Analysis of Vegetation Restoration Potential and Its Influencing Factors on the Loess Plateau: Based on the Potential Realization Model and Spatial Dubin Model

Abstract

Improvements in vegetation coverage are driven by both resource endowment conditions and policy behaviors. To accurately reflect the vegetation restoration effect after ecological policies, this study used the potential realization model to calculate the potential realization degree of vegetation restoration on the Loess Plateau and to assess the vegetation restoration effect after the Grain for Green Program from 2000 to 2020. Then, the influencing factors were explored using the spatial Dubin model. The results reveal that (1) the EVI value of the Loess Plateau in northern Shaanxi increased from below 0.25 at the beginning of the study to approximately 0.35 by the end, indicating that the green territory of the Loess Plateau gradually expanded to the northwest over the study period, and that the east and west of the Loess Plateau are key areas of vegetation cover for further improvement; (2) compared to the traditional EVI indicator, the potential realization degree can more accurately evaluate the vegetation restoration effect driven by ecological policies; (3) policy intensity is positively correlated with the growth rate of the vegetation restoration potential realization degree by 0.183 and significant at 1% level, making it the primary factor influencing the effect of vegetation restoration. Additionally, annual average precipitation and annual sunshine percentage have significant spatial positive contributions to the improvement of vegetation restoration on the Loess Plateau. The study’s findings are expected to contribute to the development of a scientific basis for adjusting the vegetation restoration policy on the Loess Plateau and enhancing ecological restoration efforts.

1. Introduction

Amidst the swift expansion of the population, urbanization, and large-scale agriculture, global ecosystems have faced enormous challenges over the past few decades, including deforestation, soil degradation, soil erosion, and resource scarcity becoming increasingly prominent [1,2]. In particular, the conflict between China’s enormous population pressure and limited agricultural land resources has led to ecological protection issues that were once neglected, causing numerous ecological issues, including deforestation, land deterioration, the destruction of vegetation, and a reduction in biodiversity [3,4]. To protect the environment and address the fragility of ecosystems in many parts of the country, China has introduced various policies aimed at restoring and managing ecosystems since the 1980s. The Grain for Green Program (GFGP) is an important one of such ecological policies [5].

China’s GFGP began early in drylands, with large-scale investment and extensive coverage, leading the world in large-scale ecological restoration in drylands [6]. The Loess Plateau is a typical example of arid and semi-arid regions worldwide. Since the 1970s, rapid population increase and economic expansion have triggered widespread deforestation and land reclamation, leading to significant environmental and ecological challenges [7]. To promote vegetation restoration and ecological improvement, the Chinese government has included the region in the first set of pilot areas in which to implement the GFGP [8].

To better formulate and implement China’s future ecological restoration and governance policies, it is increasingly important for scholars to assess the effectiveness of ecological policy implementation based on vegetation cover [9,10]. This is because vegetation restoration is crucial for achieving ecological engineering goals such as conserving soil and water, sequestering carbon, and preserving biodiversity [11,12]. Generally, there are three recognized methods for assessing the degree and effect of vegetation restoration. The first method involves site observations to measure vegetation coverage, biomass, and species diversity at the site [13]. The second method employs remote sensing technology to collect high-resolution data and analyze the data to generate indicators like vegetation coverage and biomass [14]. The third method combines site observations with remote sensing [15]. Compared to site observations, remote sensing offers faster, more efficient, objective, and reliable information on vegetation restoration [16]. Consequently, remote sensing is widely used to evaluate vegetation restoration. Remote sensing vegetation indices allow for a direct, simple, efficient, and empirical assessment of vegetation coverage in a given area. Commonly used vegetation indices include the normalized NDVI, EVI, and kNDVI [17]. However, some scholars have indicated that there is a vegetation oversaturation effect in NDVI, its ability to distinguish vegetation coverage decreases with dense vegetation [18]. There are also studies using the kNDVI index that have found a poor relationship with vegetation traits [19,20]. In comparison, EVI can solve the problems of the easy saturation of the vegetation index and the lack of a linear relationship with actual vegetation cover.

It has been over 20 years since the GFGP was first implemented, and existing studies often calculate NDVI, EVI, and kNDVI values at the county, provincial, or watershed scales to explore the regional characteristics of the ecological policy effects [21,22,23]. While the goal of ecological policies is to restore vegetation and protect ecological systems, factors influencing vegetation coverage include natural conditions, global climate change, and human-made policies [24,25]. The vegetation index values alone reflect only the level of vegetation cover and may not adequately represent the effects of the GFGP, as they do not account for the spatial differentiation of geographical characteristics. This omission makes it challenging to accurately evaluate the implementation effects of the program. Additionally, the effects of ecological projects are influenced by various factors, yet current studies often lack a comprehensive understanding of the interactions and spatial spillover effect of these factors. This gap may impede the optimization of ecological protection policies in China.

