Integrating Sentinel-2 and MODIS BRDF Imagery to Invert Canopy Fractional Vegetation Cover for Forests and Analyze the Corresponding Spatio-Temporal Evolution

Abstract

Canopy fractional vegetation cover (FVCc) is a critical indicator for evaluating the effectiveness of ecological restoration, and its accurate estimation provides valuable data for regional ecological management. In this study, Sentinel-2 and MODIS data were integrated to develop an angular relationship model for MODIS reflectance, which was then used to estimate Sentinel-2 reflectance at a 45° viewing angle. Background reflectance at a 10 m spatial resolution was derived using a four-scale model, and total and shrub-herb fractional vegetation cover were estimated using a pixel dichotomy model. Finally, an empirical model tailored to the characteristics of the study area was developed to retrieve FVCc. Cross-validation results demonstrated that the multi-angle retrieval method proposed in this study achieved higher accuracy than the single-angle approach. The spatial distribution of FVCc in Changting County is characterized by higher values in peripheral areas and lower values in the central region. Temporal transitions among fractional vegetation cover classes were predominantly upward, indicating an overall trend of continuous improvement. These findings provide important technical support and a scientific basis for estimating and monitoring dynamic changes in forest canopy fractional vegetation cover.

1. Introduction

As a fundamental component of the Earth’s biosphere, forest ecosystems play an irreplaceable ecological role in regulating the global carbon cycle, maintaining biodiversity, and mitigating climate change [1,2]. Forest canopy fractional vegetation cover (FVCc) is a key parameter for characterizing forest health, productivity, and ecological quality. Accurate estimation and long-term spatio-temporal monitoring of FVCc are essential for revealing vegetation growth patterns and evaluating the effectiveness of ecological protection policies [3,4,5]. Traditional methods for estimating fractional vegetation cover have primarily focused on inverting overall forest cover based on integrated forest structure, without distinguishing the respective contributions of the canopy and the understory shrub-herb layer. Consequently, these methods are insufficient for the fine-scale quantification required to assess the effectiveness of ecological management strategies [6,7]. Precise inversion of the fractional vegetation cover of the upper forest canopy can effectively distinguish vertical structural variations among vegetation layers, accurately quantify the ecological contribution of the canopy, and support forest resource management as well as the evaluation of ecological restoration outcomes [8].

Driven by advancements in remote sensing, the estimation of fractional vegetation cover has transitioned from traditional ground-based sampling to sophisticated remote sensing inversion techniques. Remote sensing inversion methods can generally be categorized into two types: empirical models and physical models [9]. Empirical models, typically based on vegetation indices, estimate vegetation cover by establishing statistical relationships between remote sensing vegetation indices and vegetation structural parameters. Due to their computational simplicity and high efficiency, these methods have been widely employed for large-scale monitoring of forest vegetation cover dynamics. However, they are highly dependent on regional conditions and exhibit limited generalizability. Moreover, they face challenges in performing layered inversion of forest vegetation cover under complex conditions, which frequently compromises their accuracy [10,11,12]. Spectral mixture analysis, a key physical modeling approach grounded in radiative transfer theory, addresses the mixed-pixel effect by isolating the spectral signatures of various land-cover types. Nevertheless, its performance is often limited by the variability of forest endmember spectra and interference caused by shadows and anisotropic reflectance [13,14,15]. In recent years, machine learning methods have gained prominence in remote sensing inversion due to their strong nonlinear fitting capabilities. By integrating multidimensional features such as multispectral reflectance, texture, and topographic factors, these methods enhance inversion accuracy in complex regions. However, they are often constrained by high sample dependency and a limited capacity to mitigate understory interference [16,17,18].

Accurately estimating understory shrub-herb background reflectance and achieving layered inversion to separate shrub-herb fractional vegetation cover and FVCc remain significant challenges for high-resolution FVCc monitoring in complex forest environments. Existing methods typically perform layered inversion by accounting for endmember variability, coupling radiative transfer models, and optimizing spectral unmixing algorithms. Some studies have incorporated high-precision UAV hyperspectral and LiDAR data to enhance the discrimination of spatial and spectral characteristics of vertical vegetation structure, thereby improving separation accuracy. However, these approaches often depend on data sources with high acquisition and processing costs, limiting their applicability in large-scale, long-term inversion scenarios [19,20,21]. Conventional methods frequently overlook the angular anisotropy of reflectance and are affected by spectral confusion between forest canopies and understory shrub-herb layers. This results in significant biases in estimating understory background reflectance and ultimately restricts inversion reliability in complex regions [22,23]. Additionally, some image data used for vegetation parameter inversion lack multi-angular observations. Consequently, their inversion accuracy is frequently compromised by forest shadowing and topographic effects, hindering the effective differentiation of vegetation cover across distinct vertical layers [24].

Because surface reflectance varies with viewing angle, single-angle observations cannot accurately characterize land surface reflectance anisotropy, which tends to reduce the accuracy of fractional vegetation cover inversion [25,26]. Multi-angular remote sensing observation technology has been introduced to address this issue. By exploiting reflectance differences between the forest canopy and understory across different viewing angles, this approach mitigates the effects of surface reflectance anisotropy and shadowing on inversion accuracy. Consequently, it reduces understory interference and enhances the ability to resolve complex vertical forest structures [27,28]. Multi-angular observations can effectively mitigate spectral confusion between the forest canopy and the understory background, thereby facilitating the inversion of FVCc [29]. MODIS has been widely used for regional-scale inversion of fractional vegetation cover because of its long time series and multi-angular observations [30,31]. However, its relatively coarse spatial resolution leads to strong pixel heterogeneity and severe component mixing in complex forest regions, making it difficult to accurately separate canopy cover from shrub-herb cover [32]. Therefore, MODIS multi-angular data alone are insufficient to meet the requirements of high-precision layered inversion [33]. High-spatial-resolution imagery must be integrated so that multi-angular reflectance information can be combined with fine spatial details to further improve the accuracy and reliability of layered inversion of FVCc [34,35].

