Long-Term Changes in Shelterbelt Stability Along the Taklimakan Desert Highway Revealed by Landsat Observations

Highlights

What are the main findings?

• Long-term remote sensing monitoring revealed distinct establishment, development, and stabilization stages in the evolution of the Taklimakan Desert Highway shelterbelt from 2005 to 2025.
• A Shelterbelt Stability Index (SSI) integrating vegetation condition and structural characteristics successfully identified both stable sections and localized degradation hotspots along the highway.

What are the implications of the main findings?

• The results demonstrate that large-scale desert shelterbelt systems can maintain effective protective functions over more than two decades under extremely arid conditions.
• The proposed remote sensing framework enables long-term assessment of linear ecological engineering projects and provides scientific support for adaptive maintenance and sustainable management.

Abstract

The Taklimakan Desert Highway shelterbelt is the world’s largest ecological protection system established along a highway in a shifting desert environment and plays a critical role in mitigating wind-blown sand hazards and ensuring transportation safety. However, its long-term stability and protective capacity after more than two decades of operation remain insufficiently understood. In this study, Landsat imagery from 2005 to 2025 was used to monitor the long-term evolution of the shelterbelt along the Middle Section (~180 km) of the Taklimakan Desert Highway. A Random Forest classifier was employed to extract shelterbelt distribution, and classification results were validated using high-resolution Google Earth imagery and unmanned aerial vehicle observations. To quantify shelterbelt condition, a Shelterbelt Stability Index (SSI) was developed by integrating fractional vegetation cover (FVC), connectivity index (CI), percentage of landscape (PLAND), and perimeter-area fractal dimension (FRAC). The shelterbelt experienced initial seedling decline from 2005 to 2011, followed by progressive restoration during 2011–2020 and finally entered a stable saturated stage after 2020. Affected by saline water drip irrigation, wind-sand erosion and pipeline clogging, the overall vegetation condition deteriorated continuously before 2011. After targeted irrigation regulation, optimization of planting patterns and replanting measures were implemented; the degradation trend was reversed, contributing to the sustained improvement of vegetation thereafter. Significant spatial heterogeneity was observed along the highway, with certain sections maintaining high continuity and vegetation coverage, while others exhibited fragmentation, local discontinuities, area shrinkage, and increasing structural complexity. The proposed SSI effectively captured long-term structural dynamics and identified vulnerable sections subject to degradation. This study provides new insights into the life-cycle evolution of desert highway shelterbelts and offers scientific support for the sustainable management of ecological protection systems in arid environments.

1. Introduction

In arid regions, the construction of desert highways and railways provides essential infrastructure support for regional resource exploitation and economic connectivity [1]. The Taklamakan Desert, located in the hinterland of the arid region of northwestern China, is rich in oil and gas resources. However, regional development and transportation costs have long remained high due to the widespread distribution of mobile dunes, intense aeolian activity, and limited transportation accessibility. To meet the demands of oil and gas exploration in the Tarim Basin and strengthen regional transportation connections, China began constructing highways crossing the hinterland of the Taklamakan Desert in the 1990s. The desert highways within the Taklamakan Desert not only significantly shortened the transportation distance between northern and southern Xinjiang, but also became representative examples of major transportation engineering projects in Chinese hyper-arid regions [2].

However, aeolian activity in the hinterland of the Taklamakan Desert is extremely intense, and rapidly migrating mobile dunes pose persistent threats of sand accumulation, burial, and wind erosion along the highway, severely affecting traffic safety and long-term operation [3,4]. To ensure safe highway operation and reduce maintenance costs, an ecological shelterbelt project based on drip irrigation systems was gradually established along both sides of the Taklamakan Desert Highway beginning in 2003 [5]. The shelterbelt is primarily composed of drought- and salt-tolerant plant species, such as Haloxylon ammodendron and Tamarix, forming a continuous vegetation barrier. This shelterbelt effectively reduces near-surface wind velocity, intercepts aeolian sediment transport, and stabilizes surrounding sand dunes, thereby ensuring the stable and safe operation of the highway. After years of development, the Taklamakan Desert Highway shelterbelt has become one of the world’s largest ecological protection projects for highways crossing shifting deserts, playing an important role in safeguarding desert transportation infrastructure and maintaining regional ecological stability [6]. In recent years, with the implementation of the Tarim Desert Highway zero-carbon demonstration project, groundwater pumping systems along the highway have gradually shifted from traditional diesel-powered irrigation to photovoltaic-powered pumping systems, achieving an “oil-to-electricity” upgrade of the drip irrigation system and providing a more stable and low-carbon energy supply for the long-term operation of ecological engineering projects in hyper-arid regions [7].

Since its full establishment in 2003, the Taklamakan Desert Highway shelterbelt has operated continuously for more than 20 years. Under long-term environmental stresses including extreme aridity, high evaporation, strong aeolian activity, and highly mineralized groundwater, the shelterbelt ecosystem faces problems such as vegetation degradation, local discontinuities, width shrinkage, and structural fragmentation. Consequently, its long-term stability and sustained protective effectiveness have attracted increasing attention [5]. Previous studies have mainly focused on the early stages of shelterbelt construction [6], with research topics involving irrigation regimes [8,9,10,11], plant adaptability [12,13], physiological and ecological characteristics [14,15], and short-term windbreak and sand-control effects [5,16]. In contrast, relatively limited attention has been paid to the structural evolution and stability changes in the shelterbelt during its long-term operational stage. Particularly after more than two decades of operation, it remains unclear whether the shelterbelt can still maintain effective and continuous sand-blocking functions, whether significant differences exist among different highway sections, and how aeolian processes influence shelterbelt degradation over time.

