erosion, forming a vicious cycle [22]. Additionally, coal mining leads to habitat loss, reduced species diversity, altered vegetation community structures, and significant declines in shrub coverage and density, thereby impairing ecosystem functions [23]. The Longde Coal Mine, located in Dabao Town, Yulin City, Shaanxi Province, at the northern edge of the Loess Plateau and the southeastern edge of the Mu Us Desert, primarily consists of shrubland. Its western boundary borders the Shenmu S. vulgaris Nature Reserve, and the mine area itself contains S. vulgaris shrubs, making it an ecologically sensitive and fragile zone [24]. Since its operation in 2012, the mine has caused significant groundwater level declines in the Quaternary Sarawusu aquifer above the mined-out area, as well as surface fissures due to subsidence [25]. Previous studies indicate that during the initial stages of coal mining, shrub and grassland areas decrease, and vegetation shows clear degradation trends [26]. While these findings reveal the overall impact of mining on shrubland, assessments of the ecological effects on S. vulgaris communities remain insufficient.

S. vulgaris, as a keystone ecological plant in the study area, faces severe threats to its survival due to environmental degradation, which jeopardizes the stability of the entire ecosystem. This study focuses on the growth dynamics of S. vulgaris shrublands within the Longde Coal Mine area, aiming to investigate the impacts of coal mining activities on its growth patterns. The findings are expected to provide critical insights for ecological conservation and restoration efforts in the region. The research utilized seven time periods (2013, 2015, 2017, 2019, 2021, 2023, and 2025) of 2 m resolution GF-1 remote sensing imagery spanning 12 years of mining operations. Mixed vegetation areas containing S. vulgaris were selected as representative sample sites for drone-based hyperspectral data acquisition, which served as the ground truth for S. vulgaris distribution mapping. The SegU-Net deep learning model was then applied to GF-1 imagery from corresponding time periods to identify S. vulgaris distributions. A confusion matrix analysis was conducted to evaluate the accuracy of the GF-1-based SegU-Net results against the drone-derived interpretations. Subsequently, the study systematically analyzed the spatiotemporal distribution and growth status of S. vulgaris shrublands across the entire Longde Coal Mine area using the multi-temporal GF-1 datasets. This approach aims to elucidate the underlying mechanisms through which coal mining activities may influence the growth of S. vulgaris communities. The research holds significant theoretical and practical implications for formulating scientifically sound ecological restoration strategies and safeguarding regional environmental integrity.

2. Research Method

2.1. Overview of the Study Area

The study area is located in Dabaodang Town, Shenmu County, Yulin City, Shaanxi Province, situated at the northern edge of the Loess Plateau and the southeastern edge of the Mu Us Desert (Figure 2). This region belongs to the sandy grassland landscape zone and constitutes a critical part of the ecological functional area for desertification control in the northern Yulin–Shenmu section along the Great Wall. The terrain is generally flat, with elevations ranging from 1095 to 1225 m, dominated by sand dunes and sandy land (approximately 95%) and valley landforms (about 5%) [27]. The climate is characterized as a semi-arid continental monsoon climate, transitioning from warm temperate to temperate zones, with an average annual temperature of 10.4 °C and precipitation of approximately 524.9 mm, predominantly concentrated in summer. The primary water systems include the Tuwei River, a tributary of the Yellow River, and the Helong Valley [27]. Geologically, the area falls within the Ordos Basin, featuring a simple tectonic structure with an NE-striking, NW-dipping monocline of small dip angles. No major faults or folds are present, and the strata are stable. The coal-bearing formations are mainly from the Middle Jurassic Yan’an Formation [28]. Vegetation is primarily composed of shrub communities such as S. vulgaris, Caragana korshinskii, and Salix cheilophila. S. vulgaris, a native evergreen shrub, plays a vital role in sand fixation, wind erosion resistance, and ecological restoration. As a protected plant in Shaanxi Province, it is widely distributed within the mining area and holds significant ecological importance [3]. The Longde Coal Mine consists of three mining panels: Panel 201 (operated from 2012 to 2014), Panel 202 (2014–2021), and Panel 203 (since 2021). The mining area overlaps with the Shenmu S. vulgaris Nature Reserve, making this ecologically sensitive and fragile zone a focal area for this study [29]. Despite the climatic characteristics of the region, such as aridity, strong winds, and significant temperature fluctuations, systematic observation of S. vulgaris has not been significantly disrupted. S. vulgaris demonstrates strong ecological adaptability under these climatic conditions, with stable phenological characteristics and growth patterns. Data derived from remote sensing imagery and ground surveys reliably and consistently reflect its spatial distribution and physiological dynamics. Through scientifically designed observation schedules and the integration of multi-source remote sensing technologies with ground verification, the study effectively mitigated the impact of short-term weather phenomena, such as dust storms, on data acquisition, ensuring the accuracy and consistency of observational results. Therefore, the climatic environment has not hindered the core progress of S. vulgaris observation research. The quality of the obtained data meets the precision requirements for subsequent analyses, supporting in-depth interpretation of related ecological processes.

2.2. Data Source

The remote sensing data were obtained from the China Geological Survey’s “Geological Cloud” big data system (https://geocloud.cgs.gov.cn/, accessed on 25 October 2025), which provided 2 m resolution GF-1 (Gaofen-1) satellite imagery. The dataset comprises seven scenes acquired in 2013, 2015, 2017, 2019, 2021, 2023, and 2025 (

Table 1. Remote Sensing Data Source.

