Temporal–Spatial Evolution and Driving Mechanism for an Ecosystem Health Service Based on the GD-MGWR-XGBOOT-SEM Model: A Case Study in Guangxi Region

Abstract

With the expansion of urbanization in China, ecological environments are becoming more and more prominent. Uncovering driving factors and ways of regulating ecosystem health has become a hot topic for regional sustainable development. This paper adopted the improved vigor–organization–resilience service (VORS) model to diagnose the regional ecosystem health status in Guangxi from 2000 to 2020 and verify the main factors affecting ecosystem health. Considering the influencing factors (including vegetation, terrain, climate and human activities), the mechanism of driving factors associated with regional ecosystem health was analyzed by using a geographic detector (GD), a multiscale geographically weighted regression model (MGWR), and the XGBOOTS-SHAP model. The results show that the spatial distribution of ecosystem health is characterized by low values in the central region and high values in the northern and eastern regions with higher elevations from 2000 to 2020. The spatial agglomeration evolution changes from agglomeration to dispersion, and the regional urbanization distribution and evolution are consistent. The interaction of driving factors for ecosystem health and vegetation is enhanced significantly, while the interaction of climate factors is relatively weak. And most of the impacts of human activities on the ecological environment are negative. The vegetation factor has a dominant positive effect on ecosystem health, while human activity elements have a weak negative effect on ecosystem health. Meanwhile, climate factors are complex and changeable, and their impacts on ecosystem health are changeable, leading to corresponding changes in other factors. This study provides scientific reference for the harmonious and sustainable development of humans and nature in southern China.

1. Introduction

With the advancement of urbanization, high-intensity human activities have significantly changed the structure and function of the ecosystem [1,2,3,4], resulting in serious degradation of said ecosystem and threatening the survival of human beings and their social and economic development. Ecosystem health (EH) is a crucial indicator for assessing environmental conditions, which plays a key role in environmental management and protection [5,6]. Finding ways to monitor, assess and govern ecosystem health is a hot topic for both academia and public governments. It is of significant value and importance for achieving sustainable development and formulating ecosystem restoration strategies [7].

Ecosystem health assessment is a very effective approach proposed for creating sustainable development strategies for human and environmental protection. Ecosystem health assessment has always been the focus of researchers. Research focused on ecosystem health assessment can be traced back to the 1980s [8,9,10,11]. Various advanced evaluation methods for ecosystem health are emerging due to the development of environmental science. In order to study ecosystem health, researchers at home and abroad have proposed several scientific methods and implementation approaches [12]. Among them are the famous pressure–state–response (PSR) model [13,14], driving–state–response (DFSR) model [15,16], and basic pressure–state–response (BPSR) model [17]. These models and methods discuss the relationship between ecosystem health and impact factors, providing a scientific basis for managers’ decisions. The current research hotspot model of ecosystem health is the most representative model of “vital-organization-resiliency” (VOR) proposed by Costanza [18,19], which mainly studies the internal structure and processes of ecosystems. Based on the VOR model, many scholars have proposed widespread use of the vital organization resilience service (VORS) model to measure the state of ecosystem health and quantify services [20,21,22,23]. He et al. [20] used the VORS evaluation framework to evaluate China’s ecosystem health during 2000–2015 and analyze its driving factors. Pan et al. [23] investigated the middle Yangtze River Economic Belt under different scenarios by using ecosystem health assessment models to analyze the relationship between climate and land use.

Research on the driving mechanism of ecosystem health has received much attention from different scholars [24,25]. Scholars consider different terrains and regions, including natural ecosystems such as lakes [26], forests [25], and mountainous regions [27], to analyze the impacts of natural environmental factors (climate and hydrology) [28,29] and human activity factors (environmental pollution, land for human production and living and regional policy) on ecosystem health [30,31] and the impacts of spatial heterogeneity [32].

In general, there are statistics-based research methods which are used to study the driving mechanism of ecosystem health, which mainly include gray correlation analysis, principal component analysis, geographic detectors [33], decision tree analysis, and geographically weighted regression models [34]. These traditional analysis methods could only study whether the driving factors are related, whether there is interaction, and how strong they are. However, the quantification of changes in ecosystem health due to driving changes cannot be presented [20,35]. There are also many methods based on machine learning and deep learning that can be used to study ecosystem health driving factors, and the results are better than those achieved with traditional methods [36,37,38]. However, they also have poor interpretability and struggle to define the characteristics of factors. The latest SHAP model seems to overcome the problems of these two methods, explaining not only the importance of each driving factor, but also the functional relationships (mathematical statistics) of changes between variables. This model effectively addresses the tendency of ecosystem health to change with changes in driving factors [39]. To explore the drivers of changes in ecosystem services that are crucial for maintaining ecosystem functionality, some studies used the integrated valuation of ecosystem service and trade-offs (InVEST) model and multiscale geographically weighted regression (MGWR) model to examine ecosystem services patterns [40,41]. The XGBoost–SHAP model was frequently used to reveal the key indicators related to climate change and human activities [42,43].

