Abiotic and biotic drivers of soil microbial diversity in an intensively grazed natural ecosystem

Abstract

Many ecosystems worldwide are threatened by anthropogenic causes, with high-intensity grazing by large herbivores as a significant risk factor for biodiversity. Although the drivers of α-diversity are well-studied for animal and plant communities, they are often overlooked for soil microbes, particularly in natural systems. We therefore used a novel innovative information-theoretic approach to structural equation model selection and multimodel path coefficient averaging to identify these drivers. Our findings show that abiotic soil characteristics, primarily soil pH, significantly shape the α-diversity of both bacteria and fungi. Biotic factors like vegetation Shannon diversity and aboveground biomass also significantly drive microbial α-diversity, especially for fungi. Our statistical approach adds robustness to our results and conclusions, offering valuable insights into the complex interactions shaping soil microbial communities in intensively grazed natural systems. These insights are crucial for developing more effective and comprehensive future ecosystem management and restoration strategies.

Introduction

As global environmental changes accelerate, maintaining ecosystem health and biodiversity becomes increasingly crucial^1–3. High-intensity grazing by large herbivores can severely impact ecosystems, often undermining their stability and functioning. These impacts include reduced vegetation diversity⁴, increased invasion by non-native species⁵, and diminished vegetation cover⁴. Consequently, high grazing intensity is a major risk factor for biodiversity loss and ecosystem degradation⁶.

Despite the well-documented state and recovery of plant and animal communities in systems degraded by grazing, soil microbial diversity is often overlooked⁷. A review of articles from the past two decades on Web of Science containing “Ecosystem recovery” and “Grazing” or related terms shows that only 2% also included “Soil microbial diversity” or related terms (Supplementary Table 1). Yet, because soil microbial diversity underpins many macro-scale ecosystem processes⁸, it is critical to develop a deeper understanding of its environmental drivers in systems degraded by high-intensity grazing. Such insights are essential for enhancing the effectiveness of future ecosystem restoration efforts.

Our current understanding of the drivers of soil microbial α-diversity in intensively grazed systems mainly derives from controlled experiments and/or pastoral systems^9–12. Chinese rangeland experiments have shown, for example, that high grazing intensities reduce total microbial numbers, with fungal losses linked to decreased aboveground biomass and bacterial declines tied to increased bulk density from soil compaction¹². Additionally, nutrient levels related to pH and soil organic matter (SOM) were identified as key drivers of both bacterial and fungal α-diversity under high grazing pressure¹¹. Moreover, vegetation cover and diversity appear to influence fungal diversity more than bacterial diversity in overgrazed pastoral systems^13,14.

However, a significant gap remains in our understanding of the strength and importance of these drivers in natural ecosystems grazed by wild herbivores. This distinction is crucial, as wild and domestic herbivores, while sometimes exerting similar effects¹⁵, can also have contrasting impacts on various ecosystem aspects, including the factors expected to influence soil microbial α-diversity^16,17. Natural ecosystems are inherently more heterogeneous and dynamic compared to managed pastoral systems, characterized by pronounced ecological gradients¹⁸ along which grazing exerts varying environmental effects^19,20. By examining these multifaceted interactions in natural systems, our study seeks to bridge this knowledge gap, which could enhance the effectiveness of future ecosystem restoration strategies.

We aim to identify the key abiotic and biotic environmental factors shaping the α-diversity of soil microbes in a long-term, intensively grazed natural coastal sand dune ecosystem in The Netherlands. To achieve this, we collected microbial samples and quantified their α-diversity (i.e., richness and Shannon diversity) using the Hill diversity indices framework (see Methods)²¹. Sampling locations covered various strong environmental gradients, allowing us to relate associated environmental factors to ecosystem-wide patterns in microbial diversity. The aforementioned studies suggest that abiotic factors strongly influence both bacterial and fungal α-diversity, whereas biotic factors have a stronger impact on fungal α-diversity than on bacterial α-diversity^9–14. We therefore hypothesize that: (i) abiotic factors play a predominant role in driving microbial α-diversity, and (ii) fungal α-diversity is driven more by biotic factors than bacterial α-diversity in our natural ecosystem with wild herbivores.