According to previous studies, current approaches for inverting FVCc face several challenges. Single-angle observations fail to characterize surface reflectance anisotropy and are highly susceptible to interference from forest shadows and understory backgrounds. Moreover, low-resolution multi-angle data are insufficient for capturing fine vertical forest structure. Achieving accurate layered inversion of forest canopy cover through multi-sensor synergy is the core scientific problem addressed in this study. We propose that MODIS multi-angle observations can effectively capture the anisotropic reflectance signatures that distinguish forest canopies from understory shrubs and grasses. Furthermore, integrating these observations with high-spatial-resolution data from Sentinel-2 will significantly enhance the accuracy and reliability of layered inversion. To test this hypothesis, Changting County, Fujian Province—a representative red-soil water-erosion region—was selected as the study area. By combining the advantages of MODIS multi-angle reflectance data and Sentinel-2 high-spatial-resolution imagery, this study accurately characterizes understory background reflectance while mitigating the effects of forest shadows and angular anisotropy. This approach facilitates high-precision inversion and robust spatiotemporal analysis of FVCc from 2016 to 2025. This work provides novel methodological insights and a scientific framework for evaluating the effectiveness of regional ecological restoration.

2. Materials and Methods

2.1. Study Area and Data Sources

Our study focused on Changting County in western Fujian Province, China (25°18′40″ N–26°02′05″ N, 115°59′48″ E–116°39′20″ E), covering a total area of approximately 3100 km². The county comprises 18 townships (Figure 1). The region experiences a subtropical monsoon climate characterized by abundant heat and rainfall. The mean annual temperature ranges from 15.0 to 19.5 °C, and the mean annual precipitation is approximately 1700 mm. The terrain is dominated by low mountains and hills, with elevations ranging from 238 to 1456 m. Elevations are higher in the east, west, and north, and lower in the central and southern parts, resulting in an overall slope from north to south. The soils are primarily mountain red soils developed from granite parent material, which are highly susceptible to erosion. Changting County is a national demonstration area for soil and water conservation and ecological restoration in the red-soil hilly region of southern China [31,32]. The zonal vegetation is characterized by subtropical evergreen broad-leaved forest, exhibiting a clearly stratified vertical structure comprising canopy, understory shrubs, and litter, providing an ideal setting for the accurate inversion of FVCc [36,37].

The data used in this study include field measurement data, remote sensing imagery, and basic geographic data. Field-measured FVCc data were collected on 20 April 2025. Standard forest plots were established in the red soil hilly region of Changting County, Fujian Province. Plot selection strictly adhered to explicit screening criteria that comprehensively considered key factors such as topography, stand characteristics, and species composition. Regarding topographic gradients, plot slopes ranged from 2° to 45°, with approximately 60% of the plots falling within the 15–25° slope range. Aspects were predominantly south-, west-, and northwest-facing slopes, typical of the red soil region. All plots were located at elevations between 200 and 400 m, representing the typical low-hill elevation range of the study area and ensuring accurate representation of local topographic characteristics. Stand density was primarily characterized by FVCc, with plot canopy cover ranging from 20% to 80%. Canopy height generally ranged from 6 to 12 m, and diameter at breast height (DBH) ranged from 8 to 15 cm, consistent with the typical forest community structure of the study area. Regarding stand age, all plots contained mature communities formed through long-term natural development and artificial restoration. In terms of species composition, the plots were dominated by native species typical of the red soil hilly region, including Pinus massoniana, Liquidambar formosana, and Schima superba. Data collection employed the five-point sampling method: five observation points were established at the four corners and the center of each 10 m × 10 m plot. Forest canopy photographs were taken upward using a digital camera with vertical photography; two repeated photographs were captured at each observation point to calculate the mean value. Corresponding geographic coordinates were recorded simultaneously during the measurement process.

To reduce the uncertainty associated with the photographic method, rigorous quality control and validation procedures were implemented during data collection. To ensure repeatability within each sample plot, all photographs were captured using equipment with identical specifications. Errors caused by instrumental and environmental factors were minimized by strictly regulating operational parameters such as shooting height and angle. For threshold sensitivity processing of the hue channel, the collected forest canopy images were preprocessed using Image J 1.54g. After separating the HSB (Hue–Saturation–Brightness) color channels and applying Gaussian blur denoising, the optimal threshold range for the vegetation hue channel of the canopy was determined through multiple preliminary tests. A fixed and uniform threshold was then applied to separate vegetation pixels from non-sky pixels, thereby avoiding the subjectivity associated with manual threshold selection. Additionally, manual inspection of the thresholded results was conducted to eliminate misclassified pixels caused by image noise, ensuring accurate vegetation pixel extraction. Using the procedures described above, the non-sky pixel ratio at each observation point was calculated to derive the canopy cover at that point. Finally, the average value from the five observation points within each sample plot was adopted as the field-measured FVCc for the corresponding plot.

