Fractional vegetation cover (FVC) quantitatively characterizes the horizontal density of photosynthetically active vegetation (Gutman and Ignatov, 1997). It is typically defined as the planar proportion of green vegetation to the total surface extent (Deardorff, 1978). The FVC is an essential parameter in climate and hydrologic models as it represents the spatial contribution of vegetation (Hirano et al., 2004; Gutman and Ignatov, 1998; Eriksson et al., 2006; Mölders and Olson, 2004). Accurate and high-resolution FVC products are in high demand for various studies, including climate change analysis, soil erosion assessment, land disturbance evaluation, and crop growth monitoring (Xie et al., 2011; Naqvi et al., 2013; Gan et al., 2014; Li et al., 2014; Zhang et al., 2013; Fernández-Guisuraga et al., 2021).

Remote sensing can rapidly and repeatedly observe the land surface, making estimating FVC on regional or global scales feasible. Over recent decades, tremendous efforts have been made to derive high-quality FVC from remotely sensed imagery. The published approaches for retrieving FVC can generally be summarized as follows: (i) the vegetation index (VI)-based mixture model (Gutman and Ignatov, 1998; Zeng et al., 2000; Wu et al., 2014; Mu et al., 2021; Zhao et al., 2023); (ii) spectral mixture analysis (García-Haro et al., 2005; Dimiceli et al., 2011; Guan et al., 2012); (iii) machine learning (Baret et al., 2007; Baret et al., 2013; Jia et al., 2015); and (iv) physical model (Xiao et al., 2016).

The linear mixture model is the most commonly used spectral unmixing method. It is generally utilized in surface elements evaluation such as vegetation classification, surface disturbance mapping, and evapotranspiration estimation (Li et al., 2018a; Lu et al., 2003; Cochrane and Souza, 1998). When considering only two endmembers (green vegetation and bare soil), linear mixture modelling can be employed to calculate the relative abundance of live vegetation from the mixed VI, known as the VI-based mixture model. It assumes that the VI for a particular pixel originates from a linearly weighted sum of green vegetation and bare soil, with their respective areal proportions as weighting coefficients (Gitelson et al., 2002). The mixed VI (V) of the pixel is linearly decomposed by the two endmembers, i.e., the VI of the fully vegetated (Vv) and bare soil pixel (Vs), to obtain the areal proportion of green vegetation as FVC: 1FVC=V-VsVv-Vs,

Despite being the most commonly used method for deriving FVC (Gao et al., 2020), the VI-based mixture model still requires enhancements in both accuracy and efficiency. A major limitation is the challenge of obtaining accurate values for Vv and Vs on a large scale. The “greenness” VIs that exhibit a strong linear relationship with FVC can be used in the VI-based mixture model, such as the enhanced vegetation index (EVI) and the normalized difference vegetation index (NDVI) (Song et al., 2022a). Among them, the NDVI is the primary VI used to derive FVC due to its strong correlation with vegetation structural parameters (Gutman and Ignatov, 1998). The two endmember NDVI values (hereafter referred to as pure NDVI values) are often assigned a priori, and the main methods for extracting these values, along with other pure VIs are summarized in Table .

Traditional methods for determining pure NDVI values have significant limitations. Collecting Vv and Vs from ground truth data or high-resolution remotely sensed imagery by extracting pure pixel values is time-consuming and often limited by data availability. These methods become ineffective when data are unavailable or there are no pure pixels in the usable dataset. The statistical method infers the two pure NDVI values from the data within a specific area or time series. It typically assumes that pixels with the lowest NDVI values represent bare soil, while those with the highest values represent pure vegetation (Gao et al., 2020). However, in arid and semiarid regions with few fully vegetated pixels, or evergreen forests with limited bare soil pixels, the statistically extracted endmember values can be significantly inaccurate (Song et al., 2017).