Our innovative approach to test these hypotheses lies in the use of a relatively novel information-theoretic approach to structural equation model (SEM) selection and multimodel inference (Fig. 1)^22–24. This method, so far rarely utilized in ecological studies, allows us to account for the simultaneous influence of multiple factors and their interconnectedness, while also considering the uncertainty in multiple competing candidate models²⁵. Unlike traditional regression and SEM methods, this approach enables a more nuanced understanding of the direct, indirect, and total causal effects of environmental factors on soil microbial α-diversity. Additionally, we used the same global SEM model for both bacterial and fungal α-diversity, enabling a direct comparison of how the same set of environmental factors influences each individual soil microbial group and diversity index. This standardized approach allows us to assess shared and distinct drivers of α-diversity across the two groups within a unified analytical framework.

Fig. 1. Schematic overview of the employed information theoretic approach to structural equation model (SEM) selection and multimodel inference.

The overview outlines the four-step process used to identify the drivers of soil microbial α-diversity in a intensively grazed natural ecosystem. Step 1 involves reducing the number of factors included in later steps to avoid overfitting and collinearity issues by calculating pairwise correlations and excluding highly correlated factors (Supplementary Fig. 1), resulting in seven environmental factors. In Step 2, the global SEM, incorporating all hypothesized causal pathways based on a priori knowledge of the study system, was specified (see Supplementary Table 2 for pathway justifications). Step 3 includes generating a list of SEM models with all possible additive combinations of direct effects for each microbial α-diversity index. Relationships between environmental factors were not expected to differ between models and were therefore kept constant over each individual model. A candidate model set was created by selecting only those models with a ΔAICc <4.0 compared to the top model. In Step 4, weighted model-averaged and standardized path coefficients were calculated through bootstrapping, adjusting for model-selection bias. These coefficients were used to decompose the total causal effect into direct and indirect components, evaluating the relative importance of each environmental factor as driver of soil microbial α-diversity. For a full description of each step, see Methods section.

Results

Environmental gradients

Our sampling strategy involved sampling various habitat types (i.e., open sand with pioneer vegetation, grassland, and scrub) at each location to capture full local and ecosystem-wide environmental gradients. Significant differences among these habitats in the environmental factors used in later analyses highlight the thoroughness of our sampling (Table 1). Specifically, soil organic matter (SOM) and aboveground biomass increased significantly from open sand to grassland and scrub. Moss/lichen cover was lowest in open sand, highest in grassland, and moderate in scrub habitats. pH values were lowest in grassland and scrub habitats, while grass/herb cover and vegetation richness were lowest in open sand. Vegetation Shannon diversity was highest in scrub habitats compared to the other types.

Table 1.

Differences between habitat types in the environmental factors used in later SEM analyses

	Open sand (n = 25)	Grassland (n = 31)	Scrub (n = 25)
Vegetation richness (effective species number)	8.3 ± 2.4a	17.1 ± 4.6b	20.4 ± 7.7b
Vegetation Shannon diversity (effective species number)	5.6 ± 2.1a	6.1 ± 1.9a	7.9 ± 3.1b
Moss/lichen cover (%)	2.6 ± 4.9a	67.2 ± 27.9b	30.5 ± 23.0c
Grass/herb cover (%)	29.3 ± 13.0a	56.6 ± 21.4b	59.2 ± 22.8b
Aboveground biomass (gr/m2)	30.0 ± 19.9a	139.1 ± 88.4b	1174.3 ± 709.8c
pH	7.5 ± 0.8a	5.5 ± 1.1b	5.7 ± 1.1b
SOM (kg/m2)	0.5 ± 0.3a	2.6 ± 1.2b	3.4 ± 1.0c

Presented are mean ± standard deviation values. Letters indicate significant differences between habitat types (permutational t-test, 9999 permutations, p < 0.05 after Holm correction for multiple comparisons). Vegetation diversity is expressed using Hill diversity indices, corresponding to species richness (Hill-richness; q = 0) and Shannon diversity (Hill-Shannon diversity; q = 1).