№	Satellite	Spatial Resolution (m)	Acquisition Times (Date–Month–Year)	Bands
1	GF-1	2.0	20 November 2013	4
2	GF-1	2.0	9 February 2015	4
3	GF-1	2.0	2 April 2017	4
4	GF-1	2.0	17 May 2019	4
5	GF-1	2.0	21 February 2021	4
6	GF-1	2.0	13 March 2023	4
7	GF-1	2.0	5 February 2025	4

) and was utilized for patch identification and area extraction of S. vulgaris shrublands across multiple temporal phases.

2.2.1. UAV Data Acquisition

The UAV data were collected on 20 March 2025 using a DJI Phantom 4 Multispectral drone (Technology Co., Ltd., headquartered in Shenzhen, China), which integrated one visible light camera and five multispectral cameras with a TimeSync system for microsecond-level synchronization between flight control, cameras, and RTK positioning to achieve millisecond-level imaging accuracy. A total of 18 flight missions were conducted, covering an area of approximately 10 km² and capturing 13,962 images at a flight altitude of 300 m with a resolution better than 0.2 m (Figure 3). The coordinate system used CGCS2000 with Gauss–Krüger 3-degree zone projection. The acquired image data were processed using Photoscan software (Agisoft Metashape v1.7.21), sequentially performing operations such as adding photos, calibrating reflectance, aligning photos, optimizing cameras, constructing dense point clouds, generating DEMs, and producing orthophotos. The orthophotos were then clipped using ArcMap software (v10.8).

2.2.2. Ground Truth Interpretation

Using aerial data acquired by the DJI Phantom 4 Multispectral drone, the orthophoto image processed by Photoscan software was imported into ArcMap. Leveraging the unique winter-green spectral characteristics of S. vulgaris shrubs, the near-infrared band data were analyzed to initially delineate shrub distribution through visual interpretation. Subsequently, the interpreted boundaries were refined by integrating texture features such as the clustered canopy structure of S. vulgaris in visible light imagery, ensuring high alignment between the vector polygons and field conditions (Figure 4). The resulting precise ground truth data provided reliable validation for the SegUNet satellite remote sensing recognition model (Figure 5).

2.3. Identification Method

2.3.1. Data Preprocessing

In order to quantitatively assess the impact of mining activities on Juniperus chinensis (scissor-leaf juniper) thickets, this study clearly defined the spatial units for comparative analysis.

(1)

The Mining Impact Zone, specifically mining blocks 202 and 203, whose boundaries correspond precisely to the projected underground working face areas delineated in the mine development design plans.

(2)

The non-mining control areas, the northern and southern sections of the Fork Juniper Nature Reserve. These areas are highly similar to the mining zone in terms of landform, soil and historical baseline vegetation, yet remain exempt from direct disturbance by underground mining due to legal prohibitions. By comparing regional geological maps, digital elevation models and historical land use data, we confirmed that, beyond mining, these areas experienced no significant anthropogenic disturbances (such as large-scale construction or land cover changes) during the study period (2013–2025) and generally exhibit gentle slopes (<5°). The study employed Gaofen-1 imagery from 2013 as the baseline for vegetation cover in all areas to ensure comparability of pre-mining conditions.

This study employs the SegU-Net deep convolutional neural network model for semantic segmentation recognition of S. vulgaris distribution in GF-1 (Gaofen-1) remote sensing imagery [30]. Prior to model training, preprocessing of GF-1 multispectral remote sensing imagery is conducted. Initially, radiometric calibration, atmospheric correction, orthorectification, image fusion and registration, mosaic, and clipping operations are performed on panchromatic and multispectral remote sensing imagery from different periods. Subsequently, a pixel segmentation model is established using ENVI’s Build Label Rasters tool, combining NDVI thresholding and manual visual interpretation to delineate S. vulgaris ROIs and generate binary label maps as training datasets for model training. During model training, the original imagery is first cropped using a sliding window method with a dimension of 572 × 572 pixels to ensure sample coverage across diverse terrain types and enhance spatial variability. To improve robustness under varying lighting and viewing conditions, data augmentation techniques—including rotation, horizontal/vertical flipping, and brightness perturbation—are applied to cropped images to stabilize prediction performance. The dataset is then partitioned into training, validation, and test sets at a 7:2:1 ratio, with strict avoidance of sample overlap.

Atmospheric correction employs the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) model to convert raw pixel brightness values into surface reflectance, effectively eliminating the effects of atmospheric scattering and absorption. Orthorectification employs the RPC (Rational Polynomial Coefficient) model, utilising the inherent RPC parameters within GF-1 imagery to correct geometric distortion and uniformly register all images to the same geographic coordinate system.

To enhance vegetation characteristics, we employ standard formulas to calculate the Normalised Difference Vegetation Index:

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

(1)

ρ_NIR and ρ_Red represent the atmospheric-corrected surface reflectance in the near-infrared and red light bands of the GF-1 imagery, respectively. To estimate vegetation cover quantitatively, a linear pixel decomposition model based on NDVI values was used to calculate the coverage.

$$FVC={\displaystyle \frac{NDVI-{NDVI}_{soil}}{{NDVI}_{veg}-{NDVI}_{soil}}}$$

(2)

In the equation, NDVIsoil and NDVIveg represent the NDVI values of pure bare soil pixels and pure vegetation pixels, respectively, determined empirically from the imagery.