With the increasing availability of long-term remote sensing datasets, remote sensing techniques now provide an effective means for monitoring the long-term dynamics of desert highway shelterbelts [17,18,19,20]. In particular, the long-term continuous observations provided by the Landsat satellite series, supplemented by high-resolution imagery such as Google Earth imagery, sub-meter commercial satellite imagery, and UAV (Unmanned Aerial Vehicle) imagery, provide an important data foundation for analyzing shelterbelt changes before and after construction as well as during long-term operation [21,22,23,24]. Compared with traditional field surveys and local-scale monitoring approaches, long-term remote sensing datasets enable the investigation of long-term changes in shelterbelt spatial structure and vegetation conditions at regional scales. Previous studies have mainly evaluated shelterbelt growth conditions using vegetation coverage or vegetation indices. However, for a typical linear ecological engineering system such as the Taklamakan Desert Highway shelterbelt, its protective effectiveness depends not only on vegetation growth conditions, but also on structural continuity, area retention, and boundary morphology. Local discontinuities, area shrinkage, or structural fragmentation may substantially reduce windbreak and sand-control effectiveness. Large-scale ecological engineering systems such as the Taklamakan Desert Highway shelterbelt are globally rare, if not unique, and have received limited attention regarding their long-term operational stability. Most previous studies have primarily focused on the early stages of shelterbelt construction and establishment, including irrigation strategies, plant adaptability, and short-term windbreak effects. In contrast, to the best of our knowledge, no systematic studies have been conducted to quantitatively evaluate the long-term operational stability of such desert highway shelterbelt systems over multi-decadal periods. Therefore, it is necessary to comprehensively evaluate the long-term stability of the shelterbelt from both structural and functional perspectives, and to reveal the characteristics and evolutionary processes of its spatial structural degradation [25,26,27,28,29]. To address this gap, this study focuses on the long-term stability assessment of the Taklamakan Desert Highway shelterbelt using multi-temporal Landsat observations. A Shelterbelt Stability Index (SSI) is proposed to quantify temporal changes in shelterbelt structural conditions, including continuity, area retention, and boundary characteristics. This provides a consistent framework for evaluating long-term stability dynamics of linear shelterbelt systems in hyper-arid regions.

2. Materials and Methods

2.1. Study Area

The study area is located along the Taklamakan Desert Highway in the hinterland of the Taklamakan Desert. The Taklamakan Desert, situated in the center of the Tarim Basin, is the largest shifting desert in China and the second largest shifting desert in the world, covering an area of approximately 337,600 km² [30]. The region is characterized by a typical warm temperate hyper-arid climate, with annual precipitation generally less than 50 mm and annual evaporation exceeding 2500–3000 mm [31]. The climate is extremely dry and aeolian activity is highly active. Influenced by prevailing winds and abundant sediment supply, various aeolian landforms are widely developed in the study area, including barchan dunes, compound longitudinal dunes, and dune chains. Mobile dunes migrate actively, resulting in intense wind-blown sand hazards [4,32,33]. The soils are mainly composed of shifting aeolian sandy soils, with sparse natural vegetation, simple community structure, and extremely low vegetation cover. Along the shelterbelt of the desert highway, surface runoff is generally absent; however, groundwater storage is relatively abundant. The groundwater salinity ranges from 2.6 g/L to 30 g/L, with more than 90% of the area exceeding 4 g/L.

The Taklamakan Desert Highway (Luntai–Minfeng Desert Highway) extends north–south across the hinterland of the Taklamakan Desert, with a total length of approximately 560 km, and serves as an important transportation corridor connecting the northern and southern margins of the Tarim Basin [34]. To mitigate sand burial and wind erosion hazards affecting the highway, an ecological shelterbelt project has gradually been established along both sides of the highway since 2003. The shelterbelt has an overall width of approximately 72–78 m (The data is derived from literature, but in actual observations, the width of the road is approximately 100–140 m) and is mainly composed of drought-tolerant species such as Haloxylon, Tamarix, and Populus euphratica. Its growth is maintained through a groundwater-based drip irrigation system. In recent years, photovoltaic-powered irrigation systems have gradually been adopted in some sections, providing energy support for the long-term operation of ecological engineering projects in hyper-arid regions [35]. The unique geographical conditions and irrigation water resources have resulted in extremely limited plant resources in the shelterbelt of the Taklamakan Desert Highway. The shelterbelt is dominated by highly stress-tolerant species adapted to salinity, drought, high temperature, wind erosion, and sand burial, including Haloxylon ammodendron, Calligonum spp., and Tamarix spp.

Shelterbelts of different stand ages are irrigated from March to November every 15 days, with each irrigation amounting to 30 L/m². No irrigation is applied from November to February. Groundwater is used for irrigation, with a salinity of approximately 5 g/L.

The shelterbelt in the study area is linearly distributed along the highway. Influenced by aeolian activity, groundwater conditions, and geomorphological differences, significant spatial heterogeneity exists in vegetation growth conditions and structural characteristics among different highway sections. In some areas, the shelterbelt maintains relatively high continuity and vegetation coverage, whereas in other sections local discontinuities, width shrinkage, and boundary fragmentation have occurred. Since the Taklamakan Desert Highway shelterbelt has been operating continuously for more than 20 years, its long-term stability and changes in windbreak and sand-control effectiveness have become important issues in current desert ecological engineering research. This study selects a 180 km-long middle section of the highway in the hinterland of the Taklamakan Desert as the research area(in Figure 1). Located deep in the desert and far from oases, this area suffers the severest aeolian sand hazards and covers northern, southern and Tazhong subsections with diverse environmental conditions, making it representative.

2.2. Data Sources

This study employed multi-source remote sensing data to monitor the long-term evolution of the shelterbelt along the Taklamakan Desert Highway. Long-term remote sensing observations were primarily derived from the Landsat satellite series. Owing to its long temporal coverage, Landsat imagery can capture conditions before and after shelterbelt construction as well as during long-term operation [36]. The native multispectral bands have a spatial resolution of 30 m, while pan-sharpening with the panchromatic band enables an enhanced 15 m dataset, making it suitable for long-term dynamic analysis at the regional scale. In addition, high-resolution Google Earth imagery, Esri World Imagery, Wayback imagery, and UAV imagery were used to assist in the identification and validation of structural details such as shelterbelt boundaries, local discontinuities, and width variations, thereby improving the extraction accuracy of the shelterbelt.

