Application of Bayesian Causal Inference in the Study of the Relationship Between Biodiversity and Aboveground Biomass of Subtropical Forest in Eastern China

Abstract

The relationship between biodiversity and ecosystem function is crucial for understanding the structure and processes of subtropical forest ecosystems. However, the extent to which biodiversity influences subtropical forest biomass remains unclear. This study applies Bayesian causal inference to explore causal relationships between forest Aboveground Biomass (AGB) and its potential driving factors (biodiversity factors, biotic factors and abiotic factors) based on Huangshan Forest Dynamics Plots. Furthermore, hypothetical interventions are introduced to these driving factors within the causal network to estimate their potential impact on AGB. The causal relationship network reveals that species diversity and functional diversity are the most direct factors influencing AGB, whereas phylogenetic diversity exerts only an indirect effect. Biotic and abiotic factors also contribute indirect effects on AGB, potentially by influencing other mediating indexes. Intervention analysis shows that with low-level interventions on direct influencing factors, the probability of low AGB is as high as 84%. As the intervention level increases to high, the probability of low AGB decreases by 36%. Moreover, AGB demonstrates a particularly sensitive response to changes in Rao’s quadratic entropy (RaoQ) intervention levels, more so than to other factors, highlighting its critical role in maintaining forest biomass. Therefore, we contend that functional diversity, due to its direct reflection of species’ roles in ecosystem processes, is a more accurate measure of the impact of biodiversity on biomass compared to species or phylogenetic diversity and the interplay between abiotic and biotic factors and biodiversity should not be overlooked. This approach offers a powerful tool for exploring causal relationships, thereby providing a more nuanced and accurate understanding of the relationship between biodiversity and forest ecosystem function.

1. Introduction

Ecosystem function includes ecosystem attributes, products, and services. It involves the material cycles, energy flows, and information exchanges occurring between the internal ecosystem and the external environment, which are crucial for maintaining ecosystem health and stability. Forest ecosystems, as one of the most important ecosystem types, hold significant ecological, economic, and social value [1]. Understanding forest ecosystem function is essential for interpreting its role in biodiversity conservation, climate change mitigation, and the provision of ecosystem services.

In 1994, Naeem et al. [2] first demonstrated that biodiversity could impact ecosystem function through experiments in ecological climate chambers. Following this seminal discovery, numerous controlled biodiversity experiments have been conducted to further investigate these effects. The Cedar Creek field experiment [3] explores the relationship between biodiversity and ecosystem productivity, and a European grassland experiment [4] examines how biodiversity influences ecosystem stability and function. Most experimental results indicate a positive correlation between biodiversity, community productivity, and ecosystem stability. This relationship is attributed to niche complementarity, where different species utilize different resources, and interspecies interactions, which enhance ecosystem function [5]. In contrast, previous observational and experimental studies have primarily investigated how productivity affects biodiversity [6]. These studies often noted a unimodal pattern of biodiversity along productivity gradients, suggesting that, under certain environmental conditions, high productivity is driven by a few dominant species. The inconsistency in methods and causal relationships between experimental and observed studies has led to considerable debate. This discussion centers on the effectiveness of experimental design, the underlying mechanisms through which biodiversity impacts ecosystem function, and the validity of extrapolating results from artificial experimental ecosystems to natural systems [7].