The model-averaged structural equation models (SEMs) (Figs. 2–3; Supplementary Tables 3-10) showed that nearly all hypothesized direct effects between individual environmental factors were significant. The only exceptions were the direct effects of grass/herb cover and vegetation richness on aboveground biomass, which were not significant.

Fig. 2. Structural equation model diagrams of the direct and indirect effects of biotic and abiotic factors on bacterial α-diversity.

Depicted are the bootstrapped standardized path coefficients averaged over the corresponding candidate model sets for (a) richness and (b) Shannon diversity (Supplementary Tables S3-4). Arrows indicate hypothesized causality, with widths proportional to the standardized path coefficient. Unimodal relationships are indicated with a “u”. Solid green arrows indicate significant (p < 0.05) positive effects and solid red arrows significant negative effects. Dashed arrows represent insignificant positive (green) and negative (red) effects. Bacterial and vegetation diversity are expressed using Hill diversity indices, corresponding to ASV/species richness (Hill-richness; q = 0) and Shannon diversity (Hill-Shannon diversity; q = 1).

Fig. 3. Structural equation model diagrams of the direct and indirect effects of biotic and abiotic factors on fungal α-diversity.

Depicted are the bootstrapped standardized path coefficients averaged over the corresponding candidate model sets for (a) richness and (b) Shannon diversity (Supplementary Tables 5-6). Arrows indicate hypothesized causality, with widths proportional to the standardized path coefficient. Unimodal relationships are indicated with a “u”. Solid green arrows indicate significant (p < 0.05) positive effects and solid red arrows significant negative effects. Dashed arrows represent insignificant positive (green) and negative (red) effects. Fungal and vegetation diversity are expressed using Hill diversity indices, corresponding to ASV/species richness (Hill-richness; q = 0) and Shannon diversity (Hill-Shannon diversity; q = 1).

Bacterial α-diversity drivers

When averaged across the candidate SEM set (Supplementary Table 3), pH (unimodal) and vegetation Shannon diversity (positive) emerged as significant direct drivers of bacterial Amplicon Sequence Variant (ASV) richness (pH: bootstrapped standardized direct (BSD) effect = 0.75; vegetation Shannon diversity: BSD effect = 0.14; Fig. 2a; Supplementary Table 7). Vegetation richness had a positive indirect effect on bacterial richness via vegetation Shannon diversity (bootstrapped standardized indirect (BSI) effect = 0.07). Moss/lichen cover (BSI effect = 0.08) and grass/herb cover (BSI effect = 0.01) also had significant positive indirect effects. Assessing environmental factor importance based on their significant total effects revealed pH as the most important driver of bacterial richness (bootstrapped standardized total (BST) effect = 0.72), followed by vegetation Shannon diversity (BST effect = 0.14) and vegetation richness (BST effect = 0.08; Fig. 4a).

Fig. 4. Effect sizes of each biotic and abiotic factor on microbial α-diversity.

Depicted are the bootstrapped and model-averaged standardized direct, indirect, and total effect sizes and their corresponding 95% confidence intervals of each biotic and abiotic factor on (a) bacterial and (b) fungal α-diversity. Microbial and vegetation diversity are expressed using Hill diversity indices, corresponding to ASV/species richness (Hill-richness; q = 0) and Shannon diversity (Hill-Shannon diversity; q = 1).

Results for bacterial Shannon diversity were similar (Supplementary Table 4), with significant direct effects of pH (unimodal, BSD effect = 0.66) and vegetation Shannon diversity (positive, BSD effect = 0.21), and a significant indirect effect of vegetation richness (positive, BSI effect = 0.11; Fig. 2b; Supplementary Table 8). Significant total effect ordering resulted in pH being the most important driver of bacterial Shannon diversity (BST effect = 0.58), followed by vegetation Shannon diversity (BST effect = 0.22; Fig. 4a).

