Is There Spatial Dependence or Spatial Heterogeneity in the Distribution of Vegetation Greening and Browning in Southeastern China?

Abstract

Vegetation is an indispensable component of terrestrial ecosystems and plays an irreplaceable role in mitigation of climate change. Global vegetation changes (i.e., greening and browning) still occur frequently, however, little is known about the spatial relationships between these two processes. Based on the normalized difference vegetation index (NDVI) dataset from 1998 to 2018 in Fujian Province, China. The Theil-Sen and Mann-Kendall tests were used to explore temporal changes in vegetation growing, then the spatial relationships of greening and browning was distinguished with bivariate spatial autocorrelation analysis, and the spatial variation in the relationship between vegetation changes and driving factors was explored by the geographical detector. The results showed that from 1998 to 2018, the average NDVI value increased from 0.75 to 0.83; 89.61% of the study area experienced vegetation greening, while 5.7% experienced significant browning, with active vegetation changes occurred along roads and nearby cities. The spatial autocorrelation results showed that the spatial relationships between vegetation greening and browning were dominated by spatial heterogeneity (i.e., high greening and low browning, H-L clusters accounting for 60% and low greening and high browning, L-H clusters accounting for 14%), but we also revealed that there were still quite a few places (4%) with spatial dependence (i.e., high greening and browning, H-H clusters), occurring around urban areas and along roads. The factor detector indicated that the nighttime light intensity was among the most dominant factor of vegetation changes, followed by elevation and slope. Although the individual effect of the distance to roads was relatively weak on the vegetation changes, its indirect effect was found to be the strongest by the interaction detector, which was obtained from the interactions much larger than its independent impact. Simultaneously, the risk detector revealed that the greening preferred occurring in places with lower nighttime light intensity (<1.1 nW cm⁻²sr⁻¹), higher elevation (>43.4 m) and slope (>6.3°). Moreover, we found that the vegetation changes primarily occurred within a distance of 1685.4 m from roads. Our findings could deepen the understanding of vegetation change patterns and provide advice for mitigating the impact on the vegetation changes.

1. Introduction

Vegetation plays a critical role in carbon sequestration, global warming mitigation, and biodiversity protection [1,2]. It promotes ecosystem services, which creates a more livable environment for humans [3]. However, human disturbances, such as urbanization, industrialization, and infrastructure construction, have challenged the normal function of vegetation. Due to deforestation, 420 million ha of forest has been lost globally from 1990 to 2020. Even if it was lower than in previous years, the annual loss rate was 10 million from 2015 to 2020 [4]. Vegetation variation is a threat to changes in landscape structure and consequently affects the associated ecological process [5]. Vegetation changes include both greening and browning, which are different ecological processes with different occurrence and evolution mechanisms. To achieve the goal of the overall improving trend of vegetation in a region, the occurrence and development of these two different ecological processes may be spatially related to a certain extent [6]. However, the spatial relationships between greening and browning are less understood. Few studies have focused on the process of further geographic change and the exploration of spatial relationships (i.e., spatial heterogeneity or spatial dependence) between greening and browning. Therefore, understanding the spatial connection and driving forces of greening and browning will help cope with environmental changes and promote sustainable development.

It is essential to grasp the temporal and spatial dynamics of the changes in vegetation cover, which can provide a scientific reference for the adjustment and tendency of protection policies [7]. The normalized difference vegetation index (NDVI), one of most commonly used vegetation index, which is obtained from the reflection difference between red light and near-infrared light by vegetation [8] However, due to the issue of saturation at higher biomass levels, the effect of soil brightness and the noise of atmospheric conditions, various alternative indices have been proposed, such as the enhanced vegetation index (EVI), which can improve the saturation at higher biomass levels, and the Soil Adjusted Vegetation Index (SAVI), which is a better indicator in lower vegetation cover [9,10]. But the monitoring accuracy will be reduced for the sensitivity of EVI to illumination conditions and the inefficient of SAVI to heterogeneous canopy [11,12]. NDVI is less sensitive to topographic illumination, illumination conditions and shading effect than EVI [11,13,14]. Moreover, NDVI can remove a large part of the noise of clouds and solar angle [15]. The characteristics above make NDVI more suitable for this study. The widely used long time-series NDVI products mainly derived from Advanced Very High-Resolution Radiometer (AVHRR) Global Inventory Modeling and Mapping Studies (GIMMS), Moderate Resolution Imaging Spectroradiometer (MODIS) and Système Probatoire d’Observation de la Terre (SPOT), of which GIMMS has a long data record but coarse resolution of 8 km, while MODIS products have a higher resolution of 250 m [16]. Previous comparisons of the three products noted that the differences in NDVI was various [17]. Except that GIMMS should be corrected due to inherent shortcomings, MODIS NDVI and SPOT NDVI have better consistency than GIMMS, and better performance in long-term vegetation monitoring [13,18].

