Plant diversity’s positive effect on soil respiration diminishes with increasing productivity in global forests

Abstract

Plant diversity is essential for regulating ecosystem functions, yet its global-scale relationship with soil respiration, a critical component of the carbon cycle, remains unexplored. While plant productivity modulates ecosystem processes, understanding how plant species richness regulates soil respiration across net primary production gradients is key to predicting carbon–climate feedbacks. Here, we integrate two global plant species richness datasets (tree species richness and vascular plant species richness, including woody and non-woody plants) with a global soil respiration dataset estimated using a deep learning model trained on 6355 field observations. We demonstrate that plant diversity enhances soil respiration in low- to mid-productivity forests (<1300 g C m⁻² yr⁻¹), but this effect diminishes in high-productivity forests. By stratifying forests along a net primary production gradient and controlling for climatic, soil, and vegetation covariates via structural equation modeling, we reveal a context-dependent role of biodiversity. Greater plant species richness amplifies soil respiration in resource-limited systems but contributes minimally in high-productivity forests, where abiotic factors exert stronger influence. Our findings provide insights into how biodiversity influences soil carbon fluxes, revealing its dynamic role in shaping ecosystem carbon dynamics across productivity gradients.

Plant diversity plays a key role in regulating ecosystem processes, yet its influence on global soil carbon release remains unclear. This study suggests that higher plant species richness is associated with greater soil respiration in low- to mid-productivity forests but has little effect in highly productive systems.

Introduction

Forest plant diversity is essential for sustaining ecosystem functions, making it critical to elucidate how the loss of this diversity might degrade soil and impair overall ecosystem health^1–5. Recognizing plant diversity patterns in global forests is particularly important, as it provides valuable insights into how biodiversity contributes to ecosystem resilience and carbon dynamics across diverse environmental conditions⁶. Over recent decades, interest has grown in understanding the ecological importance of plant diversity for ecosystem functions^3,5,7–9. Numerous studies have shown a strong, positive relationship between plant diversity and soil organic carbon (SOC) storage and vegetation productivity across various ecosystems^3,10–13. As a result, biodiversity declines, driven by climate and land-use changes, could lead to important global carbon losses¹⁴. While substantial progress has been made in understanding the role of plant diversity in preserving SOC storage, its effect on soil respiration (R_s), a critical process in global carbon cycling, remains elusive.

R_s, the release of carbon dioxide (CO₂) from the soil into the atmosphere, is the second largest carbon flux and a key factor in regulating global carbon cycling^15,16. However, the relationship between plant diversity and R_s remains relatively unexplored, with no consensus emerging from the few studies available^17–20. Most evidence has been derived from small-scale experiments that manipulate plant species richness, often focused on grassland ecosystems, although such experiments are increasingly being conducted in forests^{3,5,18,21–24}. While it has been suggested that biodiversity experiments can effectively reflect real-world biodiversity–ecosystem functioning relationships²⁵, others argue that such experiments, especially those of short duration, may not fully capture the complexity of natural ecosystems, where biodiversity is not randomly distributed and biotic and abiotic factors are inherently intertwined¹³. Furthermore, these experiments are often limited to unique locations and involve a comparison of monocultures with mixtures of relatively few species, impairing their generalizability^10,23. A key limitation in understanding the global plant diversity–R_s relationship is the lack of consistent R_s observations at biodiversity monitoring sites. Addressing these limitations requires an investigation of the plant diversity–R_s relationship across global forests, which is essential for predicting future carbon dynamics and guiding biodiversity conservation efforts.

Belowground, the role of biodiversity in supporting key ecosystem functions may decline as ecosystems mature, due to higher plant biomass production and the accumulation of larger carbon and nutrient pools²⁶. For example, a recent study showed that the relationship between soil biodiversity and ecosystem function is stronger in geologically younger and less productive systems²⁷. In contrast, in older and more productive ecosystems, ecosystem functioning (including soil respiration) appears to become less reliant on soil biodiversity, suggesting that the accumulation of resources may reduce biodiversity’s marginal contribution to key soil processes. Aboveground, however, it remains unclear whether plant diversity follows similar patterns in influencing ecosystem functions. In particular, the extent to which plant diversity contributes more strongly to R_s in less productive ecosystems, where resource limitation may amplify diversity effects, has not been explicitly tested. If plant diversity functions analogously to belowground diversity, its influence on R_s may diminish in highly productive ecosystems. Therefore, it is crucial to consider productivity levels when examining the biodiversity–R_s relationship, as the impact of plant diversity may differ along gradients of net primary production (NPP).

In the present study, we investigate broad-scale empirical patterns between R_s, plant species richness, and productivity across global forests by integrating multiple global datasets. Drawing on evidence that soil biodiversity–ecosystem function (BEF)^1,2,28 may be stronger in less productive systems²⁷, we hypothesized that plant diversity is positively associated with R_s, and that its influence is more pronounced in low-productivity forests. To test this, we first develop global R_s predictions using a deep learning approach²⁹ based on 6,355 R_s field observations (Supplementary Figs. 1 and 2). Deep learning models have been increasingly adopted in Earth system science for their ability to capture complex, nonlinear relationships among interacting variables, without relying on linear assumptions^30,31. We then overlay global datasets of plant diversity, represented by tree species richness (TSR)³² and vascular plant species richness (including woody and non-woody plants; VPSR)³³, onto the predicted global R_s map along with environmental variables such as climate, vegetation, and soil properties (Methods). To explore the diversity–R_s relationship, we quantify the effects of plant species richness on observed R_s. To capture broader spatial patterns, we use partial dependence plots (PDPs) from model-predicted R_s at the global scale and across NPP gradients while controlling for environmental variables. We further quantify the unique contribution of TSR and VPSR to R_s by isolating their effect ($\theta$) while controlling for other environmental factors using a multiple linear regression approach across standardized levels of NPP. Finally, to disentangle direct and indirect pathways linking environmental variables, plant species richness, and R_s across productivity gradients, we use a structural equation modeling (SEM) framework.