Fungal α-diversity drivers

For fungal richness, significant direct effects were observed for pH (unimodal, BSD effect = 0.28), vegetation Shannon diversity (positive, BSD effect = 0.20), aboveground biomass (positive, BSD effect = 0.27), and SOM (unimodal, BSD effect = 0.13) when averaged over the candidate model set (Fig. 3a; Supplementary Table 5; Supplementary Table 9). Significant indirect effects included vegetation richness (BSI effect = 0.17), moss/lichen cover (BSI effect = 0.17), grass/herb cover (BSI effect = 0.05), and SOM (BSI effect = 0.08). Aboveground biomass was the most important driver of fungal richness based on significant total effects (BST effect = 0.27), followed by vegetation Shannon diversity (BST effect = 0.22), SOM (BST effect = 0.20), and vegetation richness (BST effect = 0.15; Fig. 4b).

Regarding fungal Shannon diversity, significant direct effects were found for aboveground biomass (positive, BSD effect = 0.30) and grass/herb cover (positive, BSD effect = 0.14) when averaged over the candidate model set (Fig. 3b; Supplementary Table 6; Supplementary Table 10). Significant indirect effects included vegetation richness (BSI effect = 0.08) and SOM (BSI effect = 0.08). Ordering of significant total effects revealed aboveground biomass as the most important driver of fungal Shannon diversity (BST effect = 0.30), followed by SOM (BST effect = 0.25), and grass/herb cover (BST effect = 0.18; Fig. 4b).

Discussion

Our study identified key environmental factors driving soil microbial α-diversity in a natural ecosystem with pronounced ecological gradients and intensive grazing by wild herbivores. We found that abiotic soil characteristics play a large role in shaping the α-diversity of both bacteria and fungi, supporting our first hypothesis. Additionally, biotic factors like vegetation Shannon diversity and aboveground biomass also drive microbial α-diversity, but, in line with our second hypothesis, for fungi more so than for bacteria. To robustly identify these drivers, we employed a relatively novel information-theoretical approach to structural equation model selection with multimodel bootstrapped path coefficient averaging^23,24,26. This method addresses the challenges of strong variable correlations in observational studies, overcoming the limitations of traditional stepwise selection methods that often ignore model uncertainty and other plausible models^23,25,27. By ranking models based on an information criterion that balances fit, precision, and complexity, our approach allows for inferences drawn from the best-aligned models with the observed data. This robust statistical framework enhances the reliability and comprehensiveness of our conclusions on the drivers of soil microbial α-diversity.

Our findings confirm the pivotal role of pH as an abiotic driver of bacterial α-diversity. This aligns with broader scale observations²⁸, experimental settings²⁹, and with patterns found in pastoral/experimental grazing effect studies^10,13. The pronounced influence of pH on bacterial richness and diversity can be attributed to its direct effects on cytoplasmic homeostasis³⁰, modulation of soil phosphorus and nitrogen availability, and varying symbiotic strategies for plant nutrient acquisition along the pH gradient^31,32. For fungal α-diversity, such processes likely shape the patterns observed in our ecosystem as well. However, the total effect of pH on fungal α-diversity was insignificant, due to the negative indirect effect through interactions with soil organic matter (SOM), vegetation Shannon diversity, and aboveground biomass. Also, the broader pH-optimum ranges exhibited by various fungal species^28,33 may have weakened the direct effect of pH on fungal richness and diversity.

An alternative explanation for the lower effect of pH on fungi may be due to bacterial inhibition of fungal growth rates across the pH gradient. In an agricultural field study, inhibiting bacterial growth led to equalized fungal growth rates across the entire pH gradient³³. Thus, the weaker effect of pH on fungal compared to bacterial α-diversity may also be due to a negative competitive effect of bacteria on fungi. Since we did not include microbial biomass or growth rates in this study, we could not test whether bacterial-fungal competition played a role. Interestingly, results from Kooijman et al.³¹ suggest that the dominance of AM (Arbuscular Mycorrhizal) plants in calcareous dune parts and the dominance of NM (Non-Mycorrhizal) plants in acidic dunes disrupt the typical relationship between microbial group dominance, influencing the effects of pH on microbial diversity.