Following the first and third laws of geography [26,27], vegetation under the same habitat conditions is likely to have a similar degree of vegetation coverage. The difference between actual vegetation coverage and the average coverage for similar habitats indicates the potential for the improvement of vegetation coverage, which we define as vegetation restoration potential (VRP) [28]. Thus, differences in natural resource endowments are an important factor affecting efforts to improve vegetation coverage [29]. To accurately evaluate ecological policies and their effects on vegetation restoration, it is essential to eliminate the influence of natural differences, which requires considering local geographical features such as terrain and soil. The spatial sliding window technique is a valuable tool for overcoming spatial differentiation. It is commonly employed to build local statistical models that mitigate the limitations of missing variables in global models by controlling spatially heterogeneous in the data [30].

Based on previous research, this study focuses on a typical arid and semi-arid geomorphological area as the research unit and addresses the following aspects: (1) the spatiotemporal characteristics of the Loess Plateau vegetation restoration from 2000 to 2020 are analyzed using stabilized EVI datasets; (2) this study innovatively employs the vegetation restoration potential realization model and spatial sliding window technology to assess local VRP and its realization (VRPRD) on the Loess Plateau. Additionally, the effectiveness of GFGP implementation is accurately measured with the help of VRPRD; (3) the Moran index is calculated to discuss the spatial autocorrelation of VRP across the Loess Plateau, while the spatial Dubin model is utilized to investigate the factors influencing VRP and its spatial spillover effect. The main contributions of this study include overcoming the influence of existing resource conditions through the use of local sliding window technology, more accurately assessing GFGP’s effects on the Loess Plateau using the VRPRD index, and exploring the interactions among influencing factors and spatial spillover effect. This study aims to serve as a valuable reference for future adjustments in ecological policies and offer a method for evaluating ecological policy implementation.

2. Study Area

The Loess Plateau is situated at the junction of northern and northwest China, within the geographical coordinates of 33°43′~41°16′ N and 100°54′~114°33′ E (Figure 1), which is the world’s largest loess deposit [31]. The Loess Plateau includes seven provincial-level administrative regions, including Qinghai, Gansu, Ningxia, Inner Mongolia, Shaanxi, Shanxi, and Henan. It consists of 341 county-level units, covering approximately 640,000 square kilometers, and has a total population exceeding 82 million. The region experiences a typical temperate continental monsoon climate, characterized by hot and rainy summers and cold, dry winters. The average annual temperature ranges from 3.6 to 14.3 °C, and the average annual precipitation varies from 150 to 800 mm, decreasing from southeast to northwest. The average altitude ranges between 200 and 3000 m [32].

Since the 1970s, large-scale deforestation on the Loess Plateau, driven by population growth and economic development, has led to severe environmental and ecological issues. To foster more ecologically minded development, the Chinese government initiated the GFGP on the Loess Plateau in 1999. Over more than 20 years of implementation, there has been a significant improvement in vegetation coverage. The development of ecological civilization has yielded numerous positive outcomes, including enhanced biodiversity, improved soil conservation, and a more sustainable environment.

3. Data and Methods

3.1. Data Sources

3.1.1. Elevation and Soil Data

The DEM elevation data used in this study (SRTM DEM V4.1) were obtained from the Geospatial Data Cloud Platform (https://www.gscloud.cn, accessed on 10 November 2024). These data, provided in raster format with a spatial resolution of 90 m, were clipped to fit the study area (Figure 1). The Raster Surface tool from ArcToolBox on the Arcgis10.8 platform was then used to extract the slope (Figure 2a) and aspect (Figure 2b) data required for this study.

The soil type data were sourced from the Resource and Environmental Science and Data Center (https://www.resdc.cn, accessed on 7 January 2025). Initially provided in raster format with a spatial resolution of 1 km, these data were resampled to 90 × 90 m and clipped to the study area using Arcgis10.8 (Figure 2c).

3.1.2. Meteorological Data

The meteorological data used in this study were sourced from the 30-year average monthly/1 km climate variable dataset of China [33]. After merging and clipping to fit the study area, three raster layers were produced, representing annual average temperature, annual average precipitation, and annual sunshine percentage. These layers were then resampled to a uniform spatial resolution of 90 m, as shown in Figure 2d–f.

3.1.3. Socioeconomic Data

The road network data were obtained from the 1:1 million vector map data shared by the National Geomatics Center of China (https://www.webmap.cn, accessed on 7 January 2025), and the layer was clipped to fit the study area. Road density is one of the important factors driving policy effects. This study defines road density as the ratio of the length of roads in each county to the county’s total area [34].

The land use data for this study were sourced from the China land cover dataset produced by Yang and Huang [35] of Wuhan University. This dataset was resampled and clipped to align with the 90 m spatial resolution and the scope of the study area used in other layers. The GFGP intensity of each county was calculated based on land use data.

3.1.4. Vegetation Coverage Data

Due to the vegetation oversaturation effect associated with NDVI, dense vegetation decreases the index’s ability to distinguish vegetation coverage. Therefore, this study used Google Earth Engine to obtain the annual average EVI dataset from MODIS13Q1 for 21 periods from 2000 to 2020. A 16-day synthetic product was used for the temporal resolution, and a 250 m spatial resolution was used for the spatial resolution. To align the data with the other layers in the study, the data were resampled to 90 m spatial resolution. To truly reflect the vegetation restoration effect after the implementation of the GFGP, this study masked out areas with terrain slopes below 6°.

