Multi-index assessment and spatiotemporal dynamics of vegetation cover in desert regions of Xinjiang based on pixel binary models

Fractional vegetation cover (FVC) as a key indicator for assessing ecological degradation and recovery in arid desert regions. However, the selection of suitable vegetation indices for desert areas remains controversial because of limitations imposed by soil background interference and regional heterogeneity. This study employs a pixel-level binary classification model, integrating three commonly used vegetation indices—EVI, NDVI, and MSAVI—to systematically evaluate the spatiotemporal dynamics of FVC in Xinjiang’s desert regions from 2016 to 2025. A comparative analysis of their inversion accuracies is also conducted. The results are as follows: (1) Leveraging its multiband atmospheric and soil correction mechanisms, the EVI significantly outperformed the NDVI and MSAVI in terms of vegetation–soil discrimination capability (endpoint difference of 0.185) and inversion accuracy (R 2 = 0.79), establishing itself as the optimal index for estimating FVC in Xinjiang’s desert regions. In contrast, the NDVI and MSAVI systematically overestimated FVC in areas with low vegetation density because of their sensitivity to the soil background. (2) The Junggar Basin exhibited high FVC (>50%) with high temporal variability (CV) and a predominant decreasing trend, whereas the Tarim Basin showed low FVC with low variability and a increasing trend. This study not only validates the superiority of the EVI in monitoring vegetation in arid desert regions but also maps the spatiotemporal evolution trends of vegetation coverage in Xinjiang’s desert areas, providing scientific evidence and theoretical support for regional ecological conservation and restoration practices.

Desert regions cover approximately 41% of the global land area. They not only play crucial roles in maintaining regional ecological stability, soil and water conservation, carbon cycling, and climate regulation but also serve as vital habitats for numerous rare and endemic species (Huang et al., 2020; Maestre et al., 2016). China’s Xinjiang region encompasses 440,600 square kilometers of desert, comprising ten major desert systems (Li et al., 2014; Yao et al., 2022). Among these systems, the Taklamakan Desert (China’s largest and the world’s second-largest shifting desert) and the Gurbantunggut Desert (China’s second-largest desert, dominated by fixed and semifixed dunes) are the most representative (Yang et al., 2011; Zhang et al., 2022). Their vast desert landscapes pose severe challenges to regional ecological security and human survival. Fractional vegetation cover (FVC) serves as a core indicator of desert ecosystem structure and function, profoundly influencing key ecological processes such as soil wind erosion resistance, water retention efficiency, carbon sequestration potential, and biodiversity maintenance (Mu et al., 2021; Shi et al., 2025). The spatiotemporal dynamics of the FVC serve not only as sensitive indicators of ecosystem stability and resilience but also as critical determinants of regional windbreak and sand fixation efficacy and desertification progression (Liu et al., 2018; Sun et al., 2024). Accurately characterizing its spatial patterns and evolution trends is vital for assessing ecological health, warning of land degradation risks, and supporting the construction of national ecological security barriers (Han et al., 2025; Xie et al., 2022). However, in desert regions characterized by sparse vegetation, extremely low biomass, and highly heterogeneous surface reflectance, traditional remote sensing inversion methods face dual challenges: While strong bare soil background signals significantly interfere with vegetation information extraction, plants commonly exhibit leaf degradation and simplified canopy structures to adapt to extreme aridity, leading to significant saturation and systematic underestimation of vegetation indices in low-cover areas (Smith et al., 2019). Therefore, the development of high-precision remote sensing inversion methods for vegetation cover in desert regions remains a critical scientific issue that urgently requires resolution.

The pixel binary model, a physically meaningful and computationally straightforward remote sensing estimation method, assumes that pixel spectral information is linearly blended from vegetation and bare soil components. It is widely applied in vegetation cover estimation (Yang et al., 2025; Zhang et al., 2025). The core of this model lies in determining the vegetation index values for pure vegetation and pure soil endpoints. Early studies predominantly relied on fixed empirical values or image histograms to select endpoints, introducing significant uncertainty. Recently, methods for dynamically determining endpoint values on the basis of image statistical quantiles (e.g., 5% and 95% quantiles) have been widely adopted because of their adaptability to spatial and temporal variations (Carlson and Ripley, 1997; Guo et al., 2024). However, their applicability in desert regions requires systematic validation.