Regarding the role of soil organic matter, high-intensity grazing can alter soil microbial communities by shifting competitive exclusion among dominant microbial species, transitioning from slow-growing species reliant on recalcitrant carbon from SOM to fast-growing species primarily sustained by labile carbon sources^34,35. This phenomenon relates to the idea that rhizosphere soils are generally inhabited by symbiotic microbes that are strongly influenced by the plant community and its labile root exudates³⁶, whereas non-rhizosphere soils primarily host free-living decomposers relying on recalcitrant carbon from SOM. Our samples included both rhizosphere and non-rhizosphere soil. When analysing such bulk soil samples, SOM frequently emerges as a major driver of microbial diversity^37,38, with a stronger effect on fungi than on bacteria³⁸. This differential impact is supported by our data.

Nonetheless, the relatively limited effect of SOM on microbial diversity in this study suggests that the microbial communities consisted mainly of bacteria and fungi associated with plant roots. This could result from a decrease in SOM input and quality due to prolonged high-intensity grazing^39,40. Selective feeding by deer on palatable species reduces the amount of easily decomposable litter rich in labile carbon sources reaching the soil⁴¹, as plant palatability and litter decomposability are governed by the same traits⁴². Additionally, less fresh litter input can lower the priming effect necessary for the stimulation of recalcitrant carbon decomposition⁴³. While aboveground biomass is heavily grazed, root systems still contribute to soil carbon and other nutrients through root exudates⁴⁴, albeit at potentially reduced levels. Together, these factors can shift the soil microbial community to rely more on labile carbon sources from root exudates³⁴. Further support for this close association with plant roots comes from the importance of various biotic factors associated with root exudates as drivers of α-diversity.

Using our structural equation model averaging approach for bacterial α-diversity, we found compelling evidence that vegetation Shannon diversity, closely linked to vegetation richness, is the sole significant biotic driver. Previous studies have reported mixed results on the presence of a causal association between bacterial and vegetation diversity^13,45,46. These conflicting findings propose that soil chemical and/or physical properties may play a more critical role in shaping bacterial communities than the vegetation does³⁶. This is supported by our study, where pH was the most influential driver, far surpassing vegetation diversity or any other biotic factor. However, the importance of vegetation Shannon diversity in driving bacterial diversity suggests that high grazing intensity and strong pH gradients do not negate the role of the vegetation. Plant communities with higher diversity may provide a broader range of resources and niches for soil bacteria, thereby promoting richer and more diverse bacterial assemblages on an ecosystem-wide scale.

In a study by Kowalchuk et al. ⁴⁷ in a comparable Dutch dune area, specific plant species harboured highly distinct bacterial communities in their rhizosphere. Consequently, more diverse plant communities were associated with greater bacterial diversity. However, bulk soil samples exhibited strikingly similar bacterial diversity, which made the authors state that “sampling strategies that include a greater amount of soil not under the direct influence of the plant, may serve to dilute out any ‘rhizosphere’ effects”⁴⁷. Nevertheless, despite using bulk soil samples, our study revealed a notable influence of vegetation diversity on bacterial richness and diversity, underscoring the close connection of our bacterial communities with plant root-derived nutrient sources. High plant diversity fosters higher bacterial α-diversity due to a more varied root exudate profile^48,49, increased variation in symbiotic nutrient acquisition strategies⁵⁰ and enhanced niche availability facilitated by diverse root architecture⁵¹.

For fungal α-diversity, similar mechanisms are thought to govern its relation to vegetation richness and diversity^52,53. Our data support this for fungal richness but not for fungal Shannon diversity. Increased plant diversity leads to greater root inputs to the soil⁵⁴, boosting microbial activity³⁶, and often correlating with higher fungal richness⁵⁴. Conversely, fungal Shannon diversity is believed to depend mainly on SOM quantity and quality³⁶. As discussed, prolonged high-intensity grazing can reduce SOM quality^39,40, shifting microbial reliance from SOM toward root-derived nutrients³⁴. While vegetation Shannon diversity is therefore often considered influential for fungal diversity^49,55, its significance may vary in environments with high local habitat heterogeneity. Shannon diversity measures species evenness, and even plant communities are expected to produce even root exudate profiles⁴⁸, theoretically resulting in fungal communities with high Shannon diversity. However, high plant species evenness may not translate to high spatial evenness due to local species aggregation and vegetation patchiness, exacerbated by high-intensity grazing⁵⁶. This could explain why our analyses did not find vegetation Shannon diversity to be an important driver of fungal Shannon diversity.