3.2. Methods

3.2.1. Local Sliding Window-Based Similar Habitat Potential Model

The first law of geography states that any geographical feature has a spatial correlation [26]. Geographical features will be more similar if the environment is similar, according to the third law of geography [27]. Following this logic, the EVI of a given location should be equivalent to that of surrounding areas under similar growing conditions. Therefore, this study uses classification layers to construct similar habitats, with the difference between the maximum and actual EVI for a specific location defined as VRP.

The local window is a theoretically smaller spatial unit. Some macro-environmental variables, such as meteorological variables, exhibit less change across a local window, have negligible impacts on vegetation recovery, and can be regarded as having the same or similar values across geographic space [36]. Therefore, in constructing similar habitat units, this study only considered variables that change at the local window scale such as elevation, aspect, slope, and soil type, and classified and coded each control variable. To ensure the robustness of statistics, the classification of each control variable should be minimized. The specific coding mode for this study is as follows:

$$Code=a\ast {10}^3+b\ast {10}^2+c\ast {10}^1+d=abcd$$

(1)

using a, b, c, and d to represent the number of categories associated with each of the four variables used to construct similar habitat units, respectively. By encoding all rasters in the study area, the calculation can be reduced to one dimension, which is beneficial to the construction of similar habitat layers. Within the window radius R, we used a sliding window to find the same habitat unit as abcd, obtained the maximum EVI under similar habitat conditions, and calculated the VRP. The formula for the local sliding window-based similar habitat potential model is as follows:

$$\begin{array}{ll}VR{P}_{ij\_t}(U_1,U_2,U_3,\dots ,U_{\mathrm{N}})=& MAX\text{·}EV{I}_{ij\_t}(U_1,U_2,U_3,\dots ,U_{\mathrm{N}})-EV{I}_{ij\_t}(U_1,U_2,U_3,\dots ,U_{\mathrm{N}})\\ & 1\le i\le \mathrm{m}\\ & 1\le j \le \mathrm{n}\end{array}$$

(2)

In the formula, U_N represents the Nth similar habitat control variable; t represents the time from 2000 to 2020; MAX∙EVI_{ij_t}(U₁, U₂, U₃, ..., U_N) refers to taking the i_th row and j_th column raster as the benchmark and R as the window radius to find the theoretical maximum EVI with similar habitat conditions during the study period; and i and j are the row and column of the maximum EVI raster, respectively. Assuming m rows and n columns within the window radius R, 1 ≤ i ≤ m, 1 ≤ j ≤ n. VRP_ij(U₁, U₂, U₃, ..., U_N) is the maximum VRP of the i_th row and j_th column of the raster, which is expressed as the difference between the theoretical maximum EVI and the actual EVI for the i_th row and j_th column of the raster.

3.2.2. Potential Realization Degree Model

At the regional scale, scholars generally regard the EVI growth rate as the vegetation restoration effect [37,38]. Based on the VRP calculated in Section 3.2.1, the VRPRD is defined here as the ratio of the raster’s actual EVI to its theoretical maximum EVI, which is used to indicate the extent to which vegetation has been restored. To avoid data errors, this study used the 99% maximum quantile value when calculating the theoretical maximum EVI. As such, VRPRD may be greater than 1; thus, it is mandatory to be defined as 1. The formula for VRPRD is as follows:

$$VRPR{D}_{ij\_t}=\left\{\begin{array}{l}0, if EV{I}_{ij\_t}\le 0\\ \\ {\displaystyle \frac{EV{I}_{ij\_t}(U_1,U_2,U_3,\dots ,U_N)}{\begin{array}{l}MAX\text{·}EV{I}_{ij\_t}(U_1,U_2,U_3,\dots ,U_N)\\ \end{array}}}\\ 1, if EV{I}_{ij\_t}\ge MAX\text{·}EV{I}_{ij\_t}\end{array}\right., if 0<EV{I}_{ij\_t}\le MAX\text{·}EV{I}_{ij\_t}$$

(3)

In the formula, EVI_{ij_t} represents the raster’s actual EVI of i_th row and j_th column in t_th year; MAX∙EVI_ij represents the theoretical maximum EVI of the raster’s i_th row and j_th column under similar habitat conditions; and VRPRD_{ij_t} represents the VRPRD of the raster’s i_th row and j_th column in the t_th year.