The remote sensing imagery used in this study comprises Sentinel-2 Level-2A surface reflectance data at a 10 m spatial resolution and MODIS Bidirectional Reflectance Distribution Function (BRDF)/Albedo Model Parameter Product (MCD43A1) data at a 500 m spatial resolution, both spanning the period from 2016 to 2025. Sentinel-2 imagery was obtained from the Copernicus Data Space Ecosystem of the European Space Agency (ESA) (https://dataspace.copernicus.eu/ (accessed on 30 October 2025)), with the study area centered on Changting County, Fujian Province. For the 2016–2025 time series, two Sentinel-2 images were selected per year, resulting in a total of 20 images acquired and applied over the decade. The two images chosen each year were acquired during the same phenological period, ensured full coverage of the study area, and satisfied the quality criterion of cloud cover below 10%. Specifically, images acquired in April were selected annually to maintain phenological consistency, thereby minimizing the effects of seasonal variation on inversion results and ensuring comparability across inter-annual inversion datasets. The 2025 imagery was temporally synchronized with the field measurement campaign. This study employed a single-temporal image analysis approach without applying temporal compositing procedures; subsequent inversion was performed directly using the preprocessed imagery.

The MCD43A1 product was obtained from the Earth Observation Data Portal of the National Aeronautics and Space Administration (NASA) (https://www.earthdata.nasa.gov/ (accessed on 30 October 2025)). This product provides multi-angular, atmospherically corrected reflectance across spectral bands, kernel weight parameters of the BRDF model, and fitting coefficients of sky albedo. These features effectively eliminate the bidirectional angular effects of surface reflectance, generating spatio-temporally consistent isotropic reflectance and broadband albedo. Based on the imaging characteristics of MODIS data, this study used the Sentinel-2 acquisition date as a reference and synthesized MODIS albedo data within a 16-day temporal window, including 8 days before and 8 days after the Sentinel-2 overpass. During preprocessing of the MCD43A1 BRDF/Albedo Model Parameter product, cloud cover, cloud shadows, high aerosol contamination, and anomalous observations were masked and removed based on the embedded Quality Assurance (QA) layer. This preprocessing step ensured the spatiotemporal consistency and reliability of the time series data and effectively reduced errors caused by atmospheric conditions, viewing geometry, and other confounding factors. Additionally, the MODIS data were resampled to a 10 m spatial resolution using bilinear interpolation, ensuring that the spatial location and pixel size of the MODIS data matched those of the Sentinel-2 data.

The basic geographic data utilized land use information as a reference to mask non-forest areas, thereby focusing the analysis on the core research subject. Administrative division data, including county-level boundaries of Changting and township-level boundaries, were used for image cropping, field plot layout and positioning, as well as spatial pattern analysis.

2.2. FVCc Remote Sensing Inversion Method

2.2.1. Calculation of Background Reflectance

Based on the kernel weight parameters for each band from the MODIS BRDF data and the observation geometry provided in the product metadata, the viewing zenith angles were set to 0° (nadir observation) and 45° (oblique observation). The bidirectional reflectance factors at these specific angles were derived by combining the kernel weights with their corresponding kernel functions. Subsequently, pixel-level quality screening was conducted using the Mandatory Quality layer. Ultimately, the surface reflectance values for the red and near-infrared bands at the 0° and 45° viewing angles were obtained separately [38,39].

To overcome the limitations of single-angle observations from high-spatial-resolution Sentinel-2 imagery and to fully exploit the multi-angle observation capabilities of MODIS, the strengths of both sensors were integrated to perform FVCc inversion. The two sensors differ in their spectral response functions, central wavelengths, and bandwidths in the red and near-infrared bands. Therefore, to account for these spectral response differences between MODIS and Sentinel-2, spectral normalization was applied prior to transferring multi-angle information [40]. Since this study focuses on angular effects and canopy structure inversion, a linear transformation model was established using spatiotemporally synchronized and matched image pairs. The 0° viewing reflectance of Sentinel-2 was normalized to the MODIS-equivalent 0° reflectance, and the MODIS 45° equivalent reflectance was estimated from the angular relationship between the original MODIS 0° and 45° viewing reflectances to mitigate systematic biases caused by bandpass differences [41]. For the bandpass-corrected MODIS 0° and 45° reflectance data in the red and near-infrared bands, 302 sample points at a 10 m × 10 m pixel scale were selected in non-forest areas across the study region using spatially uniform sampling. An angular relationship model was then constructed and applied to the Sentinel-2 0° reflectance data for the corresponding year to obtain the 45° off-nadir viewing reflectance of Sentinel-2 for that year.

Based on the 0° and 45° reflectance data from Sentinel-2, the four-scale geometric optical model was employed to calculate the background reflectance at a spatial resolution of 10 m. It is assumed that the reflectance of a single pixel can be expressed as a linearly weighted sum of reflectance from four components: illuminated canopy, illuminated background, shaded canopy, and shaded background [42]:

$$R=P_T·R_T+P_G·R_G+P_{ZT}·R_{ZT}+P_{ZG}·R_{ZG}$$

(1)

In the formula, R represents the total pixel reflectance observed by the sensor. R_T, R_G, R_ZT and R_ZG denote the reflectance of the illuminated canopy, illuminated background, shaded canopy, and shaded background, respectively. P_T, P_G, P_ZT and P_ZG represent their respective proportions within the pixel under specific solar elevation and viewing angles. Due to insufficient illumination, reflectance in shaded regions is significantly lower than in sunlit regions. The reflectance of the shaded canopy and shaded background can be approximated as the product of the corresponding sunlit reflectance and a multiple-scattering factor, expressed as:

$$R_{ZT}=M·R_T$$

(2)

$$R_{ZG}=M·R_G$$

(3)

For the same pixel, the reflectance in the nadir direction (R_n) and the reflectance at an alternative viewing angle (R_a) can be expressed as:

$$R_n=P_{Tn}·R_{Tn}+P_{Gn}·R_{Gn}+P_{ZTn}·R_{ZTn}+P_{GTn}·R_{GTn}$$

(4)

$$R_a=P_{Ta}·R_{Ta}+P_{Ga}·R_{Ga}+P_{ZTa}·R_{ZTa}+P_{GTa}·R_{GTa}$$

(5)