Over the past few decades, there has been a significant increase in the demand for methods of measuring the economic and social value of ecosystem health services by environmental public institutions of government to improve the conservation and management of the natural environment. Economic assessment is one of the important considerations for ecosystem services, such as biophysical indicators and social impacts [44]. The United Nations Statistical Commission (UNSC) proposed an integrated and spatial statistical framework for measuring the flows of ecosystem services generated, tracking changes in ecosystem extent and conditions, assessing ecosystem goods and services, and linking to the measures of economic, policy and human activities [45]. Amatucci et al. [46] allowed for the establishment of a stylized economic value model for ecosystem services and provided the foundation for the development of tools to support environmental policy decisions.

At present, many scholars believe that ecosystem service assessment can contribute to economic development, policy decision making and social stability, and further guide managers to consider more far-reaching issues, such as low-carbon earth and species diversity. Ecosystem service assessments were originally considered to be a ‘panacea’. They raised the awareness of policy-makers about the importance of the services nature provides to humanity, their probable value and the significance of its protection [47]. The idea behind the valuation of ecosystem services is embodied in the notion that decision makers will understand the immense value of ecosystems and their services. There is a need for a review of the actual effectiveness of valuations of ecosystem services as a means of communicating scientific research to policy-makers [48].

Guangxi Zhuang Autonomous Region is located in the west of South China and is an important ecological barrier in southern China. With the rapid development of China’s economy, Guangxi’s economy has also made important achievements [49]. Meanwhile, rapid urbanization, industrialization and frequent man-made destruction of the ecological environment have brought unprecedented pressure on the health of the ecological environment [50]. A large number of forests, lakes, swamps and farmland have been transformed into industrial land, residential buildings and commercial buildings. This leads to soil erosion, rocky desertification, desertification, and the ecological environment of animals and plants is seriously damaged [51]. Nowadays, there are few studies on ecological environmental health assessment in Guangxi, especially the analysis of spatio-temporal influence factors of human activities on vegetation. In order to reveal the driving factors of ecosystem health assessment and the combined impact of socioeconomic and human development needs in Guangxi, and fill some gaps, we carried out this research work.

Therefore, taking Guangxi in southwest China as the research area, this study proposes a composite ecosystem health assessment method that takes into account ecological, socio-economic and human development needs. The Ecosystem Services calculated by the InVEST model were incorporated into an ecosystem health assessment system, and the VORS model was used to analyze the ecosystem health status in Guangxi. Meanwhile, a geographical detector (GD) was used to analyze the driving factors, and the reliability was improved by optimizing the data discretization method. The multiple geographical weighted regression (MGWR) model is used to express the spatial differences of driving factors considering spatially unevenly distributed ecosystem health and driving factors of spatial heterogeneity. The XGBOOTS-SHAP model was used to explore the spatio-temporal changes in the influence degree and direction of the influencing factors, and to reveal the spatial heterogeneity of the mutual process between different variables. This study not only provides a scientific basis and decision support for developing effective ecosystem health assessment strategies and promoting regional sustainable development in Guangxi, but also provides valuable experience for the sustainable development of humans and the ecosystem in China and even the world.

2. Materials and Methods

2.1. Research Area and Data Source

Guangxi Zhuang Autonomous Region is located in the southeast edge of the Yunnan-Guizhou Plateau of China, which is the second step of the national topography. It is a generally mountainous and hilly basin landform, surrounded by plateau and mountain, mainly consisting of mountains, hills, plateaus and other types of landform, with few plains, high in the northwest and low in the southeast. Sloping from northwest to southeast (as can be seen in Figure 1), the karst geomorphology is widespread and is typical of the southern karst geomorphology region. The land area of the administrative region is 237,600 square kilometers. The climate is mainly subtropical monsoon climate and tropical monsoon climate. The annual average temperature is between 16.5 and 23.1 °C, the annual precipitation is 1080~2760 mm, and the annual sunshine hours are 1169~2219 h. Benefiting from suitable hydrothermal climate conditions, the native vegetation is a typical subtropical evergreen broad-leaved forest, which is conducive to vegetation growth. The vegetation cover is better, and the net primary productivity (NPP) of vegetation is higher. This hydrothermal conditions have an important impact on agricultural production and the ecological environment, and the actual evapotranspiration is high, reflecting its humid climatic characteristics. Guangxi has various soil types, mainly including paddy soil, lateritic soil, red soil, yellow soil and so on. The soil organic matter content is higher, which is conducive to agricultural production and vegetation growth. Population density is low, but it is unevenly distributed, and economic activity is mainly concentrated in plains and basin areas. GDP per capita is relatively low, the level of economic development is significantly different, and economic conditions have a significant impact on vegetation cover and land use. There are significant differences in socio-economic development levels, and economic conditions have a significant impact on vegetation cover and land use.

Considering the geographical location, climatic characteristics and social and economic structure of Guangxi have a profound impact on the ecological environment and social and economic development. So, this study fully considers the four aspects of vegetation cover, climate conditions, topography, and human activities in order to more accurately reflect the actual situation of the regional ecological environment evolution mechanism. Meanwhile, the research pays attention to the comprehensive analysis of regional characteristics and multi-dimensional analysis to improve the rationality and practicability.