3.2.3. Spatial Econometric Model

Analysis based on VRP shows that the conditions of resource endowment such as elevation, slope, aspect, and soil are the primary factors influencing vegetation restoration. However, improvements in the vegetation restoration effect are caused by multiple interacting factors. This study analyzed the growth rate of VRPRD from 2000 to 2020 to evaluate how the VRPAD mitigates resource endowment effects on vegetation restoration under GFGP. The Ordinary Least Squares (OLS) model is commonly used for data regression [39], and is defined by the following basic form:

$$Y_i=\alpha +{\displaystyle \sum _k{\beta }_k{x}_{ik}}+\epsilon$$

(4)

In the formula, x_k represents factors in addition to resource endowment conditions that affect vegetation restoration; Y represents the growth rate of VRPRD; i represents the i_th county in the Loess Plateau; α and β are the coefficients to be estimated; and ε represents the error.

Since the VRPRD growth rate is calculated based on a similar habitat theory, we consider it to be spatially autocorrelated, and the Moran index is a widely used metric for measuring spatial autocorrelation in its basic form:

$$I={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{\displaystyle {\sum }_{j=1}^n{w}_{ij}}\left(x_i-\overline{x}\right)}\left(x_j-\overline{x}\right)}{s^2{\displaystyle {\sum }_{i=1}^n{\displaystyle {\sum }_{j=1}^n{w}_{ij}}}}}$$

(5)

$$s^2={\displaystyle \frac{{{\displaystyle {\sum }_{i=1}^n\left(x_i-\overline{x}\right)}}^2}{n}}$$

(6)

In the formula, W_ij represents the spatial adjacency matrix, x_i and x_j represent the VRPRD growth rate for spatial locations i and j, respectively, and S² represents the sample variance. I ∈ [−1,1], I > 0 and I < 0 indicate spatial positive and spatial negative correlation, respectively, and I = 0 indicates spatial uncorrelation. Values of I closer to 1 indicate stronger spatial clustering, whereas values near −1 suggest more pronounced spatial convergence.

Considering the spatial heterogeneity and spatial dependency of the influencing factors, this study further investigates the drivers of VRP using a spatial econometric model. Commonly used spatial econometric models include the spatial Dubin model (SDM), the spatial Lag model (SLM), and the spatial Error model (SEM) [40]. SEM accounts for spatial correlation in the error terms of both independent and dependent variables; SLM addresses spatial dependence in the independent variables that affect the dependent variable. The SDM integrates elements of both SLM and SEM by incorporating spatial lag terms for independent and dependent variables, thus providing a more comprehensive estimation of spatial effects. Its basic formulation is as follows:

$$Y=\rho WY+X\beta +WX\theta +\epsilon ,\epsilon ~N(0,{\delta }^2)$$

(7)

In the formula, ρWY represents the effect of the neighboring region’s dependent variable on the region’s dependent variable; WXθ represents the effect of the neighboring region’s independent variable on the region’s dependent variable; ρ is the spatial autoregressive coefficient; β is the coefficient of the influence factor; and ε is the random perturbation term. When θ = 0, the model becomes SLM; when θ + ρβ = 0, the model becomes SEM.

4. Results and Analysis

4.1. Spatiotemporal Changes in Vegetation Coverage

This study generated maps showing EVI distribution on the Loess Plateau for the years 2000 and 2020 to assess the changes in vegetation coverage over the study period. Figure 3a reveals that in 2000, vegetation growth was poor throughout the central and northwest parts of the Loess Plateau, particularly in northern Shaanxi, the junction of Shaanxi and Shanxi, and the junction of Shaanxi, Gansu, and Ningxia. The annual average EVI in these areas was below 0.25. As shown in Figure 3b, the Loess Plateau’s EVI has increased significantly from 2000 to 2020. This shows that decision makers in the area are gradually prioritizing ecological protection, regional vegetation has been effectively restored, and the green area has expanded toward the northwest.

Figure 4 shows a significant increase in vegetation coverage on the Loess Plateau during the study period. By 2010, EVI values had risen for 60.74% of the rasters in the Loess Plateau compared to 2000. The number of rasters with EVI values in the range of 0.65 and above had the highest growth rate of over 145.41%. However, only 7.14% of the rasters in the Loess Plateau in 2020 showed an increase in EVI values compared to 2010. Within this period, rasters with EVI values in the range of 0.45 to 0.65 experienced the highest growth rate, exceeding 10.16%. Comparing different periods, it is evident that the EVI growth rate was more pronounced during the first half of the period (2000–2010) than in the second half (2010–2020). Despite the GFGP having been implemented in the Loess Plateau since 2000, the current results are insufficient to attribute the increase in EVI solely to human policy factors. Therefore, isolating the influence of resource endowment conditions is crucial for accurately evaluating the GFGP’s effects on vegetation restoration.