Results

R_s-modeled predictions

To predict global R_s, we used a multi-layer perceptron (MLP), a deep learning approach, trained on 6355 monthly R_s observations from the MGRsD dataset³⁴ (Supplementary Fig. 1), covering the period 1982–2018 at the global scale. To ensure the model captured fine-scale spatiotemporal variability, we assembled a set of environmental variables at a resolution of 0.05° and at monthly intervals to better capture intra-annual variability, ensuring that the predicted annual mean R_s reflects underlying seasonal dynamics²⁹. These predictors, including climate, soil, and vegetation variables (Methods), were extracted to match each R_s observation in the training dataset. After performing a variance inflation factor (VIF) test to check for multicollinearity, feature selection yielded a comprehensive set of 12 variables used as input to train the model. Monthly global R_s predictions were then generated for each month from 1982 to 2018 and were subsequently aggregated to derive a mean annual R_s map (Supplementary Fig. 2a). The model shows strong predictive performance, achieving a coefficient of determination (R²) of 0.76 and a relative root mean square error (RRMSE) of 0.38 using 10-fold cross-validation (Supplementary Fig. 2c). The model’s robustness was validated through leave-one-year-out (LOYO) (R² = 0.74) and leave-one-site-out (LOSO) (R² = 0.65) cross-validations (Supplementary Fig. 2d, e), demonstrating its ability to accurately predict R_s while capturing both spatial and temporal patterns.

Global patterns of the forest species richness–R_s relationship

To quantify the marginal contributions of species richness to R_s, we trained Random Forest (RF) models based on observed R_s, while controlling for the effects of environmental variables (Supplementary Table 1). The purpose of this analysis was not to generate new predictions, but to interpret how species richness contributes most to observed variation in R_s after controlling for confounding factors. Contributions of plant species richness were quantified using Shapley Additive Explanations (SHAP) values (Methods). SHAP values provide an interpretable measure of how each variable influences R_s across its observed range. For species richness, higher SHAP values indicate a positive contribution to R_s, meaning that locations with greater VPSR or TSR tend to have higher R_s. The analysis revealed a positive relationship between species richness and R_s, with SHAP values increasing with VPSR (R² = 0.36, P < 0.001) and TSR (R² = 0.30, P < 0.001) (Fig. 1, main panels), indicating that species richness consistently enhances R_s.

Fig. 1. Marginal effects of plant species richness on soil respiration (R_s).

a SHAP-derived contribution of vascular plant species richness (VPSR) to observed R_s (Model efficiency: R² = 0.75). b SHAP-derived contribution of tree species richness (TSR) to observed R_s (Model efficiency: R² = 0.73). SHAP values (y-axis) indicate the marginal contribution of species richness to R_s, plotted against the corresponding richness values (x-axis). SHAP values were calculated for all R_s observations using 1000 Monte Carlo repetitions. Each blue dot represents one modeled R_s estimate (SHAP value) for an individual observation. Black lines show the estimated mean response from linear regression, with R² and P-values indicated. Shaded gray bands indicate 95% confidence intervals of the mean. Insets show partial dependence plots (PDPs) of R_s predicted by the MLP model. The red curves show the marginal effect of VPSR and TSR on R_s in global forests. VPSR includes all plants (i.e., herbs, shrubs, and trees), while TSR refers specifically to tree species richness. Note: ${\widehat{R}}_s$ does not show the predicted values, instead it shows how the values are changing with the change in VPSR or TSR, while averaging out the effect of other independent controls. Source data are provided as a Source Data file.

Further, to examine how species richness influences global patterns of predicted R_s, we included PDPs based on R_s predicted by the MLP model (Fig. 1, insets). PDPs were generated to visualize the marginal effect of VPSR and TSR on predicted R_s. We observed a generally positive influence of species richness on R_s, with a pronounced effect at lower richness levels that tapered off at higher levels. Importantly, the environmental covariates used in the PDPs were not included for variable selection or independent prediction of R_s, but for controlling for key environmental drivers, allowing the marginal effect of species richness to be better isolated at the global scale.

Plant diversity is an important driver of forest productivity^7,10,35, though its influence is often weaker than that of abiotic factors and functional composition in naturally assembled communities^36,37. Given that R_s tends to be higher in more productive forests³⁸, analyzing this relationship along an NPP gradient can provide an effective framework for understanding these dynamics. To evaluate whether species richness influences R_s differently across distinct forest productivity levels, we stratified our analysis by dividing NPP into 10 quantile-based bins, with an equal number of data points (pixels) for each bin (Supplementary Fig. 3). Our findings reveal that species richness exhibits a more pronounced positive relationship with R_s in low-productivity forests, whereas this relationship flattens in high-productivity forests, indicating reduced additional gains from species diversity (Fig. 2). Similar results were found for 20 and 50 quantile-based NPP bins (Supplementary Fig. 4 and Supplementary Tables 2 and 3).

Fig. 2. Partial dependence plots of soil respiration and species richness across net primary production (NPP) quantile-based bins.

a Relationship with vascular plant species richness (VPSR). b Relationship with tree species richness (TSR). VPSR includes all plants (i.e., herbs, shrubs, and trees), while TSR refers specifically to tree species richness. Note: ${\widehat{R}}_s$ does not show the predicted values, instead it shows how the values are changing with the change in VPSR or TSR, while averaging out the effect of other variables. Bins were divided into ranges of NPP values (10 quantiles) (Supplementary Fig. 3), ensuring each interval contains approximately the same number of data points.