Over the past two decades, many classic ecological hypotheses have been proposed concerning the biodiversity-productivity relationship. Grime et al. [8] introduced the ‘quality ratio hypothesis’, which posits that the functional traits and diversity of dominant species are crucial for determining ecosystem function. This hypothesis emphasizes that the key characteristics of these species, rather than the overall species richness, drive ecosystem processes. Tilman et al. [9] suggested the ‘sampling effect’ and ‘complementarity effect’ to elucidate how biodiversity positively impacts forest biomass. The ‘sampling effect’ hypothesis posits that higher species richness increases the likelihood of including high-productivity species, thereby boosting community biomass. In contrast, the ‘complementarity effect’ hypothesis argues that increased biomass results from positive interactions among species and reduced competition due to niche differentiation [10]. Additionally, abiotic factors play a crucial role in determining biomass distribution [11]. Topographic factors such as altitude, slope, and roughness significantly affect the distribution of plant Aboveground Biomass (AGB) and are critical in influencing basal area [12]. These factors alter microclimates and resource availability, which in turn, impact plant growth and biomass. Climate factors have both direct and indirect effects on AGB and its dynamics [13]. Generally, forest primary productivity is positively correlated with temperature and moisture [14]. However, this relationship can be reversed in some tropical regions, where increased temperatures are associated with reduced forest aboveground biomass due to heat stress and moisture limitations [15].

Historically, many studies on the relationship between biodiversity and biomass have focused on temperate and boreal forests [16]. These regions were often chosen due to their relatively simpler species compositions and well-documented environmental variability. However, these forests, are species-poor and exhibit significant environmental variability, which led to conclusions that species richness significantly impacts biomass in forests from northern China to warm temperate regions, but has a minimal effect on subtropical forests [17]. Compared to boreal or temperate forests, subtropical forests occur in ecological transition zones with complex ecosystem structures, offering significant potential for studying the relationship between biodiversity and biomass from different perspectives. Relying solely on species diversity indicators is inadequate for capturing the full impact of biodiversity on ecosystem function.

Recent studies have shown that functional diversity and phylogenetic diversity have a more significant relationship with subtropical forest biomass than species diversity, making them better indicators for predicting ecosystem function [18]. Functional diversity emphasizes the unique functional traits of species within a community, which are directly linked to ecosystem processes [19]. Tilman et al. [20] suggested that ecosystems with higher functional diversity can use resources more complementarily, thus making the system operate more efficiently. Phylogenetic diversity provides a comprehensive perspective by incorporating the evolutionary history and relationships among species [21]. Srivastava et al. [22] highlighted that many ecological and physiological traits are conserved through evolution, so closely related species tend to share similar traits. This allows phylogenetic diversity to indirectly estimate the functional trait space of a community, offering insights into ecosystem function and responses to environmental changes [23].

It is important to note that ecosystem function prediction is a significant interdisciplinary research area involving ecology, environmental science, and statistics [24]. Traditionally, studies have utilized straightforward statistical regression methods. Chen et al. [25] used a generalized linear model (GLM) to explore the relationship between plant diversity and ecosystem function. However, these models are often inadequate for capturing the complex nonlinear relationships and interactions between forest biomass and various abiotic and biotic factors [26]. Jonsson et al. [27] used a structural equation model (SEM) to reveal plant-community drivers of carbon storage in boreal forest ecosystems. However, SEM does not technically go beyond the realm of correlation. Bayesian causal inference model is a normative method that can make testable predictions; however, it has not been widely used in the study of biodiversity and ecosystem function.

This study utilizes data from the Huangshan Forest Dynamics Plots, which provide detailed information on forest composition and biomass dynamics. The aim of this study is to reveal the causal relationship between biodiversity factors, abiotic factors, biotic factors and AGB using Bayesian causal inference. By intervening on factors that have direct and indirect effects on AGB, the study assesses the impact of these interventions on AGB and predicts future biomass dynamics. To conduct this, the following two hypotheses were tested. Bayesian causal inference was used to test hypothesis 1: biodiversity factors significantly and directly affect AGB, while biotic and abiotic factors have covariant effects and indirectly affect AGB; the intervention analysis was used to test hypothesis 2: functional diversity has a more significant effect on AGB than species diversity and phylogenetic diversity.