4.2. Vegetation Restoration Potential Based on Similar Habitat Units

The degree of VRP in the Loess Plateau can be better understood by constructing similar habitat units for vegetation growth in the region. This study merged the EVI layers from 2000 and 2020 using the formula (2) to calculate the VRP of each raster and retained the maximum value to represent optimal vegetation growth. The purpose of constructing similar habitat units is to isolate the contribution of resource endowment conditions to improved vegetation coverage and to facilitate a more accurate assessment of the effects of GFGP by calculating the growth rate of VRPRD. Figure 5 illustrates that the VRP of the Loess Plateau generally decreases from southeast to northwest. The maximum value of VRP in the southeast ranges from 0.65~1, and the number of rasters accounts for 26.7%. The minimum value of VRP in the northwest ranges from 0~0.25, and the number of rasters accounts for only 1.4%. This pattern is consistent with the water and heat distribution in the Loess Plateau. Similar habitat units were constructed using four natural conditions including elevation, slope, aspect, and soil type, which show that the resource endowment conditions in the southeast are more suitable than those in the northwest for vegetation growth. The proportion of different potential values shows the number of rasters with the VRP exceeding 0.45 to account for more than 77% of the entire study area, indicating that vegetation on the Loess Plateau has high growth potential. The VRP of the Loess Plateau is also higher than the EVI in 2020 (Figure 3b), indicating the capacity for further improvement in vegetation restoration in the region.

4.3. Assessment of Vegetation Restoration Potential Realization Degree

To better understand the VRPRD in the Loess Plateau at different periods, this study used formula (3) to calculate the ratio of EVI to VRP in 2000 and 2020. Figure 6 shows that the distribution of VRPRD is consistent between the two periods and decreases from southeast to northwest. Specifically, there were low VRPRD values across the Loess Plateau in 2000, with more than half of the rasters having VRPRD values below 0.65. That is, the proportion of EVI to VRP in 2000 was less than 65% (Figure 6a). This indicates that the improvement capacity of EVI is at least 35%, especially in parts of Qinghai, and that the VRPRD is lower than 0.45, indicating a greater improvement capacity of vegetation cover. The VRPRD of the region increased significantly in 2020 compared to 2000. The vast majority of the GFGP implementation area within the Loess Plateau shows VRPRD values of 0.75 and above (Figure 6b). Figure 6c further demonstrates that 40.19% of the raster VRPRD values have been boosted. The highest number of rasters with VRPRD values in the range of 0.45 to 0.65 were transferred out and the highest number of rasters with VRPRD values in the range of 0.75 to 0.85 were transferred in. As shown in Figure 6d, the central part of the Loess Plateau exhibits the most significant enhancement in VRPRD, with growth rates exceeding 15%. Comparing the VRP of the Loess Plateau with the 2020 VRPRD layer reveals that the theoretical maximum potential for vegetation restoration in southern Qinghai and western Shanxi exceeds 0.45, but the potential realization degree is lower than 0.65. These findings suggest that there is still considerable room for improving vegetation cover on the Loess Plateau, particularly in the above highlighted areas, which will be crucial for future restoration efforts.

4.4. Analysis of Vegetation Restoration Potential Influencing Factors and Spatial Spillover Effect

4.4.1. Analysis of Influencing Factors

Several factors have contributed to the increase in vegetation cover. To determine whether VRPRD effectively overcomes the influence of natural endowment conditions and accurately measures the implementation effects of GFGP, this study uses the formula (4) for quantitative analysis. Taking counties as sample units, we examine the correlations between the EVI growth rate and variables such as GFGP intensity in the Loess Plateau, as well as between the VRPRD growth rate and GFGP intensity.

Table 1. Correlation between each variable and EVI growth rate and VRPRD growth rate.

Variable Type	Variable	EVI Growth Rate	VRPRD Growth Rate
Meteorological variables	Temp	0.160 ***	0.080 ***
	Prec	−0.026	0.035
	Sunp	0.062 **	0.060 ***
Socioeconomic variables	R&D	−0.022 **	−0.020 ***
	Intensity	0.139 ***	0.183 ***

Note: Temp indicates annual average temperature; Prec indicates annual average precipitation; Sunp indicates annual sunshine percentage; R&D indicates road density; Intensity indicates the intensity of GFGP implementation, specifically, the ratio of GFGP area to total forest land area at the study’s end. Significance levels are indicated as *** for 1%, and ** for 5%.

presents the magnitude and significance of the resulting correlation coefficients.

Overall, the positive and negative correlations of each variable with the VRPRD growth rate are consistent with the EVI growth rate, except for the annual average precipitation. Additionally, each variable’s absolute correlation coefficient with the VRPRD growth rate is smaller than that with the EVI growth rate. Notably, the policy intensity positively correlates with the VRPRD growth rate, having a coefficient of 0.183. This value is greater than the correlation coefficient between policy intensity and the EVI growth rate and is significant at the 1% level. The results of this study indicate that the EVI growth rate is more dependent on regional resource endowment conditions, while the VRPRD growth rate is less dependent on resource endowment conditions and responds more strongly to GFGP intensity. This finding aligns with the study’s initial hypothesis. Therefore, it can be concluded that the VRPRD growth rate index effectively isolates the role of resource endowment from the vegetation restoration effect, providing a more accurate evaluation of the GFGP’s impact on vegetation restoration.