Various methods are used for analysis of vegetation change dynamics in long time-series, including linear regression, the Theil-Sen and Mann-Kendall test, among which the latter has been widely used, due to its advantages in outliers’ avoidance [19,20]. For example, Ren et al. [21] used Theil-Sen and Mann-Kendall test to observe the NDVI trend in the Yellow River Basin from 2000 to 2020, concluding that the spatial distribution of NDVI increased from northwest to southeast; Dagnachew et al. [22] used the test and revealed that significant increase and/or decrease in annual NDVI and rainfall in Gojeb River Catchment from 1982 to 2015; using the Theil-Sen and Mann-Kendall test, Chen et al. [23] also found that the vegetation improvement zones were larger than those of the degradation in the farming-pastoral ecotone in northern China from 1982 to2017. Therefore, the Theil-Sen and Mann-Kendall test were effective tools to examine vegetation temporal change patterns owing to the robust against non-normal and heteroskedastic data [24].

Greening and browning are complicatedly affected by various factors with strong spatial heterogeneity, so the driving mechanism may vary across different regions and over periods [25,26]. In the response of vegetation changes to climatic conditions, most studies employed temperature (TMP) and precipitation (PRE) as proxies. At the national scale, changes in vegetation in southern and eastern China are dominated by TMP, while PRE occurs in the north [27,28,29]. In ecologically fragile areas, such as the Loess Plateau and the Tibetan Plateau, climate disturbances are more complicated [30,31]. Topographic factors also directly or indirectly affect vegetation changes, including slope, aspect, and elevation [32]. It is also undeniable that human activities play an important role in vegetation changes. It has been reported that the impact of human interference can be effectively reflected by nighttime light (NTL) data, which is also an indicator of urban sprawling [33].

In the analysis of the driving forces of vegetation change, correlation and partial correlation analysis as well as multivariate regression analysis are often employed [34,35]. However, due to the complicated influence process caused by spatial heterogeneity, there is no strict linear relationship between NDVI changes and impact factors; that is, it is not an independent effect of a single factor but even a synergistic effect of multiple factors. In this context, geographical detector (GD), a tool for detecting and exploiting spatial differentiation, was proposed to explore spatial variation in the relationship between vegetation changes and driving factors in this study [36]. GD is a statistical method to identify the relationship between independent (driving factors) and dependent variables (greening or browning) by similarity of spatial distribution, which fully considers geographic spatial association between variables [37,38]. In addition to analyzing the explanatory power of a single factor, it can also identify the gradient characteristics of each factor on vegetation changes and the interaction of various factors [39]. Furthermore, GD has been applied maturely in the fields of human health, social sciences and ecological environment. For example, Huang et al. [40] used GD to identify the influencing factors of hand, foot and mouth disease in China. They found that the dominant factor in the occurrence of the disease was GDP, and the interaction of temperature and precipitation caused the highest disease risk. Wang et al. [41] evaluated the determinants of housing prices in China based on GD. They concluded that at the county level, the cost of land had the strongest explanatory power for China’s housing prices, and the interaction between the proportion of renters and the state of the housing market had a greater effect on prices. Yuan et al. [42] explored the relationship between NDVI changes and environmental factors in the Heihe River Basin with the help of GD. They believed that precipitation in the Heihe River Basin played a major role in NDVI change, while the leading interaction in different subbasins was different. Hence, geographic detector with the advantage of two-factor mutual detection is suitable for the exploration of the internal driving mechanism of various fields, which is achieved by quantifying the spatially stratified heterogeneity.

Fujian Province is located in a subtropical region with abundant vegetation resources and rich biodiversity. As one of the four major forest areas in China, it has accumulated a vast amount of carbon storage [43]. As with other provinces in China, it has implemented restoration projects constantly, such as the “National Forestation Program”, “Natural Forest Conservation Program” and “Sloping Cropland Conversion Program”, which is beneficial to the local area [44]. Even if the forest cover increases continuously, there are still active changes to an extent (that is, greening and browning occur synchronously) [45,46]. Different from the existing research focusing on key protected, environmentally fragile and economic central areas, this study supplements the vacancy of areas with a certain ecological foundation and ecological orientation that are easily overlooked and strengthens the comprehension of the impact of vegetation changes on the southeast coast of China [47,48,49]. Understanding the change process and driving factors is rewarding for decision-makers to improve or adjust protection measures.

Spatial autocorrelation provides a tool for denoting the correlation and geospatial characteristics between variables, which can realize the visualization of change characteristics on a certain scale [50]. It can reveal the clustering feature of vegetation changes from a global and local perspective [51]. To fill the deficiency in the knowledge on the spatial heterogeneity or spatial dependence between greening and browning, the Theil-Sen and Mann-Kendall test were used to explore temporal changes in vegetation greening and browning, the spatial relationships of greening and browning were explored by the bivariate spatial autocorrelation analysis. Then, the geographical detector was applied to explore the problem of spatial heterogeneity and the interaction of influencing factors. Therefore, the purposes of this study are to: (1) locate greening and browning; (2) identify spatial relationships between greening and browning; and (3) identify spatial variation in the relationship between driving factors and vegetation changes. Our study can provide implications for vegetation protection and restoration.