Spatial variability of the effect of species richness on R_s

To estimate $\theta$ (i.e., the effect of plant diversity on R_s), we adopted a multiple regression approach to control for other potential factors (see Methods for details). We found that $\theta$ estimates were mostly positive in regions with lower productivity for both VPSR and TSR (Fig. 3), suggesting that increased species richness could enhance R_s in low- and moderate-productivity forests (normalized NPP from 0 to 50). Conversely, forests with the highest productivity (normalized NPP > 50) showed small or negative effects of species richness on R_s (Fig. 3a, b). The decline of this effect along the NPP gradient was significant for both VPSR (R² = 0.48, P < 0.001) and TSR (R² = 0.49, P < 0.001) (Fig. 3b, d). To propagate uncertainty in R_s estimates through our analysis, we conducted a supplementary sensitivity test using generalized least squares (GLS) regression. In this approach, each observation was weighted by the variance derived from the coefficient of variation (CV) of the R_s predictions (Methods). The resulting estimates of $\theta$ were highly consistent with the original unweighted estimates (R² = 0.95 for VPSR and R² = 0.82 for TSR, for weighted versus unweighted $\theta$), indicating that incorporating R_s uncertainty did not substantially affect the inferred effect sizes and the main observed patterns (Supplementary Fig. 5). Furthermore, we compared the estimates of $\theta$ across different forest types and climate zones (Supplementary Fig. 6). The effect of species richness on R_s varied significantly across forest types and climate zones (Kruskal-Wallis test and post-hoc test, P < 0.05) and was significantly different from zero in all groups (Wilcoxon test, P < 0.05).

Fig. 3. Spatial estimates of the species richness effect on soil respiration (R_s).

a, c Global distribution of the effect ($\theta$) of vascular plant species richness (VPSR) and tree species richness (TSR) on R_s, respectively. The $\theta$ values were calculated for each 100 bins of NPP. Values of $\theta$ > 0 indicate a positive effect of species richness on R_s, while values of $\theta$ < 0 indicate a negative effect. Only regions where $\theta$ values are significant at the 0.05 level are displayed. b, d Trend of the species richness effect across an NPP gradient (normalized NPP) for VPSR and TSR, respectively. Each dot represents the effect of species richness on R_s in the NPP bin. The black solid lines denote the mean $\theta$ from linear regression (P < 0.001) and the gray bands indicate 95% confidence intervals of the mean. Source data are provided as a Source Data file.

We validated the relationship between VPSR, TSR, and R_s using two methods (Supplementary Figs. 7 and 8). Significant positive relationships were found, particularly in low NPP regions, corroborating the initial patterns observed (Supplementary Fig. 7a, b). Additionally, we examined predicted R_s against observed plot-level VPSR and TSR. While there was a consistently positive trend across all productivity levels, the slope was steeper for plot-level data in low-NPP regions (Supplementary Fig. 7d). In contrast, some of the relationships between predicted R_s and plot-level VPSR were non-significant, notably for high and low NPP levels (Supplementary Fig. 7c).

The direct and indirect effects of environmental factors on R_s

In addition to the direct effect of species richness on R_s (i.e., its influence not mediated by other variables), environmental factors can affect R_s both directly and indirectly, by altering species richness, which in turn influences R_s. To evaluate the relative contributions of these pathways, we used a SEM approach. The direct effect of environmental variables on R_s was strong in low-productivity forests (${\beta }_1$ > 0.75), weakened in mid-productivity forests, and increased again in high-productivity forests (Fig. 4c, e). The indirect effect of environmental factors via VPSR was most pronounced in low- to mid-productivity forests (with a coefficient of 0.04 at normalized NPP level 30) and gradually decreased along the NPP gradient approaching zero, while remaining relatively stable for TSR (Fig. 4a, d). Additionally, $\theta$ for VPSR peaked in mid-productivity forests (normalized NPP level 30) but declined as NPP increased, suggesting that VPSR has its strongest influence on R_s at intermediate productivity levels. In contrast, $\theta$ for TSR was highest in low-productivity forests and diminished in mid-productivity forests. The SEM output was validated by incorporating weights derived from R_s uncertainty, confirming the robustness of the findings (Methods and Supplementary Fig. 5e–h).

Fig. 4. Direct and indirect effects of environmental factors and species richness on soil respiration (R_s) across net primary production (NPP) gradients.

a, d The direct effect of VPSR and TSR ($\theta$) (in orange) and the indirect effect of environmental factors (${\beta }_2\times \theta$) (in blue), respectively. b Structural equation model illustrating the relationship between a composite latent variable (environment) built from the 14 environmental variables (X₁–X₁₄), species richness, and the response variable (R_s). c, e The direct effect of environmental factors on R_s (${\beta }_1$) for VPSR and TSR, respectively (in red). The colored solid lines in (a–e) represent the loess regression of the coefficients to capture the trends. The shaded areas show the 95% confidence interval (CI) of the mean.

Discussion

Our analyses demonstrate that greater plant diversity is generally positively correlated with higher R_s in low to moderate levels of NPP. However, this effect shows diminishing returns as species richness increases, consistent with the concept of functional similarity (also referred to as functional redundancy)³⁹, whereby additional species may contribute less to ecosystem functioning due to overlapping ecological roles. This relationship is particularly pronounced in low- to moderate-productivity forests (Fig. 2), where the contribution of diverse plant communities to soil organic matter and microbial activity can significantly enhance R_s rates. These findings suggest that the positive effect of plant diversity on R_s is context-dependent, and corroborate previous studies that have unveiled the importance of biodiversity in maintaining ecosystem functions^36,40, particularly in carbon cycling and climate regulation^2,12.

Our results reveal a positive and significant relationship between plant diversity (both VPSR and TSR) and R_s across diverse forest types, supporting the BEF theory², which posits that higher species richness enhances ecosystem functioning. This relationship can be attributed to several mechanisms. One central mechanism is that diverse plant communities are more likely to include species with complementary belowground functional traits, such as variation in root depth⁴¹, root exudation rates⁴², or types of mycorrhizal associations⁴³, which can enhance nutrient cycling and stimulate soil microbial activity, ultimately promoting R_s and overall ecosystem productivity^7,44,45. Furthermore, increased productivity often results in higher rates of R_s due to greater inputs of soil organic matter⁴⁶ and possibly via nutrient resource maximization⁴⁷. However, the strength of the diversity–R_s relationship varies across biomes, suggesting that local environmental conditions and climate modulate this effect^13,48. For example, the tree diversity–R_s relationship was more pronounced in cold and boreal forests, where climatic conditions often limit aboveground plant growth (Supplementary Fig. 6d). In these regions, the positive impact of plant diversity on R_s may be amplified, as each additional species can enhance ecosystem function through both above- and belowground pathways, particularly under resource-limited and stressed conditions⁴⁹ where root activity and belowground productivity play a crucial role. This occurs because greater plant productivity from higher diversity promotes the accumulation of dead plant materials, providing organic resources for soil microorganisms and boosting soil microbial respiration¹⁷. In contrast, tropical forests, characterized by high species richness, exhibited a more heterogeneous effect on R_s spatially, possibly due to saturation effects where additional species contribute marginally to ecosystem function^2,50. Additionally, favorable conditions such as hot and wet climates can accelerate microbial activity, establishing a high baseline of decomposition and respiration¹⁶ that may optimize R_s (Fig. 4c, e) while reducing the influence of additional species on R_s, as observed in tropical systems where abiotic controls dominate over tree diversity effects⁵¹.