Additionally, statistical methods typically assign a single value of Vv or Vs for a specific region or land type. However, endmember values can vary significantly from pixel to pixel due to differences in species composition, vegetation health, moisture levels, and other factors (Jensen, 2000). In many studies, a fixed value of Vs is adopted for various soil types, as shown in Table , which overlooks the spatial variations in soil moisture, texture, mineralogy, organic matter, and other characteristics (Yang and Yang, 2006; Zeng et al., 2000). The NDVI values of soil samples exhibit considerable variation, ranging from 0 to 0.4, with a mean value of 0.2, which is significantly higher than the commonly used Vs value (of approximately 0.05) (Montandon and Small, 2008). Notably, the accuracy of FVC estimation is highly sensitive to variations of Vs, particularly in sparse-vegetated areas (Asrar et al., 1984; Montandon and Small, 2008). Underestimating Vs may lead to an overestimation of FVC, with errors reaching up to 20 % in grassland and shrubland regions (Montandon and Small, 2008; Ding et al., 2016). It has been demonstrated that locally derived pure NDVI values provide higher accuracy for FVC estimation, compared to using a fixed global Vs value (Donohue et al., 2014; Montandon and Small, 2008). Therefore, pixel-wise maps of endmember NDVI are essential for effectively addressing the spatial variability of plant and soil reflectance. However, such products are currently unavailable.

Recent studies have proposed an alternative method that uses multi-angle datasets from the Moderate Resolution Imaging Spectroradiometer (MODIS) to derive pixel-wise Vv and Vs (MultiVI algorithm) (Mu et al., 2018; Song et al., 2022a). The discrepancies in directional NDVI values observed from multiple viewing geometries imply vegetation structural information and soil characteristics (Chen et al., 2005; Diner et al., 1999; Deering, 1999; Verrelst et al., 2008). The MultiVI algorithm utilizes the variations from two large viewing angles to establish equations for simultaneously retrieving Vv and Vs, without assuming invariant endmember values within a scene or biome. Feasibility analysis indicated its potential to estimate high-quality, high-resolution FVC products (Song et al., 2022b). The Vv and Vs derived from the MultiVI algorithm have been used to generate 30 m/15 d FVC products, which are reported to possess satisfactory accuracy (Zhao et al., 2023). However, the accuracy of the Vv and Vs maps still requires evaluation, and strategies to optimize the MultiVI algorithm for large-scale Vv and Vs mapping are essential. Providing high-quality pure NDVI values for the VI-based mixture model will significantly improve the accuracy and efficiency of FVC estimation.

This study aims to generate and validate 30 m resolution pixel-wise Vv and Vs maps across China. These datasets can be applied to accurately calculate FVC at various resolutions on regional or national scales. The Vv and Vs maps were derived from MODIS reflectance data using the MultiVI algorithm. Subsequently, the 500 m MODIS Vv and Vs were downscaled to 30 m resolution based on land cover types (hereafter referred to as MultiVI Vv and Vs). Traditional statistical methods were employed to extract Vv and Vs from Landsat data (hereafter referred to as statistical Vv and Vs). The two sets of pure NDVI values were then compared and analysed. The generated Vs maps were validated using soil NDVI values calculated from a soil spectral library. Finally, the FVC values derived from the MultiVI algorithm and the statistical method were validated against field-measured FVC obtained from various experimental sites across China.

The VI-based mixture model is widely used to estimate FVC due to its ease of implementation. The model's simplicity is primarily attributed to the preselection of two critical parameters: the NDVI of bare soil (Vs) and that of pure vegetation (Vv). These parameters are essential for the model's performance and have a significant impact on its accuracy. The traditional statistical method for obtaining Vv and Vs assumes the existence of pure pixels, either spatially or temporally. This method employed extreme values from regional or temporal datasets to represent Vv and Vs. However, this approach has two significant limitations: (1) pure pixels may be absent in certain ecosystems, such as evergreen forests, where bare soil pixels are lacking, or in bare lands, where pure vegetation pixels are absent; (2) the Vv and Vs can vary from pixel to pixel, yet the traditional statistical method often assigns a single value for an entire region or land type.