2. Materials and Methods

2.1. Study Site

The UNESCO Mount Huangshan World Heritage Site is a National Park located in Huangshan City, Anhui Province, and spans approximately 1200 square kilometers [https://whc.unesco.org/en/list/547/ (accessed on 5 October 2024)]. Renowned for its natural beauty and its critical role in biodiversity conservation, Huangshan is positioned at the northern edge of the subtropical zone. The average temperature of Huangshan in summer is 25 °C, the average temperature in winter is above 0 °C, and the average annual temperature is 7.8 °C. The average annual rainfall days are 183 days, mostly concentrated from April to June, and the annual precipitation on the mountain is 2395 mm, which varies significantly with altitude. The region is characterized by red soil and yellow soil, and the soil’s physical properties are close to neutral, with a pH value of 5.5–5.9. This climate diversity creates many microhabitats that foster a wide variety of plant and animal species. Huangshan is located in the transitional zone between the northern and southern flora, with diverse vegetation types, mainly evergreen broad-leaved forests. The vertical distribution of plant communities from lowlands to mountains is evident, further enhancing the rich biodiversity of the region.

2.2. Data Source and Processing of Sample Plots

The data used in this study are sourced from the Huangshan Forest Dynamics Plots [28], covering an area of approximately 10.24 hectares (320 m × 320 m). The plot was established in 2014, with the first survey completed that year near Xiaolingjiao, west of the Huangshan Scenic Area [28]. The data sample for this study comes from the second survey conducted in 2019, which was five years after the initial survey. The coordinates of the plot’s reference point are (30°8′26″ N, 118°6′38″ E), with an elevation range of 430 to 560 m, and an azimuth of approximately 20° west of north. The topographic map, shown in Figure 1, was generated by secondary interpolation of elevation data. The main ridge and primary valley in the plot run generally east–west and are intersected by several smaller north–south ridges and valleys. The sunlit slopes cover about one-third of the total plot area, with the remainder consisting of shaded slopes.

During the field survey, the plot was subdivided into a grid of $16\times 16$ plots, each measuring 20 m × 20 m. Each plot was further divided into $4\times 4$ subplots, each measuring 5 m × 5 m, with relative coordinates assigned to each subplot, as shown in Figure A1. We conducted a census of individual trees with DBH ≥ 1 cm in the plot and recorded their species, relative location, DBH, tree height, and height under the branch.

2.3. Biomass Data

This study employed allometric equations to estimate the biomass of woody plants with a diameter at breast height (DBH) ≥ 1 cm within the plot, as allometric equations provide a reliable method for relating plant dimensions to biomass. For each plant species, the appropriate allometric equation parameters were selected. Based on the measured DBH of plants in the plot, the biomass of stems, bark, branches, and trunks was calculated separately, and the sum of these biomass components was used to determine the aboveground biomass of individual plants. The specific allometric equations and their parameters used in this study are detailed in the R package “allodb” (v.0.0.1.9000) [29], which is widely used for accurate biomass estimation.

2.4. Biotic, Abiotic and Biodiversity Factors

In this study, for each subplot (5 m × 5 m), stand density was calculated as a biotic factor, while elevation, slope, aspect, and roughness were used to represent abiotic factors. Indices of species diversity, functional diversity, and phylogenetic diversity were also calculated to examine the distinct and combined effects of biotic and abiotic factors, alongside biodiversity, on AGB. See

Table 1. Descriptive statistics of forest AGB and its driving factors. FRic: Function richness, FEve: Functional evenness, FDiv: Functional divergence, FDis: Functional dispersion index, RaoQ: Rao’s quadratic entropy, PD: Faith’s Phylogenetic Diversity.