2. Materials and Methods

2.1. Study Area

Located on the southeast coast of China, west of Taiwan, Fujian Province (23°30’−28°19’ N and 115°50’−120°47’ E), covers an area of over 120,000 km² (Figure 1). The province generally presents the characteristics of high terrain in the northwest and low terrain in the southeast. Eighty percent of the whole area is covered by mountains and basins, and the eastern coastal plains are densely distributed in central cities. Fujian has a temperate oceanic climate, with sufficient sunlight and abundant rainfall. The annual average temperature is 18–26 °C, and the average rainfall is 1300–2000 mm, which is suitable for vegetation growth. As a key demonstration area for the construction of ecological civilization, it is the provincial administrative region with the highest forest coverage in China, accounting for as much as 66.8%.

2.2. Data Sources and Processing

2.2.1. NDVI

Considering the large scope of the study area and the consistency of pixel accuracy with driving factors, which avoids errors caused by resampling, SPOT NDVI dataset was selected. The source of NDVI datasets (1998–2018) (1 km) was the Resource and Environment Science and Data Center [52]. Based on SPOT/VEGETATION satellite remote sensing data, maximum value composites (MVC) were used to generate annual vegetation index datasets, which means that the image was represented by the maximum NDVI value in each year. This method can reduce the influence of clouds and atmospheric conditions [53].

2.2.2. Driving Factors

Geomorphic, meteorological and anthropogenic factors are all related to vegetation changes [54]. Based on the above three categories, 10 variables of 7 elements were selected (

Table 1. Data sources and details.

Type	Data	Time Scale	Data Sources	Preprocessing	Variable	Resolution
Vegetation data	NDVI	1998–2018	Resource and Environment Science and Data Center (https://www.resdc.cn/, accessed on 21 January 2021)	The image acquired by maximum value composites (MVC)	NDVI	1 km
Geomorphic factors	DEM	/	Geospatial Data Cloud site, Computer Network Information Center, Chinese Academy of Sciences (http://www.gscloud.cn, accessed on 26 September 2021)	Surface analysed by GIS	Elevation	30 m
					Slope
					Aspect
Meteorological factors	Monthly average temperatures	1998–2018	National Earth System Science Data Center (http://www.geodata.cn, accessed on 26 September 2021)	Annual change rate computed by Theil-Sen and Mann-Kendall test	Annual average temperature (TMP_A)	1 km
					Temperature change trend (TMP_T)
	Monthly precipitation	1998–2018			Annual average precipitation (PRE_A)
					Precipitation change trend (PRE_T)
Anthropogenic factors	Nighttime light intensity	2000–2018	National Earth System Science Data Center (http://www.geodata.cn, accessed on 26 September 2021)	Annual change rate computed by Theil-Sen and Mann-Kendall test	Annual average nighttime light (NTL_A)	500 m
					Nighttime light change trend (NTL_T)
	Road	2018	OpenStreetMap (http://www.openstreetmap.org, accessed on 26 September 2021)	Euclidean distance computed by GIS	Distance to road (Distance)	Vector

). All of the acquired data were extracted with the administrative district vector boundary. Monthly TMP and PRE datasets were provided by the National Earth System Science Data Center, National Science & Technology Infrastructure of China. The dataset was generated using delta spatial downscaling from the 30′ Climatic Research Unit (CRU) time series dataset and the climatology dataset of WorldClim (https://www.worldclim.org/, accessed on 26 September 2021), which was verified by the observations of 496 weather stations in 1951–2016 across China [55]. Annual TMP and PRE data were obtained through extraction and processing of monthly data by R: a free software environment for statistical computing and graphics and ArcGIS (Esri, CA, USA). Additionally, from the above platform, we acquired the extended timeseries of NTL data (2000–2018) produced by Chen et al. [56]. This dataset was generated using a new cross-sensor calibration with the Defense Meteorological Satellite Program Operational Linescan System stable nighttime light (DMSP-OLS NTL) data (2000–2012) and the composition of monthly Suomi National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite nighttime light (NPP-VIIRS NTL) data (2013–2018), which presented good accuracy at pixel, city and regional scales [57]. Additionally, TMP, PRE, and NTL intensity were presented in two forms, including average value and growth rate, respectively. The latter was the calculated change rate (TMP_T, PRE_T and NTL_T) using the Theil-Sen and Mann-Kendall tests, which were introduced in the next section, and the former was the calculated multiyear average (TMP_A, PRE_A and NTL_A). The digital elevation model (DEM) data collected from the Resource and Environment Science and Data Center were preprocessed by ArcGIS to obtain elevation, slope and aspect data. The road network data were derived from OpenStreetMap, and the distance to road (Distance) was calculated by the Euclidean Distance tool in ArcGIS. All of the above data were converted to the coordinate system CGCS2000 and resampled to a spatial resolution of 1 km × 1 km.