All remote sensing data processing, including atmospheric correction, orthorectification, and vegetation index calculation, was performed using ENVI software (version 6.1). Training and inference of the SegU-Net model were implemented within the ENVI Deep Learning Module (version 3.0.1), with computational acceleration provided by an NVIDIA GeForce RTX 4060 GPU.

2.3.2. Seg U-Net Model

The Seg U-Net model consists of an encoder, a bottleneck layer, and a decoder [31]. As illustrated in the network diagram (Figure 6), the framework follows a U-shaped layout. The left side represents the contracting path (encoder), where preprocessed 4-channel GF-1 image data is input. The encoder first undergoes two sequential 3 × 3 convolutional operations with batch normalization (BN), each followed by a ReLU activation function [31]. A 2 × 2 max-pooling operation is then applied for downsampling, progressively extracting hierarchical feature vectors. With each downsampling step, the feature map size halves while the channel count doubles. After four pooling operations, the bottleneck layer aggregates global features. The right side constitutes the expansive path (decoder), which gradually restores spatial resolution through upsampling [31]. Unlike U-Net, SegU-Net avoids transposed convolutions. Instead, it preserves feature map information during downsampling via pooling indices and concatenates corresponding encoder-level feature maps to the current upsampled output, enhancing local detail recovery [32]. Post-concatenation, the feature maps undergo two additional 3 × 3 convolutions with ReLU to extract fused deep features. This process repeats four times, incrementally reconstructing image resolution. The output layer employs a 1 × 1 convolution to map channels to class numbers, enabling pixel-wise semantic segmentation prediction. The final output comprises class activation grids and classification grids. The model training is based on a systematic configuration scheme. The training employs 320 × 320 pixel image patches as input units, with the background patch ratio set at 20% and the effective feature patch ratio controlled at approximately 60%. During the training process, 80% of the data is used for model learning, with each batch processing 4 image patches, and the total number of training epochs is set to 50. To enhance the model’s adaptability to different scenarios, two data augmentation strategies—random scale transformation and multi-angle rotation—are enabled during the training phase. This project performs a pixel-level semantic segmentation task, and the relevant parameter configurations include: setting the solid boundary distance to 1–2 pixels, using a progressive blur distance in the range of 0.0 to 3.0 to handle edge transitions, assigning differentiated loss weights of 1.0 and 2.0 to different geographic feature categories, and setting the overall loss function weight to 1.0. All detailed hyperparameters and training configurations are summarized in

Table 2. Deep learning model parameters.

Code Name	Usage Parameters
model_architecture	“Seg U-Net”
patch_size	320
background_patch_ratio	0.2 (20%)
loss_weight	1.0
training_split	80
patches_per_batch	4
epochs	50
feature_patch_percentage	0.60000002 (≈60%)
augment_scale	True (Boolean value)
augment_rotation	True (Boolean value)
project_type	“pixel_segmentation”
├─ solid_distance	[1, 2] (List of integers)
├─ blur_distance	[0.0, 1.0, 2.0, 3.0] (Floating-point list)
└─ class_weight	[1.0, 2.0] (Floating-point list)
n_bands	4

.

2.3.3. Data Processing

Post-model training, image post-processing is performed. Initially, the class activation grids output by the model undergo denoising using morphological filtering to eliminate isolated noise points, followed by Gaussian filtering to smooth classification boundaries, significantly enhancing the spatial continuity of classification results. Subsequently, the processed class activation grids are converted into binary raster images, focusing on extracting distribution information of S. vulgaris vegetation and outputting S. vulgaris classification grids. To further analyze the dynamic changes in S. vulgaris vegetation, the S. vulgaris distribution raster data obtained from different years are imported into ArcGIS Pro 2024 for processing. Spatial overlay analysis is employed to identify vegetation coverage change areas, and computational geometry tools are utilized to precisely analyze the distribution area of S. vulgaris across years. A time-series database is constructed to provide a reliable data foundation for subsequent studies on vegetation change trend analysis and classification accuracy evaluation (

Table 3. Coverage areas of S. vulgaris within Panel 202, Panel 203, the mining area, the S. vulgaris Protection Area (Northern Section) and the S. vulgaris Protection Area (Southern Section).

	Panel 202 (m2)	Panel 203 (m2)	The Mining Area (m2)	The S. vulgaris Protection Area (Northern Section) (m2)	The S. vulgaris Protection Area (Southern Section) (m2)
2013	58,447.7	3155.5	61,603.2	669,705.5	2,609,797.0
2015	69,627.4	3184.6	72,878.0	1,446,838.6	3,566,351.6
2017	61,650.7	3297.6	64,948.4	1,566,220.7	3,884,349.9
2019	48,122.3	3629.8	51,752.1	1,756,044.8	5,193,425.1
2021	41,432.8	3888.8	45,321.6	1,859,556.9	6,893,901.3
2023	43,943.1	3439.7	47,382.7	2,052,731.5	7,750,668.0
2025	49,465.0	2572.7	52,037.7	2,403,638.3	7,880,707.4

).