This study selects potential driving factors from four aspects, including vegetation cover, climatic conditions, topography and human activities, as shown in

Table 1. Indicators and data sources.

Factor Type	Data Type	Data Description	Data Source
Ecology	Land utilization	2000–2020 resolution 1 km × 1 km	National Tibetan Plateau Scientific Data Center
	NPP	Spatial resolution 500 m × 500 m	Earth resource data cloud platform
	Vegetational type	China 1:1 million vegetation dataset	National Data Center for Glaciology and Permafrost Desert Science
	Agrotype	Spatial resolution 1 km × 1 km	Resource Environmental Science and Data Center
Terrain	DEM	Spatial resolution 450 m	SRTM15
Climate	Potential evapotranspiration	Spatial resolution 500 m × 500 m	National Tibetan Plateau Scientific Data Center
	Mean annual temperature	Spatial resolution 1 km × 1 km	Resource Environmental Science and Data Center
Human activity	Population	Spatial resolution 1 km × 1 km	Resource Environmental Science and Data Center
	Construction land	Spatial resolution 1 km × 1 km	Resource Environmental Science and Data Center

. The factors are vegetation cover, net primary productivity of vegetation, precipitation, temperature, actual evapotranspiration, drying index, elevation, slope, soil organic matter, population density and GDP (gross domestic product). Considering the size of the study area and computational efficiency, a 10 km × 10 km fishing net is selected as the basic analysis unit. The landscape pattern index is mainly calculated by FRAGSTAT 4.2, and the others are mainly extracted by fishing net statistics by ARCGIS for subsequent ecosystem health assessment. Finally, the collinearity analysis is removed by variance inflation factor (VIF). And variables GPP, NDVI, l_pd, l_cnnct, DEM, SOIL, GDP, NLT, P, T with VIF less than 10 were selected, as shown in

Table 2. Variable declaration.

Variable Abbreviation	Variable Name	Variable Description
GPP	Gross primary productivity	The total amount of carbon fixed by plants through photosynthesis per unit time, reflecting the primary productive capacity of the ecosystem (gC/m2·yr).
NDVI	Normalized difference vegetation index	The vegetation coverage index (range: −1~1) calculated based on the reflectance of red and near-infrared bands. The higher the value, the more dense the vegetation.
l_pd	Landscape patch density	The number of landscape patches per unit area (unit: units/km2), which represents the degree of landscape fragmentation (the higher the value, the more severe the fragmentation).
l_cnnct	Landscape connectivity	Describes the spatial connectivity between patches in a landscape (common index: probabilistic connectivity index) that affects species dispersal and ecological processes.
DEM	Digital elevation model	Digital representation of surface elevation (in meters) used to analyze topographic features such as slope and slope direction and their impact on ecological processes.
SOIL	Soil organic matter content	Soil key attributes (such as organic matter content, pH, texture, etc.), directly affecting vegetation growth and nutrient cycling.
GDP	Gross domestic product	Regional economic aggregate index (unit: CNY/person), reflecting the intensity of human economic activities and their pressure on the ecosystem.
NLT	Nighttime light	Night light brightness based on satellite remote sensing as an indicator of human activity intensity.
P	Precipitation	The total amount of precipitation per unit time (unit: mm) affects the water budget and vegetation distribution pattern of ecosystem.
T	Temperature	Temperature index (unit: °C), regulating plant photosynthesis, respiration and species distribution.

.

2.2. Research Method

The designed technical workflow of this work is as follows (Figure 2):

(1) Considering terms of human activities, natural vegetation, topographic elements, and climatic elements from 2000 to 2020, the ecosystem health of the study area was revised and estimated, and the spatial evolution characteristics of regional ecosystem health were further analyzed. In this paper, the indicators related to human activities are selected, including population density and GDP factor. Natural factors mainly consider vegetation factors such as NDVI and GPP. Terrain elements are mainly from high level, slope, aspect and soil properties. Climate elements mainly include temperature, precipitation and evapotranspiration. Moreover, the factors with a large relationship are further removed by collinearity. The data are derived from the connection center in Table 1. The main variables are listed in Table 2.

(2) The impact intensity of driving factors and their interactions on ecosystem health was analyzed by means of geographic detectors and ecosystem health assessment model.

(3) The multiscale geographic weighted regression (MGWR) model was used to further analyze the spatial characteristics of the impacts of each driving factor on ecosystem health.

(4) The XGBOOTS-SHAP model was used to analyze the marginal contribution of the interaction effects among the driving factors to the model output and its response characteristics.

(5) The structural characteristics of the impact of driving factors on ecosystem health were analyzed using the SEM model.