Factor	Index	Mean Std	Min	Max
Species Diversity	Richness	$6.79\pm 3.07$	1.00	19.00
	Shannon–Weaver	$5.00\pm 2.34$	1.00	15.67
	Simpson	$0.72\pm 0.18$	0.00	0.93
	Pielou	$0.85\pm 0.18$	0.00	1.00
Functional Diversity	FRic	$6.79\pm 3.07$	1.00	19.00
	FEve	$0.65\pm 0.21$	0.00	0.99
	FDiv	$0.64\pm 0.21$	0.00	0.98
	FDis	$0.88\pm 0.29$	0.00	2.58
	RaoQ	$1.05\pm 0.6$	0.00	7.90
Phylogenetic Diversity	PD	$670.56\pm 338.61$	0.00	1853.63
Biotic	Stand Density (per 25 m2)	$19.62\pm 13.68$	1.00	103.00
Abiotic	Elevation (m)	$496.93\pm 34.56$	425.58	554.63
	Slope (%)	$27.25\pm 8.36$	5.08	55.38
	Aspect (°)	$189.82\pm 106.82$	0.67	358.49
	Roughness (°)	$0.21\pm 2.93$	−10.82	14.60
Biomass	AGB (kg/25 m2)	$414.33\pm 634.68$	0.08	9144.97

.

Species diversity is represented by species richness, the Shannon–Weaver index (SW), the Simpson (Sp), and the Pielou evenness index. Species richness indicates the number of different species present in a specific area. The Shannon–Weaver and Simpson indices both consider species richness and evenness, with the latter focusing more on the relative abundance of species. The Pielou evenness index specifically measures the evenness of species distribution. The calculation formulas are as follows [30]:

$$SW=-\sum (n_i/N){log}_2(n_i/N),$$

(1)

$$Sp=1-{\displaystyle \frac{n_i(n_i-1)}{N(N-1)}},$$

(2)

N is the number of individuals observed in the plot; $n_i$ is the number of individuals of the i-th species.

$$Pielou={\displaystyle \frac{SW}{lnSp}},$$

(3)

Functional diversity is represented by the Functional Richness Index (FRic), Functional Evenness Index (FEve), Functional Divergence Index (FDiv), Functional Dispersion Index (FDis), and Rao’s Quadratic Entropy Index (RaoQ). FRic measures the volume of the convex hull in an n-dimensional functional trait space occupied by the species in the community, reflecting the level of n-dimensional functional traits. FEve indicates the evenness of the distribution of functional traits of organisms within the ecological space of the community. FDiv reflects the degree of dispersion of species’ functional traits within the convex hull. FDis represents the average distance of each species’ n-dimensional functional traits to the centroid of the functional trait space of all species within the community. RaoQ combines information on species richness and functional differences between species, primarily reflecting the variation in distances between species. Data for calculating functional diversity were from Wang et al. [31] and Lv et al. [32], and based on total carbon, total nitrogen, total phosphorus, leaf area and specific leaf area. The calculation formulas are as follows [33,34,35]:

$$FRic={\displaystyle \frac{S{F}_{ci}}{R_c}},$$

(4)

$S{F}_{ci}$ is the niche space occupied by species in community i, and $R_c$ is the absolute range of trait c.

$${\displaystyle FEve=\frac{{\sum }_{i=1}^{S-1}min\left(\frac{\frac{dist(i,j)}{w_i+w_j}}{{\sum }_{i=1}^{S-1}\frac{dist(i,j)}{w_i+w_j}}\frac{1}{S-1}\right)-\frac{1}{S-1}}{1-\frac{1}{S-1}},}$$

(5)

$dist(i,j)$ is the Euclidean distance between species i and species j; $w_i$ is the relative abundance of species i and i is the branch length.

$$FDiv={\displaystyle \frac{\Delta d+\overline{dG}}{\Delta \left|d\right|+dG}},$$

(6)

$dG$: mean distance to the center of gravity. $\Delta d$: sum of abundance-weighted deviances. $\Delta \left|d\right|$: absolute abundance-weighted deviances from the center of gravity.

$$FDis={\displaystyle \frac{\sum a_j{z}_j}{aj}},$$

(7)

$a_j$ is the abundance of species j; $z_j$ is the distance from the weighted center of mass of species j.

$$RaoQ=\sum _{i=1}^{S-1}\sum _{i=i+1}^S{d}_{ij}{p}_i{p}_j,$$

(8)

S is community species richness; $d_{ij}$ is the overlap of the probability density function of trait values of species i and species j; $p_i$ and $p_j$ are the proportion of the individuals of species i and species j to the total number of species in the community, respectively.