2.3. Method

2.3.1. Theil-Sen and Mann-Kendall Test

The Theil-Sen and Mann-Kendall test, two nonparametric methods, are widely used in the study of vegetation long-term series due to their advantages in handling outliers and no requirement on the normal distribution of data [19]. Theil-Sen can be calculated by the following median function:

$$\begin{gathered} \mathrm{s}=Median\left(\frac{x_j-x_i}{j-i}\right) \\ 1998<i<j<2018 \end{gathered}$$

(1)

where $x_j$ and $x_i$ represent the NDVI of the $j$ th and $i$ th years, respectively. When $s$ > 0, the NDVI in this time series shows an increasing trend. Otherwise, it is a decreasing trend.

The Mann-Kendall test was used to evaluate the significance, which Theil-Sen cannot detect by itself. To perform the test, the null hypothesis (H₀) and alternative hypothesis (H₁) are designed [58]. The calculation of the test is specified as follows:

Statistic $S$ is first calculated as follows:

$$S={\displaystyle \sum }_{i=1}^{n-1}{\displaystyle \sum }_{j=i+1}^{n}sgn\left(x_j-x_i\right)$$

(2)

$$sgn\left(x_j-x_i\right)=\left\{\begin{array}{c} 1, x_j-x_i>0\\ 0, x_j-x_i=0\\ -1, x_j-x_i<0\end{array}\right.$$

(3)

when $n$ ≥ 10, the $S$ statistic approximately follows a normal distribution. Then, $S$ is standardized to $Z$, which is applied for the significance test. $Z$ value is defined as follows:

$$Z=\left\{\begin{array}{c}\frac{S-1}{\sqrt{VAR\left(S\right)}}, S>0\\ 0, S=0\\ \frac{S+1}{\sqrt{VAR\left(S\right)}}, S<0\end{array}\right.$$

(4)

$$VAR\left(S\right)=\frac{n\left(n-1\right)\left(2n+5\right)-{\displaystyle \sum _{i=1}^m}{t}_i\left(t_i-1\right)\left(2t_i+5\right)}{18}$$

(5)

where $n$ is the number of study years, $m$ is the number of tied groups (the number of repeated data points), and $t_i$ indicates the number of ties of Group $i$ (the number of repeated data points in Group $i$).

In this study, significance level α = 0.05 was chosen. Therefore, if | $Z$ | > 1.96, the trend is significant; otherwise, it is not.

Theil-Sen and Mann-Kendall test were conducted by the Python package pyMannKendall (version 1.2) to analyse the change trend of NDVI(s) at the pixel level [59]. Moreover, the three variables (TMP_T, PRE_T and NTL_T) mentioned above were calculated in the same way.

2.3.2. Spatial Autocorrelation Analysis

Spatial autocorrelation can be used to determine the geographical relevance of attribute values between a spatial unit and neighboring spatial units, which can describe a single or multiple variables [60]. Its magnitude is measured by global Moran’s I value and local Moran’s I value, which evaluate global and local correlations, respectively. Global Moran’s I revealed the correlation of vegetation changes across the study area, while the local Moran’s I denoted the aggregation characteristics of local vegetation changes. To evaluate the relevance between vegetation greening and browning, bivariate spatial autocorrelation was selected [61]. Since the spatial autocorrelation analysis in the same area tends to be affected by the “scale effect”, 1 km × 1 km, 2 km × 2 km, 3 km × 3 km, 4 km × 4 km, 5 km × 5 km fishnets were created as spatial analysis units. Then, the area proportion of vegetation greening (s > 0) and browning (s < 0) in each unit was calculated using the Zonal Statistics tool of ArcGIS. Thus, the bivariate spatial autocorrelation was processed using GeoDa software (The Center for Spatial Data Science, Chicago, IL, USA) [62]. The global Moran’s I value obtained from GeoDa can determine the appropriate spatial units, and spatial dependence (i.e., high greening and high browning, H-H clusters, and low greening and low browning, L-L clusters) and spatial heterogeneity (i.e., high greening and low browning, H-L clusters, and low greening and high browning, L-H clusters) of vegetation changes can be visualized by LISA map.

2.3.3. Geographical Detector

The conceptual framework of GD is to reveal the driving mechanism by quantifying the spatially stratified heterogeneity between the vegetation changes and the explanatory factors, which consists of four types of detectors: factor detectors, interaction detectors, risk detectors, and ecological detectors. The former three types of detectors were analyzed in this study, which was performed by R [63].

The factor detector is the core component, which reveals the relative importance of the explanatory variables. The q value is computed by the following formula for quantitative determination as follows:

$$q=1-\frac{{\displaystyle \sum }_{j=1}^M{N}_j{\sigma }_j^2}{N{\sigma }^2}$$

(6)

where $N_j$ and ${\sigma }_j^2$ are the number of samples and variance of the independent variables (explanatory factors) within j strata, and N and ${\sigma }^2$ are the number of samples and variance of dependent variable (s) within the entire study area. The larger the q value is, the greater the importance of the explanatory variable.