To reduce the influence of seasonal variation on vegetation information extraction, remote sensing images with relatively low cloud cover during the growing season (June–September) were preferentially selected for each year. All datasets from different years underwent consistent preprocessing procedures, including radiometric calibration, atmospheric correction, image mosaicking, clipping, and cloud masking. Considering that differences in spectral responses among sensors and acquisition years may introduce biases into vegetation indices, long-term time series data were further normalized to improve interannual comparability [37,38].

Normalized Difference Vegetation Index (NDVI), Modified Soil-Adjusted Vegetation Index (MSAVI), and related spectral features were calculated from the remote sensing imagery and combined with texture information to construct the feature variable set [39]. Subsequently, the Random Forest (RF) classification algorithm was employed to extract the shelterbelt distribution. Training samples were mainly obtained through manual interpretation based on high-resolution Google Earth imagery and UAV imagery, while independent validation samples were used to assess classification accuracy.

Wind speed and wind direction observations were obtained from the meteorological station located near the Taklamakan Desert Highway (Xiaotang, Tazhong Town, and Andier River stations). Hourly records were used to calculate the frequency distribution of wind directions and wind speed classes, and the wind rose diagram was generated to characterize the dominant wind regimes in the study area.

2.3. Methods

2.3.1. Random Forest-Based Shelterbelt Extraction

The study area is located in the interior of the Taklimakan Desert, where land cover types are relatively simple and primarily consist of desert surfaces and highway shelterbelt vegetation. Owing to the distinct spectral differences between shelterbelt vegetation and surrounding bare sand, a machine-learning-based classification approach was considered sufficient for shelterbelt extraction. Therefore, a Random Forest (RF) classifier [40,41] was employed to map the spatial distribution of the shelterbelt from multi-temporal Landsat imagery. Landsat images from eight representative years (2005, 2008, 2011, 2014, 2017, 2020, 2023, and 2025) were selected to characterize the long-term evolution of the Taklamakan Desert Highway shelterbelt. Each year’s imagery was independently classified using a year-specific Random Forest model, and the resulting land-cover maps were used for subsequent SSI calculation.

Initially, the seven Landsat spectral bands and the Modified Soil-Adjusted Vegetation Index (MSAVI) were combined to form an eight-band feature stack. Training samples of shelterbelt vegetation and desert surfaces were collected through visual interpretation of high-resolution Google Earth imagery and UAV observations. The RF classifier was then applied to generate a preliminary shelterbelt distribution map.

To further improve the discrimination of sparse vegetation in the desert environment, fractional vegetation cover (FVC) was subsequently incorporated into the classification framework. Based on the preliminary classification results, MSAVI values of the identified shelterbelt and desert pixels were extracted separately. Considering the relatively high background MSAVI values commonly observed in desert regions, the 98th percentile MSAVI values of the shelterbelt and desert classes were selected as representative values of pure vegetation (MSAVI_veg) and bare soil (MSAVI_soil), respectively. FVC was then estimated using the pixel dichotomy model [42]:

$$FVC={\displaystyle \frac{(\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}-{\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}_{soil})}{({\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}_{veg}-{\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}I}_{soil})}}$$

(1)

where MSAVI is the MSAVI value of an individual pixel, while MSAVI_veg and MSAVI_soil represent the MSAVI values of pure vegetation and bare soil, respectively.

Subsequently, the derived FVC layer was combined with the seven Landsat spectral bands to construct a new feature stack, which was used as the input for a second RF classification. By integrating vegetation cover information into the classification process, this two-stage RF framework effectively enhanced the identification of shelterbelt vegetation and reduced the influence of desert background noise. The resulting shelterbelt distribution maps were used for subsequent analyses of shelterbelt stability and structural evolution.

The Random Forest (RF) classification was conducted independently for each year to ensure temporal consistency in land cover mapping. For each time period, separate training and validation datasets were generated based on the corresponding annual imagery, rather than using a unified model across multiple years.

A total of 30,351 samples were collected in this study, including 12,657 shelterbelt samples and 17,694 bare sand samples. These samples were evenly distributed across different years to ensure temporal representativeness. For each annual classification, 70% of the samples were randomly selected for model training, while the remaining 30% were used as independent validation data for accuracy assessment.

The sample selection followed a spatially even random sampling strategy across the entire study area to reduce spatial bias and ensure representative coverage of both land cover types. The RF classifier was implemented using the ENVI (v5.6.2) Random Forest module with default parameter settings, and no hyperparameter tuning was applied. Although FVC is derived from MSAVI, it is calculated using a pixel dichotomy model with percentile-based vegetation and bare-soil endmembers (98th percentile) and therefore is not a simple linear transformation of the original spectral bands. A VIF analysis based on the training samples yielded values of 46.96, 40.16, and 6.33 for the Red, NIR, and FVC variables, respectively, indicating that FVC exhibits substantially lower collinearity than the original spectral bands. Furthermore, Random Forest is generally insensitive to multicollinearity because it is based on recursive tree splitting rather than parameter estimation [40]. Therefore, combining FVC with the original spectral bands provides complementary ecological information on vegetation cover without adversely affecting the classification performance.

2.3.2. Development of the Shelterbelt Stability Index (SSI)

To quantitatively evaluate the long-term stability of the Taklimakan Desert Highway shelterbelt, a Shelterbelt Stability Index (SSI) was developed by integrating vegetation condition, structural continuity, area retention, and boundary complexity. Considering the linear characteristics of the shelterbelt, a road-oriented assessment framework was established.

First, a 250 m buffer was generated on both sides of the highway centerline to delineate the shelterbelt corridor, resulting in a corridor width of 500 m. The corridor was subsequently divided into consecutive 500 m × 500 m analysis units along the highway alignment. Each unit was treated as an independent assessment cell for the calculation of shelterbelt stability indicators. This design ensured consistency with the linear structure of the shelterbelt while allowing detailed spatial analysis along the entire highway.