On the other hand, increasing grass/herb cover does lead to a more even distribution of roots and a more uniform spatial input of root exudates and fresh litter^57,58. This increases the number of available niches and food sources for soil fungi, positively influencing fungal abundances⁵⁹. This agrees with the relatively high importance of grass/herb cover as a driver of fungal Shannon diversity rather than of fungal richness. Additionally, rising grass/herb coverage and the concurrent decrease in open sand cover reduces temperature fluctuations, enhances soil water holding capacity, and lowers bulk density⁶⁰. Lower bulk density, strongly correlated with grass/herb cover, benefits soil fungi by increasing pore space and connectivity, improving air circulation⁶¹, and by providing more room for hyphal growth⁶². This may also explain why grass/herb cover was not a significant driver of bacterial α-diversity. Bacteria generally require less oxygen for respiration and less space to grow and disperse^63,64, allowing them to persist over a wider range of soil bulk densities⁶⁵.

The predominant driver of both fungal α-diversity indices was aboveground biomass. We included aboveground biomass because of its use as a proxy for belowground biomass, given the typically isometric relationship between the two⁶⁶. As root biomass increases, niche availability and nutrient supply to soil microbes are enhanced, thereby positively impacting soil microbial α-diversity⁶⁷. This influence appears to be more pronounced for fungal biomass compared to bacterial biomass, a finding supported by our results on amplicon sequence variant (ASV) species richness and diversity. These results are also in line with a study on grazing intensity in a Chinese steppe ecosystem, which identified pH as a dominant driver of bacterial α-diversity, while aboveground biomass was the predominant driver of fungal α-diversity¹³.

It is important to acknowledge that our study did not directly measure belowground biomass, potentially leading to a misestimation of its direct impact on soil microbial α-diversity. Within the coastal sand dunes, calcareous areas are characterized by high pH and low P-availability. Under P-limitation, plants can adopt a strategy of “outsourcing” nutrient acquisition by prioritizing mutualistic mycorrhizal relationships over root growth⁶⁸. Conversely, in more decalcified areas, plants tend to adopt a “do-it-yourself” approach, investing more in root biomass. Hence, the assumed isometric relationship between above- and belowground biomass may not apply to our study system. Moreover, grazing may further influence this relationship, as high-intensity grazing tends to favour short-stature plants with a high root-to-shoot ratio^4,69. Future research should therefore examine the exact relationship between above- and belowground biomass within the coastal dune system to verify the validity of using of aboveground biomass as a proxy for belowground biomass.

We anticipate that with a complete cessation of large herbivore grazing, a rapid increase in vegetation cover, aboveground biomass, and vegetation diversity will occur⁷⁰. Based on the results presented here, these changes are expected to enhance microbial α-diversity, particularly for fungi. However, the potential negative impact of high levels of nitrogen deposition in Western Europe must also be considered. Increased nitrogen deposition can lead to soil acidification⁷¹, promotion of the growth of a few dominant plant species^72,73, and a reduction in overall vegetation diversity⁷⁴, negatively impacting bacterial diversity. In contrast, fungal α-diversity may be less affected or even benefit from the overall increase in vegetation cover and biomass, possibly resulting in a net positive effect on fungal richness and diversity.

This also emphasizes the significance of large herbivore grazing, albeit at lower levels. High-intensity grazing, as observed in our study system, reduces overall plant diversity and biomass. Conversely, without or with very low levels of grazing, highly competitive and often tall species can start to dominate, ultimately reducing plant diversity and increasing total biomass. Intermediate grazing intensity is expected to result in peak overall biodiversity⁷⁵, aligning with the intermediate disturbance hypothesis⁷⁶. However, it is important to note that each ecosystem is unique, and universally applicable methods for restoration do not exist. Even at smaller scales, the heterogeneity of natural systems limits the development of a singular approach to ecosystem restoration. Therefore, we advocate for future research to not only examine overall biodiversity recovery following reductions in grazing intensity but also to investigate how specific local communities are influenced by their environment.