Interaction detector is a tool for judging the role of any two factors. The q value of the interaction and independence were compared to illustrate whether the effect of two factors on NDVI was enhanced, weakened or independent [64]. Therefore, the results are composed of the q value and the types of interaction (

Table 2. Description of interaction type.

Criterion	Interaction Types
$q\left(X1\cap X2\right)>Max\left(q\left(X1\right),q\left(X2\right)\right)$	bi-variable enhance
$q\left(X1\cap X2\right)>q\left(X1\right)+q\left(X2\right)$	nonlinear-enhance
$q\left(X1\cap X2\right)=q\left(X1\right)+q\left(X2\right)$	independent
$q\left(X1\cap X2\right)<Min\left(q\left(X1\right),q\left(X2\right)\right)$	nonlinear-weaken
$Min\left(q\left(X1\right),q\left(X2\right)\right)<q\left(X1\cap X2\right)<Max\left(q\left(X1\right),q\left(X2\right)\right)$	uni-variable weaken

).

Risk detection can verify whether there is a significant difference in the mean s between any two subregions of each factor. The greater the difference in the mean s between subregions was, the greater the impact on the trend of greening or browning that occurred of the subregion [38]. On the one hand, the mean s in each subregion can be calculated to further understand the response of browning and greening to the change in a specific factor. On the other hand, the significance statistic of the mean s is achieved by t test [65]:

$$t=\frac{{\overline{Y}}_{h=1}-{\overline{Y}}_{h=2}}{{\left[\frac{Var\left({\overline{Y}}_{h=1}\right)}{n_{h=1}}-\frac{Var\left({\overline{Y}}_{h=2}\right)}{n_{h=2}}\right]}^{1/2}}$$

(7)

where ${\overline{Y}}_h$ represents the mean value of NDVI change (mean s) in subregion h; $n_h$ is the number of samples in subregion h; and Var means variance.

The calculation of degree of freedom is as follows:

$$df=\frac{\frac{Var\left({\overline{Y}}_{h=1}\right)}{n_{h=1}}+\frac{Var\left({\overline{Y}}_{h=2}\right)}{n_{h=2}}}{\frac{1}{n_{h=1}-1}{\left[\frac{Var\left({\overline{Y}}_{h=1}\right)}{n_{h=1}}\right]}^2+\frac{1}{n_{h=2}-1}{\left[\frac{Var\left({\overline{Y}}_{h=2}\right)}{n_{h=2}}\right]}^2}$$

(8)

To test the significance, a null hypothesis (H₀) of ${\overline{Y}}_{h=1}$ = ${\overline{Y}}_{h=2}$ is made. Given the significance level α = 5%, if |t| > t_α/2, H₀ is rejected, which means there is a significant difference between subregion 1 and subregion 2.

3. Results

3.1. Spatiotemporal Change Dynamics of the NDVI

The average NDVI values increased from 0.75 in 1998 to 0.83 in 2018, increasing by 10.7%, with a slope of 0.00412 (Figure 2). This illustrated that the vegetation coverage fluctuated to improve and maintained at a relatively high level during the past 21 years in Fujian Province. The change trend of NDVI can be divided into three stages: the first stage was composed from 1998 to 2002, with the NDVI value oscillating in this stage. From 2002 to 2003, the NDVI value increased sharply, and then it was relatively stable from 2003 to 2012. Since 2012, the NDVI has increased, with high growth rates predicted for the next nine years.

Theil-Sen and Mann-Kendall test were combined to calculate the change trend of NDVI from 1998 to 2018.

Table 3. Trend analysis statistics of NDVI.

Trends	Description	The Proportion (%)
$S<0,\left|Z\right|>1.96$	Significant browning	5.70
$S<0,\left|Z\right|<1.96$	Slight browning	4.69
$S>0,\left|Z\right|<1.96$	Slight greening	6.77
$S>0,\left|Z\right|>1.96$	Significant greening	82.84

revealed that 89.61% of the study area presented a positive trend in vegetation coverage, including 82.84% significant greening and 6.77% slight greening, while 10.39% of the study area was characterized by browning, including 5.70% significant browning and 4.69% slight browning. The outcome was in line with the ecological positioning of Fujian, which is the most greening province in China. As shown in Figure 3, browning areas were concentrated in the eastern coastal area, with a higher developed level. Meanwhile, we can also see that the browning areas were star-shaped distributed in the urbanization area and the belt-shaped distributed on both sides of the road networks.

3.2. Spatial Autocorrelation Analysis of Vegetation Greening and Browning

Based on the spatial distribution features of greening and browning mentioned earlier, we took a step forward to use bivariate spatial autocorrelation to explore the spatial relationships between these two processes. The global Moran’s I value of five analysis units from 1 km to 5 km was −0.67,−0.5, −0.56, −0.61 and −0.62, respectively. The results indicated that the 1 km × 1 km scale had the most significant cluster, with the largest absolute value of Moran’s I at all scales. Combined with the local correlation features of the five spatial units which are shown in Figure 4, 1 km × 1 km was chosen as the study scale. The results showed that the spatial relationships between greening and browning were dominated by spatial heterogeneity, with H-L accounting for 60% and L-H for 14%, however, a few places also showed spatial dependence, with H-H accounting for 4%. Additionally, H-L was widely distributed throughout the study area, and there was a synchronized phenomenon of the L-H and H-H clusters, which was the law of H-H surrounding L-H, along the road and around the eastern coastal area with a high level of urbanization. In other words, the characteristics of high greening and high browning occurred on both sides of the road and edge of the coastal cities.