2.4. Classification Accuracy Assessment

To evaluate classification accuracy, ArcGIS can be utilized to statistically analyze the classified results of the study area by constructing a confusion matrix. This approach quantifies classification performance using metrics such as overall accuracy, expected accuracy, and the Kappa coefficient, followed by further discussion (Figure 7). The workflow begins by applying overlay analysis tools to categorize the study area into four scenarios: true positives (correctly identified S. vulgaris), true negatives (correctly identified non-S. vulgaris areas), false positives (misclassified as S. vulgaris), and false negatives (unidentified S. vulgaris). Area statistics tools are then used to summarize each category, enabling the construction of a confusion matrix. Subsequently, the Kappa coefficient is calculated to assess the agreement between classification results and reference data, providing a standardized measure of accuracy while accounting for chance agreement. This quantitative analysis facilitates rigorous discussion of classification reliability and potential improvement strategies.

2.4.1. Confusion Matrix

In image accuracy evaluation, confusion matrices are commonly used to compare classification results against ground truth data (Figure 8). Each column represents the predicted class, with the total count indicating the number of samples predicted as that class. Each row represents the actual class, with the total per row indicating the number of samples truly belonging to that class. Here, TP (True Positive), FN (False Negative), FP (False Positive), and TN (True Negative) correspond to true positives, false negatives, false positives, and true negatives, respectively.

2.4.2. Assessment Index

(1)

Kappa Index

The Kappa coefficient is a classification accuracy metric derived from the confusion matrix, commonly used for agreement analysis. By quantifying the divergence between classification results and random chance, it provides an objective evaluation of classification accuracy. The calculated values range from −1 to 1, though typically observed between 0 and 1 [33]. The specific formula is

$$\mathsf{\kappa }={\displaystyle \frac{{\mathrm{P}}_{\mathrm{o}}-{\mathrm{P}}_{\mathrm{e}}}{1-{\mathrm{P}}_{\mathrm{e}}}}$$

(3)

$$P_o={\displaystyle \frac{{\sum }_{i=1}^C{T}_i}{n}}$$

(4)

$$P_e={\displaystyle \frac{{\sum }_{i=1}^C{a}_i\ast b_i}{n^2}}$$

(5)

where

• K—Classification agreement (Kappa coefficient);
• P_o—Overall classification accuracy;
• P_e—Expected classification accuracy (chance agreement);
• T_i—Number of correctly classified samples;
• C—Number of classes;
• n—Total sample size.

When quantifying the degree of classification accuracy, the results can be categorized into five groups to represent different levels of agreement: 0.0–0.20, indicating very low agreement, 0.21–0.40 slight agreement, 0.41–0.60 moderate agreement, 0.61–0.80 substantial agreement, and 0.81–1.00 almost perfect agreement [33].

(2)

Precision

Precision is determined based on prediction results; hence it is also termed precision rate. It evaluates the accuracy of predictive behavior relative to the predicted outcomes [34].

$$\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(6)

(3)

Recall

Recall is determined based on actual results; hence it is also referred to as recall rate. It evaluates the proportion of all positive cases that are successfully predicted [35].

$$recall={\displaystyle \frac{TP}{TP+FN}}$$

(7)

(4)

F1

The F1 score evaluates the performance of a classifier across different categories, with values ranging from 0 to 1. A higher score indicates that the classifier performs better.

$$F1={\displaystyle \frac{2\times precision\times recall}{precision+recal}}$$

(8)

(5)

Confidence interval

Confidence intervals are used to quantify the uncertainty in parameter estimates derived from sample datasets. They provide a numerical range within which the true parameter value is expected to fall with a specified level of confidence (typically 95%). The narrower the confidence interval, the higher the degree of precision in the estimate. The calculation process is as follows:

True proportion estimation

$$\widehat{P}={\displaystyle \frac{TP+FN}{N}}={\displaystyle \frac{N_1}{N}}$$

(9)

Area point estimation

$$\widehat{A}=A_{total}\times \widehat{P}$$

(10)

Proportional Variance

$$Var(\widehat{P})={\displaystyle \frac{\widehat{P}(1-\widehat{P})}{N-1}}$$

(11)

Area standard error

$$SE(\widehat{A})=A_{taota}\times \sqrt{Var(\widehat{P})}=A_{total}\times \sqrt{{\displaystyle \frac{\widehat{P}(1-\widehat{P})}{N-1}}}$$

(12)

95% confidence interval

$$CI=\widehat{A}\pm 1.96\times SE(\widehat{A})$$

(13)

3. Results

The geographical distribution of S. vulgaris in the study area primarily consists of coal mining zones and nature reserves designated for its conservation. The seven-year imagery acquisition dates were November 2013, February 2015, April 2017, May 2019, February 2021, March 2023, and February 2025. Due to seasonal variations in image acquisition, the observed differences in S. vulgaris coverage between autumn–winter and spring–summer periods fall within acceptable error margins, as the species’ foliage density naturally fluctuates with seasons.

3.1. The Spatiotemporal Distribution of S. vulgaris in Coal Mining Areas

Within the coal mining area, S. vulgaris is primarily distributed in the central-western region, concentrated in panel 202, with sporadic occurrences along the southwestern boundary of panel 203. No significant S. vulgaris distribution was observed in panel 201 (Figure 9). Coal mining directly alters surface topography, destroying the growth substrate of S. vulgaris shrubs. Statistical analysis of seven-year imagery data revealed temporal variations in S. vulgaris coverage, showing a significant decline within the pit field from 2015 to 2021, peaking at 72,878 m² and reaching a minimum of 45,321 m² in 2021 (Figure 10). Mining activities in Area 202 (2014–2021) were identified as the primary cause of the overall reduction in S. vulgaris within the pit field. Although mining commenced in panel 203 in 2021, its limited initial S. vulgaris distribution resulted in a relatively minor impact on the total area-wide coverage.