Four indicators were selected to characterize different aspects of shelterbelt condition, including fractional vegetation cover (FVC), continuity index (CI), percentage of landscape (PLAND), and perimeter-area fractal dimension (FRAC).

FVC was used to represent vegetation growth conditions within each analysis unit. PLAND [43] was used to quantify the proportion of shelterbelt area within the analysis unit and was calculated as:

$$PLAND=A_s/A_t$$

(2)

where A_s is the shelterbelt area and A_t is the total area of the analysis unit.

FRAC was used to characterize the complexity and irregularity of shelterbelt boundaries [44]:

$$FRAC={\displaystyle \frac{2\ln \left(0.25P\right)}{\ln A}}$$

(3)

where P is the perimeter of the shelterbelt patch and A is the corresponding patch area. Higher FRAC values indicate more irregular patch shapes and a higher degree of fragmentation.

To characterize shelterbelt continuity, a Continuity Index (CI) was specifically designed for the linear shelterbelt system. Within each analysis unit, the extracted shelterbelt polygons were projected onto the centerline of the highway. If the projected shelterbelt completely covered the highway centerline segment within the analysis unit, the shelterbelt was considered continuous and assigned a value of 1. Otherwise, the shelterbelt was considered interrupted and assigned a value of 0. The CI was therefore defined as:

$$CI=\left\{\begin{array}{c}1,\hspace{1em} |\hspace{1em} L_p=L_r\\ 0,\hspace{1em} |\hspace{1em} L_p<L_r\end{array}\right.$$

(4)

where L_p is the projected shelterbelt length on the highway centerline and L_r is the length of the highway centerline segment within the analysis unit.

Prior to index construction, all indicators were normalized to a range of 0–1 using min–max normalization. Since FRAC represents a negative indicator, it was normalized using the Min–Max method, and its negative contribution was incorporated through a negative coefficient in the SSI formulation. Based on the relative importance of vegetation condition, structural continuity, area retention, and boundary integrity in maintaining shelterbelt functionality, weights of 0.2, 0.4, 0.3, and 0.1 were assigned to FVC, CI, PLAND, and FRAC, respectively [45,46]. The weighting scheme was determined according to the relative importance of each indicator in maintaining the protective function of a linear shelterbelt. Structural continuity was considered the most critical factor because interruptions in the shelterbelt may create pathways for sand transport even when vegetation coverage remains high. Area retention was assigned the second highest weight as it reflects the preservation of shelterbelt width and spatial extent. Vegetation condition contributes to shelterbelt functionality but cannot fully compensate for structural discontinuities. Boundary complexity primarily reflects fragmentation characteristics, indicating the proportion of direct contact between the shelterbelt and external wind–sand intrusion. Therefore, it was assigned the lowest weight. The SSI was calculated as:

$$SSI=0.2\mathrm{F}\mathrm{V}\mathrm{C}+0.4\mathrm{C}\mathrm{I}+0.3\mathrm{P}\mathrm{L}\mathrm{A}\mathrm{N}\mathrm{D}-0.1\mathrm{F}\mathrm{R}\mathrm{A}\mathrm{C}$$

(5)

where higher SSI values indicate better vegetation growth conditions, stronger structural continuity, greater shelterbelt area retention, lower fragmentation, and consequently higher long-term stability and potential protective capacity.

2.3.3. COSI-Corr-Based Dune Migration Analysis

To investigate the influence of aeolian processes on shelterbelt stability, dune migration was quantified using the Co-registration of Optically Sensed Images and Correlation (COSI-Corr) technique. COSI-Corr is a sub-pixel image correlation method originally developed for measuring horizontal ground displacement from multi-temporal optical imagery. By tracking the movement of stable image texture patterns between image pairs, the method can derive spatially continuous displacement fields with sub-pixel accuracy [47]. Owing to its ability to detect subtle surface movements over large areas, COSI-Corr has been widely applied in studies of tectonic deformation, glacier motion, landslides, and dune migration. Previous studies have demonstrated that image correlation techniques can effectively capture the migration of active dunes in arid environments and provide reliable estimates of dune movement rates [48,49,50,51].

COSI-Corr analysis was performed using two Landsat 8 Collection 2 Level-2 (C2L2) images acquired on 28 June 2023 and 19 July 2025. Both images were pan-sharpened to a spatial resolution of 15 m prior to image correlation analysis, ensuring consistent radiometric and geometric characteristics. Prior to displacement estimation, all images were geometrically corrected and accurately co-registered to minimize geometric distortions and registration errors. The frequency-domain correlation algorithm implemented in COSI-Corr was then applied to derive east–west (U_x) and north–south (U_y) displacement fields between image pairs.

The horizontal displacement magnitude (D) was calculated as:

$$D=\sqrt{U_x^2+U_y^2}$$

(6)

where U_x and U_y represent the east–west and north–south displacement components, respectively.

To facilitate comparisons among image pairs with different observation intervals, annual dune migration rate R was calculated as:

$$R={\displaystyle \frac{D}{∆t}}$$

(7)

where $∆$t is the time interval between image acquisitions in years.

As dune migration is primarily driven by wind erosion and sediment transport processes, the migration rate was used as a proxy indicator of aeolian activity intensity. Higher migration rates indicate stronger wind-driven sediment transport and more active dune dynamics. The resulting spatial patterns of aeolian activity were subsequently compared with SSI distributions to explore the potential influence of wind-blown sand processes on long-term shelterbelt stability.

3. Results

3.1. Classification Accuracy of Shelterbelt

The Random Forest classification successfully extracted the spatial distribution of the shelterbelt along the Taklimakan Desert Highway. Visual comparison with high-resolution World Imagery Wayback images demonstrated that the extracted shelterbelt boundaries were generally consistent with the actual shelterbelt distribution, accurately capturing variations in shelterbelt width, local discontinuities, and fragmented sections (Figure 2). The classification results effectively distinguished shelterbelt vegetation from the surrounding desert background despite the sparse vegetation conditions characteristic of the study area.