A key finding from this study is the substantial variation in the impact of plant diversity on R_s across different productivity levels. In low- and mid-productivity forests, plant diversity plays a crucial role in driving the resource flow into the soil system, stimulating nutrient cycling and microbial activity²⁷. This diversity may foster heterogeneous root structures and microbial communities, which could in turn amplify R_s⁵². Therefore, the lower availability of resources in these systems suggests that each additional plant species may have a significant effect on ecosystem processes. Conversely, the observed effect of species richness decreased or even became negative in mid- to high-productivity forests. Here, we propose plausible mechanisms that could underlie these patterns, while recognizing that these remain hypotheses beyond the direct evidence provided by our analysis. Underlying this trend are likely mechanisms associated with resource flows, competitive dominance, and functional similarity. First, the abundant organic matter inputs from a few dominant plant species in these forests lead to high microbial activity and decomposition rates³. Second, competition for resources intensifies and dominant species often monopolize available nutrients in high-productivity natural forests⁵³, potentially reducing the contribution of other species to ecosystem processes. Thus, the specific contributions of individual species to R_s are likely overshadowed by the high volume of organic matter from dominant species, leading to a reduced influence of plant diversity on R_s. Third, functional similarity, where species perform similar ecological roles, increases with species richness and can become more pronounced as productivity rises⁵⁴. When multiple species share similar functional traits, the loss or addition of species may therefore have a smaller impact on R_s, decreasing the ecosystem’s reliance on plant diversity for regulating soil processes and potentially diminishing or even reversing the positive relationship between species richness and R_s. Thus, this implies greater resilience to global change drivers in high-diversity forest ecosystems^55,56.

While our main analysis focused on the effects of plant diversity and productivity on R_s, we also explored whether specific plant functional groups might contribute to the observed patterns, as plant species with unique traits may impact microbial communities and R_s rates^57,58. Arbuscular mycorrhizal (AM) plants, for instance, are widely recognized for their role in carbon fluxes and may play a crucial role in modulating R_s rates^59,60, raising the question of whether their presence could control the biodiversity–R_s relationship. To investigate this, we examined how the biomass fraction of AM plants influenced the effects of species richness on R_s. Our findings suggest that this influence is minimal, as there was no consistent amplification of the TSR and VPSR effects on R_s with increased AM plant biomass (Supplementary Fig. 9). Instead, the influence of AM plants on this relationship was variable and context-dependent, with some forests showing weak or negligible effects. Therefore, the positive effect of plant diversity on R_s appears to be more closely tied to the functional diversity of the entire plant community rather than to a functional group effect⁶¹ driven by specific beneficial species such as AM plants. It is worth noting that while we focused on AM plants in this analysis, other plant-microbe interactions, such as ectomycorrhizal (EM) associations or soil rhizobia associations, could also play significant roles in influencing R_s⁶², potentially contributing to variability in our findings. Additionally, soil C/N ratios are known to influence microbial activity and nutrient cycling⁶³, and their interactions with plant diversity may further shape the biodiversity–R_s relationship. The decision to focus on AM plants was primarily driven by their established role in nutrient cycling, but we acknowledge that testing these other variables, such as EM, N-fixing plants, and soil C/N ratios, would provide a more comprehensive understanding of the factors driving R_s patterns.

This study provides valuable insights into the relationship between plant diversity and R_s across global forests, yet several limitations should be acknowledged. First, our analyses rely on global modeled datasets for VPSR, TSR, and R_s, which vary in quality, resolution, and collection methods across different data sources. To address this heterogeneity, we evaluated the impact of resampling and found high correlation and minimal distortion between original and resampled plant diversity datasets (Supplementary Fig. 10), indicating negligible effects on our conclusions. Second, our focus on forest ecosystems excludes other ecosystems, such as shrublands, grasslands, and croplands. Expanding future research to include these ecosystems would help clarify whether the observed diversity–R_s relationships hold across a wider range of vegetation and ecological conditions. Third, the non-significant trends in Supplementary Fig. 7c indicate that plot-level VPSR does not consistently explain R_s variability, revealing a limitation of the predictive model in capturing fine-scale biodiversity effects on R_s, which could be improved with more localized data or finer model resolution. We recommend that future R_s research systematically collects local plant diversity data to enable more robust paired data analysis. Finally, our study primarily captures a snapshot of the plant diversity–R_s relationship without accounting for temporal variations. Given that R_s is influenced by seasonal changes, climatic fluctuations, and long-term ecological dynamics will be essential for assessing how these relationships evolve over time, particularly in the context of ongoing environmental changes. Long-term studies monitoring changes in plant diversity and R_s are crucial for understanding the impacts of biodiversity on soil carbon dynamics.