From Equations (2) and (3), we obtain:

$$R_{ZTn} =M·R_{Tn}$$

(6)

$$R_{ZGn} =M·R_{Gn}$$

(7)

$$R_{ZTa}=M·R_{Ta}$$

(8)

$$R_{ZGa} =M·R_{Ga}$$

(9)

Substituting Equations (6)–(9) into Equations (4) and (5), respectively, yields the following simplified form:

$$R_G={\displaystyle \frac{R_n\left(P_{Ta}+P_{ZTa}\cdot M\right)-R_a\left(P_{Tn}+P_{ZTn}\cdot M\right)}{-P_{Tn}\cdot P_{Ga}+P_{Gn}\cdot P_{Ta}+M\left(-P_{Tn}\cdot P_{ZGa}+P_{Gn}\cdot P_{ZTa}-P_{Ga}{P}_{ZTn}+P_{Ta}\cdot P_{ZGn}\right)+M^2\left(-P_{ZTn}\cdot P_{ZGa}+P_{ZGn}\cdot P_{ZTa}\right)}}$$

(10)

In this study, R_n and R_a represent the bidirectional reflectance in the nadir and an alternative viewing direction, respectively, calculated from Sentinel-2 bidirectional reflectance data according to Equation (2). The multiple-scattering factor M was set to an empirical value of 0.2 [43]. The component proportions (R_T, R_G, R_ZT and R_ZG) were derived using the four-scale geometric optical model. The main input parameters of the model include canopy structural parameters, viewing geometry parameters, and spectral parameters.

2.2.2. Estimation of Shrub-Herb Fractional Vegetation Cover (FVCs)

Using the derived background reflectance data at a 10 m spatial resolution for the red and near-infrared bands from 2016 to 2025, the normalized difference vegetation index of shrub-herb vegetation (NDVI_s) was calculated annually. The Pixel Dichotomy Model was employed to separately derive the fractional vegetation cover of shrub-herb fractional vegetation cover (FVCs) for each year [44].

$${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}}={\displaystyle \frac{{\mathrm{N}\mathrm{I}\mathrm{R}}_{\mathrm{s}}-{\mathrm{R}}_{\mathrm{s}}}{{\mathrm{N}\mathrm{I}\mathrm{R}}_{\mathrm{s}}+{\mathrm{R}}_{\mathrm{s}}}}$$

(11)

$$\mathrm{F}\mathrm{V}\mathrm{C}\mathrm{s}={\displaystyle \frac{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}}{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}\mathrm{v}\mathrm{e}\mathrm{g}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}}}$$

(12)

Among them, NIR_s represents the background reflectance of the near-infrared band, and R_s represents that of the red band. NDVI_ssoil denotes the pure bare soil value, corresponding to the pixel value at the 5% cumulative frequency of NDVI_s in that year; NDVI_sveg denotes the pure vegetation value, corresponding to the pixel value at the 95% cumulative frequency of NDVI_s in that year.

2.2.3. Estimation of Total Fractional Vegetation Cover (FVCt)

Based on the original Sentinel-2 imagery from 2016 to 2025, the annual total Normalized Difference Vegetation Index (NDVI_t) was calculated separately. Subsequently, the annual total vegetation fractional coverage was derived using the pixel dichotomy model, resulting in the total fractional vegetation cover (FVCt).

$${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{t}}={\displaystyle \frac{{\mathrm{N}\mathrm{I}\mathrm{R}}_{\mathrm{t}}-{\mathrm{R}}_{\mathrm{t}}}{{\mathrm{N}\mathrm{I}\mathrm{R}}_{\mathrm{t}}+{\mathrm{R}}_{\mathrm{t}}}}$$

(13)

$$\mathrm{F}\mathrm{V}\mathrm{C}\mathrm{t}={\displaystyle \frac{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{t}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{t}\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}}{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{t}\mathrm{v}\mathrm{e}\mathrm{g}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{t}\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}}}\mathrm{ }\mathrm{ }\mathrm{ }$$

(14)

Among these, NIR_t represents the reflectance of the Sentinel-2 near-infrared band, while R_t represents the reflectance of the Sentinel-2 red band. NDVI_tsoil denotes the pure bare soil value, corresponding to the pixel value at the 5% cumulative frequency of NDVI_t in that year; NDVI_tveg denotes the pure vegetation value, corresponding to the pixel value at the 95% cumulative frequency of NDVI_t in that year.

2.2.4. Empirical Model Inversion of Canopy Fractional Vegetation Cover

In this study, an empirical model was developed to invert FVCc by integrating three key elements: the superposition principle of vertically stratified vegetation cover, the methodological framework for estimating vegetation cover from ground-based digital photography, and the logic of stratified representation of vertical vegetation structure. Previous research has demonstrated that FVCt is not simply the sum of canopy and shrub-herb cover, but rather a coupled outcome of these two components and their overlapping area under vertical projection. Moreover, the contribution of this overlapping portion is strongly correlated with the product of canopy and shrub-herb cover [45]. Accordingly, based on these theoretical foundations and the definition of fractional vegetation cover, the FVCt in the study area is less than the sum of canopy cover and shrub-herb cover, with the difference representing the overlapping cove (L) of the two layers.

Thus, the relationship among FVCt ($T$), field-measured FVCc ($C\prime )$, shrub-herb fractional vegetation cover S, and overlapping cover L can be expressed as follows: $T=C\prime +S-L$. Because L represents the overlapping cover between the canopy and shrub-herb layers under vertical projection, it can theoretically be assumed to correlate with $C\prime$ and $S$. To capture this relationship, a variable P was introduced to represent the product of $C\prime$ and S. The values of $C\prime$ were obtained from 20 representative field sample plots in the study area. The results showed that L and P exhibit a significant linear relationship, with a correlation coefficient R² = 0.9673. Using the least-squares method, the regression equation was fitted as L = 1.0546 × P + 0.0533, with a coefficient of determination R² = 0.9356.