The classification accuracy was further evaluated using an independent validation dataset. Confusion matrix and Kappa coefficient were selected to quantitatively assess classification precision [41,52]. This error matrix establishes a cross-tabulation between pixel-wise predicted categories from model simulations and corresponding field-verified reference labels [53,54]. The confusion matrix and corresponding accuracy metrics are presented in

Table 1. Confusion matrix and classification accuracy metrics for bare land and shelterbelt.

Ground Truth/ Prediction	Bare	Shelter	Ground Truth Total	Producer’s Accuracy (PA)	User’s Accuracy (UA)	F1-Score
Bare	3117	269	3386	92.06%	97.62%	94.76%
Shelter	76	1783	1859	95.91%	86.89%	91.18%
Prediction Total	3193	2052	5245
Overall Accuracy (OA)						93.42%
Kappa Coefficient						0.8596

. The overall classification accuracy (OA) reached 93.42%, with a Kappa coefficient of 0.8596, indicating a high level of agreement between the classified results and reference samples.

For the shelterbelt class, the producer’s accuracy (PA) and user’s accuracy (UA) were 95.91% and 86.89%, respectively, resulting in an F1-score of 91.18%. The high producer’s accuracy indicates that most shelterbelt pixels were successfully identified, whereas the slightly lower user’s accuracy suggests that a small proportion of desert pixels were misclassified as shelterbelt vegetation. The relatively lower user’s accuracy of the shelterbelt class can be primarily attributed to the spatial resolution of Landsat imagery. Compared with high-resolution World Imagery Wayback images, Landsat pixels are considerably coarser and often contain mixed land-cover information. Along shelterbelt boundaries, some pixels may include only a small proportion of vegetation or be located adjacent to shelterbelt patches. Although these pixels do not represent pure shelterbelt vegetation in the reference imagery, they may still exhibit vegetation-related spectral characteristics and therefore be classified as shelterbelt pixels. Consequently, a limited number of surrounding desert pixels were misclassified as shelterbelt vegetation, leading to a slight reduction in user’s accuracy. This phenomenon is particularly common in narrow linear vegetation belts distributed within sparsely vegetated desert environments. For the bare desert class, the producer’s accuracy, user’s accuracy, and F1-score reached 92.06%, 97.62%, and 94.76%, respectively, demonstrating reliable discrimination between shelterbelt vegetation and surrounding desert surfaces.

Overall, the classification results achieved high accuracy and effectively captured the spatial characteristics of the shelterbelt. The extracted shelterbelt maps therefore provide a reliable basis for subsequent analyses of shelterbelt stability and long-term structural evolution.

3.2. Spatiotemporal Evolution of Shelterbelt Stability

The Shelterbelt Stability Index (SSI) was retrieved from time-series Landsat satellite imagery to quantify the spatiotemporal dynamics of vegetation growth and overall integrity of the Tarim Desert Highway shelterbelt. Eight representative years covering a 20-year span (2005, 2008, 2011, 2014, 2017, 2020, 2023, 2025) were selected to generate SSI longitudinal profiles (Figure 3). The horizontal axis of each subplot corresponds to sequential sampling cell IDs (P0–P359) evenly distributed across the full highway corridor, while the vertical axis represents the SSI value. The theoretical range of SSI is −0.1 to 0.9, although all calculated SSI values in this study were positive. Higher SSI values indicate greater shelterbelt integrity, characterized by higher vegetation coverage, stronger spatial continuity, larger shelterbelt proportion, and lower boundary complexity. Notably, sampling cells P56–P63 were deliberately left unplanted during the initial shelterbelt construction as blank control plots, which persistently exhibited SSI values close to zero across all observation years and should be excluded when interpreting vegetation degradation signals. The segment P155–P170 covers the range of Tazhong Town, a core human settlement along the desert highway [55]. Benefiting from convenient logistics and resident workforce, the construction and daily maintenance costs of local shelterbelts are significantly lower than remote desert sections, supporting long-term stable irrigation management.

3.2.1. Interannual Variation in Regional Mean SSI

According to the longitudinal SSI profiles, 2005 marked the initial stage right after seedling planting. The overall vegetation foundation was relatively weak, with continuous high-SSI patches only concentrated around Tazhong Town. Most other sections along the highway exhibited extensive low-value patches due to the limited growth vigor of newly planted seedlings. By 2008, the vegetation condition of the entire highway deteriorated further compared with 2005, accompanied by a comprehensive decline in the average SSI value. The upper envelope of peak SSI values dropped sharply in most sampling cells, and both the quantity and coverage of low-value depressions within the corridor expanded simultaneously. Visual comparison of the profile curves from the two years reveals that low-value areas were only widely distributed in marginal sections in 2005, whereas fragmented low-value patches spread toward the highway mid-section in 2008, further shrinking the range of continuous high-value zones.

After the full-line average SSI bottomed out in 2011, the shelterbelt entered a slow yet sustained recovery phase from 2014 to 2020. The spatial extent of vegetated cells with persistently low SSI values shrank year by year, the width of natural zero-value bare gaps along the highway corridor narrowed continuously, and the upper limit of SSI curves rebounded steadily.

The period after 2020 (2023–2025) witnessed a new stage of saturated and high-level shelterbelt integrity. Nearly all vegetated cells across the full range P0–P359 formed continuous high-SSI coverage, long strips of natural zero-value bare gaps almost disappeared completely, and the fluctuation amplitude of SSI curves decreased significantly. Two large-scale infrastructure optimization projects completed after 2022 further consolidated the high and stable growth status of the shelterbelts: photovoltaic retrofitting of all diesel-powered irrigation wells along the full highway, and the full deployment of intelligent precision drip irrigation equipped with real-time soil moisture and salinity sensors [56,57]. The 2025 longitudinal SSI profiles demonstrate that the spatial uniformity of vegetation vitality across all planting zones reached its highest level during the 20-year observation period. These results describe the observed spatiotemporal patterns of SSI, while potential driving mechanisms are further discussed in Section 4.