Despite these limitations, our findings have significant implications for understanding how plant diversity influences R_s. The strong effect of plant diversity on R_s in low-productivity forests suggests that biodiversity plays a critical role in less productive ecosystems. Conversely, the weaker effect in higher-productivity forests implies that simply increasing plant diversity may not lead to higher R_s. These insights emphasize the importance of considering ecosystem productivity when designing biodiversity-focused conservation and management strategies. With ongoing global biodiversity loss, low-productivity forests face degradation risks, which could compromise their role in carbon sequestration. Our results emphasize that preserving biodiversity in these ecosystems is essential for maintaining R_s and supporting carbon cycling. Consequently, global conservation policies should recognize the importance of low-productivity forests as essential carbon sinks, crucial for climate change mitigation by supporting carbon cycling and potentially reducing atmospheric CO₂ levels. Integrating plant diversity metrics, such as species richness and functional traits, into carbon cycling models will improve our understanding of the feedback of ecosystem carbon dynamics to ongoing climate warming.

Methods

Soil respiration dataset

R_s observations used to train and predict global R_s were collected from the MGRsD dataset³⁴, which provides measurements at monthly intervals. Numerous studies have predicted R_s using the SRDB observation data^64,65; however, SRDB provides data at annual intervals, potentially overlooking seasonal variability in R_s driven by climate factors and thereby reducing prediction accuracy²⁹. Specifically, when studies in the SRDB provided detailed daily or monthly R_s data, this finer-scale information was extracted and included in MGRsD. Therefore, the use of MGRsD enhances temporal granularity and can better capture seasonal variability and climate-driven dynamics of R_s. In MGRsD, a unique SiteID is assigned if an observation’s spatial location (i.e., unique latitude–longitude coordinates) or experimental treatment (e.g., warming, nitrogen addition, or precipitation manipulation) is different. As a result, a site generally corresponds to a specific geographic point, although different treatments at the same location are assigned distinct SiteIDs. Moreover, the same location may appear in multiple studies, leading to potential overlap in coordinates across SiteIDs from different sources. The spatial extent of each site typically represents a small, localized area defined by the measurement coordinates. The number of observations per site varied widely, ranging from a single monthly value to more than 80, as many sites were sampled repeatedly across multiple months and years. Monthly R_s values were usually aggregated from daily or sub-daily measurements. In many cases, only one day of measurement was available per month, but a substantial portion of the dataset includes multiple daily observations per month, which were averaged to generate monthly values. Although most SiteIDs contain measurements for all 12 months of the year, some records have temporal gaps, particularly during winter months in cold regions, due to logistical challenges in field sampling. Despite missing data in some months, MGRsD retains sites with limited temporal coverage, including those with only a single month of data, which helps expand spatial and ecosystem representation in the dataset. Despite the extensive spatial coverage of the MGRsD, it is worth noting that biases persist in certain regions, particularly in the Northern Hemisphere and temperate ecosystems^29,66.

Our inclusion criteria were as follows: (1) Records lacking clear location, missing coordinates, or R_s information were not included, (2) measurements obtained through alkali absorption or soda lime methods were omitted, as they tend to underestimate soil CO₂ fluxes^67,68, (3) data from manipulative experiments such as nitrogen manipulation or fertilization were also excluded, and (4) potential duplicates and all other observations flagged as suspected problem or similar were excluded in the dataset. The final R_s dataset comprised 6355 observations spanning the period 1982 to 2018 (Supplementary Fig. 1).

Species-richness, arbuscular-mycorrhizal-tree, and vegetation datasets

To analyze species richness, we utilized datasets detailing TSR and VPSR. TSR data were obtained from the work by Liang et al.³², providing a high-resolution map (0.025°) of estimated local tree species richness (mean tree species richness per hectare) derived from a global forest inventory database from ~1.3 million sample plots. The authors used plot data from ~700 terrestrial ecoregions across the world, with each ecoregion encompassing 1800 sample plots on average, providing a robust and representative basis for their model. The dataset spans a wide range of forest types and climatic zones, making it one of the most comprehensive resources for studying global tree diversity. Moreover, its high spatial resolution and extensive coverage enable detailed analyses of species richness patterns and their links to forest ecosystem functioning across diverse biomes. While extensive, the authors acknowledged potential uncertainties in the dataset, which were mitigated to some extent. Measurement errors in forest inventories were addressed through strict taxonomic validation and data standardization, while uncertainties from the imputation of diversity maps were minimized using a three cross-validation approach. VPSR data were sourced from Sabatini et al.³³, providing a global-scale assessment of vascular plant alpha diversity, specifically local species richness at 2.5 arcminute resolution. This dataset was developed using the sPlot database, which incorporates more than one million vegetation plots and covers natural ecosystems on all continents and in all biomes, ensuring that richness estimates are consistent and comparable globally. For our analysis, we used the subset built using a fine spatial grain of 1 hectare (for forests only), which is necessary for capturing the variability of plant biodiversity at local scales, particularly in diverse and complex forest ecosystems. Nonetheless, the VPSR dataset has certain limitations. As noted by the authors, plot coverage is uneven, with dense samples in Europe but sparse data from tropical regions such as the Amazon and equatorial Africa. Furthermore, several tropical forest records focused mainly on woody species, leaving gaps in overall plant composition. Although targeted data searches and stratified resampling were employed to address these issues, data gaps and associated uncertainties still exist. For AM plants, we used data from ref. ⁴³, which provides global, high-resolution maps of aboveground vegetation biomass distribution by dominant mycorrhizal associations. In the study, the authors assembled the maps of mycorrhizal vegetation based on multiple published datasets, using a number of conversion factors to obtain per pixel values of mycorrhizal plant biomass fractions. We selected the AM plant biomass (expressed as the percentage of aboveground plant biomass of arbuscular mycorrhizal plant fraction), with a resolution of 10 arcminute. We used the MODIS NPP product (https://glass.hku.hk/archive/LAI/MODIS/0.05D/) to represent the forest productivity gradient. This dataset provides reliable productivity estimates at a resolution of 0.05°.

All variables were uniformly resampled to a resolution of 0.05° for VPSR and TSR analysis and ~0.16° for AM plant analysis, using bilinear interpolation. Forest areas were selected using a global forest dataset at a resolution of 0.05°⁶⁹. Non-forest areas were masked, retaining global forests with canopy cover greater than 10%, following the Food and Agriculture Organization definition⁷⁰.