3.1.1. The Spatiotemporal Distribution of S. vulgaris in Panel 202

The mining activities in panel 202 of Longde Mining Area commenced in 2014 and ceased in 2021. The S. vulgaris remote sensing identification and statistical results based on SegU-Net revealed that during 2013–2015, the vegetation coverage of S. vulgaris exhibited a growth trend, with an average annual increase rate of approximately 19% (Figure 11). However, with the continuation of mining activities, the distribution area of S. vulgaris showed a year-on-year decline from 2015 to 2021, decreasing from 69,627.41 m² in 2015 to 41,432.8 m² in 2021 (Figure 12), representing a cumulative reduction of 40%. Notably, a significant decline in S. vulgaris area occurred in 2019, which was closely associated with the vegetation destruction caused by the construction of a coal transportation railway crossing the primary distribution area of S. vulgaris in the 202 Panel. According to the latest imagery data from 2025, since the cessation of mining in 2021, the S. vulgaris vegetation in the area has shown a clear recovery trend.

3.1.2. The Spatiotemporal Distribution of S. vulgaris in Panel 203

Mining activities in panel 203 of Longde Mining Area officially commenced in 2021. The distribution area of S. vulgaris in this region was significantly smaller than that in panel 202, with the initial coverage being merely 1/20 of panel 202’s (Figure 13). Pre-mining monitoring data indicated that S. vulgaris exhibited a natural growth trend, with an average annual growth rate of approximately 2%, reaching a peak area of 3888.82 m². However, since the commencement of mining in 2021, indirect impacts from mining operations have led to significant changes in surface environments, resulting in continuous shrinkage of S. vulgaris coverage. By 2025, the distribution area had decreased to 2572.67 m², with an average annual shrinkage rate of 18% (Figure 14). This trend clearly demonstrates that mining activities in Panel 203 have exerted a pronounced negative impact on the growth of S. vulgaris vegetation in the region.

3.2. The Spatiotemporal Distribution of S. vulgaris in the Protection Zone

The S. vulgaris protection zone in the mining area can be divided into two sections: the north area of the S. vulgaris protection zone and the south area of the S. vulgaris protection zone. The north area is located northwest of the coal mining area, while the south area lies southeast. During the study period, no significant human or engineering disturbances were observed within the reserve.

3.2.1. The North Area of S. vulgaris Protection Zone

The north area of the S. vulgaris protection zone covers approximately 55 km², serving as a critical vegetation conservation area in the region. Remote sensing data from 2013 to 2025 indicate that S. vulgaris within the reserve exhibited robust growth with high coverage density, displaying distinct identifiable features in imagery (Figure 15). During the monitoring period, the distribution area of S. vulgaris increased cumulatively by 1.73 million m² (259% growth), averaging an annual expansion of 144,500 m² and demonstrating a compound annual growth rate (CAGR) of 11.8% (Figure 16). As per the latest monitoring data, the total area of S. vulgaris vegetation in the reserve has reached approximately 2.40 million m².

3.2.2. The South Area of the S. vulgaris Protection Zone

The South Area of the S. vulgaris protection zone covers approximately 55 km², serving as a vital vegetation conservation area in the region. Compared to the Northern Zone, the Southern Zone exhibits denser S. vulgaris coverage. Remote sensing data from 2013 to 2025 indicate robust vegetation growth with high coverage density, displaying distinct identifiable features in imagery (Figure 17). The reserve’s S. vulgaris distribution area expanded from 2.61 million m² in 2013 to 7.88 million m² in 2025, a net increase of 5.27 million m² (202% growth). The monitoring period revealed distinct growth phases: 2013–2017 showed an average annual expansion of 318,400 m², while post-2017 accelerated to 499,500 m²/year. As per the latest 2025 data, the total S. vulgaris area has reached 7.88 million m², demonstrating the reserve’s exceptional ecological benefits and vegetation recovery capacity (Figure 18).

3.3. Classification Result Reliability

Before we can check how accurate the maps are, we need to make sure that the deep learning model we used is reliable. Figure 19, Figure 20 and Figure 21 show you the key performance metric curves of the SegU-Net model that we used in this study during the training cycle. They have got all the important stuff like training loss, validation loss, recall, and accuracy for the Juniperus chinensis category.

I’m so happy to tell you that during model training, the loss function value showed a nice, steady decline, which is a really good sign. It’s great to see that it’s working so well and getting closer to the right answer. It’s really great to see that both class accuracy and class recall for the Thuja occidentalis category showed significant synchronous improvement. This shows that the model made classification more accurate overall, and it also improved its ability to identify and recall each category more accurately, especially for Thuja occidentalis samples. The reduction in loss value had a really positive effect, making the model much better at making sense of things.

This study processed the satellite remote sensing classification results from February 2025 and the UAV ground truth interpretation results from March 2025 through clipping, overlay, intersection, and inversion operations. The four required scenarios were then statistically analyzed and are summarized in

Table 4. The Confusion Matrix of the S. vulgaris classification reliability.