2.2.1. Ecosystem Health Assessment Model

Ecosystem vitality, ecosystem organization and ecosystem resilience are fundamental elements that reflect ecosystem health; they focus on assessing the integrity and sustainability of the ecosystem itself. This study selected landscape indices such as Shannon diversity index and area-weighted average patch fractal index, determined the weights, and calculated the ecosystem organization force with reference to relevant studies and theories. The establishment of an ecosystem health assessment system is a vital service function for maintaining ecosystem health. We selected the statistical data of the study area from 2000 to 2020. Ecosystem service capacity was assessed using soil conservation and water volume calculated by the InVEST model [40,41]. An improved vigor–organization–resilience–service (VORS) model is adopted to calculate the ecosystem health index (EHI) based on ecosystem integrity and ecosystem services [15,22,26]. The assessment formula for the ecosystem health is as follows:

$$EHI={\left(EV\times EO\times ER\times ES\right)}^{{\displaystyle \frac{1}{4}}},$$

(1)

where EHI is ecosystem health index, and its value range is [0, 1]. EV, EO, ER, ES denote, respectively, ecosystem vigor, ecosystem organization, ecosystem resilience, and ecosystem services. EHI is divided into five levels: low (0–0.2), relatively low (0.2–0.4), general (0.4–0.6), relatively high (0.6–0.8), high (0.8–1.0).

2.2.2. XGBOOTS-SHAP Model

The XGBOOST-SHAP model had demonstrated considerable effectiveness in nonlinear fitting applications [42,43]. The XGBOOST-SHAP model has many advantages over classical models such as GBDT and Bayesian estimation. The objective function of the XGBOOST-SHAP model is as follows:

$${\stackrel{⌢}{y}}_l^t={\displaystyle \sum _t^{k-1}{f}_k(x_l)}={\stackrel{⌢}{y}}_l^{t-1}+f_t(x_i),$$

(2)

where ${\stackrel{⌢}{y}}_l^t$ denotes the cumulative counts after the t-th iteration. ${\stackrel{⌢}{y}}_l^{t-1}$ is the predicted value from the ensemble of t − 1 decision trees. $f_t(x_i)$ is the t-th decision tree. $x_i$ is the explanatory variable. The construction of this model involves solving for $f_t(x_i)$ to minimize the residuals of the t − 1 iteration. The objective function of solving the residual of this model is as follows:

$$O^t={\displaystyle \sum _{t=1}^{i-1}[L(y_i,{\stackrel{⌢}{y}}_l^{t-1})+g_i{f}_t(x_i)+\frac{1}{2}{h}_i{f}_t{(x_i)}^2]}+\Omega (f_t),$$

(3)

where $g_i$ and $h_i$ are first- and second-order derivatives of the loss function $L(y_i,{\stackrel{⌢}{y}}_l^{t-1})$. $y_i$ is the actual count value. $\Omega (f_t)$ is the regularization term for the t-th decision tree; its expression is

$$\Omega (f_t)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {\displaystyle \sum _T^{j-1}{w}_j^2},$$

(4)

where T represents the number of leaf nodes in the decision tree and $w_j^2$ is the squared vector of the output values corresponding to the leaf nodes. γ and λ are the parameters to be estimated. The core of the XGBOOST-SHAP model takes full advantage of the residuals from the j-th round decision tree to fit the base decision tree for the (j − 1)st round. Upon reaching the set number of iterations, it integrates the outcomes of all decision trees.

2.2.3. MGWR Model

Considering the limitations of the traditional geographical weighted regression (GWR) model, we further improved the multiscale geographic weighted regression model, or the MGWR model for short. Different from the GWR model, the MGWR model considers the variation difference of different covariates on the spatial scale [33,46], and allows different covariates to have different bandwidths, which is more sensitive to the spatial heterogeneity of geographical phenomena [40]. In this study, the MGWR model was adopted to obtain the optimal ecosystem service combination mode. The expression of the MGWR model is as follows.

$$y_i={\beta }_0(u_i,v_i)+{\displaystyle \sum _{i=1}^n{\beta }_{bwj}(u_i,v_i)x_{ij}}+{\epsilon }_i,$$

(5)

where ${\beta }_0$ is the intercept constant of the regression equation, n is the total observed quantity, ${\beta }_0(u_i,v_i)$ is the spatial position of the i-th observation, ${\beta }_{bwj}(u_i,v_i)$ is the regression coefficient of the j-th covariate, $x_{ij}$ is covariant, ${\epsilon }_i$ is the error term.

2.2.4. Geodetector (GD)

Geodetector is a tool for analyzing the relationship between independent and dependent variables in spatial data. It is particularly suitable for revealing the spatial heterogeneity of geographical phenomena and the driving factors behind it. Its main idea is that if a certain independent variable has a significant influence on the dependent variable, then the two should have a consistent spatial distribution. The GD can process numerical and categorical data without linear assumptions and has a clear physical meaning. It is highly applicable to a variety of fields, including land use, public health, regional economics, and more. Its advantages are that it can deal with categorical data and it is immune to collinearity of multiple independent variables. So it is especially suitable for analyzing nonlinear relationships in complex spatial data.

In this study, GD is used to analyze the spatial and temporal changes in ecosystem health and its driving factors in Guangxi. GD is suitable for investigating driving factors of annual average temperature, annual average precipitation, elevation, soil organic matter, per capita GDP, population density, nighttime light intensity, and landscape pattern rate, and quantifying the impact of these factors on ecosystem health.