Environmental variables and feature selection for R_s prediction

We initially selected 29 variables encompassing climatic, soil, and vegetation factors to train the R_s model. Climatic variables were derived from the WorldClim dataset (1901–2018) at a spatial resolution of 2.5 minutes, which included monthly minimum and maximum temperatures and total precipitation⁷¹ (https://worldclim.org/data/monthlywth.html, accessed May 2023). These high-resolution data were downscaled from CRU-TS 4.06 by the Climatic Research Unit at the University of East Anglia⁷². Additionally, evapotranspiration data from 1981 to 2018 were sourced from the Global Land Surface Satellite (GLASS) product at a spatial resolution of 0.05°⁷³ (http://www.glass.umd.edu/index.html, accessed May 2023). Soil variables were obtained from SoilGrid V2.0, covering depths of 0–5 cm and 5–15 cm, and with a spatial resolution of 250 m⁷⁴ (https://soilgrids.org/, accessed May 2023). These variables included soil bulk density (g cm⁻³), cation-exchange capacity (mmol(c) kg⁻¹), clay content (g kg⁻¹), coarse fragments (cm³ dm⁻³), pH in soil water, sand content (g kg⁻¹), silt content (g kg⁻¹), soil nitrogen (dg kg⁻¹), soil organic carbon (dg kg⁻¹), and soil organic carbon stock (t ha⁻¹, depth 0–30 cm). Vegetation variables, sourced from the GLASS product, included the fraction of absorbed photosynthetically active radiation (FAPAR), leaf area index (LAI), gross primary productivity (GPP), and fractional vegetation cover (FVC). These were available at a spatial resolution of 0.05° for 1981–2018^75,76 (http://www.glass.umd.edu/index.html, accessed March 2022). For each R_s observation site, environmental variables were extracted based on the site’s longitude and latitude coordinates. When necessary, these variables were uniformly resampled to a resolution of 0.05° using the bilinear interpolation method. The final model period, incorporating all variables, spanned from 1982 to 2018.

Deep learning model description for R_s predictions

The MLP model, a fundamental deep learning structure, extends the capabilities of the feedforward neural network⁷⁷. This model is renowned for its ability to capture complex, non-linear relationships, mapping capability, high parallelism, and global optimization^78,79. Incorporating an excessive number of environmental factors in machine learning models can disrupt data integrity, increase multicollinearity among variables, lead to overfitting, and ultimately result in model inefficiency and deviation. To address this, we employed the VIF test to assess multicollinearity among the environmental variables. Environmental factors with high multicollinearity were eliminated, ensuring that only independent variables were included in the MLP model. This approach enhances the model’s interpretability and effectiveness by avoiding overfitting and reducing computational costs⁸⁰. The VIF test was conducted using the usdm package in R, with a threshold set at 7. To compute the mean RMSE loss and determine variable importance, we utilized the DALEX package in R⁸¹. During the feature selection process, all input variables were initially ranked based on their importance and included in the MLP modeling process. The least important indicators were sequentially removed to create new feature sets. For each iteration, a new MLP model was constructed, and this process continued until the optimal feature set was identified. The final set of variables selected for the model included total precipitation, minimum temperature, bulk density (at 5–15 cm depth), cation-exchange capacity (at 5–15 cm depth), coarse fragments (at 0–5 cm depth), clay content (at 0–5 cm depth), soil nitrogen (at 0–5 cm depth), soil organic carbon stock (at 0–30 cm depth), soil organic carbon (at 5–15 cm depth), sand content (at 0–5 cm depth), silt content (at 5–15 cm depth), and LAI (monthly data at www.glass.umd.edu).

Our MLP model comprises an input layer, two hidden layers, and an output layer. The input layer corresponds to the number of input variables. The first hidden layer contains 200 neurons, while the second hidden layer has 120 neurons. To prevent overfitting, we implemented dropout regularization, randomly dropping 10% of neurons in each hidden layer during training. The output layer consists of a single neuron dedicated to regression predictions. Two critical parameters for MLP algorithms are the number of epochs (iterations of model training) and the batch size (number of samples per training session). These parameters were optimized for model performance. The epoch count was determined based on the root mean square error (RMSE) from 10-fold cross-validation of the MLP model over an initial set of 500 epochs. The batch size was set to 64 to mitigate potential overfitting issues, given the uneven distribution and limited size of the data samples. To prevent overfitting, we employed an early stopping technique, which halts training when performance on a validation dataset begins to degrade. Although the initial training was set for 500 epochs, early stopping identified the optimal number of epochs as 117. The MLP model was trained using the Keras library (version 2.9.0) in R. Subsequently, predictions were generated for each grid cell and for each month at a spatial resolution of 0.05° for the period 1982–2018. Monthly R_s predictions were chosen to better capture intra-annual variability, ensuring that the predicted annual mean R_s reflects underlying seasonal dynamics²⁹. We then aggregated monthly predictions to derive annual mean estimates for the period 1982 to 2018. While some environmental variables were spatially upscaled to coarser resolutions for analysis, R_s estimates were retained at their original prediction scale.

Evaluation of the MLP model performance for R_s predictions

The MLP model performance was evaluated using two metrics, the R² and the RMSE. R² quantifies the linear relationship between observed and predicted values, with values ranging from 0 to 1. A value closer to 1 indicates greater model efficiency. The equation for the calculation of R² is presented as follows (Eq. 1):

$$R^2=1-\frac{{\sum }_{i=1}^n{\left(y_i-\widehat{y_{\stackrel{°}{i}}}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y_{\stackrel{°}{i}}}\right)}^2}$$ 1

where $y_i$ represents the observed value, $\widehat{y_{\stackrel{°}{i}}}$ represents the predicted value, $\overline{y_{\stackrel{°}{i}}}$ represents the mean of observed values, and $n$ represents the number of observations.