4.4.2. Exploration of Spatial Autocorrelation

Because the VRPRD growth rate is calculated based on similar habitat theory and is spatially autocorrelated, the OLS model does not account for the spatial autocorrelation between VRPRD growth rates across locations. To present the changes in the spatial correlation of VRPRD, this study chooses 2005, 2010, 2015, and 2020 as observation points to explore the evolution of spatial correlation through temporal and spatial changes. Figure 7 shows that the values of Moran’s index in 2005, 2010, 2015, and 2020 are 0.557, 0.614, 0.649, and 0.585, respectively, which are all significant at the 1% level, and most of the counties are located in the first and the third quadrant, which form either High–High (H-H) agglomerations or Low–Low (L-L) agglomerations. Therefore, there is a significant spatial autocorrelation among the counties in the Loess Plateau as far as the VRPRD growth rate is concerned.

Specifically (Figure 8), within the four selected time nodes, 50~58 counties in the Loess Plateau’s central region showed H-H aggregation, indicating that vegetation restoration in this region is strong, with surrounding areas also exhibiting strong vegetation restoration. In contrast, 51~60 counties in the Loess Plateau’s northwest and northeast showed L-L aggregation, indicating that vegetation restoration in these areas is poor, as well as in the surrounding areas.

4.4.3. Model Selection for Spatial Econometric

Based on the identified spatial correlation of the VRPRD growth rate in the Loess Plateau, we further explored the spatial spillover effect of factors such as policy intensity on vegetation restoration using a spatial econometric model. To determine the optimal spatial econometric model, we employed the Lagrange multiplier (LM) test to assess the presence of spatial lag and spatial error terms. For both LM-error and Robust LM-error, the hypothesis of no spatial error term was rejected at the 1% significance level. Similarly, for LM-lag and Robust LM-lag, the hypothesis of no spatial lag term was also rejected at the 1% significance level. Additionally, the LR statistic values were 75.63 and 14.22, respectively, with significant p-values at the 5% level, rejecting the hypothesis that the SDM could be simplified to either the SEM or SLM model. The Hausman test statistic for this study was 134.75, with a p-value significance at the 1% level. Combining these various tests, the fixed-effect spatial Dubin model was selected for this study to investigate the factors influencing the effectiveness of vegetation restoration. To ensure robustness, we also compared the results with those from the SLM and SEM models, with the corresponding estimates provided in

Table 2. Regression results of SDM, SEM, and SLM models.

Variable	SDM VRPRD Growth	SEM VRPRD Growth	SLM VRPRD Growth
Intensity	0.160 ***	0.166 ***	0.164 ***
R&D	−0.001	0.003	−0.001
Temp	0.183 ***	0.169 ***	0.055 ***
Prec	−0.161 ***	−0.114 ***	−0.018
Sunp	0.057 **	0.050 **	0.050 ***
W × Intensity	−0.105 ***
W × R&D	−0.017 **
W × Temp	−0.194 ***
W × Prec	0.180 ***
W × Sunp	−0.054
rho	0.809 ***	0.824 ***	0.805 ***
sigma2	0.006 ***	0.006 ***	0.006 ***
N	341	341	341

Note: Significance levels are indicated as *** for 1%, and ** for 5%.

.

4.4.4. Results Analysis of the Spatial Dubin Model

Table 2 shows the fitting results of different models, with the estimation results of SDM showing minimal differences from those of other models, indicating that SDM has a certain degree of robustness. The spatial correlation coefficient was 0.809, significant at the 1% level, suggesting that improvements in vegetation cover in neighboring areas contribute to improvements in this area’s vegetation cover, leading to a clustering effect in the spatial layout. A spatial spillover effect is also evident in vegetation restoration across the Loess Plateau. All variables, except for road density, passed the 1% or 5% significance tests. However, each variable exerts a different effect, direction, and mode of influence. The marginal effect of specific variables needs to be further decomposed by solving for partial differentials (

Table 3. Decomposition of direct and indirect effects of SDM.

Variable	Direct Effect	Indirect Effect	Total Effect
Intensity	0.376 ***	−0.607 ***	−0.231
R&D	−0.035 **	−0.180 ***	−0.215 ***
Temp	−0.187 **	0.265 **	0.078
Prec	1.205 ***	−0.182	1.023 ***
Sunp	0.670 ***	−0.487 ***	0.182 **

Note: Significance levels are indicated as *** for 1%, and ** for 5%.

).

In terms of total effect, both the annual average precipitation and annual sunshine percentage significantly contributed to the enhancement of the vegetation restoration effect across the Loess Plateau. The annual average precipitation had the highest degree of influence on the improvement of the vegetation restoration effect, while the road density showed a significant negative effect. As the increase in road density will lead to land fragmentation, it does not contribute to the improvement of vegetation cover. Thus, when implementing GFGP on the Loess Plateau, it is important to avoid high-density road traffic on the surface, reduce land fragmentation, and provide a good green space environment for vegetation growth.