3.2.2. Spatial Heterogeneity of SSI Along the Highway

Obvious spatial heterogeneity of shelterbelt stability was observed across all eight observation years. This spatial pattern is co-regulated by multiple environmental stress factors, including regional aeolian sand intensity, surface soil salt accumulation, and shallow groundwater burial depth. Three planting segments (P0–P50, P200–P250 and P350–P359) consistently displayed low SSI values throughout the observation years. Each segment suffers from superimposed abiotic stresses, yet the dominant limiting factor varies spatially: wind erosion and frequent sand burial constitute the primary constraints in windward dune zones; low-lying lands face severe salt accumulation driven by capillary rise in shallow groundwater. Remote hinterland near the terminal P350–P359 features deep groundwater inaccessible to plant roots and inconvenient maintenance of irrigation pipelines due to remoteness from residential areas. These three sections acted as persistent weak zones of the shelterbelt system over the 20-year monitoring period.

Mid-corridor sampling cells P100–P180 maintained relatively high baseline SSI values in all study years, among which P155–P170 corresponds to Tazhong Town. The gentle terrain in these zones greatly alleviates sand burial stress; seasonal winter salt-leaching irrigation keeps root zone salinity at a moderate level. Furthermore, Tazhong Town (P155–P170) boasts convenient material transportation and stable on-site maintenance personnel, which cuts construction and routine operation costs of shelterbelts and guarantees uninterrupted irrigation supply to sustain continuous shrub growth. Although these high-integrity core segments also experienced successive SSI declines from 2005 to 2011, their vegetation recovery rate after 2011 significantly outperformed marginal cells subject to dual wind-sand and salt stresses, narrowing the longitudinal spatial disparities in shelterbelt stability year by year.

From 2005 to 2025, the stability gap between high-SSI core sections and low-SSI vulnerable marginal vegetated segments narrowed gradually. By 2025, even sampling cells that suffered severe historical degradation and long-term dual wind-sand and salt stresses sustained baseline SSI values above 0.4. This result confirms that long-term optimized irrigation regulation, periodic supplementary afforestation, and integrated ecological remediation effectively mitigated the severe spatial heterogeneity of vegetation growth in the early stage after shelterbelt construction. This long-term narrowing of spatial stability disparities also provides direct evidence that the artificial desert shelterbelt represents a self-reinforcing protective system, whose sand-fixing capacity and ecosystem stability continuously improve with operation time under effective human intervention.

3.2.3. Summary of Temporal Evolution Patterns

Synthesizing the interannual and spatial SSI patterns derived from Landsat observations in this study, three long-term evolutionary phases of shelterbelt stability were identified based on all planting cells within P0–P359 (excluding artificial control plots P56–P63):

• Early Salinity Degradation Phase (2005–2011): 2005 fell in the initial seedling establishment stage, where continuous high-SSI bands only appeared locally around Tazhong, and extensive low-value patches spread across most areas of the full highway due to weak seedling vigor. Salt stress intensified in 2008, expanding the scope of vegetation degradation; combined with the poor resistance of young seedlings to aeolian sand damage, the full-line SSI status was inferior to that in 2005. The average full-line vegetation status hit the cycle trough in 2011, marking the peak of overall degradation, while partial vegetation recovery emerged in formerly barren large-scale low-value areas of the southern highway. The damage to vegetation integrity intensified continuously during this phase, and the range of low-value degraded patches along the highway kept expanding.
• Artificial Restoration and Gradual Recovery Phase (2011–2020): Targeted artificial management practices were widely implemented after the 2011 vegetation degradation survey, driving steady rebounds in SSI values over the subsequent decade. The overall vegetation coverage and canopy integrity rebounded year by year, and the scope of low-value degraded areas was continuously compressed. Benefiting from lower construction and operation costs as well as uninterrupted irrigation supply, this central segment of Tazhong Town (P155–P170) recovered faster than surrounding remote desert sections.
• Saturated Stable Maintenance Phase (2020–2025): Full-line planting cells covering P0–P359 maintained high regional average SSI values, most natural vegetation gaps were completely eliminated, and the shelterbelt delivered long-term stable sand-fixing functions supported by upgraded intelligent irrigation infrastructure. With the gradual growth of Haloxylon ammodendron and other shrubs, their resistance to wind and sand improved remarkably, further promoting the overall functionality and stability of the shelterbelt ecosystem.

Overall, the three-stage evolution trajectory fully demonstrates the self-improving property of this saline-water irrigated desert shelterbelt. After overcoming the mid-term salt and water stress crisis via human regulation, the ecosystem enters a stable state with continuously rising anti-interference capacity, verifying that artificial shrub shelterbelts can achieve progressively stronger and more reliable sand fixation effects over decades of operation.

3.3. Spatiotemporal Evolution of SSI Component Indicators

Four remote sensing-derived indicators, namely FVC, CI, PLAND and FRAC, were adopted to describe vegetation coverage, corridor continuity, shrub patch proportion and patch shape complexity, respectively. Longitudinal heatmaps (Figure 4) take sampling cells P0–P359 as the horizontal axis and eight monitoring years as the vertical axis, with gradient colors representing metric magnitudes. The segment P155–P170 corresponds to Tazhong Town, a key Tarim Oilfield base. Sufficient ecological investment from oil operations guarantees year-round maintenance and stable irrigation, making this section retain superior vegetation status through all monitoring years.