2.2.5. Structural Equation Model

Structural equation modeling (SEM) is a method to establish, estimate and test a causality model, and can replace multiple regression, path analysis, factor analysis, covariance analysis and other methods. So, it could analyze the role of individual indicators on the whole and the mutual relationship between individual indicators.

3. Results

3.1. Characteristics of Ecosystem Health Changes

3.1.1. Spatial Distribution of Ecosystem Health, 2000–2020

According to the natural breakpoint, ecosystem health was divided into five levels (low, relatively low, medium, relatively high, and high, corresponding lv_1, lv_2, lv_3, lv_4, lv_5), as shown in Figure 3. The spatial distribution characteristics of ecosystem health in the study area were similar, with low values mainly distributed in the middle of the study area, and high values mainly distributed in the north and east of the study area. The areas with very low ecosystem health are mainly concentrated in the central region, which is also the main area of economic development. The change of the proportion of lv_1 was 8.26%, 9.27%, 9.36%, 7.88% and 8.18% from 2000 to 2020. The results showed that the phenomenon of the proportion rises first and then decreases, indicating that the urbanization process affects the ecosystem health, and the rapid development at the beginning causes the ecosystem health to decline. Later, the urbanization focuses to the construction of the ecological environment, and the ecosystem health becomes better. The proportion of low level (lv_2) was 28.95%, 16.35%, 14.96%, 13.86% and 16.47%, respectively. The proportion of middle level (lv_3) was 31.94%, 20.98%, 20.69%, 20.02% and 22.04%. The proportion of high level (lv_5) was 7.21%, 27.90%, 28.40%, 31.02% and 27.39%, respectively. The relatively high level (lv_4) was mainly distributed in the northern and eastern regions, which belong to the higher-elevation marginal areas, and the human activities are relatively weak, accounting for 23.64%, 25.50%, 26.59%, 27.22%, 25.92%.

From 2000 to 2005, the transition from relatively low level to low level was significantly greater than the transition from low level to relatively low level. The transition from middle level to relatively low level was more obvious, and the transition from high level to middle level was present, indicating that ecosystem health was declining during this period, and the transition of ecosystem health level from 2005 to 2020 was relatively stable.

Research results indicate that the changes of ecosystem health in the study area showed a trend of first decreasing and then increasing. The changes of low and relatively low level fluctuated, mainly because these levels were mainly distributed in the main areas of urbanization, and the impact of further surface areas with intensive human activities on the ecological environment was greater. However, the ecological environment recovery was better in the remote mountainous areas with higher elevation and less impact of human activities.

3.1.2. Characteristics of Ecosystem Service Value Spatial Differentiation

Cluster analysis was performed for the ecosystem health services studied, as shown in Figure 4. The Moreland index values of regional ecosystem health aggregation from 2000 to 2020 were 0.6699, 0.6744, 0.6692, 0.6462 and 0.6375, respectively. The overall trend of decline was not obvious from 2000 to 2010, and the change was particularly obvious from 2010 to 2015. This also indicates that in the process of urbanization construction in the study area, the ecological environment is divided or destroyed due to large-scale development, which leads to the decline of the healthy spatial aggregation of the regional ecosystem.

The evolution of ecological services’ healthy spatial aggregation is shown in Figure 5. For Figure 4 and Figure 5, HH (red) stands for high-high aggregation of ecological health. HL (orange) stands for high-low aggregation of ecological health. LH (light blue) stands for low-high aggregation of ecological health. LL (blue) represents low-low aggregation of ecological health. ns (gray) represents not significant. High concentration is mainly distributed in Baise City and Hechi City in the western region of Guangxi, Guilin City and Liuzhou City in the north of Guangxi, while low concentration is mainly distributed in Nanning City, Laibin City and Chongzuo City in the middle of Guangxi, and Qinzhou City and Beihai City in the south of Guangxi. From the perspective of the agglomeration process, each agglomeration is gradually dispersed. For example, the high agglomeration in the northwest of Guangxi is obviously dispersed, and the low agglomeration in the center of Guangxi is also gradually dispersed, and the agglomeration area is also shrinking.

To sum up, the spatial agglomeration distribution characteristics of ecological service health are as follows. High aggregation is mainly distributed in mountain areas with high posters and relatively weak human activities. However, low aggregation is mainly distributed in central developed areas with relatively frequent human activities. The evolution of agglomeration is gradually dispersed, which is consistent with local development. Along with the destruction and segmentation of the regional ecological environment caused by urbanization development, the ecological area is fragmented.