In terms of direct effect, the increase in annual average precipitation can better promote the vegetation restoration effect for the region. There is an arid and semi-arid climate on the Loess Plateau, and water scarcity has been recognized as a limiting factor for vegetation restoration. Therefore, an increase in annual average precipitation can alleviate the current water scarcity and promote vegetation restoration on the Loess Plateau. Light is an important basis for photosynthesis in vegetation, so an increase in annual percent sunlight has a significant effect on vegetation recovery. Policy intensity represents the level of local government investment in GFGP implementation, and increased policy intensity will also result in the increased restoration of vegetation. An excessive number of roads can lead to the fragmentation of land, thus destroying the growing environment of vegetation. There is a significant negative impact of road density on vegetation restoration. The increase in annual average temperature will lead to more evapotranspiration, which will make the Loess Plateau, which is already in an arid and semi-arid region, face even more severe water shortages, which will not be conducive to the recovery of vegetation.

In terms of indirect effect, under the adjacency matrix, policy intensity, road density, and annual sunshine percentage in neighboring areas had a significant negative effect on native vegetation recovery. This is because neighboring areas that implement a strong GFGP may result in competition for limited water resources and soil nutrients with the growth of vegetation in the area. An increase in annual sunshine percentage will prolong the growth time of the vegetation, also competing for water and soil nutrients from the surrounding area, all of which are detrimental to the recovery of vegetation in the region. The connectivity of roads in neighboring areas, often forming cross-regional road networks, also leads to the fragmentation of the land on which vegetation grows in the region, thus limiting vegetation growth. An increase in annual average temperature usually changes the local climate, increasing evapotranspiration while bringing rainfall to the local area through air circulation, thus to some extent alleviating water stress for the growth of vegetation in the region.

5. Discussion

5.1. Evaluation of the Grain for Green Program’s Effects on the Loess Plateau

During the study period, the vegetation coverage level of the Loess Plateau increased significantly, and the vegetation restoration effect was the best in the central region. Due to resource pressure from population growth and economic development, before the implementation of GFGP in this region, a large amount of vegetation was cut down for production materials and domestic fuel, and a large amount of sloped land was reclaimed for farming. This resulted in serious damage to vegetation and low vegetation coverage [7]. In order to curb soil erosion in the region, the Chinese government took the lead in implementing GFGP in the region. Due to the longer and larger area of GFGP implementation, the vegetation restoration effect is most significant here, and the EVI growth rate is also faster [41]. However, the vegetation index is only a simple representation of vegetation cover, which is influenced by natural resource endowment conditions and human policy measures on vegetation restoration [29]. Only focusing on EVI growth rate cannot fully eliminate the role of resource endowment, which may lead to the incorrect estimation of GFGP’s effects on the Loess Plateau. This study showed that the EVI growth rate is more affected by natural resource endowment compared to the VRPRD growth rate. Consequently, the EVI growth rate shows stronger absolute correlation coefficients with meteorological variables. Furthermore, the GFGP intensity variable was able to explain 18.3% of the VRPRD growth rate, which is 4.4% higher than the EVI growth rate. The implementation of GFGP to promote vegetation restoration in the Loess Plateau has been widely recognized by the academic community [42,43,44]. However, the analysis results of influencing factors in most studies show that natural conditions such as climate are the main factors affecting the improvement of vegetation coverage, rather than GFGP input [45]. This shows that EVI or NDVI indicators are greatly affected by resource endowment conditions, and the policy effect is submerged in them, which is difficult to reflect. Therefore, compared with other scholars’ research, the GFGP’s effects evaluation index (VRPRD) proposed in this study is more scientific and credible.

5.2. Spatial Dependence of Vegetation Restoration Effect

The global Moran index statistics show that the vegetation restoration effect of the Loess Plateau has a strong positive spatial correlation, which means that the vegetation restoration of each county is not isolated. Changes in vegetation restoration in the target county may significantly affect the effect of vegetation restoration in neighboring counties. Some scholars verified this conclusion when studying the vegetation restoration of the Loess Plateau [46], and Chen et al. (2017) calculated that each increase in the vegetation restoration level by 1 unit in the target county would increase the vegetation restoration effect of neighboring counties by 0.2 [47]. Based on this positive correlation, there may be a positive aggregation effect during the implementation of GFGP in various counties. From 2005 to 2015, the global Moran index showed an increasing trend because neighboring counties would learn from each other, share the successful experience of GFGP implementation, and carry out healthy competition, which effectively restored the vegetation on the Loess Plateau [48]. At the same time, the target county and neighboring counties cooperate and interact with each other in the process of policy implementation, which will have important guiding significance for optimizing the design and implementation of GFGP on the Loess Plateau. From 2015 to 2020, the global Moran index decreased, which may be related to the fact that GFGP entered the consolidation period, large-scale afforestation was no longer implemented in various places, and the spatial dependence of vegetation restoration was weakened [49]. From the perspective of spatial Dubin results, this spatial dependence is also reflected in the influencing factors. Relevant studies show that the natural and socio-economic factors in the target area will have an impact on the vegetation restoration effect in adjacent areas [50]. This further confirms the results of this study, namely that the annual average precipitation and annual sunshine percentage in the target county have a positive effect on the vegetation restoration of neighboring counties, while the road density in the target county has a negative effect on the vegetation restoration of neighboring counties.