		Class		Unit (m2)
		S. vulgaris	Others	Total
Validation set	S. vulgaris	50,879.17	17,664.50	68,543.67
	Others	32,812.43	9,524,762.51	9,557,574.94
	Total	83,691.60	9,542,427.01	9,626,118.62

.

It should be noted that the test set used for accuracy validation in this study has a significant imbalance in its sample. Specifically, the target category (forked juniper) constitutes only around 0.87% (83,691.60 m²/9,626,118.62 m²) of the total test area, while the background category (non-forked juniper) dominates. This imbalance leads to an overly high overall accuracy score during evaluation, which primarily reflects the correct classification of the predominant background category. Therefore, overall accuracy is not a reliable metric for evaluating the model’s ability to identify the target species. To accurately evaluate the model’s ability to recognise Juniperus chinensis, it is necessary to analyse metrics specific to this category, such as precision, recall and F1 score. Based on the confusion matrix analysis, the following results were obtained:

(1)

Overall classification accuracy (Po) = 0.99;

(2)

Expected classification accuracy (Pe) = 0.98;

(3)

Kappa coefficient = 0.66;

(4)

For positive examples (Juniperus sibirica), precision (P) = 0.61 and recall (R) = 0.74;

(5)

For negative examples (non-Juniperus sibirica), precision (P) = 0.99 and recall (R) = 0.99.

(6)

The F1 score is the harmonic mean of precision and recall for the positive class, calculated as 0.668.

(7)

The 95% confidence interval for the total area of the fork-branched juniper was calculated to be [83,126.88; 84,256.32] m², which encompasses the estimated area of 83,691.60 m².

Therefore, for this study, the overall classification accuracy exceeds 99%, with the expected classification accuracy surpassing 98%. The Kappa coefficient of 0.66 indicates a high level of consistency, demonstrating that the classification results meet expectations. However, the Kappa coefficient is not only closely related to overall accuracy but also influenced by complex factors such as the difference in the ratio of positive to negative examples and measurement bias in the study. Consequently, when evaluating consistency, relying solely on the Kappa coefficient may lead to incorrect conclusions if the number of categories is small. Therefore, we combined precision and recall for comprehensive analysis. The analysis of precision reveals that both the precision and recall rate for negative examples exceed 99%, indicating a low probability of misclassifying non-S. vulgaris as S. vulgaris. For positive examples, the precision exceeds 60% and the recall rate surpasses 74%, demonstrating accurate identification of true S. vulgaris cases. Furthermore, the F1 score of 0.668 and the narrow confidence interval for area estimation demonstrate that the SegU-Net model provides reliable and consistent identification outcomes for the large-scale ecological evaluation of the distribution of Juniperus chinensis within this challenging mining environment. This is a significant improvement over traditional methods, which are hindered by spectral mixing issues. Overall, the model can both accurately distinguish S. vulgaris and non-S. vulgaris classes.

3.4. Sensitivity Analysis of Seasonal Variations

This study selected three Gaofen-1 satellite image sets acquired during distinct phenological phases within the Southern Protected Area to verify whether seasonal variations in image acquisition introduce systematic biases into remote sensing monitoring results for Juniperus chinensis coverage, thereby compromising the accuracy of long-term ecological trend assessments. The image sets were acquired in: November 2013 (late autumn/early winter), May 2019 (late spring/early summer), and February 2025 (late winter/early spring). These three periods cover the entire study cycle, representing the end of vegetation dormancy, the peak growth period and the initial budding stage, respectively.

We extracted the coverage areas of fork juniper from these three seasonal images by applying the trained SegU-Net model. The results show that the fork juniper coverage area in the southern reserve measured 2.61 km² in November 2013, increasing to 5.19 km² by May 2019 and expanding further to 7.88 km² by February 2025. Despite the differing acquisition seasons, the sequence of areas exhibits sustained and near-linear growth (Figure 22). Calculations of the average annual growth rate between adjacent periods revealed that the average annual increase in area was approximately 0.43 km²/year from 2013 to 2019 (spanning winter and summer) and approximately 0.45 km²/year from 2019 to 2025 (spanning summer and spring). These rates are highly comparable, with a relative difference of less than 4%, indicating that the expansion rate of Juniperus chinensis coverage remains consistent across different seasonal combinations in the observations. The stability of the growth trend is unaffected by the seasonal transition in data acquisition. Furthermore, to isolate long-term trends and visually represent potential seasonal fluctuations, we calculated the absolute area increments (yield and ΔArea) between adjacent periods: ΔArea_(2013–2019) and ΔArea_(2019–2025). Both increment values remained at the order of magnitude of 2.6 × 10⁶ m². This suggests that variability in area estimates caused by phenological differences is much smaller than the signals of systematic change driven by the ecosystem’s evolution, such as natural recovery and conservation measures.

4. Discussion

4.1. The Impact of Coal Mining on the Growth of S. vulgaris Shrublands

This study systematically revealed the dual inhibitory mechanisms of coal mining on the growth of S. vulgaris shrublands through vegetation remote sensing dynamic monitoring data. During mining operations, the accompanying subsidence significantly altered the soil physical structure, leading to reduced clay content, increased porosity, and a higher proportion of non-capillary pores [36,37]. This structural degradation further accelerated leakage rates [38], expanded vertical and lateral evaporation surfaces [39], and ultimately significantly reduced soil water content in subsidence areas [40]. As underground coal mining progressed, the expanding subsidence areas led to prolonged soil water deficit, directly limiting the water absorption capacity of S. vulgaris roots, forming the primary effect of water stress. Consequently, noticeable growth shrinkage was observed in S. vulgaris within subsidence zones during the mining process.