The global soil spectral library was used to calibrate Vs and to account for its spatial variability (Montandon and Small, 2008). Although this method improves the accuracy of FVC estimation compared to using a single Vs value across large regions, it is limited in that it only considers soil variability within the spectral library and fails to address the spectral differences among individual pixels. An alternative approach is to use each pixel's historical lowest NDVI values as Vs to ensure pixel-wise variability. In practice, Vs is influenced not only by mineral soil properties but also by non-photosynthetic vegetation (NPV), biological residues, and surface litter. A previous study reported that the NDVI of NPV endmembers is generally higher than that of bare soil, with differences reaching up to 0.4 in some cases (Tian et al., 2021). This suggests that Vs may vary even within the same soil type. Land cover classification can partially account for such heterogeneity, and land cover data were utilized as a practical proxy to disaggregate 500 m Vs to a 30 m resolution in this study.

Figure  further illustrates that the retrieved Vs values deviate from the soil NDVI in humid regions. The most significant discrepancy is observed in the soil types of Anthrosols and Ferralisols, which are prevalent in the humid areas of southern China. This discrepancy is likely attributed to their relatively high NDVI values, and the presence of surface litter and biological residues. However, the overestimation of MultiVI Vs in these areas has a limited impact on the FVC estimation (Fig. a), with the resulting FVC bias generally less than 0.1. Previous studies have also indicated that the influence of Vs on FVC estimation is more pronounced in areas with low NDVI values, such as grasslands and croplands/natural vegetation areas, compared to regions with high NDVI values (Ding et al., 2016).

In this study, statistical maps of Vv and Vs were generated using a common statistical criterion (Sect. ). Empirical NDVI for fully vegetated pixels is typically reported to exceed 0.5 in most studies (Gao et al., 2020; Zeng et al., 2000; Montandon and Small, 2008), whereas the statistical Vv are observed to be below 0.3 in arid regions of China (Fig. a). Except for these extreme areas, the statistical method demonstrates comparable accuracy to the MultiVI algorithm (Figs. , , and c, d). The simple statistical methods face challenges in identifying appropriate NDVI values when pure pixels are lacking. In contrast, several studies have adopted more sophisticated statistical approaches to produce accurate FVC estimates across various ecosystems (Donohue and Renzullo, 2025; Donohue et al., 2018). These enhanced statistical methods take into account the inherent characteristics of vegetation in different ecosystems, employing separate thresholds to determine endmembers in arid and vegetated areas, respectively. The use of spatially adaptive thresholding facilitates the acquisition of reasonable endmembers and achieves high FVC accuracy, with a RMSE of approximately 0.1 (Donohue and Renzullo, 2025).

The errors in the FVC estimated from the VI-based mixture model vary across seasons and are more pronounced during periods of low vegetation cover. Figure  shows that FVC errors are generally larger in early spring and winter, particularly for low FVC values. This seasonal pattern can be explained by the sensitivity of FVC estimation to NDVI values in the 0.2–0.4 range (Montandon and Small, 2008). Within this interval, small errors in Vs can lead to systematic overestimations of FVC, especially over grasslands and shrublands. During peak growing seasons, when NDVI values exceed 0.7, the model estimates become more stable and less sensitive to endmember NDVI values. Furthermore, the saturation of FVC estimates, resulting from an underestimation of Vv and an overestimation of Vs, typically influences winter and summer periods (Fig. f).