5.3. Policy Recommendations

The results show that the annual average precipitation and annual sunshine percentage significantly contributed to vegetation recovery, while road density had a significant spatial inhibitory effect. In addition, some scholars have counted the tree species planted by GFGP on the Loess Plateau, including robinia pseudoacacia, populus simonii, pinus tabulaeformis, armeniaca sibirica, hippophae rhamnoides, and other trees, among which robinia pseudoacacia is a common afforestation tree [51]. Chen et al. (2022) also confirmed that robinia pseudoacacia was the main planted tree species in Yan’an during field investigation [52]. As can be seen from Figure 6, the vegetation restoration effect is the most significant in the central part of the Loess Plateau, especially in Yan’an. Combined with previous studies, it can be concluded that robinia pseudoacacia is the main contributing tree species to the improvement of local vegetation cover. Building on these results, the following policy recommendations are proposed to enhance vegetation restoration on the Loess Plateau. Firstly, in cases where it is difficult to change the local climate conditions, areas with poor vegetation restoration but better water resource conditions should improve the implementation of the GFGP to avoid the phenomenon of only retiring but not returning. Secondly, choose suitable tree species to reduce vegetation water consumption and evaporation. In general, the construction of artificial vegetation improves the natural environment to a certain extent, but also poses greater challenges to the utilization and allocation of local water resources. Facing the current situation of water shortage in the Loess Plateau, the successful experience of implementing GFGP in Yan’an is the selection of drought-tolerant tree species of robinia pseudoacacia. Therefore, when planting artificial forests in the above areas, adhere to the principle of matching trees to the sites and prioritize drought-tolerant species with lower water requirements rather than high-water-consuming species. Thirdly, protect the environment for vegetation growth and plan road construction scientifically. Excessive road density can negatively affect the growth of vegetation; the scientific basis and necessity of road construction in forest areas must be demonstrated to reduce land fragmentation and protect forest ecosystems.

5.4. Limitations and Prospects

This study used a similar habitat potential model based on a local sliding window to overcome the influence of natural endowment conditions on vegetation restoration, thereby enabling a more accurate assessment of ecological policy effects. This approach is particularly valuable for evaluating VRP in arid and semi-arid regions. Additionally, the spatial spillover effect of vegetation restoration on the Loess Plateau was analyzed using a spatial Dubin model. However, there are some limitations in our study that warrant further investigation. Firstly, this study explored the spatial spillover effect of vegetation restoration based on the geographic adjacency matrix, but the economic level of counties also impact policy intensity. Future research should incorporate an economic distance matrix to explore the spatial correlation of vegetation restoration on the Loess Plateau. Secondly, the influencing factors of vegetation restoration still need to be improved. Water carrying capacity is a crucial factor in vegetation restoration [53]. Although the construction of artificial vegetation can improve the natural environment to some extent, it also increases the challenge of utilizing and allocating local water resources. Therefore, future research should include data on water resource carrying capacity to enhance the reliability of regression results. Thirdly, because it is difficult to obtain specific information about afforestation species, the vegetation restoration effect obtained in this study is actually the result of all tree species working together. However, the vegetation restoration effect of different tree species is obviously different [54]. Therefore, when evaluating the effect of GFGP implementation in the future, as much information about tree species as possible should be collected to evaluate the vegetation restoration effect under different tree species. This will not only improve the evaluation of the effectiveness of the GFGP implementation, but will also provide more detailed and specific guidance for the future implementation or design of the GFGP.

6. Conclusions

Firstly, this research analyzed vegetation restoration spatiotemporal patterns on the Loess Plateau using the EVI dataset from 2000 to 2020. Secondly, the VRPRD of the Loess Plateau was calculated using similar habitat theory and the sliding window model to evaluate the effectiveness of GFGP implementation, and used empirical analysis to test the accuracy of the new metrics. Finally, the spatial Dubin model was employed to assess the spatial spillover effect of VRP on the Loess Plateau. The following are the main conclusions drawn from this study: (1) the green territory of the Loess Plateau gradually expanded to the northwest between 2000 and 2020, with the central parts of the region experiencing the greatest increase. In northern Shaanxi, the junction of Shaanxi and Shanxi, and the junction of Shaanxi, Gansu, and Ningxia, EVI increased by more than 40%, rising from 0.25 to 0.35; (2) the Loess Plateau VRP still has room for further improvement. Specifically, the Loess Plateau’s eastern and western regions have a VRP higher than 0.45 but a VRPRD lower than 0.65, indicating that these areas should be the focus of future vegetation restoration efforts; (3) by using the local sliding window model, it is possible to isolate the effects of resource endowment conditions, allowing for an accurate assessment of the vegetation restoration improvements driven by GFGP. This contributes to the evaluation of future ecological policies by proving both theoretical and methodological support; (4) policy intensity, annual average temperature, and annual sunshine percentage significantly promote VRP on the Loess Plateau. Additionally, annual average precipitation and annual sunshine percentage have a significant positive spatial spillover effect on the VRP.