The selection of vegetation indices represents another critical step in the application of pixel-level binary classification models. As a classic vegetation index, the NDVI is susceptible to soil background interference and prone to saturation in areas with low vegetation cover (Kwan et al., 2020). To overcome these limitations, a series of improved indices has been proposed. For instance, MSAVI incorporates a soil adjustment factor to mitigate background effects, making it more suitable for sparsely vegetated regions (Bai et al., 2022); EVI effectively mitigates NDVI saturation in high-vegetation-cover conditions by incorporating blue-band correction for aerosol scattering alongside soil and atmospheric background correction terms, thereby demonstrating superior sensitivity and dynamic range in high-biomass regions (Phompila et al., 2015; Wallace et al., 2008). However, in the unique environment of the Xinjiang’s desert regions—which are characterized by low coverage and high heterogeneity—the performance of these vegetation indices in terms of estimating vegetation cover still requires systematic comparison and validation on the basis of field observation data.

Currently, studies on vegetation cover in Xinjiang’s desert regions focus primarily on macroscale analyses and lack high-precision, long-term vegetation cover monitoring data. Furthermore, the optimal vegetation index selection remains unclear. Therefore, this study uses Landsat-8 imagery to compute the NDVI, MSAVI, and EVI through the Google Earth Engine platform, which is combined with a pixel-level binary model to estimate the FVC. The objectives are to (1) identify the most suitable vegetation index for estimating the FVC in desert regions; (2) reveal the spatial patterns of the FVC in Xinjiang’s desert areas; and (3) analyze the dynamic changes in the FVC over the past decade. These findings provide precise data support for regional ecological restoration and desertification control, facilitating the transition from extensive to precision-based management approaches.

Study area

The Xinjiang Uygur Autonomous Region (73°40′–96°18′E, 34°25′–48°10′N) lies at the heart of the Eurasian continent and is situated within China’s arid–semiarid ecological transition zone in the northwest, characterized by a distinct geomorphological pattern of “three mountain ranges enclosing two basins”. Deserts cover approximately 44% of the region’s total land area, accounting for nearly half of China’s total desert area. The region hosts ten major deserts, including the Taklamakan Desert, Gurbantunggut Desert, Kumtag Desert, and Huocheng Desert (Yu et al., 2022). Among these deserts, the Taklamakan Desert (China’s largest and the world’s second-largest shifting desert) and the Gurbantunggut Desert (China’s second-largest desert, characterized by fixed and semifixed dunes) are the most typical and representative (Amar et al., 2023; Sun and Gao, 2022). Ecologically, the vegetation in this region is sparse and highly heterogeneous, primarily consisting of drought-resistant shrubs (e.g., Haloxylon ammodendron, Tamarix spp.) and ephemeral herbs, which are predominantly distributed along riverbanks and piedmont alluvial fans. This complex spatial pattern, where high-reflectance backgrounds (such as shifting sands and Gobi) are interspersed with low-coverage vegetation, poses significant challenges for remote sensing retrieval accuracy (Guo et al., 2024; Han et al., 2025). Consequently, This study targets the dominant landform types across Xinjiang’s desert regions—namely shifting sand dunes, semishifting/semifixed sand dunes, fixed sand dunes, Gobi terrain, and saline‒alkali land—and presents a systematic analysis of spring FVC dynamics, as illustrated in Figure 1. The study area has a typical temperate continental hyperarid climate, with annual precipitation ranging from 15 to 150 mm, while potential evapotranspiration generally exceeds 2000 mm. Annual and daily temperature variations are pronounced, with summer highs exceeding 40 °C and winter lows plunging below −30 °C. Surface water resources are severely scarce and primarily rely on seasonal meltwater from surrounding mountain ranges such as Tianshan and Kunlun. The coupled dynamics of water, heat, and vegetation are highly sensitive, rendering ecosystems exceptionally fragile. The desert distribution data for Xinjiang were obtained from the National Cryosphere Desert Data Center (http://www.ncdc.ac.cn).