While our results align well with findings from experimental and pastoral studies with domestic herbivores, our statistical approach adds further robustness to the conclusions, offering more reliable insights into the relationships between environmental factors and soil microbial communities. Our approach enhances the understanding of the drivers of soil microbial α-diversity under high-intensity grazing, which can inform more effective and comprehensive future ecosystem management and restoration strategies.

Methods

Research area and sampling sites

Our study was carried out in the Amsterdam Water Supply Dunes (AWD; 3495 ha), situated along the west coast of The Netherlands (52°20’04.4”N; 4°31’29.9”E; Fig. 5). This coastal sand dune and water catchment area is characterized by both calcareous (lime-rich) and decalcified (lime-poor) dune zones of different ages that are under strong past and present aeolian influence. The landscape is a mosaic of Grey dune grasslands (EU-habitat type H2130), and scrub with sea-buckthorn (Hippophae rhamnoides; H2160) interspersed with patches of (almost) open sand. The climate is temperate humid, with mean annual precipitation over the past two decades of 890 mm and mean annual temperature of 11 °C.

Fig. 5. Map of the study area.

Sampling locations within the Amsterdam Water Supply Dunes are marked with red dots.

The AWD are further characterized by high-intensity fallow deer (Dama dama) grazing since the early 2000s, peaking at nearly 4000 counted individuals in 2016. Since then, their population size has been gradually reduced through culling to approximately 2700 counted animals in 2020, the year of data collection for this study. Other grazers present in small numbers in the AWD are European rabbit (Oryctolagus cuniculus), European roe deer (Capreolus capreolus) and European hare (Lepus europaeus).

The past two decades of high grazing intensity has resulted in a collapse of insect and plant richness and diversity in the area⁷⁷. To protect the still present biodiversity, 14 grazing exclosures and accompanying adjacent grazed reference locations were constructed in different dune zones throughout AWD in the winter of 2020. The exclosures, ranging in size from 0.5 to 5.5 ha, allow entry of small herbivores but exclude deer grazing. At each exclosure and reference area, one permanent vegetation quadrat (PQ; 2 × 2 m) was installed per habitat type available (i.e., grassland and/or scrub). These PQs functioned as the sampling sites for our environmental and microbial samples. To capture the full successional gradient from pioneer vegetation in open sand to scrub, together with the associated environmental gradients, additional sampling sites for the habitat type open sand were selected at each exclosure and reference location when present. In total, 35 grassland and 20 scrub PQs, along with an additional 25 open sand sites, were sampled in this study.

Environmental and microbial sample collection

For each PQ, percentage cover of different vegetation layers (i.e., open sand, moss/lichen, grass/herb, shrub, and total cover) and of individual species were estimated at the peak of vegetation development (July/August) in 2020. Assessment of vegetation layer and species cover data for open sand sampling sites was done for an area of 2 × 2 m around the sampling site center point.

For each grassland and scrub sampling site, we created a 25 × 25 cm sample plot next to the PQ, in vegetation closely resembling its species composition and structure. For open sand sites, we used 50 × 50 cm plots due to the patchier vegetation, ensuring both structural elements (i.e., open sand and pioneer vegetation) were captured. For each plot, aboveground biomass was collected, including standing dead material, by clipping all vegetation to the soil level. After this, all litter material lying on the soil was collected. Subsequently, three soil cores were taken from the upper 5 cm soil layer using metal rings of 100 cm3. Microbial samples were taken from each of the three holes created by soil core extraction by filling a 1.5 mL Eppendorf tube^® with bulk soil material using a sterile spatula. The samples were stored in a cool box at sub-zero temperatures in the field for transportation to the Molecular Biology and Ecology laboratory at the University of Amsterdam. Here the samples were stored at −80 °C until DNA extraction. We focused specifically on the upper 5 cm soil layer for environmental and microbial sampling, as previous research has indicated that microbial communities in this layer are most affected by grazing⁷⁸. Bulk soil for this study was defined as a mixture of both rhizosphere soil (soil in direct contact with plant roots) and non-rhizosphere soil (soil not in direct contact with plant roots and their exudates).