The MultiVI algorithm, which uses multi-angle data to retrieve Vv and Vs, has demonstrated effectiveness in estimating FVC and has been applied to generate FVC time series on a national scale (Song et al., 2022a; Zhao et al., 2023). This algorithm retrieves angle-invariant NDVI values for endmembers using observations from two large VZAs. In this study, the retrieval strategy of the MultiVI algorithm was optimized through the construction of well-posed equations based on NDVI time series. Additionally, the statistical values of Vv and Vs were incorporated as boundary constraints during the equation-solving process. The downscaling procedure integrates a 30 m land cover product, which introduces finer spatial detail to describe the sub-pixel heterogeneity within MODIS pixels. Our downscaling approach applies the assumption that the 30 m pixels of the same land cover type share constant Vv and Vs values within a localized 3 × 3 window of 500 m MODIS pixels. This assumption may introduce uncertainty when significant soil type variation exists within the 3 × 3 MODIS window (1.5 km × 1.5 km). Despite the simplifications involved, the comparison with the 500 m results shows that the downscaled 30 m Vs values achieve comparable accuracy (Song et al., 2022b). The downscaling process introduces minimal changes to the endmember values (Song et al., 2022b). Considering the increasing demand for high spatial and temporal resolution applications, providing 30 m endmember products is of practical significance.

As a critical vegetation parameter, FVC is frequently demanded in various studies, serving as an input for models or as fundamental data for ecological analyses. The VI-based mixture is one of the simplest methods for converting remotely sensed images into FVC products, provided that the pure NDVI values are obtained beforehand. However, the two essential parameters in the VI-based model, Vv and Vs, currently lack standardized and reliable data sources. The newly generated 30 m Vv and Vs maps address this gap. Users can calculate accurate FVC values flexibly using the MultiVI Vv and Vs with their NDVI to meet specific requirements.

The 30 m MultiVI Vv and Vs maps for 2014 are available at 10.5281/zenodo.15720622 (Zhao et al., 2024a). The Supplement produced the 30 m MultiVI Vv and Vs for 2018 and 2022 (available at 10.5281/zenodo.17463344, Zhao et al., 2024b) and evaluates their interannual stability. All Vv and Vs data are stored in GeoTIFF files using the WGS-84 (World Geodetic System 1984) coordinate system with UTM (Universal Transverse Mercator) projection. The data are categorized into tiles based on a latitude size of 5° and a longitude size of 6°. The file names consist of 18 characters following these rules: North-South latitude abbreviation (1 digit) + 6° zone number (2 digits) + “_” + starting latitude (2 digits) + “_” + product time (4 digits) + “_” + resolution (3 digits) + “_” + data attribute (2 digits). Additionally, each image contains a pure NDVI value band ranging from 0 to 100, with invalid values or water surfaces labelled as NaN. The 500 m MultiVI Vv and Vs maps are also publicly available at 10.5281/zenodo.15597968 (Zhao et al., 2024c).

This study demonstrates the feasibility of generating pixel-wise NDVI for fully vegetated area (Vv) and bare soil (Vs) for the VI-based mixture model. In this study, 30 m resolution maps of pure NDVI values for the year 2014 were produced for China using multi-angle remotely sensed data from MODIS and land cover datasets from Globeland30. The assessment and validation of the Vv and Vs maps were conducted from three aspects: (1) comparing the MultiVI Vv and Vs maps with those generated using the statistical method; (2) comparing the derived Vs with reference soil NDVI from the ICRAF soil spectral library; and (3) validating the accuracy of FVC calculated from the pure NDVI values against field-measured FVC. Our findings reveal the urgent need for reliable Vv and Vs per pixel for large-area FVC production. Traditional statistical methods are impractical to achieve this goal due to their reliance on pure pixels. The MultiVI algorithm has proven to be a viable solution, yielding Vv and Vs with a more reliable spatial pattern and magnitude than statistical methods. The MultiVI Vs is closely aligned with soil NDVI across various soil types in China. Furthermore, the FVC estimated using the MultiVI Vv and Vs demonstrated improved accuracy in comparison to those derived from the statistical Vv and Vs, with RMSD values around 0.1 and R2 values near 0.8 for all validation sites. Moreover, the products generated in this study show broad applicability across a variety of climate zones and soil types.

The MultiVI Vv and Vs maps provide essential parameters for FVC estimation using the widely adopted VI-based mixture model, which is known for its ease of use and reasonable accuracy. Hence, these derived Vv and Vs maps are anticipated to facilitate the estimation of fine-resolution, high-frequency FVC with reliable quality at large scales.