Vegetation index calculation

Data sources and preprocessing

In this study, Landsat 8 OLI surface reflectance data provided by the United States Geological Survey (USGS) are used, covering the time span from 2016 to 2025. To minimize the effects of phenological variations and specifically focus on the peak growth period of ephemeral plants following spring snowmelt, we selected all available imagery from May to June each year (Fan et al., 2022). All data processing was conducted within the Google Earth Engine (GEE) platform. This processing included applying masks for clouds and cloud shadows using the image’s native pixel_qa band to ensure the reliability of the pixel quality.

Vegetation index calculation

Within the GEE platform, the annual image collection from May to June is first subjected to median composite processing to further mitigate residual cloud contamination, atmospheric interference, and noise effects, thereby generating representative annual imagery. Three commonly used vegetation indices are subsequently calculated: Normalized difference vegetation index (NDVI)

2. Improved soil regulation vegetation index (MSAVI)

3. Enhanced vegetation index (EVI)

Among these data, NIR, red, and blue represent the atmospheric-corrected reflectance of the Earth’s surface in the near-infrared band (∼800–900 nm), red light band (∼630–690 nm), and blue light band (∼450–520 nm), respectively.

Pixel binary model

Calculation principle

The pixel binary model assumes that the spectral signal of a remote sensing pixel is a linear mixture of vegetation and soil components (Carlson and Ripley, 1997). The formula for calculating the FVC is as follows:

Among these data, the VI represents the vegetation index value (NDVI, MSAVI, or EVI) of the current pixel. VIsoil and VIveg denote the vegetation index values of pure bare soil and pure vegetation endpoints, respectively. The FVC ranges from 0 to 1.

Method for determining quantiles

On the GEE platform, statistical percentiles were employed to determine endpoint values: The 5th percentile of the annual composite vegetation index imagery was assigned to VIsoil (representing nearly bare soil with minimal vegetation), while the 95th percentile was assigned to VIveg (representing the most densely vegetated areas within the study region) (Gitelson et al., 2002). This approach accommodates the sparse vegetation and high spatial heterogeneity characteristics of Xinjiang’s desert regions.

Vegetation cover calculation process

On the basis of the pixel-based binary model, the FVC values corresponding to the NDVI, MSAVI, and EVI were calculated using annually estimated terminal values for vegetation and bare soil. To eliminate outliers (e.g., <0 or >1) caused by terminal value estimation errors, the FVC results were constrained within a reasonable range [0, 1]. This process ultimately generated a 10-year FVC dataset (2016–2025) based on three vegetation indices. Jenks’ natural breaks classification was then applied to categorize FVC into five tiers: 0%–14%, 14%–30%, 30%–52%, 52%–76%, and 76%–100%, reflecting the gradient variations in vegetation cover.

Ground-based measured coverage

In this study, a total of 86 ground survey plots were established. These plots were selected to cover low, medium, and high vegetation cover gradients across the study area to ensure sample representativeness and avoid systematic biases in the validation results. During field surveys, 30 m × 30 m sample plots were established. Drones captured vertical photographs of the plots at a fixed altitude. The raw images were subsequently imported into the ACGISmap platform. The images were cropped to precisely retain only the areas within the sample plots. A supervised classification method was employed. Typical vegetation and soil background samples were selected as training areas to construct a classification model, precisely dividing the image pixels into two categories: vegetation and nonvegetation (soil background). Next, using the software’s spatial statistics function, the percentage of pixels in the vegetation category relative to the total number of pixels within the plot was calculated. Finally, the precise FVC value for the plot was calculated using the following formula:

FVC (%) = (number of vegetation pixels/total number of pixels) × 100%.

Model accuracy validation

To evaluate the accuracy of fractional vegetation cover (FVC) retrieval based on three vegetation indices (NDVI, EVI, and MSAVI), field-measured FVC data from the same period in 2025 were used as reference values for validation. Field-measured sample point coordinates were spatially matched with the 2025 FVC imagery generated by pixel-by-pixel models driven by each vegetation index. The corresponding pixel-level inversion values were extracted, and the inversion performance of each model was quantitatively evaluated using the coefficient of determination (R 2) and root mean square error (RMSE). The calculation formulas are as follows:

Method for analyzing trends in vegetation cover changes

The Theil-Sen Median and Mann-Kendall methods (Sen, 1968; Mann, 1945; Kendall, 1990) were used in combination to calculate the spatiotemporal trends of the FVC in the study area from 2016 to 2025 (Mann, 1945; Sen, 1968). Based on the slope values and the standard normal variate (Z), the FVC trends were categorized into five levels: extremely significant decrease (slope < 0, |Z| ≥ 2.58), significant decrease (slope < 0, 1.96 ≤ |Z| < 2.58), no significant trend (−1.96 < |Z| < 1.96), significant increase (slope > 0, 1.96 ≤ |Z| < 2.58), and extremely significant increase (slope > 0, |Z| ≥ 2.58). Additionally, to evaluate the temporal stability of the FVC, the coefficient of variation (CV) was calculated as the ratio of the standard deviation (σ) to the multi-year mean value (μ) of the FVC time series over the study period. All these processes were implemented using the R programming language (version 4.5.2), specifically utilizing the trend package for Theil-Sen slope estimation and Mann-Kendall trend testing (Gutierrez-Hernández and García, 2024), and custom functions for calculating the coefficient of variation. The research flowchart of the paper is depicted in Figure 2.

Spatial patterns of vegetation indices

The spatial patterns of the three vegetation indices (NDVI, MSAVI, and EVI) are relatively consistent, as shown in Figure 3. High-value areas are concentrated primarily in the alluvial fan zones at the foothills of the Altai Mountains, Tianshan Mountains, and Kunlun Mountains. In contrast, the interior of the Tarim Basin has extensive low-value areas, indicating extremely sparse or completely absent vegetation. Considering the numerical ranges and classification proportions of the three indices, the EVI has the widest dynamic range (−0.872 to 0.796). Its high-value zones feature the highest color saturation, making it the most sensitive in capturing vegetation information. MSAVI has a low-value zone covering 53.1% of the area, reflecting its stricter threshold for distinguishing vegetation from soil. This strictness may cause sparse vegetation to be classified as soil background, leading to the underestimation of vegetation areas. The NDVI distribution falls between those of the other two indices. Overall, all three indices effectively reflect the spatial heterogeneity of vegetation in Xinjiang’s desert regions, but the EVI demonstrates superior performance in distinguishing vegetation abundance.

Spatial patterns of vegetation cover

As shown in Figure 4, the spatial distribution of the FVC in the desert region of Xinjiang, which is derived from pixel-based binary models using the NDVI, MSAVI, and EVI, significantly differs among the three vegetation indices in terms of its detailed representation and quantitative accuracy. These differences primarily stem from the varying sensitivities of these indices to background noise, particularly soil reflectance. In the EVI-derived results, the highest canopy cover class (76%–95%) accounted for 6% of the total area, which was lower than that of the NDVI (7%) and MSAVI (10%). However, in terms of identifying low-cover areas (0%–14%), the EVI captured 52% of the low-cover zones, which was significantly greater than the values obtained by the NDVI (43%) and MSAVI (20%). These findings indicate that the NDVI and MSAVI are insensitive to bare soil backgrounds and are prone to confusion between low vegetation cover and bare soil, thereby leading to the overestimation of low-cover areas. Endpoint analysis further validated this conclusion: The vegetation endpoint of the EVI was 0.226, and the soil endpoint was 0.041, with a difference of 0.185—the greatest difference among the three indices. This finding indicates that the EVI most clearly distinguishes vegetation from soil, providing optimal input parameters for the pixel binary classification model. In contrast, the vegetation threshold (NDVI, 0.199; soil threshold, 0.035; difference, 0.164) and vegetation threshold (MSAVI, 0.144; soil threshold, 0.020; difference, 0.124) exhibit smaller vegetation–soil threshold differences. This attribute makes them more susceptible to soil interference during low-coverage identification, leading to reduced accuracy.

Model accuracy validation

The linear fitting relationships between the FVC estimated by the three vegetation indices (EVI, NDVI, and MSAVI) and the field observations are shown in Figure 5. The results indicate that all three indices effectively reflect actual vegetation cover, with R 2 values of 0.79 (EVI), 0.74 (NDVI), and 0.76 (MSAVI), demonstrating good model explanatory power. Among these indices, the EVI performed best, with a slope (0.82) closest to 1 and the smallest intercept (3.82), indicating relatively minor estimation biases in high- and low-cover areas. The NDVI had the lowest slope (0.78) and a higher intercept (5.54), suggesting a systematic tendency to overestimate the FVC in low-cover regions. The performance of the MSAVI falls between the two, with a slope of 0.79 and an intercept of 7.01, revealing a certain overestimation trend in low-cover areas. Overall, the EVI is more accurate and stable for estimating the FVC, making it the preferred index for remote sensing-based FVC inversion in this region.