FVC reflects pixel-level canopy coverage and seedling growth vigor. High FVC zones concentrated around Tazhong in 2005, while most other sections had sparse seedlings and low coverage. Vegetation conditions degraded entirely in 2008, and further deteriorated to the full-line trough in 2011, yet scattered local FVC improvements emerged in southern segments as some seedlings adapted to saline soil. Large-scale vegetation recovery took place across the corridor in 2014, and FVC kept rising afterwards; southern areas still exhibited periodic FVC fluctuations, possibly attributable to unstable local groundwater recharge and sporadic wind-sand damage. CI only takes values of 0 and 1, where 1 stands for gapless continuous shrub belts and 0 refers to broken canopy. The number of cells with CI = 0 declined continuously starting from 2014, showing progressive recovery of corridor connectivity. PLAND describes landscape-scale area proportion of shrub patches, which was lower in 2008 compared with 2005, and witnessed consistent growth year by year after hitting the lowest value in 2011. FRAC quantifies patch edge irregularity; higher FRAC values represent more fragmented, complex patch outlines and inferior landscape integrity, implying shelterbelts with identical total area expose a larger proportion of their periphery to external aeolian sand invasion. Distinct from the other three indicators, FRAC rose steadily across the whole corridor after 2014. This trend originates from two factors: shifting mobile dunes cut complete shrub patches into scattered blocks, and overgrown shrubs spread beyond the originally planned planting belts, forming jagged patch edges. Meanwhile, FRAC values of all sampling cells gradually converged, indicating that the spatial differentiation of patch morphological complexity along the highway was continuously weakened.

4. Discussion

4.1. Interpretation of Temporal Patterns in Shelterbelt Stability

Spatially, the SSI exhibited a pronounced corridor-wide decline during the early monitoring period, with low values distributed continuously along most sections of the highway shelterbelt. This phenomenon suggests that the observed vegetation degradation during this period may be associated with long-term soil salinization following several years of saline irrigation after afforestation. The primary driver of this early-stage degradation is surface soil salt accumulation induced by long-term saline groundwater irrigation [58]. Intense desert evaporation drives soluble soil salts upward to topsoil layers, inhibiting shrub photosynthesis and root water absorption. Meanwhile, partial mortality of salt-sensitive Calligonum mongolicum and Tamarix further pulled down the regional average SSI of the entire shelterbelt [59,60]. This period represents a system-wide low stability stage characterized by uniformly reduced SSI values across the majority of analysis units.

Although the overall SSI remained low in 2011, spatial heterogeneity began to emerge, with localized areas showing partial recovery signals while most sections continued to exhibit low stability values. The overall vegetation degradation trend persisted, and the average vegetation status of the full highway hit its lowest point across the whole monitoring period in 2011. From a spatial perspective, although the overall SSI level of the highway decreased universally in 2011, partial improvements appeared in the southern highway sections that had been covered by extensive continuous low-value zones in previous years. Some seedlings gradually adapted to saline-alkali soil environments and expanded their canopy size, forming scattered small peaks of rebounded SSI values. As clearly illustrated in the profiles, apart from the partial recovery in the southern segments, persistent low SSI values prevailed across most highway areas; the upper envelope of SSI curves in core planting zones reached the minimum of the monitoring cycle, and large scattered low-value depressions emerged within formerly intact shelterbelt segments. Large-scale shrub decline during the middle growth stage was driven by multiple interacting biotic and abiotic stresses following six years of saline irrigation. These patterns are likely driven by multiple interacting environmental and ecological processes, including progressive soil salinization, reduced irrigation efficiency due to sand-induced pipeline blockage, intensified aeolian erosion, and increased intraspecific competition within dense shrub canopies. Tamarix ramosissima and Calligonum mongolicum suffered the most severe impacts, exhibiting decreased canopy vitality and extensive plant mortality [61], which eventually pulled the regional average SSI down to the minimum across the entire monitoring period (Figure 5).

From 2011 onwards, SSI values exhibited gradual spatial redistribution, with an increasing number of medium-to-high stability patches appearing along the corridor. The gradual vegetation recovery was associated with a series of field management measures implemented after the large-scale vegetation degradation survey in 2011. These measures included seasonal winter salt-leaching irrigation, selective thinning of dense shrub stands, biological pest control, and supplementary replanting in salt-affected and wind-eroded areas. Together, these interventions significantly improved soil conditions and vegetation survival, thereby promoting the long-term stabilization of the shelterbelt system [63,64]. By 2020, the shelterbelt stability of most highway planting sections tended to be balanced, with only a small number of isolated short sampling cells presenting low SSI values (Figure 6).

4.2. Environmental Controls on the Spatial Heterogeneity and Long-Term Evolution of Shelterbelt Stability

The spatial heterogeneity of shelterbelt stability observed along the Taklimakan Desert Highway is the result of long-term interactions among aeolian processes, groundwater conditions, soil salinity, and human management. Although the shelterbelt has been continuously maintained for more than two decades, substantial differences in SSI remained among planting sections, indicating that local environmental constraints continue to exert a strong influence on vegetation performance.

The spatial consistency between the 2025 SSI distribution and the COSI-Corr-derived surface displacement field provides suggestive evidence that aeolian processes may influence shelterbelt stability (Figure 7). Figure 8 presents wind roses from three meteorological stations along the desert highway: Xiaotang at the northern starting section, Tazhong in the middle reach, and Andir River in the southern segment. The drift potentials (DP) at the three stations were 147.7, 64.4, and 0.41, respectively, showing a clear decreasing trend from north to south (“strong in the north and weak in the south”). The wind roses indicate that the prevailing regional wind directions are consistent with dune migration patterns. The displacement vectors reveal a highly organized regional sediment transport pattern across the desert surface. Several low-SSI sections coincide with zones experiencing relatively intense sediment transport and frequent sand accumulation. Under such conditions, continuous sand burial increases mechanical damage to shrubs, suppresses seedling recruitment, and raises irrigation demand. Consequently, vegetation recovery in these areas is slower than in sections subjected to weaker aeolian disturbance.

The magnitude of sand erosion on the shelterbelt is jointly governed by aeolian transport intensity and the intersection angle between dune migration paths and the highway. Larger angles between prevailing sand-moving winds and the highway corridor bring more lateral sand influx into the planted zone, aggravating sand burial and canopy abrasion; mild erosion occurs where wind directions run nearly parallel to the shelterbelt. Dune migration across the inner Taklamakan Desert generally follows a northeast-to-southwest trajectory [4], largely parallel to the linear stretch of the desert highway, yet sharp bends along the highway alter the local wind field, triggering concentrated intense sand erosion at these turning points. The southern corridor segments sit within extensive linear dune fields where regional sand transport intensity peaks [33,65], producing persistent heavy sand pressure and the lowest local SSI values across the study area.