3.2. Spatial Influence Analysis of Driving Factors

In Figure 6, red indicates that the correlation of driving factors is significant (p < 0.05), and black indicates that the correlation is insignificant. To make Figure 6 more compact, we perform some simplifications, such as 0.23 instead of 0.23. Through the analysis of geographic detector from 2000 to 2020, as shown in Figure 6, the interaction of driving factor is significantly enhanced, and gradually weakened. In particular, the interaction of natural vegetation factors GPP and NDVI, as well as DEM factors are significantly enhanced. There are significant differences in the driving factors affecting the spatial heterogeneity of ecosystem health, and there are different degrees of interaction among the driving factors. It is strongly influenced by the natural environment. Compared with a single driving factor, the interaction between driving factors effectively improves the explanatory power, and the natural factor is dominant.

The multi-scale geographical weighted regression (MGWR) model is used to analyze the impact characteristics of driving factor on ecosystem health. The MGWR model has better fitting than the GWR model, as shown in

Table 3. The comparison of GWR and MGWR model fit from 2000 to 2020.

	2000		2005		2010		2015		2020
	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR
AIC:	3234.86	1915.94	3926.00	2311.69	3815.37	2306.74	3576.94	2252.14	4086.72	2591.34
AICc:	3236.99	2050.59	3928.13	2410.94	3817.50	2458.77	3579.07	2392.13	4088.86	2758.86
BIC:	−17,819.20	4033.76	−17,637.28	4151.23	−17,670.08	4545.46	−17,735.77	4407.98	−17,586.83	4931.04
R2	0.77	0.90	0.70	0.88	0.71	0.89	0.74	0.89	0.68	0.88
Ad. R2	0.77	0.89	0.70	0.86	0.71	0.87	0.74	0.87	0.67	0.85

. The R² values of GWR were 0.77, 0.70, 0.71, 0.74 and 0.68, respectively, and the R² values of MGWR were 0.90, 0.88, 0.89, 0.89 and 0.88, respectively. The fit degree of the MGWR model is basically greater than 0.85. This shows that the MGWR model can better explain the spatial influence characteristics of each driving factor.

Meanwhile, the coefficients of the MGWR model reflect the spatial differences of the impact of driving factor on ecosystem health. As shown in Figure 7, the influence coefficients of l_pd and l_connct, GDP and NLT factors have negative effects in most regions, and the spatial–temporal evolution differences are obvious. The influence coefficients of GPP and NDVI are positive in most regions, and the spatial distribution of GPP is decreasing in the southeast and northwest, while that of NDVI is increasing in the southeast and northwest. The spatial distribution of DEM effects is decreasing in the southeast and northwest, while the SOIL impact factors are decreasing in the east and west, and the negative effects are mainly concentrated in the eastern region. The spatial difference of T influence distribution is obvious, and the effect is negative in most areas. The influence of P-factor is decreasing from east to west, and most of them are positive.

With the development of society and the acceleration of urbanization, the damage of human activities to the ecological environment is becoming more and more obvious, and most of its effects are negative. However, there is no obvious rule of climate variability, and the spatial characteristics of its influence are obviously different. The terrain of the study area is complex, there are hills and mountains, and the spatial characteristics of the influence are spatially specific.

3.3. The Analysis of Driving Factor Influence Mechanism

The XGBOOTS-SHAP model was used to analyze the influence feature structure of driving factors, as shown in Figure 8. The order of the influence of driving factors in 2000 was GPP > l_pd > NDVI >> T > DEM > l_cnnct > P > GDP > SOIL > NLT. In this period, the influence of natural factors was obviously in the front and most of them had a positive effect, while the influence of human activity factors GDP and NLT was relatively behind. In 2005, the order of influence of driving factors was GPP > NDBI > l_pd > DEM > GDP > T > NLT > SOIL > l_cnnct > P. During this period, the influence of vegetation factors GPP and NDVI was still ahead. However, the influence of human activity factors GDP and NLT moved forward, which indicates that human activities became more intense. In 2010, the ranking of the driving factors was GPP > l_pd > DEM > NDVI > T > GDP > NLT > l_ cnnct > SOIL > P, which was still the top of the natural factors, and the ranking of human activity factors had no significant change. In 2015, the order of the influence of driving factors was GPP > NDVI > l_pd > DEM > T > l_ cnnct > GDP > NLT > SOIL > P, the order of the influence of natural factors was higher, and the influence of human activity factors was lower than that of 2010. In 2020, the order of driving factors was GPP > NDVI > l_pd > DEM > T > NLT > l_ cnnct > GDP > SOIL > P, and the influence of vegetation factors ranked first. The influence of human activities NLT factor increased, and the influence of GDP factor decreased. In summary, the natural factors play a dominant role, especially the vegetation factors GPP and NDVI, and most of them have a positive effect. The influence of human activity factors fluctuates, rising first and then becoming stable, which corresponds to regional development needs. In the development process of the study area, the urbanization may significantly damage the ecological environment. Managers and the public realized the importance of the ecological environment, so they also paid attention to the construction of ecological protection in the later construction and development process.