Trends in vegetation cover dynamics

The dynamic changes in FVC across Xinjiang’s desert regions from 2016 to 2025 are shown in Figure 6a. Specifically, 41.3% of the total study area showed a significant increase and 6.5% showed an extremely significant increase, primarily concentrated in the Tarim Basin. In contrast, 33.3% of the total area exhibited a significant decrease and 3.9% showed an extremely significant decrease, mainly distributed in the Junggar Basin. The remaining 15.1% of the area showed no significant trend. Overall, the FVC in Xinjiang’s desert regions exhibited a dynamic pattern characterized by predominant increases in the Tarim Basin and decreases in the Junggar Basin. In terms of temporal stability, the Coefficient of Variation (CV) of FVC ranged from 0.40 to 1.43 (Figure 6b, right panel). Areas with high CV values (0.82–1.43), indicating high interannual fluctuation and ecological vulnerability, were primarily scattered in the oasis-desert ecotones and regions experiencing significant FVC trends. Conversely, regions with low CV values (0.40 – 0.66) exhibited relatively stable vegetation dynamics, predominantly located in the stable oasis interiors or deep desert hinterlands.

Mechanism underlying the superiority of the EVI in desert area cover estimation

In this study, the performances of three vegetation indices—EVI, NDVI, and MSAVI—in the desert regions of Xinjiang were compared. The results indicate that the EVI achieves optimal estimation accuracy (R 2 = 0.79) for FVC inversion while demonstrating superior performance in distinguishing vegetation from soil endpoints, with an endpoint difference of 0.185. This advantage stems primarily from its unique multispectral design: The EVI incorporates a blue light band (approximately 450–480 nm) and employs empirical parameters (C 1 = 6, C 2 = 7.5) alongside a soil adjustment coefficient (L = 1) to effectively suppress aerosol scattering and soil background brightness (Fang et al., 2024; Matsushita et al., 2007; Obata et al., 2016). This approach enhances the distinguishability between vegetation signals and nonvegetation backgrounds. Specifically, the difference between the vegetation endpoint and soil endpoint is significantly greater for the EVI (0.185) than for the NDVI (0.164) and MSAVI (0.124), which results in the improved identification accuracy of the EVI in low-coverage areas (0%–14%) (Yan K. et al., 2022). In contrast, the NDVI lacks atmospheric correction terms, resulting in low slopes (0.78) and high intercepts (5.54) because of soil background interference in low-coverage areas (Li et al., 2012; Yan J. J. et al., 2022). Although the MSAVI reduces soil noise through its dynamic soil adjustment parameter (L), its nonlinear enhancement algorithm tends to misinterpret bare soil reflections as vegetation signals, causing systematic overestimation in low-coverage areas (20% coverage) and resulting in inflation in high-coverage areas (76%–95%) because of overcompensation for soil noise (Gilabert et al., 2002; Han et al., 2025).

The findings of this study differ from some previous results. For instance, Guo and Wen (2020)demonstrated through desert region research that the MSAVI is the optimal vegetation index for desertification inversion. In a recent multi-scale desertification assessment conducted in the Eastern Mediterranean, Alghababsheh (2024) utilized MSAVI precisely because of its known robustness in sparsely vegetated areas, which characterise the study region. This divergence stems primarily from regional environmental specificity: Xinjiang’s desert regions exhibit extreme vegetation cover gradients (from sandy areas to oases) and elevated atmospheric aerosol concentrations, making EVI’s multiband correction mechanism more suitable for this complex environment (Reynolds et al., 2007; Smith et al., 2019). Furthermore, the wide dynamic range of the EVI (−0.872 to 0.796) renders it highly sensitive to vegetation gradient changes, as it exhibits a near-linear response (R 2 = 0.79) in sparsely vegetated areas (Gao et al., 2020). Conversely, MSAVI suffers from signal distortion in low-coverage zones because of overcorrection. The quantification of terminal values further confirms the superior ability of the EVI to distinguish vegetation from soil (terminal difference of 0.185). In summary, the EVI provides a reliable basis for endpoint separation in vegetation monitoring in arid desert regions through its blue-band atmospheric correction and soil noise suppression mechanisms, making it the preferred choice for vegetation index applications in these areas.