Phylogenetic diversity is measured using Faith’s Phylogenetic Diversity (PD) [36], which quantifies diversity based on the evolutionary relationships among species. Unlike traditional species richness, PD accounts for evolutionary differences between species, offering a more comprehensive measure of biodiversity. Initially, the scientific names of plants were standardized using the “plantlist” software package (v.0.8.0) [37]. Subsequently, mega-trees were used as phylogenetic support to generate an evolutionary tree to measure plant phylogenetic diversity (PD) [38], by “V. PhyloMaker2” package (v.4.1.2) [39]. The calculation formula is as follows [40]:

$$P{D}_i=\sum _{j\in T}{I}_{ij}\times branchLe{n}_j\left(T\right),$$

(9)

where $P{D}_i$ is Faith’s PD for sample i; $I_{i}j$ indicates if sample i has any features that descend from node j, and $branchLe{n}_j\left(T\right)$ indicates the length of the branch to node j in the tree T.

2.5. Statistical Analysis

In this study, Bayesian causal inference was employed to reveal the causal relationships between biotic, abiotic, and biodiversity-related factors (including species diversity, functional diversity, and phylogenetic diversity) and AGB in Huangshan Forest Dynamics Plots. The method also identifies the most direct drivers of AGB. Hypothetical interventions are then applied to these drivers within causal networks to estimate their potential impact on AGB. This approach helps to pinpoint the drivers most likely to significantly affect AGB, providing insights into how changes in biodiversity and other factors influence AGB.

In the Bayesian causal inference phase, the R package “bnlearn” [41] is utilized to learn the Bayesian network structure [42]. This algorithm first determines the dependency structure between nodes and constructs an undirected graph, known as the skeleton. It then uses Fisher Z-tests [43] to assess D-separation [44] and infer the direction of edges in the Bayesian network, converting the undirected graph into a completed partially directed acyclic graph (CPDAG) to establish causal relationships between factors. Following this, intervention analysis is performed by assuming different values for the factors to estimate their potential impact on biomass. K-means clustering is used to categorize the factors into three levels: 0 for low, 1 for medium, and 2 for high. This approach further explores the impact of each factor on forest biomass at varying levels.

3. Results

3.1. Species Composition of the Plot

A total of 53 families, 105 genera, and 171 species of trees were recorded, resulting in 79,303 individual records.

As shown in

Table 2. Families with a sum of importance value greater than 90.00% in the plot.

Families	Importance (%)	Accumulative Total (%)
Fagaceae	29.10	29.10
Ericaceae	21.91	51.01
Theaceae	14.84	65.85
Rhamnaceae	6.56	72.41
Pinaceae	5.69	78.10
Hamamelidaceae	4.22	82.32
Ilexaceae	3.98	86.30
Lauraceae	2.92	89.22
Theaceae	2.36	91.58

, the importance values of the families Fagaceae, Ericaceae, Theaceae, Rhamnaceae, Pinaceae, Hamamelidaceae, Ilexaceae, Lauraceae, and Theaceae in the plot together exceed 90.00%. Among these, the Fagaceae has the highest importance value at 29.10%, followed by the Ericaceae (21.91%) and the Theaceae (14.84%).

As shown in

Table 3. Species with a sum of importance values greater than 50.00% in the plot.

Species	Importance (%)	Accumulative Total (%)
Castanopsis eyrei	29.46	29.46
Eurya nitida	10.03	39.49
Rhododendron ovatum	9.05	48.54
Pinus massoniana	5.68	54.22

, there are 20 tree species in the plot with importance values greater than 1.00%. The genus and species with the highest importance value is Castanopsis eyrei (29.46%), which is the dominant species, followed by Eurya nitida (10.03%), Rhododendron ovatum (9.05%), and Pinus massoniana (5.68%). The combined importance values of these four tree species exceed 50%.