The vegetation in the study area is dominated by arbor species with broad crowns and heights significantly greater than those of the shrub-herb layer, resulting in a relatively stable pattern of vertical projection overlap between the two layers. Based on these findings and the vegetation structural characteristics of the region, the empirical model was refined so that L explicitly quantifies the contribution of the vertically overlapping region, as shown in Equation (15). Through algebraic transformation, the inversion formula for FVCc was derived and is presented in Equation (16). Since fractional vegetation cover is defined as the proportion of the pixel area occupied by the vertical projection of woodland vegetation, its valid range is [0, 1]. Accordingly, constraints were imposed during the inversion process: values less than 0 were set to 0, and values greater than 1 were set to 1.

$$T=C+S-1.0546 \times C \times S-0.0533$$

(15)

$$C={\displaystyle \frac{T-S+0.053}{1 - 1.0546 \times S}}, (1 - 1.0546 \times S) \ne 0$$

(16)

Among them, $T$ denotes the total fractional vegetation cover (FVCt), $C$ denotes the canopy fractional vegetation cover (FVCc), and $S$ denotes the shrub-herb fractional vegetation cover (FVCs).

2.2.5. Accuracy Evaluation

In this study, the accuracy of the 2025 FVCc inversion was evaluated using leave-one-out cross-validation (LOOCV) based on 20 field-measured sample plots [46]. In each iteration, one sample was used as the validation set, while the remaining 19 samples were used to establish the fitting equation. This procedure was repeated 20 times, ensuring that each sample was validated once. The root mean square error (RMSE) and mean absolute error (MAE) were employed as error metrics to comprehensively evaluate model performance and minimize evaluation bias. The correlation and inversion accuracy were assessed using the fitted equation and the R² between the field-measured and inverted FVCc values for the 20 sample plots.

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-x_i\right)}^2},(i\le n)$$

(17)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n|y_i-x_i|$$

(18)

$$R^2=1 - {\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-y’\right)}^2}{{\sum }_{i=1}^n{\left(y_i-Y\right)}^2}},(i\le n)$$

(19)

Among them, n denotes the number of sample sites, $y_i$ denotes the field-measured FVCc value at sample site $i$,$y’$ denotes the fitted value at sample site $i$, $Y$ denotes the mean value of the field-measured FVCc across the 20 sample sites, and $x_i$ denotes the inverted value of FVCc at sample site $i$.

2.3. Spatio-Temporal Variation Analysis

Based on the remote sensing inversion results of FVCc from 2016 to 2025, this study calculated the annual mean values, annual standard deviations, and the overall 10-year mean to analyze the magnitude, spatial variability, and long-term average level of FVCc. This approach further elucidated its interannual variation patterns and evolutionary trends. FVCc was classified into five categories: Low Coverage (0 ≤ FVCc < 0.2), Medium-Low Coverage (0.2 ≤ FVCc < 0.4), Moderate Coverage (0.4 ≤ FVCc < 0.6), Medium-High Coverage (0.6 ≤ FVCc < 0.8), and High Coverage (0.8 ≤ FVCc ≤ 1.0) [47]. Subsequently, the spatio-temporal variation and distribution characteristics of each category were analyzed. Additionally, based on pixel-wise changes in FVCc levels between 2016 and 2025, a transition matrix was constructed to quantify conversions among different FVCc levels over the 10-year period, characterize the direction and magnitude of class shifts, and further reveal their spatial evolutionary processes and characteristics.

The coefficient of variation (CV) is a statistical metric used to assess the fluctuation status of long-term time series data, which can effectively reflect the stability and relative dispersion of environmental changes within a region [48].

$$\mathrm{CV}={\displaystyle \frac{\sigma }{X}}$$

(20)

Among them, CV denotes the coefficient of variation of FVCc, σ denotes its standard deviation, $X$ and denotes its mean value, A higher CV value indicates a greater degree of dispersion and increased interannual variation, whereas a lower CV suggests less dispersion and more stable interannual variation.

The Theil-Sen Median trend analysis and Mann–Kendall significance test were employed to systematically assess the dynamic variation characteristics of FVCc in the study area from 2016 to 2025 [49]. The Theil-Sen Median method describes the temporal variation trend of the time series using the slope β, where a positive value indicates an improving trend and a negative value represents a degrading trend. The Mann–Kendall test determines the significance of the trend via the test statistic $\mathrm{Z}$; $|\mathrm{Z}|>1.96$ indicates a statistically significant trend at the 95% confidence level. The classification of trend categories is presented in

Table 1. Trend Feature Category Classification.

$\mathit{\beta }$	$\left|\mathit{Z}\right|$	Trend Characteristics
$\beta <$0	$\left|Z\right|$ > 1.96	Significant Deterioration
	$\left|Z\right|\le$ 1.96	Mild Degeneration
$\beta >$0	$\left|Z\right|$> 1.96	Significant Improvement
	$\left|Z\right|\le$ 1.96	Mild Improvement

.