3.3. Driving Force Analysis of Vegetation Cover Changes

3.3.1. Individual Effect

The factor detector is a key detector used in a single factor importance evaluation, and the variables indicated in Figure 5 were all statistically significant (p < 0.01) for the vegetation changes. Figure 5 reveals that anthropogenic and geomorphic factors had stronger explanatory power than meteorological factors. NTL_T and NTL_A had the highest q values (0.454 and 0.453, respectively), followed by elevation, slope and meteorological factors, indicating that NTL was the dominant factor in greening and browning. Among the topographical factors, elevation and slope had a greater impact on NDVI changes than aspect, with q values of the former being 0.311 and 0.273 and q values of the latter being only 0.007. In addition, although the q values of TMP_A, PRE_A and TMP_T were relatively small, they still had a significant impact on the NDVI changes in the study area. Relative to TMP, PRE_T had little effect on the NDVI in the period (q = 0.023).

3.3.2. Interaction Effect

Interaction detection is used to evaluate the contribution of any two factors to the change in NDVI. There was bi-variable enhance and nonlinear-enhance between any two factors, and bi-variable enhance was the majority, with the interaction of any two factors on NDVI changes being much stronger than their respective effects (Figure 6a). The interaction of the NTL and the slope was among the strongest (q = 0.489), indicating that the two variables were the primary driving factors. Meanwhile, we compared the results of the interaction effects of each factor and their individual effects and found that although the individual effect of Distance was weak (q = 0.136), its interactions with other factors were much greater than its independent effect (Figure 6b). The interaction of Distance with NTL_A and Slope was much larger than the independent role of Distance, while the difference between the interaction of Distance and the meteorological factors and the independent effect of Distance was relatively small.

3.3.3. Spatial Stratification Effect

Risk detection can reveal the response of the mean s to the changes in subregions of each factor and measure the significance of the difference between the mean s of two subregions (Figure 7). With the increase in NTL_T and NTL_A, the greening trend weakened or even degraded, and the turnover point of NTL_A occurred between the ranges of 0.5 to 1.1 nW cm⁻²sr⁻¹ and 1.1 to 3.4 nW cm⁻²sr⁻¹. When NTL_T, the dominant factor, was small (0 < NTL_T < 0.018), and there were no significant NDVI changes, indicating that significant browning was associated with larger NTL_T and NTL_A. Browning mainly occurred in the east, as seen in Figure 7a,b, with the high browning areas concentrated in Fuzhou, Putian, Quanzhou, and Xiamen, which had a higher level of urbanization. Additionally, a belt-like distribution of browning and minor greening patches was noticed around roads. In terms of road effects, the greening trend that had the least influence was closest to the roads (0–235.9 m). As the distance from roads increased, the greening trend was more evident, and it reached stability when the distance exceeded 1685.4 m. Likewise, significant differences in the mean s were found near the road but not further away, implying that significant changes occurred within 1685.4 m. Furthermore, the road density was higher in the east than in the west, so the latter had a stronger greening trend than the former. (Figure 7c).

In terms of the geomorphic factors, the vegetation changed from browning to greening as the elevation increased, implying that the vegetation in lower elevation areas was degraded, and vice versa, with 43.4 m serving as the boundary. Statistical tests showed that the differences in the mean s were significant except for the areas with lower elevations (0~11.5 m), which exhibited a significant greening pattern at higher elevations. As a result, the western areas with higher elevations greened, whereas the eastern areas with lower elevations browned during the study period (Figure 7d). Similarly, as the slope increased, the mean s increased as well, reaching a maximum in the steep slope area (slope > 26.7°). The mean s was significantly different between all of the slope subregions. Figure 7e depicts the occurrence of browning in the eastern areas with relatively gentle slopes. Aside from the most significant browning trend in the plain area, the other aspects showed a greening trend, with the shady slopes displaying the largest greening trend (Figure 7f).

Regarding the climatic parameters, it was discovered that a notable greening trend occurred in regions with lower TMP_A and TMP_T. The greening trend was maximum in the areas with TMP_A between 16.6 and 17.4 °C. In contrast, the excessively high TMP_A and TMP_T made the mean s reach the minimum. Figure 7g,h reveals that the lowest greening trend was found in the southeastern part of the study area with the highest TMP_A and the central and eastern regions with the highest TMP_T, while the highest greening trend occurred in the western and northern regions with relatively low TMP_A and the lowest TMP_T. Except for the browning trend when PRE_A was relatively low (1004.4 mm~1223.8 mm), sufficient precipitation played a positive role in vegetation change. The mean s reached a peak value when the PRE_A was between 1882.1 and 2101.5 mm. Figure 7i indicates that browning occurred in the southeastern part of the study area with the lowest PRE_A, while the highest greening occurred in the northern part of the study area with a relatively high PRE_A. The areas with the lowest PRE_T were distributed from southeast to northwest, which showed the lowest greening trend, while the highest greening trend occurred in the southwest and northern regions with relatively high PRE_T. It should be noted that there was no significant difference in the mean s of most subregions of PRE_T with a lower q value (Figure 7j).