3.2. Causal Relationships Between Biotic, Abiotic, Biodiversity Factors and AGB

This study performs Bayesian causal inference on 15 indexes across five types of factors at the 5 m × 5 m plot scale. The names and definitions of these indexes are provided in Table 1.

As shown in Figure 2, the Shannon–Weaver index from the species diversity factors and the RaoQ index from the functional diversity factors have the most direct impact on AGB.

Other species diversity indexes (Richness, Pielou, and Simpson) indirectly affect AGB by influencing Shannon–Weaver. The functional diversity index (FDis, FDiv, and FDic) indirectly affects AGB by influencing RaoQ, while FEve has no significant impact on AGB.

Neither biotic, abiotic, nor phylogenetic diversity factors directly affect AGB. The biotic factor (Stand Density) influences AGB indirectly through its effect on the Shannon–Weaver index. The abiotic index Elevation directly affects the biotic factor (Stand Density) and indirectly impacts AGB through Shannon–Weaver. The phylogenetic diversity factor (PD) has no significant direct effect on AGB.

3.3. Analysis of Factor Interventions That Have a Direct Impact on AGB

Interventions were applied to Shannon–Weaver and RaoQ that directly affect AGB, and they were clustered into Level 0 (Low Degree); Level 1 (Medium); Level 2 (High Degree), see Figure 3.

Table 4. Conditional probability distribution of the level to which AGB may behave after intervention with Shannon–Weaver and RaoQ.

Shannon–Weaver		0			1			2
RaoQ		0	1	2	0	1	2	0	1	2
AGB	0	0.84	0.82	0.53	0.72	0.76	0.61	0.58	0.66	0.48
	1	0.12	0.11	0.25	0.22	0.18	0.22	0.28	0.26	0.28
	2	0.04	0.07	0.22	0.06	0.06	0.17	0.14	0.08	0.24

shows that with low-level interventions on Shannon–Weaver and RaoQ, the probability of AGB being at a low level exceeds 80%, while the probability of AGB being at a high level approaches 0. As the level of intervention increases, the likelihood of AGB being at a medium or high level increases slightly. Specifically, when interventions on both Shannon–Weaver and RaoQ are set to level 2, the probability of AGB being at level 2 is 24%, which is a 20% increase compared to when interventions are both at level 0. Simultaneously, the probability of AGB being at level 0 decreases by 36%.

3.4. Analysis of Factor Interventions with Indirect Effects on AGB

Interventions were applied to biodiversity indexes that directly affect AGB, and they were clustered into Level 0 (low degree); Level 1 (medium); Level 2 (High Degree), see Figure 4.

Table 5. Conditional probability distribution of the level to which AGB may behave after intervention with biodiversity indexes which indirectly affect AGB.

Intervention Index 1		0	1	2
AGB	0	0.75	0.76	0.62
	1	0.18	0.18	0.25
	2	0.07	0.06	0.13

¹ Richness, Pielou, Simpson, PD, FDis, Fdiv, FDic.

shows that with low-level interventions on these seven indexes, the probability of AGB being at a low level is high but does not exceed 80%. As the level of intervention increases, the likelihood of AGB being at a medium or high level increases slightly. Specifically, when interventions on all indexes are at level 2, the probability of AGB being at level 2 is 13%, an 8% increase compared to when interventions are at level 0; concurrently, the probability of AGB being at level 0 decreases by 13%. However, compared to interventions on indexes with a direct impact on AGB, the effect of these indirect interventions is not significant.

Interventions were applied to Stand Density and Elevation, which have an indirect impact on AGB within the abiotic and biotic factors and they were clustered into Level 0 (Low Degree); Level 1 (Medium); Level 2 (High Degree), see Figure 5.