The secondary effects of coal mining subsidence on plant physiological processes are manifested at the level of transpiration inhibition. On one hand, the decline in soil water content directly reduces the available water for plants; on the other hand, the mechanical damage to roots during subsidence (e.g., root breakage) further impairs water absorption efficiency [41]. This dual stress prevents S. vulgaris from maintaining normal transpiration demands, which aligns with the mechanism observed in Liu Ying’s study, where Caragana korshinskii in subsidence areas exhibited reduced transpiration rates due to decreased stomatal conductance [42]. This water–physiology coupled inhibition mechanism ultimately results in the decline in growth vigor and limited biomass accumulation in S. vulgaris shrublands.

The sequential monitoring data from 202 and 203 panels of Longde Mining Area provide empirical support for the aforementioned mechanisms. Following the commencement of mining operations in Panel 202 in 2014, significant surface deformation was observed in the mining area (Figure 23), with the S. vulgaris coverage area sharply decreasing from 69,627.41 m² in 2015 to 41,432.8 m² by 2021 (a cumulative reduction of 40%), reversing the pre-mining average annual growth rate of 19%. Notably, Panel 203 exhibited a similar growth inflection point after mining began in 2021: the pre-mining natural growth rate averaged approximately 2% (peak area of 3888.82 m²), but within four years of mining, the area shrank to 2572.67 m², with the annual decline rate reaching 18%. Both panels demonstrate a significant spatiotemporal correlation between the initiation of mining activities and S. vulgaris decline, with the severity of decline positively correlated with mining duration. In contrast, the southern and northern zones of the adjacent natural reserve exhibited gradually increasing S. vulgaris coverage over time. This further underscores the close association between the shrinkage of S. vulgaris coverage in subsidence areas and coal mining activities during the mining period. Encouragingly, after mining completion, the subsidence areas began to show a reversal of the declining trend, with S. vulgaris coverage gradually increasing again.

This study has shown us some really interesting things about how Juniperus chinensis scrub reacts to being disturbed by mining. But to really understand this, we need to think about how it has adapted to live in its special environment. Juniperus chinensis is a really special plant that has helped to build the community in the Mu Us Desert. It has developed some amazing qualities to help it deal with drought and being buried in sand. Its extensive shallow root network is really effective at capturing sporadic rainfall and surface soil moisture, while its deep taproots are like its lifeline, keeping it safe during times of extreme drought by accessing deep soil water or groundwater. This clever way of getting water is really important for survival in stable sandy habitats. But mining can have two negative effects on this: it can cause groundwater to drain away and the land to sink into the ground. It’s really sad, but the levels of groundwater are dropping all the time, and this is having a direct effect on the survival of its deep taproots. At the same time, the soil structure can be disrupted by subsidence, which can lead to things like fissure formation and increased porosity. This can then speed up surface soil moisture loss, which unfortunately can impair the functionality of the efficient shallow root network. It’s like they’re having a hard time finding water, whether it’s in the shallow parts or the deeper parts, and it’s happening over different timescales, too, like short-term and long-term. This makes it really hard for them to keep their leaves and branches healthy.

4.2. Protection of S. vulgaris Natural Forests Promotes Growth of S. vulgaris

The continuous advancement of S. vulgaris natural forest protection in Yulin City has gradually established a systematic and multi-tiered conservation framework across multiple policy phases, significantly enhancing the natural regeneration and growth conditions of S. vulgaris populations. Beginning with the establishment of county-level nature reserves in the 1980s [43], progressing to the construction of provincial-level conservation systems in the 2000s [44], and further integrating them into regional ecological projects and desertification control strategies during the 13th Five-Year Plan period [45], the protection efforts for S. vulgaris have been continuously strengthened, with the conservation scope expanding from isolated local areas to cross-county and cross-basin ecological networks. Notably, the Yulin Forestry Construction Five-Year Promotion Implementation Plan (2016–2020) designated S. vulgaris as a key protected species in 2017, underscoring its critical role in the regional ecological security framework. Concurrently, the initiative encouraged resource-based enterprises to participate in mine-site greening, effectively mitigating anthropogenic disturbances and land degradation that threatened S. vulgaris habitats [46]. These institutional arrangements have not only expanded the living space for S. vulgaris but also provided more stable habitat conditions for its natural dispersal and population recovery.

The protection of S. vulgaris entered a phase of scientific refinement and precision during the second decade of the 21st century, following the promulgation of the Shaanxi Province List of Locally Key Protected Plants and the establishment of regional technical standards [47]. Local standards such as the Technical Regulations for Cutting Seedling Production and Afforestation of S. vulgaris Using Nutrient Bags (DB 6108/T 28-2021) provided technical safeguards for the conservation and sustainable utilization of S. vulgaris germplasm resources by standardizing procedures including branch collection, cutting propagation, and planting [48]. This integrated “conservation + propagation + utilization” model not only enhanced the growth quality and stress resistance of individual S. vulgaris plants but also improved the population’s overall adaptability to extreme environmental pressures such as drought and wind erosion. Furthermore, the implementation of the Yulin City Forest Chief System Implementation Plan and the introduction of the digital management platform “Yulin Ecological Cloud” [49,50] enabled dynamic monitoring of S. vulgaris reserves and rapid responses to violations, strengthening the “closed management + cultivation” protection of its native habitats. These measures significantly reduced threats from human activities and illegal encroachment, providing dual institutional and technological safeguards for the natural growth and community stability of S. vulgaris.