4. Discussion

4.1. The Spatial Relationships between Greening and Browning

As a key link in mitigating climate warming, it is necessary to pay more attention to vegetation changes [66]. With the growing awareness of environmental protection and stringent enforcement of legislation, China’s contribution has accounted for a considerable share of to the total world’s greening [67]. Fujian Province, with the highest forest coverage in China, showed an upwards trend in NDVI during the study period. The sharp increase in NDVI from 2002 to 2003 and 2012 to 2013 was derived from the execution of the ecological province strategy and the control of soil erosion, respectively, suggesting that a great deal of effort was committed to ecological conservation [68]. However, under the background of global greening, hidden browning should be used as a wake-up call [69]. Previous studies have indicated that active forest changes still exist in the study area, despite it being the province with the highest forest coverage rate of China for 42 years [45,46]. Our study is consistent with these previous studies, and we also found that active vegetation changes, both greening and browning, occurred simultaneously along roads and around cities in the eastern coastal areas.

Previous studies have shown that forest displacement occurred in China; that is, forest loss was compensated by forest restoration in other regions. For example, the loss of forest area on forestland was compensated by forestation on newly reclaimed land and a rise in plantations [70]. The location of the displacement is our original purpose to explore the spatial relationships between greening and browning; therefore, we can trace the cause of greening and browning in depth. The spatial heterogeneity and spatial dependence in the distribution of vegetation greening and browning allowed us to dig further into the deeper reasons for vegetation changes. The characteristics of the L-H clusters distributed within and the H-H clusters appearing on the edges of roads and cities were a manifestation of the steady internal development and the rapid expansion of the cities. The spatial distribution of NTL intensity showed that the eastern coast had a higher level of urbanization than the west and north, which, combined with higher levels of economic development and population density in the east, meant that human activities were more frequent in the east. Furthermore, the eastern plain made human activities more accessible. While urban growth damaged existing forests owing to land scarcity, increasing ecological awareness resulted in vegetation restoration [71]. However, the specific locations of forest loss and gain do not coincide, and this forest displacement cannot compensate for the impact on the forest landscape pattern—forest fragmentation, which is a problem that may be masked by the appearance of increasing forest cover. In addition, by comparing the displacement location with urban regions extracted from NTL data and land use change, the relationship between urbanization expansion and forest displacement can be further intuitively derived [72].

4.2. The Driving Pattern of Vegetation Changes

Vegetation changes closely related to human activities are consistent with previous research [73,74]. In this study, NTL and Distance were signs of human disturbances, but the explanatory power of the two variables was not the same. The NTL with the highest q value indicated that human activities were dominant in greening and browning of the study area. We compared the locations of browning trend areas affected by NTL (NTL_A > 1.1 nW cm⁻²sr⁻¹and NTL_T > 0.14), where the browning was significant. These areas were concentrated in Fuzhou, Putian, Quanzhou, and Xiamen along the eastern coast with a higher level of urbanization, which was consistent with the location of the significant browning obtained from the trend analysis. Meanwhile, previous studies used NTL data to obtain the optimal thresholds to obtain urban areas, which implies that the persuasiveness of urbanization impact on vegetation changes can be improved by comparing greening to browning thresholds in the gradient of the mean s of factors with the urban area threshold [75,76]. Relative to the direct impact of the road, the indirect impacts of roads were clearer. We found that the interaction of the road with NTL_A on NDVI was much larger than the independent role of roads, which can be explained by the fact that road development, as a manifestation of urbanization, causes land use changes and facilitates the accessibility of human interference, such as infrastructure construction and forest activities, thus exacerbating vegetation changes [77,78,79]. The difference between the interaction of roads and slope and the independent effect of roads reveals the higher road density in plain areas than in mountainous areas, which has an inverse effect on vegetation changes [80,81]. The vegetation changed actively within 1685.4 m from the road and that the farther the distance was more stable which were results that are supported by previous work in which vegetation changes mainly occurred within 2000 m around the road [45,82]. However, the smaller q value of distance is likely since the impact of roads on vegetation is phased, and the negative side will be greater during the construction phase. After the construction is completed, the browning will be replaced by vegetation restoration during the operation phase [83,84]. The period of road construction and operation was included simultaneously from 1998 to 2018, indicating that the vegetation browning of the construction phase will be alleviated during the operation phase. Likewise, the coexistence of greening and browning and H-H aggregation appeared along roads, which probably produced an “offset” effect during the study period. Therefore, this may be the reason why roads have lower explanations.