RMSE measures the discrepancies between observed and predicted values, providing an average estimate of the model’s error. Lower RMSE values indicate better model performance. Here, we used the RRMSE, which represents the RMSE normalized by the average of the actual values. The equation for RRMSE is as follows (Eq. 2):

$$RRMSE=\sqrt{\frac{\frac{1}{n}{\sum }_{i=1}^n{\left(y_i-\widehat{y_{\stackrel{°}{i}}}\right)}^2}{{\sum }_{i=1}^n{\left(\widehat{y_{\stackrel{°}{i}}}\right)}^2}}$$ 2

where $y_i$ represents the observed value, $\widehat{y_{\stackrel{°}{i}}}$ represents the predicted value, and $n$ represents the number of observations.

To assess the model’s ability to capture temporal and spatial patterns, we utilized two cross-validation methods: LOYO and LOSO^82,83. In these methods, data from one year (or one site) are used for testing, while data from the remaining years (or sites) are used for training, ensuring comprehensive coverage across all years and locations. Additionally, we trained the MLP model 200 times using a Monte Carlo approach. Each time, we used a different random seed to initialize the model weights. For each grid cell, we calculated the CV of the predicted R_s values across these runs. CV values were predominantly below 100% (99.7% below CV of 100), indicating robust and consistent predictive performance. Only 0.0008% of terrestrial grid cells exhibited high uncertainty (CV > 200). The spatial pattern of uncertainty is shown in Supplementary Fig. 11.

SHAP and partial dependence plot analyses

To interpret the RF models, we applied SHAP analysis, a game-theoretic framework that quantifies the contribution of each predictor to the modeled variation in R_s while controlling for interactions among variables. SHAP values were calculated for all R_s observations using the fastshap package in R, specifically using the explain function and 1,000 Monte Carlo repetitions. To illustrate the influence of species richness on R_s, we generated SHAP dependence plots showing the relationship between richness and its corresponding SHAP values.

To visualize large-scale relationships between species richness and R_s, we generated PDPs based on R_s predicted by the MLP model. PDPs were computed using the pdp⁸⁴ package in R, which estimates the marginal effect of a selected variable on the modeled outcome, while averaging over the influence of all other covariates^85,86. Specifically, PDPs provide insights into the average marginal effects of variables across their entire range, revealing whether the relationships are linear, monotonic, or more complex. After conducting the PDP analysis at a global level, we performed PDP at different levels of NPP. We divided the NPP data into ten equal quantile-based bins. This method allows us to observe how the diversity–R_s relationship varies across distinct NPP ranges. Each bin contains N = 25,486 data points.

It is important to note that we consistently included a comprehensive set of environmental covariates (e.g., climate, soil properties, and vegetation) across all modeling steps. While the MLP model was used to predict global R_s, some of the same covariates were also incorporated in the subsequent PDP analysis. These overlapping variables were not included for variable selection or independent prediction, but to control for key environmental influences, allowing a clearer interpretation of the marginal influence of species richness after adjusting for the effects of key environmental drivers.

Effects of species richness (θ) on R_s and mapping

Forests with high species richness, typically found in more humid environments with elevated temperatures (e.g., tropical regions), might naturally exhibit higher R_s due to these favorable conditions rather than species richness alone. To remove the potential confounding effect from environmental covariates, we estimated the effect of species richness on R_s ($\theta$) using a multiple regression approach⁸⁷. We selected 14 predictor variables, including climate variables, soil physical properties, soil nutrients, and vegetation characteristics (Supplementary Table 1). Forests were pooled into different levels of NPP (normalized NPP gradient), and the effect was calculated for each bin. All data points within each NPP bin were used to fit regression models. Because we included 14 parameters in the models, a minimum sample size of ten times the number of parameters (at least 140 pixels) was required for accurate estimation of $\theta$. We excluded NPP bins with fewer than 140 pixels to ensure robust analysis across the 100 NPP bins.

The multiple regression model was developed to estimate $\theta$ while accounting for the potential confounded effects from environmental covariates. The multiple regression model was structured as follows (Eq. 3):

$$\log \left({Rs}_i\right)={\alpha }_0+\theta \times \log \left({\widehat{S}}_i\right)+{\sum }_{j=1}^p{\alpha }_j{X}_{i,j}+{\epsilon }_i$$ 3

where ${\epsilon }_i~$ i.i.d. $N\left(0\right),$σ²), ${Rs}_i$ is the soil respiration at location $i$, ${\widehat{S}}_i$ is the relative species richness at location $i$, calculated as the percentage of species richness relative to the maximum observed species richness within each NPP bin. $X_{i,j}$ are environmental variables for $j=1,\dots ,p$, ${\alpha }_0$ is the intercept term, ${\alpha }_j$ are regression coefficients, and ${\epsilon }_i$ is the error term. Both $Rs$ and $\widehat{S}$ terms were log-transformed to satisfy the normality assumptions of the linear model.

To assess the robustness of our results to uncertainty in R_s estimates, we conducted a supplementary sensitivity analysis in which we explicitly propagated R_s uncertainties into the regression framework. For each pixel, we used the CV raster derived from the R_s prediction uncertainties (Supplementary Fig. 11) as a measure of observation-specific variance. We then fitted our multiple regression models using GLS in the nlme package, weighting each observation by its variance. This downweighted higher-uncertainty observations and propagated R_s uncertainty into the downstream models. The GLS-based parameter estimates and p-values were compared with those from the main analysis (multiple regression and SEM) (Supplementary Fig. 5). Although R² values were slightly lower, the qualitative conclusions were unchanged, indicating that our results are robust to plausible levels of uncertainty in R_s predictions.