However, despite continuous improvements in conservation policies, the growth-promoting effects of S. vulgaris natural forests still exhibit significant regional disparities, particularly in mineral resource-intensive areas such as the Panel 203 region of Longde Coal Mine, where S. vulgaris coverage has been shrinking at an annual rate of 18%. This phenomenon highlights the inherent tension between ecological conservation and economic development goals. The challenges stem from inadequate implementation of corporate environmental responsibilities and imperfect ecological compensation mechanisms within current protection frameworks. To address these issues, future conservation efforts should integrate “legalized supervision” with “community co-management” models under existing policy frameworks. Additionally, exploring the synergy between carbon sink forest construction under the “dual-carbon” goals and ecosystem service value assessment of S. vulgaris could extend protection policies toward mechanisms for realizing ecological product value, thereby achieving synergistic benefits between S. vulgaris natural forest conservation and regional sustainable development.

4.3. The Importance of Multi-Temporal High-Resolution Remote Sensing and UAV Remote Sensing in Monitoring S. vulgaris Shrub Growth in Coal Mining Areas

Traditional vegetation monitoring in mining areas primarily relies on field quadrat surveys and medium–low-resolution satellite imagery (e.g., Landsat series), which suffer from insufficient spatial resolution (30 m) and difficulties in capturing fine-scale changes in S. vulgaris shrubs [29]. In recent years, the rapid development and widespread application of high-resolution remote sensing (e.g., 2 m level Gaofen satellites) and UAV technology have provided new technical pathways for ecological monitoring in mining areas. This study integrates seven-phase GF-1 multispectral data (2013–2025) with 0.2 m resolution UAV multispectral imagery from small-scale experimental areas to construct a satellite–UAV collaborative observation method. Compared to previous studies relying on MODIS or Landsat data for large-area observation [51,52], the spatial resolution of the data in this study has been significantly improved, greatly enhancing the ability to capture dynamic changes in S. vulgaris shrubs.

In the identification methods of S. vulgaris, early studies predominantly employed NDVI threshold segmentation or traditional machine learning techniques (e.g., random forest) [53]. However, constrained by the issue of spectral mixed pixels, the recognition errors for shrub boundaries often exceeded 20%. This study introduces the SegU-Net deep learning model, trained with UAV ground truth data, achieving an overall accuracy of 99.48%, a Kappa coefficient of 0.67, and both precision and recall exceeding 99% for non-Sabina categories. This performance significantly surpasses the results of Zhang et al. (2023) based on U-Net for desert vegetation classification (90.8% accuracy) [54], demonstrating particularly stronger robustness in high-heterogeneity mining environments. Multi-temporal analysis further quantified growth disparities in protected areas: the expansion rate of Sabina in the southern zone (499,500 m²/year) was 3.5 times that of the northern zone (144,500 m²/year), complementing Wang et al.’s (2022) conclusion that “clonal propagation rates of Sabina are influenced by microtopography” [55].

5. Conclusions

This study systematically analyzed the impact mechanisms of mining activities on S. vulgaris shrub growth in Longde Coal Mine from 2013 to 2025 through multi-temporal remote sensing monitoring and evaluated the effectiveness of conservation policies. The main findings are as follows:

(1)

The study effectively identified the stress effects of coal mining on S. vulgaris. Mining activities directly inhibited shrub growth by altering hydrogeological conditions. During the Panel 202 mining period (2014–2021), Sabina area decreased by 40% cumulatively, and after the Panel 203 mining (2021–2025), the annual shrinkage rate reached 18%, indicating a positive correlation between mining intensity and vegetation decline. Subsidence-induced soil structural degradation (reduced clay content and increased porosity) and groundwater loss created a dual inhibition mechanism of water stress–root damage, leading to decreased transpiration efficiency and limited photosynthetic capacity in Sabina.

(2)

The statistical results demonstrated the ecological restoration effectiveness of the protected areas. The Sabina Nature Reserve (Northern and Southern Zones) showed continuous growth without mining disturbances, with area increases of 259% and 202% over 12 years, respectively. The Southern Zone’s coverage reached 7.88 million m² in 2025, confirming the effect of significant promotion of enclosure measures on natural regeneration. Systematic implementation of policy evolution (from county-level reserves to forest chief systems) and technical standards (e.g., seed propagation protocols) provided stable habitat protection and germplasm resources for Sabina.

(3)

The technical methodology proved reliable. The SegU-Net model achieved an overall accuracy of 99.48% and a Kappa coefficient of 0.67 (high consistency) in identifying Sabina distribution, with precision and recall exceeding 99% for non-Sabina categories, validating the effectiveness of remote sensing data for large-scale ecological assessments.

By synergizing multi-temporal high-resolution remote sensing and UAV technologies, this study overcame the spatiotemporal resolution bottleneck of traditional mining ecological monitoring, offering a methodological paradigm for balancing mineral resource development and ecological conservation in arid regions. The technical framework can be extended to long-term dynamic assessments of similar sensitive ecosystems (e.g., desert oases and wetlands).