3. Results

3.1. Accuracy Verification

3.1.1. Reflection Angle Relationship Model

A multi-angle reflectance conversion model was developed for forested areas within the study region, utilizing the 2025 MODIS BRDF product as the data source to extract surface reflectance in the red and near-infrared bands at viewing zenith angles of 0° and 45°, respectively. A total of 302 samples, each with a spatial resolution of 10 m × 10 m were employed for model construction. During data preprocessing, pixels contaminated by clouds, cloud shadows, high aerosol loading, and anomalous observations were removed based on the quality assurance (QA) band of the MCD43A1 BRDF product, thereby reducing errors induced by atmospheric conditions and viewing geometry. Subsequently,, linear regression fitting was performed using the ordinary least squares (OLS) method, with reflectance at 0° was used as the independent variable and reflectance at 45° as the dependent variable. The regression coefficients were determined by minimizing the sum of squared residuals between the observed surface reflectance and the fitted values. During the fitting process, the 3σ criterion was applied to eliminate outliers whose reflectance deviated from the mean by more than three standard deviations, enhancing the stability of the fitting results. The results indicate that both the red and near-infrared bands exhibit a significant linear correlation between reflectance at 0° and 45° viewing angles, and the fitted models demonstrate high reliability. For the red band, the coefficient of determination was R² = 0.7026, with the fitted equation $y=2.3223x-0.0195$ (Figure 2a), while for the near-infrared band, R² = 0.7709, with the fitted equation $y=1.9117x-0.0154$ (Figure 2b). This linear model effectively characterizes the quantitative conversion relationship between reflectance at 0° and 45° viewing angles. Therefore, when applied to Sentinel-2 reflectance data observed at 0°, this model enables reasonable estimation of Sentinel-2 surface reflectance at the 45° oblique viewing angle.

3.1.2. Verification of Inversion Result Accuracy

Accuracy validation was conducted for the 2025 multi-angular FVCc inversion results. Using field-measured FVCc values from 20 representative sample plots as references, correlation analysis was performed with the temporally consistent 2025 multi-angular inversion products. The correlation analysis and linear regression revealed a significant positive linear relationship between the multi-angular inversion values and field measurements. The fitting equation is y = 0.9196x + 0.0212, with a slope of 0.9196, an intercept of 0.0212, and a coefficient of determination (R²) of 0.8583, indicating a strong linear relationship. As illustrated by the scatter distribution in Figure 3, the inversion and measured values are generally distributed along the 1:1 line. Slight deviations exist in a small number of sample points, with no obvious systematic bias. Leave-one-out cross-validation produced a root-mean-square error (RMSE) of 0.0629 and a mean absolute error (MAE) of 0.0496. The mean absolute deviation of individual samples remained stable at approximately 0.05, indicating a low error level. Although direct accuracy validation was not performed for the long-term inversion results from 2016 to 2024, the accuracy assessment based on the 2025 multi-angular inversion confirms the reliability and applicability of the method proposed in this study. Accordingly, this method can be applied to the long-term inversion of FVCc from 2016 to 2024, supporting quantitative inversion and spatiotemporal dynamic analysis of regional stratified vegetation cover.

Meanwhile, to fully verify the effectiveness of the research procedure in which the MODIS multi-angular relationship was applied to obtain Sentinel-2 reflectance at a 45° viewing angle for calculating background reflectance at 10 m spatial resolution, thereby achieving stratified inversion of FVCc, also performed canopy cover inversion using 2025 Sentinel-2 0° data and validated the results against field measurements. As shown by the fitting results in Figure 4, the single-angle inversion values are linearly correlated with the measured values, with a slope of 1.1261, an intercept of −0.0294, and a coefficient of determination (R²) of 0.5195. After Leave-one-out cross-validation, the RMSE of 0.1436 and MAE of 0.1087. A comparison between the single-angle and multi-angle FVCc inversion results indicates that the multi-angle approach achieves higher accuracy. Collectively, these results demonstrate that the multi-angle remote sensing inversion is reliable and can effectively characterize the actual canopy cover conditions in the study area, making it suitable for long-term FVCc inversion from 2016 to 2024.

3.2. FVCc Spatial Variation Analysis

According to the analysis of the spatial distribution characteristics of the annual average forest FVCc in Changting County from 2016 to 2025, the canopy vegetation cover in the study area generally exhibits a spatial pattern characterized by higher values in the southwest and lower values in the central region, as illustrated in Figure 5a. Based on the multi-year average canopy vegetation cover of each township, southwestern townships such as Hongshan (HS) and Sidu (SD) have maintained relatively high coverage levels for many years, whereas central and northern townships, including Tingzhou (TZ), Sanzhou (SZ), and Datong (DT), show relatively low values, as shown in Figure 6b. At the township scale, the range of annual average FVCc increased from 0.44 to 0.60 in 2016 to 0.65–0.85 in 2025, indicating an overall upward trend in canopy vegetation cover across the study area, as depicted in Figure 5c,d). Although Tingzhou (TZ) and Sanzhou (SZ) have long been in the relatively low-value range of canopy vegetation cover within the study area, their coverage has still increased. Tiechang (TC) and Anjie (AJ) also show a positive growth trend in canopy vegetation cover, but their growth rates are relatively slow compared with those of townships. Meanwhile, the global variance of canopy vegetation cover in the study area decreased from 0.05 in 2016 to 0.03 in 2025. This slight decline in global variance suggests that the spatial heterogeneity of canopy vegetation cover has been marginally reduced.

Based on an analysis of the spatial variation characteristics of FVCc gradation in Changting County, the overall regional vegetation cover exhibits a continuous improving trend, as shown in Figure 6. Low-coverage areas are primarily concentrated in the woodland–non-woodland transition zones of central townships such as Tingzhou (TZ), Hetian (HT), and Cewu (CW), displaying a relatively stable spatial distribution pattern with only minor local fluctuations. The area of medium-low coverage zones shows a gradual decreasing trend, accompanied by a corresponding expansion of moderate coverage zones. In contrast, medium–high and high coverage zones have increased significantly, evolving from an initially scattered distribution to a more concentrated one. Overall, the spatial disparity in FVCc across Changting County has gradually decreased, accompanied by a substantial improvement in vegetation cover at the regional scale. Although the rate of increase is relatively modest in some individual townships, a stable upward trend is evident throughout the entire county.