3.4. The Impact Characteristics of Driving Factors

From 2000 to 2020, the critical value changes of the impact of driving factors on the health of ecological services are analyzed, as shown in Figure 9. The influence of l_pd factor showed a downward trend, and when it was greater than the critical value, it was interpreted as a negative effect. The upward trend of the critical value changed less, successively 6.19, 6.19, 6.34, 6.87, 6.38. The influence of l_cnnct factor also showed a downward trend. When the critical value was exceeded, the influence was negative, and the critical value changed little, successively 0.24, 0.26, 0.26, 0.27, 0.28. The effects of human activity factors GDP and NLT are similar, showing a negative decline. There are fluctuations that first decrease rapidly and then stabilize. And some have multiple critical values, indicating that human activity had a negative effect on ecosystem health. It was complex and variable, and the effect was relatively weak in areas with weak human activity. The influence of vegetation index elements GPP and NDVI showed an upward trend, and most of them had a positive effect. The critical value of GPP was 1558.15, 1440.87, 1603.71, 1739.84, 1722.84. The critical value of NDVI was as follows: 0.77, 0.80, 0.81, 0.84, 0.84, which also indicated that vegetation factors had the same trend as ecosystem health. The DEM influence factors increased first and then decreased. The critical values were 262.37, 266.51, 258.86, 263.63 and 246.89, indicating that different poster heights had different impacts on ecosystem health, which had positive effects within a certain range. However, it had negative effects beyond a critical value. The effects of soil organic matter factors decreased first and then increased, with negative effects at 3.51~7.18, 3.63~6.52, 3.55~6.03, 3.57~6.49 and greater than 3.67. This condition indicated that different soil organic matter contents had different effects on ecological impacts. The influence of T-factor showed a downward trend, and its critical values were 20.34, 21.02, 20.82, 21.32, 21.28, indicating that the effect is positive at a certain critical value. The effect is negative beyond this critical value. The influence trend of P on ecosystem health was variable, because the main P factors were complex and changeable, and the influence trend was different in different periods. For example, its influence showed an upward trend in 2000, a downward trend in 2005 and 2010, and then an upward trend in 2015, and an upward trend in 2020.

In summary, vegetation factors have a positive effect, human activities have a negative effect, topographic factors have an inverted U-shaped trend, and soil organic matter has a U-shaped trend. Precipitation has a large change range, among which the intensity of human activities has become more intense, making its influence rise. At the same time, they cause the critical values of other influencing factors to change. The influence trend of vegetation factors is relatively simple, and the impact is the largest. Therefore, in the process of urbanization, we should focus on protecting ecological vegetation. There is no obvious change in the topographic factor. So, in the process of development and construction, we pay attention to the reasonable distribution and construction of the critical value range to achieve the harmonious and sustainable development of man and land. In view of the multiple changes of climate factors and the different characteristics of climate factors in different regions, we should carry out scientific and reasonable ecological development construction according to the climate characteristics of different regions.

3.5. The Dependent Characteristics of Driving Factors

The structural characteristics of driving factors were analyzed by structural equation modeling, as shown in Figure 10. From 2000 to 2020, the fit degree of the model was 0.6911, 0.6402, 0.6563, 0.6494 and 0.6464 successively, indicating a good fit degree. The influence structure was the influence of index factors (NDVI, GPP), human activity factors (NLT, GDP), landscape pattern factors (L_pd, L_cnnct), topographic factors (DEM, SOIL), and climate factors (P, T) on ecosystem health.

In 2000, vegetation factors and climate factors had positive effects on ecosystem health, and their effects were 0.5936 and 0.2902, respectively. The effects of topographic elements, human activity elements and landscape pattern elements on ecosystem health were negative, and the effect values were −0.0488, −0.0086 and −0.2451, respectively. On the whole, the positive effect was obviously greater than the negative effect, the effect of natural factors was more obvious, and the negative effect of human activities was less. The effect of topographic factors on climate factors was positive, and the effect value was 0.8701.

In 2005, the positive effects of vegetative elements, landscape pattern elements, and topographic elements on ecosystem health were 0.0445, 0.6525, 0.0071, respectively. Landscape pattern elements were the most important. Terrain factors, climate factors and human activity factors had negative effects on ecosystem health. And the negative effects of human activities had increased compared with 2000. The effect of topographic factors on climate factors was negative, and the effect value was −0.7841.

In 2010, the effect of vegetation elements on ecosystem health was positive, and the effect value was 0.6785. Other factors had negative effects, and the effects of topographic factors, climate factors, landscape pattern factors and human activity factors on ecosystem health were −0.0383, −0.3997, −0.0594 and −0.0155, respectively. The effect of topographic factors on climate factors was negative, and the effect value was −0.8448.

In 2020, vegetation elements, landscape pattern elements and climate elements had positive effects on ecosystem health, and their effects were 0.6101, 0.1201 and 0.3812, respectively. The effects of terrain elements and human activity elements on ecosystem health were negative, and the effects were −0.0819 and −0.0408, respectively. The effect of topographic factors on climate factors was positive, and the effect value was 008738.

In 2015, vegetation elements, topographic elements and landscape pattern elements had positive effects on ecosystem health. Their effects were 0.6946, 0.0128 and 0.0434, respectively. Human activity factors and climate factors had negative effects on ecosystem health, and the effect values were −0.3255 and −0.0053, respectively. The effect of topographic factors on climate factors was negative, and the effect value was −0.7867.