Table 6. Conditional probability distribution of the level to which AGB may behave after intervention with Stand Density and Elevation.

Intervention Index 1		0	1	2
AGB	0	0.75	0.70	0.68
	1	0.18	0.21	0.22
	2	0.07	0.09	0.10

¹ Stand Density, Elevation.

shows that low-level interventions yield results similar to those in Table 5. However, increasing the intervention level from 0 to 2 only results in a 3% increase in the probability of AGB being in Level 2. This indicates that the intervention effect on AGB is not significant.

4. Discussion

4.1. Causal Analysis of Forest Biomass and Its Driving Factors

Bayesian causal inference provides a deeper understanding of the causal relationships between factors. The results of our causal inference indicate that only the Shannon–Weaver index from the species diversity factors and the RaoQ index from the functional diversity factors directly influence forest biomass. This finding is consistent with Ouyang et al.’s [45] conclusion that functional and species diversity significantly impact biomass in subtropical forest plots. Additionally, we identified causal relationships among the variables within these factors. Shannon–Weaver is directly influenced by other indexes within the species diversity factors and by the biotic factor (Stand Density). This is because the Shannon–Weaver index integrates both species richness and evenness, making it a more comprehensive and representative measure of species diversity [46]. Moreover, the direct impact of stand density on Shannon–Weaver supports the complementarity effect hypothesis, which suggests that in denser stands, more intense competition between species may lead to niche differentiation and reduced direct competition, thus increasing diversity [47]. The RaoQ index is directly influenced by certain indexes within the functional diversity factors (FDis, FDiv, FRic) and Richness. RaoQ integrates information on species richness and functional differences among species, encompassing both aspects of functional diversity: functional richness and functional divergence, making it a comprehensive indicator of functional diversity [48].

However, our study found that phylogenetic diversity does not directly affect forest biomass. Instead, it influences forest biomass indirectly by affecting species diversity and functional diversity factors. Since phylogenetic diversity reflects the evolutionary history of species, these causal relationships suggest that evolutionary history may impact species distribution and functional performance within communities [21]. Additionally, phylogenetic diversity is directly influenced by environmental factors such as elevation. Different elevation zones provide varying climatic conditions, soil types, and vegetation types, which may promote species differentiation and adaptation, thus increasing phylogenetic diversity [22].

4.2. Compared with Species and Phylogenetic Diversity, Functional Diversity Had a More Significant Impact on Forest Biomass

The estimated results of the intervention show that applying a low degree of intervention to Shannon–Weaver and RaoQ results in a high probability of AGB being at a low level, reaching up to 84%. As the intervention level increases to high, the probability of AGB being at a high level rises by 20%, while the probability of it being at a low level decreases by 36%. In contrast, intervening in variables with indirect effects on AGB from low to high levels only increases the probability of AGB being at a high level by 8% and decreases the probability of it being at a low level by 13%. This indicates that species diversity and functional diversity have a more significant impact on forest biomass compared to phylogenetic diversity, as well as abiotic and biotic factors.

Furthermore, regarding the impact of species and functional diversity on forest biomass, when the level of intervention on Shannon–Weaver remains constant while the level of intervention on RaoQ increases, the probability of AGB being at a high level increases. This observation supports Petchey et al.’s [49] conclusion that functional diversity is a more significant predictor of forest biomass compared to species diversity. This may be because (1) species diversity does not fully account for functional similarity between species. If multiple species perform the same ecological function, species diversity might not reflect this functional redundancy [50]; (2) species diversity may not distinguish ecological niche differences between species. Even with high species richness, if many species occupy similar ecological niches, their functional diversity might be limited [51].