Topography and superimposed dune morphologies further modulate aeolian erosion intensity. When the highway traverses the crests of massive compound sand mountains, abundant small secondary superimposed dunes gather on windward slopes (Figure 9). Wind energy dissipates substantially as airflow climbs the steep dune surface, weakening near-surface sand carrying capacity and reducing sand deposition within the shelterbelt, which delivers relatively mild aeolian stress and improved vegetation stability compared with linear dune zones in the south.

4.3. Implications for Long-Term Sustainability of Artificial Desert Shelterbelt Ecosystems

The Taklimakan Desert Highway shelterbelt represents one of the largest and longest-operating artificial vegetation systems established in an extremely arid environment. The twenty-year evolution revealed in this study demonstrates that long-term ecological stability can be achieved under continuous management; however, the results also highlight the inherent vulnerability of artificial desert ecosystems and the necessity of sustained human intervention.

Unlike natural desert ecosystems that have evolved under long-term environmental selection, the shelterbelt relies heavily on artificial irrigation and maintenance. The vegetation system is therefore strongly coupled to external resource inputs, particularly groundwater extraction and irrigation infrastructure. Although SSI values gradually recovered after 2011 and remained relatively high after 2020, the observed degradation phase indicates that ecosystem stability is not guaranteed even after successful establishment. Relatively small changes in water availability, salinity conditions, or maintenance intensity may trigger significant ecological responses.

The spatial distribution of SSI further demonstrates that environmental constraints continue to regulate shelterbelt performance despite more than two decades of operation. Low-stability sections persist in areas affected by strong aeolian activity, severe salinity stress, or unfavorable groundwater conditions. These findings suggest that ecological vulnerability remains embedded within the system and may become increasingly important under future environmental change. Consequently, the apparent stabilization observed in recent years should not be interpreted as complete ecological independence from external management.

Climate change may further amplify these challenges. Rising temperatures and increasing evaporative demand could intensify soil salinization and increase irrigation requirements, while changes in wind regimes may alter patterns of sediment transport and sand burial. Under such circumstances, vulnerable sections of the shelterbelt may experience renewed degradation if management strategies remain unchanged. Therefore, long-term monitoring should remain an essential component of shelterbelt management to detect early warning signals of ecological decline.

Another important implication concerns the sustainability of saline-water irrigation. The successful recovery observed after 2011 demonstrates that ecological risks associated with highly mineralized groundwater can be mitigated through optimized irrigation scheduling, winter salt-leaching practices, and supplementary vegetation management. However, the long-term accumulation of salts within the soil profile remains a potential threat that requires continuous attention. Future management should therefore focus on maintaining a balance between water supply, salt removal, and vegetation demand to avoid repeating the degradation cycle observed during the early operational period.

The results of this study suggest that the long-term success of artificial desert shelterbelts depends not only on vegetation establishment but also on the development of adaptive management systems capable of responding to environmental change. Particular attention should be directed toward sections characterized by strong sand transport activity, shallow saline groundwater, or deep groundwater tables, as these areas are likely to remain the most vulnerable components of the shelterbelt network. Targeted management measures in these high-risk zones may substantially improve the overall resilience of the system.

Overall, the Taklimakan Desert Highway shelterbelt demonstrates that large-scale ecological engineering projects can maintain stable ecological functions for decades in hyper-arid environments. Nevertheless, the system remains fundamentally different from natural ecosystems because its stability is continuously supported by human intervention. The future sustainability of the shelterbelt will therefore depend on the ability to balance ecological processes, water resources, and management inputs under changing environmental conditions.

4.4. Limitations

Several limitations should be acknowledged to improve the interpretation of the results.

First, although Landsat imagery with 15 m spatial resolution (after spectral fusion) provides consistent long-term observations, the detection of fine-scale structural changes in narrow shelterbelt systems may still be affected by mixed-pixel effects, particularly at shelterbelt–desert boundaries. This may introduce uncertainty in delineating small-scale fragmentation patterns.

Second, the SSI is a composite indicator designed to characterize the structural and functional stability of the shelterbelt system. It does not directly measure ecological protection effectiveness, but serves as a proxy to represent long-term stability trends derived from multiple structural attributes.

Third, uncertainties are associated with COSI-Corr-derived surface displacement products due to image registration accuracy, surface texture heterogeneity, and the inherent limitations of optical image correlation in extremely arid environments with low-contrast surfaces.

Fourth, the interpretation of shelterbelt evolution is based on available remote sensing observations and published literature, while some detailed field-based management records (e.g., irrigation pipeline condition, pest occurrence, and maintenance logs) are not fully available at the same temporal resolution, limiting direct attribution of observed changes to specific management actions.

Finally, the findings are derived from a highly specific engineered ecological system in the central Taklamakan Desert, where the landscape is dominated by a binary structure of artificial shelterbelt corridors and mobile desert dunes. Therefore, caution should be taken when generalizing the results to other ecological or geographical contexts with more complex vegetation structures.

5. Conclusions

This study uses long-term Landsat satellite observation data to obtain the coverage of shelterbelts along the Middle Section (~180 km) of the Tarim Desert Highway, and constructs the shelterbelt stability index (SSI) by combining shelterbelt functional attributes, vegetation characteristics and landscape ecological indicators. It reveals the spatiotemporal evolution of the shelterbelts since their establishment (2005–2025). The results show that the shelterbelts have experienced successive stages of vegetation maturation, degradation induced by ecological problems, saline water irrigation and plant diseases and insect pests, and gradual improvement and stabilization after artificial restoration. This study provides empirical references for long-term monitoring, targeted maintenance and stability improvement of desert highway shelterbelts.