Topography has a significant impact on the spatial distribution of vegetation [32]. The response of the mean s to the change in elevation and slope had similar features, and the greening trend was significant in areas with higher elevation and slope (elevation > 43.4 m and slope > 6.3°). The impact of human activities lessened as the slope and elevation increased, and the browning trend occurred in plain areas due to intensive construction [85,86]. Existing research has verified the benefits of shaded slopes for vegetation growth [87].

TMP and PRE are essential elements for vegetation growth, but they were not the primary factors influencing vegetation changes in the area. The change trend of PRE was not the restriction of NDVI changes in Fujian, since sufficient water conditions for vegetation growth were available, as seen by the mean s with no significant difference between PRE_T subregions, explaining the lower explanatory power of PRE_T [88]. Furthermore, the terrain with a majority of mountains and hills, as well as elevation variances, led to differences in PRE, with the rainfall increasing from southeast to northwest and northeast [89]. Therefore, the trend of browning occurred in the southeast region where PRE_A was relatively low. With respect to TMP, the greening trend was more pronounced when the average and growth rate were lower, but it was severely weakened with a higher average temperature and growth rate, which is a warning of vegetation change under the background of climate warming. Additionally, except TMP and PRE were not the dominant factors causing vegetation changes, interannual data rather than growing seasons were used, which could be a reason for the lower q value of the two variables.

The impact of driving forces on NDVI is complicated, as it is influenced by numerous elements rather than just one. Furthermore, it is demonstrated that the effect is not a simple summation from the composition of bivariable enhancement and nonlinear enhancement obtained by the interaction detector. The strongest interaction between NTL and slope is likely due to the frequency of human activities in plain areas, which is intensified by vegetation changes [90].

4.3. Implications and Limitations

Benefiting from the implementation of the ecological strategy in Fujian Province, NDVI has steadily increased in the past 20 years. However, this study also identified the geographical locations of forest fragmentation and forest displacement, where is mainly influenced by urbanization and roads construction. In order to improve the ecological environment comprehensively, on the one hand, forest landscape structure improvement and biodiversity protection should be listed as the primary target, which can be alleviated by constructing ecological corridors. On the other hand, it is a priority to improve the quality of urbanization and prevent inefficient land use. The delineation of primary function area should be promoted and the disorderly expansion of construction land should be strictly limited. It is suggested that one should focus on the effective management of urban ecosystems and harmonization of the relationship between development and protection. Ecological civilization planning and territorial spatial planning are reasonable means to build a better livable environment and contribute to the harmonious coexistence and sustainable development of man and nature.

The geographical detectors are effective tools to quantify spatial heterogeneity in the causal relationships between the potential factors and NDVI changes. However, there are still certain flaws in our analysis.

(1) Although SPOT NDVI dataset preprocessed by MVC can minimize the disturbance of atmosphere conditions, cloud and solar angle, residual noise still remains, which can be reduced by several methods, such as temporal-based methods, frequency-based methods and hybrid methods [8]. Regarding the 1km resolution of NDVI, smaller patches (such as streets trees and small greensward) will be neglected, while the products derived from Landsat and Sentinel can improve the observation accuracy [91,92].

(2) Active vegetation changes were found in the study area, but the effect of vegetation changes on forest landscape structure was not quantitatively evaluated in this study. The changes in the landscape patterns affected by vegetation changes should be furtherly investigated, in order to better propose countermeasures for forest protection and restoration for the study area.

(3) Limited by the availability of the road network dataset, only the road network of 2018 was used to examine the impacts of roads on vegetation changes. Although the nighttime light intensity and the distance to roads were used to describe the effects of human activities on vegetation changes, it is still not fully representative of the diversity and complexity of human interference. Moreover, the investigation of interaction was confined to two factors instead of three or more factors, which also urgently need to be further explored.

5. Conclusions

In this study, we identified the spatial patterns in the greening and browning of vegetation and their relationships, then explored the driving patterns of the change dynamics in vegetation. We found that most study areas showed a significant greening trend, except for the eastern coastal areas, indicating a significant browning trend. The spatial relationships between greening and browning were dominated by the spatial heterogeneous (i.e., H-L accounting for 60% and L-H accounting for 14%), but there were still a quite part of places showing spatial dependencies (i.e., H-H accounting for 4%), mainly in the edges of cities and paralleling to the roads. The above phenomenon can be explained by the leading role of the NTL and strong indirect effect of road indicators. The explanatory power of elevation and slope was second only to the NTL, while the impact of the climatic factors was lower. Additionally, the interaction of any two factors was stronger than that of each single factor, which was an enhancement to vegetation changes. Specifically, the greening trend occurred in areas with lower NTL_A (<1.1 nW cm⁻²sr⁻¹), higher elevation (>43.4 m) and slope (>6.3°) and beyond a distance of 1685.4 m from roads. Our findings raise awareness of vegetation change patterns and the influence of driving factors on greening and browning in Fujian Province, which can provide a scientific reference for further optimizing vegetation protection and restoration.