To test the robustness of our estimates obtained from the multiple regression approach, we also applied alternative modeling techniques, including two types of sequential regression models^87,88. These methods were chosen because they are specifically designed to accurately separate or remove the confounding effects of environmental covariates on R_s, allowing us to validate the consistency and reliability of our results. The first sequential regression model (Eqs. 4, 5) removes the influence of environmental variables on species richness by regressing log-transformed species richness against the environmental covariates. The residuals from this regression represent the variation in species richness unexplained by the environment. These residuals are then used as a predictor in a second regression of log-transformed soil respiration, alongside the environmental variables. This approach isolates the effect of species richness on soil respiration independently from environmental drivers, while still controlling for environmental influences on R_s.

$$\log \left({\widehat{S}}_i\right)={\alpha }_0+{\sum }_{j=1}^p{\alpha }_j{X}_{i,j}+{\epsilon }_i^{\left(1\right)}$$ 4

$$\log \left({Rs}_i\right)={\gamma }_0+\theta \times \mathrm{residual}\left(\log \left({\widehat{S}}_i\right)\right)+{\sum }_{j=1}^p{\gamma }_j{X}_{i,j}+{\epsilon }_i^{\left(2\right)}$$ 5

The second sequential regression model reverses the order by first regressing log-transformed soil respiration on environmental variables, then regressing the residuals from this model on log-transformed species richness. This approach estimates the effect of species richness on the component of R_s variation not explained by environmental covariates (Eqs. 6, 7):

$$\log \left({Rs}_i\right)={\alpha }_0+{\sum }_{j=1}^p{\alpha }_j{X}_{i,j}+{\epsilon }_i^{\left(1\right)}$$ 6

$$\mathrm{residual}\left(\log \left({Rs}_i\right)\right)=\theta \times \log \left({\widehat{S}}_i\right)+{\epsilon }_i^{\left(2\right)}$$ 7

The estimated values of $\theta$ from the first sequential regression model were identical to those obtained from the multiple regression approach (Supplementary Fig. 12a,c). Additionally, estimates of $\theta$ derived from the second sequential regression model were also highly consistent with those from the multiple regression model (VPSR R² = 0.95, P < 0.001 and TSR R² = 0.79, P < 0.001) (Supplementary Fig. 12b,d), supporting the robustness of $\theta$ estimation.

To map the $\theta$, we ensured the species richness effect was evaluated across different productivity levels. We mapped the values across the global landscape by assigning the corresponding $\theta$ to each grid cell based on its NPP value. Regressions that did not reach a significant level of 0.05 were not represented on the maps. The forest types were defined following the Global Land Cover⁸⁹ and classified into five categories including deciduous broadleaf forests, deciduous needleleaf forests, evergreen broadleaf forests, evergreen needleleaf forests, and mixed forests. Climate types were defined following the Köppen-Geiger climate classification⁹⁰, including arid, cold, polar, temperate, and tropical climates.

We validated the relationship between VPSR, TSR, and R_s using a two-way validation approach. This method helped account for potential spatial mismatches in the datasets and verify the consistency of our results. First, we selected observed R_s values and extracted the corresponding locations for predicted TSR and VPSR data. Second, we conducted a reciprocal analysis by examining observed VPSR and TSR values and extracting the corresponding predicted R_s data (Supplementary Fig. 7). For clarity, this validation was conducted across three defined NPP levels, defined as low level (13.1 to 555.9 g C m⁻² yr⁻¹), medium level (561.2 to 709.6 g C m⁻² yr⁻¹), and high level (710.9 to 1995.0 g C m⁻² yr⁻¹).

Structural equation modeling analysis

To explore the relationships between environmental variables, species richness, and R_s, we employed an SEM approach. Unlike multiple regression, which estimates marginal effects while controlling for covariates, SEM allows us to explicitly model the causal pathways and partition direct and indirect effects among correlated predictors. This is particularly useful in ecological systems where environmental variables simultaneously influence both biodiversity and ecosystem functioning. Therefore, we modeled the direct and indirect effects of multiple environmental drivers and plant diversity on R_s across varying productivity levels. We created a composite latent variable (environment), to represent the collective influence of climate, soil, and vegetation variables (including NPP), to reduce model complexity and account for shared variance among environmental predictors. To ensure normal distribution species richness variables (VPSR and TSR) and R_s were log-transformed. The SEM model involved two primary relationships⁸⁷. First, we modeled the direct effect of the environmental composite variable on species richness (${\beta }_2$), hypothesizing that environmental conditions directly influence biodiversity levels. Second, we assessed how the environment directly affects R_s (${\beta }_1$), while also considering the indirect effects of species richness on R_s (${\beta }_2\times \theta$). This allowed us to estimate both the direct influence of environmental variables and the contribution of species richness as a mediator. We then applied the SEM model to each NPP level. The SEM analysis was also complemented by weighted linear regression models, where observations were weighted by the inverse variance derived from R_s prediction uncertainty (Supplementary Fig. 5e–h).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

B.L. conceptualized, curated the data, performed the formal analysis, conducted the investigation, developed the methodology, developed the software, visualized the data, and wrote the original draft. Z.Y. performed the formal analysis, developed the methodology, developed the software, and reviewed and edited the manuscript. J.J., P.C., L.C., Y.Y., X.C., and N.H. reviewed and edited the manuscript. Z.T. developed the software, developed the methodology, and reviewed and edited the manuscript. X.P. conceptualized the study and acquired the funding. X.T. conceptualized the study, supervised the research, acquired the funding, and reviewed and edited the manuscript.

Peer review

Peer review information

Nature Communications thanks Syeda Ferdous and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The monthly soil respiration database is freely available at (https://data.lib.vt.edu/articles/dataset/Monthly_Global_Soil_Respiration_Database_MGRsD_/14098967)(monthly data; MGRsD). The soil data are freely available at (https://soilgrids.org/). The climate data used in this study can be accessed from WorldClim at (https://www.worldclim.org/data/monthlywth.html). Vegetation and evapotranspiration data were accessed from GLASS website at (http://www.glass.umd.edu/index.html). Forest cover data are available freely at ref. ⁶⁹. TSR data are available at: 10.1038/s41559-022-01831-x and at: 10.6084/m9.figshare.17232491. The VPSR data are available at: 10.25829/idiv.3506-p4c0mo. The AM plant maps used in this study are provided as a Source Data file (ref. ⁴³) and are also deposited at (https://github.com/nasoudzilovskaia/Soudzilovskaia_NatureComm_MycoMaps). The global predicted R_s map from this study is available at Figshare [10.6084/m9.figshare.29371901]. Source data are provided with this paper.

Code availability

The R codes used to produce the results can be accessed at Figshare (10.6084/m9.figshare.29371901).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-69594-8.