Based on the standard deviation and mean values of FVCc at the township level in Changting County for 2016 and 2025, the spatial variation in the coefficient of variation (CV) at the township scale was calculated separately, as illustrated in Figure 7. The results indicate that the CV range of township-level canopy FVCc decreased from 0.39 to 0.47 in 2016 to 0.20–0.30 in 2025, suggesting an overall improvement in the stability of regional vegetation cover during the study period. Specifically, northern townships, represented by Tingzhou (TZ) and Anjie (AJ), exhibited a pronounced decline in CV, reflecting enhanced stability of canopy FVCc in these areas. Central and eastern townships showed a relatively slight but consistent downward trend in CV, indicating that canopy FVCc in these regions changed gradually and remained generally stable. Mountainous townships in the southwest, including Sidu (SD), Hongshan (HS), and Zhuotian (ZT), transitioned from relatively high to low CV values, also demonstrating increased stability of canopy FVCc. Overall, spatial differences existed in the stability of township-level canopy FVCc across Changting County during the study period, but a general improving trend was observed throughout the entire region.

3.3. FVCc Temporal Variation Analysis

Changes in the area proportions of each FVCc class in Changting County from 2016 to 2025 are illustrated in Figure 8. The area proportions of different FVCc classes exhibited a pattern of gradient optimization and overall improvement. Low FVCc remained at an extremely low level throughout the study period. The area proportion of medium-low FVCc decreased significantly, shrinking gradually from 25% in 2016 to less than 2% in 2025. The area proportion of moderate FVCc initially increased, peaking in 2018, and then declined progressively to approximately 11% in 2025, reflecting a continuous transition of moderate coverage areas toward higher FVCc classes. The area proportions of medium-high and high FVCc showed a consistent increasing trend. Specifically, medium-high FVCc rose from 11% in 2016 to about 48% in 2025, while high FVCc increased from 19% in 2016 to 39% in 2025. Their combined proportion grew from 30% in 2016 to 87% in 2025, becoming the dominant coverage classes in the region. The mean FVCc in the study area increased from 0.54 in 2016 to 0.77 in 2025, directly demonstrating the continuous improvement and overall enhancement of canopy vegetation cover in Changting County. Overall, during the study period, canopy FVCc in Changting County exhibited a successional trend characterized by reduced proportions of lower classes and expanded proportions of higher classes.

According to the pixel-level transition matrix analysis of FVCc classes in Changting County from 2016 to 2025, positive transitions dominated the regional FVCc class changes during the study period. The specific transition paths for each class are illustrated in Figure 9. Only 8% of the low-coverage areas remained unchanged, while 92% transitioned to higher classes: 10% shifted to medium-low coverage, 26% to moderate coverage, 36% to medium-high coverage, and 20% to high coverage. Among medium-low coverage areas, 3% remained stable, 1% degraded to low coverage, and 96% transitioned to higher classes: 24% to moderate coverage, 54% to medium-high coverage, and 18% to high coverage. For moderate coverage areas, 8% remained unchanged, 2% degraded to lower classes, and 90% upgraded to higher coverage classes. In medium-high coverage areas, 43% remained stable, 52% transitioned to high coverage, and 5% degraded to moderate coverage or below. In high coverage areas, 66% remained unchanged, 30% degraded to medium-high coverage, and 4% fell to moderate coverage or below. Overall, FVCc class transitions from 2016 to 2025 were characterized by upgrades from moderate and lower classes to medium-high and high classes, while the proportion of degradation to lower classes was extremely low, indicating a relatively low risk of vegetation coverage degradation.

Based on the coefficient of variation, Theil-Sen Median trend analysis, and Mann–Kendall test, a comprehensive assessment was conducted on the temporal variation characteristics of FVCc in Changting County from 2016 to 2025, focusing on stability and trend, as illustrated in Figure 10. Over the decade, areas with high stability accounted for the largest proportion, reaching 48%; the proportions of medium-high and moderate stability areas comprised 21% and 9%, respectively; while medium-low and low stability areas accounted for 5% and 17%. Regarding interannual trends, FVCc showed significant improvement as the dominant feature, covering 65% of the area, with slightly improved areas accounting for 26%, totaling 91%. Approximately 4% of the area exhibited no obvious trend, and only 5% showed degradation, primarily slight degradation, with significantly degraded areas representing less than 1%. This comprehensive analysis indicates that FVCc in the study area experienced a sustained and stable improvement over the decade. High stability provided a solid foundation for the continuous enhancement of canopy vegetation cover, whereas dynamic fluctuations in low stability areas were linked to local degradation. These findings suggest that ecological restoration efforts should focus not only on increasing vegetation coverage but also on strengthening ecosystem stability and resilience to disturbances.

4. Discussion

4.1. Comparison with Existing Inversion Methods

Traditional remote sensing inversion methods for FVCc have notable limitations. Single-angle Sentinel-2 observations are significantly affected by angular effects caused by topographic relief, which compromises the stability and robustness of inversion results. Conversely, although MODIS offers multi-angle observations, its coarse spatial resolution is insufficient to resolve fine-scale canopy structures in complex forest stands. Consequently, neither sensor alone can simultaneously provide both high spatial resolution and multi-angle observations [50]. To overcome this limitation, MODIS multi-angle reflectance data were integrated with the high spatial resolution of Sentinel-2 imagery to retrieve FVCc in Changting County. Validation against field-measured canopy cover data demonstrates that the multi-angle inversion results obtained here are more accurate than those derived from single-angle Sentinel-2 data and better capture the complex vertical forest structure in the study area. The proposed inversion method effectively compensates for the shortcomings of single-sensor data sources in terms of viewing geometry and spatial resolution, making it well-suited for accurate FVCc inversion in regions with complex stand structures. Thus, the method used in this study is feasible for retrieving FVCc in Changting County from 2016 to 2024 and provides technical support for analyzing its spatiotemporal dynamics.