To sum up, vegetation factors have a large and positive impact on ecosystem health. Human activities have a negative impact on ecosystem health. Topographic factors are fixed, but their impacts change, mainly because climate factors are complex and changeable. So their impacts on ecosystem health are also changeable, resulting in changes in the impact of topographic factors. At the same time, the influence of other factors will also change accordingly. This further shows that vegetation factors play a dominant role in ecosystem health, while human activity factors have more negative impacts.

4. Discussions and Conclusions

This paper introduces the VORS framework to assess ecosystem health and reveal the main driving factors affecting ecosystem health from 2000 to 2020 in Guangxi. Combining GD, the MGWR model and the XGBOOTS-SHAP model, we investigate the mechanism of driving factors on regional ecosystem health (including vegetation and human activities). The study yields the following discussions:

(1) From 2000 to 2020, the spatial distribution of ecosystem health is characterized by low values in the central region and high values in the northern and eastern regions at higher elevations. The spatial clustering evolution is gradually dispersed from cluster to cluster. This indicates that human activity factors have a greater impact on the ecological environment in the regions with frequent human activities, which results in a relatively low ecosystem health. However, the human activities in the remote areas are relatively weak, and the impact on the ecological environment is small, and the ecosystem health value is high. In the process of urbanization, a large amount of development and construction makes the ecological environment divided and destroyed, resulting in its agglomeration gradually dispersing.

(2) In the analysis of driving factors, the interaction of driving factor is enhanced, in which the interaction of vegetation index and DEM factor is enhanced significantly. However the interaction of climate factor is relatively weak. There are differences in the spatial effects of various factors. The damage of human activities to the ecological environment is becoming more and more obvious, and most of them have negative effects. The influence characteristics of climate variability are obviously different, and the influence spatial characteristics of terrain complexity are spatial differences.

(3) In the mechanism analysis of driving factor, the influence of natural factors plays a dominant role, especially the vegetation factors GPP and NDVI. And most of them have a positive effect. The influence intensity of human activity factors increases first and then becomes stable, indicating that the development process of the study area may cause obvious damage to the ecological environment in the early stage of urbanization. When managers and the public are aware of the importance of the ecological environment in the later stage, they pay attention to the construction of ecological protection in the later stage of construction and development. The proximity values of human activities change greatly, which indicates that the influence of human activities is becoming more and more obvious, resulting in the corresponding change of the influence critical values of other factors.

(4) Vegetation elements have a dominant positive effect on ecosystem health, while human activity elements have a weak negative effect on ecosystem health. The influence of climate factors on ecosystem health varies greatly, mainly because climate factors are complex and changeable. Although the topographic elements are unchanged, their impact on climate elements and the impact of the variability of climate elements on ecosystem health also lead to changes in their impact on ecosystem health. The impacts of landscape pattern elements are also variable, mainly due to the changes of human activities and climate factors on ecosystem health, which lead to corresponding changes in their impacts.

The research on Guangxi based on the VORS model can help to understand regional ecosystem health, and better serve ecological and economic sustainable development through ecosystem health assessment. Guangxi is located in South China, which is close to Vietnam, Guangdong, Guizhou and Yunnan. This study only focuses on the differences between different ecosystem health services. According to the above discussions, we obtained the following conclusions.

(1) Spatial–temporal changes of ecosystem health. The ecosystem health in Guangxi showed a fluctuating upward trend from 2000 to 2020, but the spatial distribution was significantly different. This finding reveals the complexity and dynamics of Guangxi’s ecological environment and provides an important basis for regional ecological environmental protection.

(2) The role of driving factors. Land, elevation, population density, slope, soil, vegetation, etc., are the main factors affecting the health of regional ecosystems. These results highlight the important role of natural and human factors in ecosystem health and provide a scientific basis for formulating targeted ecological protection policies.

(3) Impact of interaction. The interaction of natural and human factors has a stronger explanatory power for ecosystem health, suggesting that ecosystem health is the result of a combination of factors. This finding highlights the importance of integrated management and the need to consider the synergistic effects of multiple factors in policy making.

However, this paper also has some limitations and shortcomings. This study was only analyzed on a small number of scales, and the findings may be influenced by spatial and temporal scales. For example, data at different scales may reveal different characteristics of spatial differentiation, while the choice of time scale may affect the understanding of dynamic changes in ecosystem health. On the other hand, the driving factors selected in this study may have failed to cover all factors affecting ecosystem health. For example, factors such as climate change, biodiversity, and policy interventions may have important effects on ecosystem health but were not fully considered in this study. Therefore, in future research work, we will focus on the proximity effect of ecosystem health services. Considering the positive or negative effects on the provision of ecosystem health services when a particular ecosystem is adjacent to different ecosystems, the transfer mechanism of ecosystem health services to neighboring regions is unclear and not fully understood. Future research could carry out multiscale analyses to reveal the characteristics and driving mechanisms of ecosystem health at different spatial and temporal scales. For example, remote sensing data and ground observation data can be combined to perform multi-scale integrated analysis.