Spatiotemporal dynamics of vegetation cover

In this study, a pixel-based binary model is employed to systematically analyze the spatiotemporal dynamics of the FVC in Xinjiang’s desert regions from 2016 to 2025. The results indicate that the Junggar Basin has an overall higher FVC than the Tarim Basin does, with vegetation coverage ranging between 30% and 95%. However, over the past decade, the FVC has tended to decrease overall, with a significant proportion of areas experiencing reduced coverage. These findings align with the trends reported by Miao et al. (2022)based on data from 2000 to 2021. Two primary factors contributed to this decline: First, intensified land reclamation activities led to vegetation degradation in some areas (Jiang et al., 2019; Li et al., 2024); second, a rebound in grazing pressure in localized regions further suppressed vegetation recovery (Amuti and Luo, 2014; Yuan and Han, 2026). In contrast, in the Tarim Basin, the FVC was generally low and virtually no vegetation was distributed in its core areas (Figure 3); however, the FVC has recently shown a relatively pronounced upward trend. These results align with the conclusions of Wang et al. (2023) and Jiao et al. (2022), further validating the positive role of the Tarim River ecological water diversion project in promoting vegetation recovery across the Tarim Basin, particularly in the oasis–desert transition zone.

Limitations and future directions

Despite the robust validation provided by UAV field data and the comprehensive assessment of vegetation indices, this study has several limitations that warrant further investigation. First, the spatiotemporal dynamic analysis was primarily based on trend detection without a comprehensive quantitative attribution analysis. Although land reclamation, grazing pressure, and ecological water diversion were qualitatively discussed as primary drivers, the exact relative contributions of climatic factors (e.g., precipitation, temperature) and anthropogenic activities to the FVC variations were not quantitatively decoupled (Lin et al., 2025). Second, the reliance on Landsat 8 imagery (30 m resolution) may lead to mixed pixel effects in highly heterogeneous oasis–desert transition zones, potentially affecting the estimation accuracy of sparse vegetation and fine-scale landscape fragmentation (Henriques et al., 2024; Song and Zhang, 2025).

To address these limitations, future research should focus on the following directions. First, advanced quantitative models, such as the Geodetector, structural equation modeling (SEM), or machine learning algorithms, should be introduced to precisely separate and quantify the impacts of climate change and human activities on vegetation dynamics (Cheng et al., 2024; Wang et al., 2025). Second, integrating predictive models, such as the Hurst index or Cellular Automata-Markov (CA-Markov) models, will be crucial for forecasting future FVC trends and assessing ecological security under different climate and management scenarios (Huang et al., 2024; Luo et al., 2024). Finally, building upon the UAV field data utilized in this study, future work could incorporate UAV-LiDAR or multi-angle hyperspectral imagery. This would facilitate a transition from 2D coverage estimation to 3D vegetation structural parameter extraction (e.g., canopy height and aboveground biomass), thereby providing deeper insights into the ecological functions and carbon sequestration potential of arid desert ecosystems (Ferreira et al., 2024; Yang, 2025).

This study systematically reveals the spatiotemporal dynamics of the forest vitality coefficient (FVC) in the desert region of Xinjiang from 2016 to 2025. It employs a pixel-based binary model and compares the performance of three vegetation indices—EVI, NDVI, and MSAVI—in this area. The key findings are as follows: The EVI demonstrated optimal performance in FVC retrieval for desert areas (R 2 = 0.79). Its multispectral structure effectively mitigated atmospheric and soil interference, with a vegetation–soil terminal difference (0.185) significantly greater than that of the NDVI and MSAVI. The EVI exhibited superior accuracy in identifying low-coverage areas, making it the optimal index for FVC estimation in this region.

From 2016 to 2025, the Junggar Basin exhibited an overall high FVC (30%–95%) with a decreasing trend and high temporal instability (high CV), reflecting ecological vulnerability, whereas the Tarim Basin demonstrated a low FVC with a pronounced upward trend and low temporal variability (low CV), indicating relatively stable vegetation dynamics.

This study provides methodological support and a scientific basis for vegetation remote sensing monitoring and ecological restoration assessment in arid desert regions.