In this study, differences between MODIS and Sentinel-2 in the red and near-infrared bands were characterized, and spectral normalization was performed by establishing linear relationships between the bands of the two sensors prior to multi-angle information transfer [51]. Although this processing strategy is relatively simple, validation of the inversion results demonstrates that it maintains physical consistency while remaining computationally efficient. It effectively suppresses systematic biases arising from bandpass differences and meets the accuracy requirements for subsequent canopy cover inversion. Nonetheless, this simplified approach has potential for further refinement. Future research could enhance canopy cover inversion accuracy and reliability by optimizing bandpass difference correction strategies. However, the fusion of bidirectional reflectance distribution function models with multispectral data presents certain limitations in highly heterogeneous environments. The forest stands in the study area exhibit significant variations in stand density, which inevitably increase surface heterogeneity to some extent [52,53]. Future studies will focus on further optimizing angular normalization techniques and data fusion strategies to address the challenges posed by heterogeneous forest regions, thereby improving the accuracy and robustness of subsequent FVCc inversion. Additionally, the multi-angle relationship model developed in this study adopts a linear form, and the underlying nonlinear transformation mechanisms warrant further investigation. Overall, the inversion results demonstrate high reliability and accuracy, providing novel insights with important implications for the accurate inversion of FVCc in similarly complex forest ecosystems.

4.2. Spatio-Temporal Variation Characteristics and Causal Analysis

Based on an analysis of the spatiotemporal evolution patterns derived from the FVCc inversion results in Changting County from 2016 to 2025, these dynamics can be attributed to the synergistic effects of topographic conditions, human activity intensity, and long-term ecological restoration projects in the study area. Spatially, canopy vegetation cover exhibits an overall distribution pattern characterized by high values in the surrounding areas and relatively low values in the central region. The southwestern part of Changting County is dominated by mountainous landforms with low-intensity human disturbance, where FVCc has remained consistently high over the long term [54]. The central and northern parts are dominated by urban construction zones, where the expansion of built-up land has encroached upon vegetated areas. Coupled with the impacts of agricultural activities, FVCc in these regions remains relatively low overall [55]. In 2025, the coefficient of variation of FVCc in the northern and western townships decreased compared with 2016, indicating an improvement in the internal stability of canopy vegetation cover in these regions. In most parts of the central and eastern areas, the coefficient of variation changed relatively slowly, suggesting that canopy vegetation cover remained relatively stable.

From a temporal perspective, based on the inversion results of this study, canopy FVCc in the study area exhibited a sustained and steady improvement trend from 2016 to 2025. Vegetation cover classes were dominated by positive transitions, with large-scale conversions from moderate and lower classes to medium-high and high classes, indicating an overall increase in regional canopy cover. Differences in classification schemes can affect the characterization of spatial patterns, dynamic transition processes, and the conclusions of stability assessments for canopy FVCc [56]. Therefore, this study adopted a vegetation cover classification standard tailored to the study area, which fundamentally ensures the reliability of inversion results and the comparability of spatiotemporal change analyses. FVCc was evenly divided into five levels. This unified classification standard objectively reflects its continuous variation and scientifically and intuitively reveals the spatio-temporal dynamics of vegetation cover from 2016 to 2025 [57]. Meanwhile, the pixel-wise Mann–Kendall test used in this study may introduce multiple hypothesis testing problems due to the extremely large number of pixels, resulting in significance inflation and thus overestimating the proportion of areas with significant changes [58]. Accordingly, future research can further improve the reliability of results by adopting multiple testing corrections such as the false discovery rate (FDR) or by aggregating analyses into coarser spatial units prior to significance testing [59].

In summary, the spatiotemporal evolution of FVCc in Changting County is likely influenced by multiple interacting factors. To date, this study has provided only a qualitative description of FVCc dynamics based on existing literature. Future research could conduct further quantitative analyses to explore the relationships between driving factors and changes in FVCc. Although this study did not perform direct accuracy validation for the long-term inversion results from 2016 to 2024, the accuracy assessment based on the 2025 multi-angular inversion results confirms the reliability and applicability of the proposed method. Therefore, it can be reasonably inferred that FVCc across the study area has exhibited a significant overall increasing trend. Townships with relatively high coefficients of variation can be identified as ecologically sensitive and key monitoring areas for long-term dynamic observation, providing scientific support for the sustainable and stable development of regional ecosystems.

5. Conclusions

This study combines the advantages of MODIS multi-angle observations with Sentinel-2 high-spatial-resolution data to effectively distinguish anisotropic reflectance differences between forest canopy and shrub-herb layers, thereby enabling the inversion of forest FVCc. We systematically analyzed the vertical structural relationship between forest canopy and shrub-herb vegetation coverage in Changting County, as well as the spatio-temporal variation characteristics of forest FVCc from 2016 to 2025. The results confirm the reliability of incorporating multi-angle reflectance observations in forest FVCc inversion. The proposed method compensates for the limited inversion accuracy caused by single-angle observations and the coarse resolution of MODIS multi-angle data, allowing quantitative estimation of dynamic changes in FVCc. This provides a methodological reference for the stratified inversion of forest vegetation coverage. Future research can optimize angular normalization and data fusion strategies in heterogeneous forest areas and further explore the complex transformation mechanisms of multi-angle relationships among remote sensing images with varying spatial resolutions.