Our results contrast with Cadotte et al.’s [52] conclusion that phylogenetic diversity (PD) significantly influences biomass. In our analysis of subtropical forests, PD does not appear to be a strong driver of biomass. This discrepancy may be because the effectiveness of PD largely depends on whether traits exhibit a strong phylogenetic signal. If functional traits are highly variable evolutionarily, PD may not be a reliable indicator for predicting ecosystem functions [22]. Additionally, the effectiveness of PD in explaining ecosystem functions might depend on whether functional traits are related to competitive traits. If there is a lack of correlation between these traits, PD may not effectively predict ecosystem functions [23].

Functional diversity, by integrating key functional traits of different species, offers a more comprehensive perspective for assessing ecosystem diversity. These traits directly influence ecosystem processes such as productivity, decomposition, and nutrient cycling. Functional diversity can reveal niche differentiation among species, which may enhance complementarity, improve resource use efficiency, and increase ecosystem stability, potentially leading to higher biomass.

4.3. Advantages of Bayesian Causal Inference in Ecosystem Research

Traditional studies often use various linear models that describe only linear correlations between variables, which makes it difficult to clearly explain the reasons behind these associations due to the complex interactions and dependencies among variables [53]. Bayesian causal inference, through Bayesian networks, can represent causal relationships between variables using a directed acyclic graph (DAG). This structure allows the model to explicitly capture both direct and indirect relationships [54]. Our findings reveal that biodiversity factors, as well as biotic and abiotic factors, not only directly or indirectly impact biomass but also have intricate causal relationships among them. By employing Bayesian networks, we can comprehensively analyze the synergistic effects of these factors on biomass, rather than focusing solely on the impact of individual factors.

Although Barrufol et al. [55] found that structural equation modeling (SEM) reveals both the direct and indirect effects of environmental factors, stand density, succession stage, and management practices on tree diversity and biomass (productivity), complex models with multiple potential pathways can lead to overfitting or reduced explanatory power. Additionally, SEM often involves multiple hypothesis tests, which may introduce multiple testing issues and increase the risk of Type I errors (incorrectly rejecting a true null hypothesis) [56]. Bayesian methods, however, provide posterior probability distributions reflecting the effects of factor interventions (e.g., how different factor values might influence AGB). This allows for a comprehensive evaluation of model predictions, rather than relying on a single result. The distribution reflects the uncertainty in parameter estimates, enhancing the transparency and interpretability of the model.

Furthermore, Thomas [57] highlighted that multicollinearity in biodiversity research can lead to distorted or unreliable model estimates. Specifically, in the presence of non-perfect collinearity, ordinary least squares (OLS) estimates, while available, are no longer optimal because collinearity increases the standard errors of parameter estimates, affecting the stability and reliability of the model. Linear models are also highly sensitive to outliers and often require large datasets for robust estimates; even a small outlier can significantly impact the results. Bayesian causal inference, however, can incorporate domain or expert knowledge through prior information, allowing for reasoning even with incomplete or missing data. This approach reduces the model’s sensitivity to highly correlated explanatory variables and enhances robustness [58]. In this study, some plots in the Huangshan site are located on steep cliffs with significant terrain variation and some missing or imprecise data. The Bayesian causal inference used in this study showed stable performance with field data, without producing anomalous results.

5. Conclusions

Significantly different from previous studies, this paper constructs a causal network of biodiversity, biotic factors, and abiotic factors through Bayesian causal inference. The results indicate that functional diversity and species diversity are the most direct factors influencing forest biomass, while phylogenetic diversity only has an indirect effect. Furthermore, our intervention analysis of both direct and indirect factors shows that functional diversity has the most significant impact on forest biomass compared to species and phylogenetic diversity. Stand density and elevation also influence forest biomass, highlighting the importance of considering the synergistic effects of biotic and abiotic factors. This research provides new evidence on how biodiversity affects biomass in subtropical forests, complementing previous findings. The Bayesian causal inference algorithm used in this study offers a new perspective on understanding forest ecosystems. We recommend establishing long-term monitoring and evaluation mechanisms to regularly assess the effectiveness of forest management practices and ensure the continued improvement of forest ecosystems and biodiversity conservation.