Aboveground biomass (gr/m²) and litter samples (gr/m²) were oven-dried for a minimum of 48 hours at 70 °C and weighed. This process was also applied to one of the three soil cores to obtain a bulk density (gr/cm³) value for each sampling plot. From this same soil core, a subsample was grounded to determine soil organic matter (SOM) content by loss of ignition (LOI) at 375 °C for 16 hours. Using the corresponding bulk density values, SOM content was calculated as kg per m² of the upper 5 cm soil layer. Note that SOM is of biotic origin but is generally considered an abiotic environmental factor and we therefore treat it accordingly in this study. The other two soil cores were mixed by hand and used for pH and electrical conductivity (EC) measurements using a 1:25 weight:volume ratio (deionised water:soil; Consort C831 multi-channel analyzer, Consort NV, Turnhout, Belgium).

Microbial DNA extraction and sequencing

Microbial genomic DNA was extracted from 0.75 g of bulk soil mixture, obtained by thoroughly mixing equal parts of each of the three microbial soil samples per sampling plot, using the DNeasy^® Powersoil Kit (Qiagen, CA, USA) and the corresponding instructions. Extracted DNA quality was checked with a Nanodrop ND-1000 (Thermo Fisher Scientific, MA, USA) and verified by gel electrophoresis.

Before sequencing, the bacterial 16S rRNA (V3-V4) and fungal ITS gene target regions were amplified by means of polymerase chain reaction (PCR) using the universal primer sets 341F-785R⁷⁹ for the bacterial 16S rRNA (V3-V4) region and ITS3-ITS4⁸⁰ for the fungal ITS region. Paired-end sequencing was performed with an Illumina MiSeq platform (Illumina, CA, USA) by BaseClear B.V., Leiden, The Netherlands.

Microbial sequence data processing

We employed the QIIME2 system (v.2021.2)⁸¹ for processing the raw sequence data and used the Deblur denoising tool, which corrects amplicon sequence errors and produces high-resolution Amplicon Sequence Variants (ASVs) that can resolve differences of as little as one nucleotide^82,83. First, forward and reverse primers were removed with the cutadapt plugin⁸⁴ and paired-end sequences were joined by using the vsearch plugin⁸⁵. Subsequently, sequences were quality-trimmed and -filtered. As the last step, we used a naïve Bayesian taxonomy classifier⁸⁶ and trained SILVA Database (v.138.99)⁸⁷ for bacteria and UNITE Database (v.8.99)⁸⁸ for fungi to perform taxonomic annotation of the ASVs⁸⁹.

Sequencing failed for one open sand sample for both bacteria and fungi (different sample for each). This resulted in a reduction of the total open sand sample size for both microbial groups to 24.

Hill diversity indices

Traditionally, α-diversity has been assessed using well-known indices such as species richness and the Shannon-Wiener index. However, these indices have been subject to criticism in recent decades, mainly due to their inability to measure the same quantities and their use of different units^90,91. Furthermore, they do not scale in the same way when species are added or removed, making cross-community comparisons challenging. To address these issues, we used the Hill diversity indices framework to quantify both vegetation and microbial ASV α-diversity at the species level (see Roswell et al. ⁹², for a detailed review).

In short, Hill diversity is a spectrum of community diversity indices developed by Hill (1973)²¹ and reintroduced to ecology by Jost (2006)⁹³ that takes into account both the number and the relative abundance of species. The Hill diversity spectrum solves the problem of incompatible numerical constraints by converting traditional diversity index values to effective numbers of species, which can be defined as “the number of equally abundant species that would be needed to give the same value of a diversity measure”⁹⁴. The Hill diversity indices are obtained by a single equation, defined as ${\left({\sum }_{i=1}^S{p}_i^q\right)}^{1/\left(1-q\right)}$, that varies depending on the value of its scaling parameter q. When q is set to 0, relative abundances are not taken into account and all species are therefore given equal weight, so q = 0 equivalises to species richness. At q = 1, all species are given weight proportional to their relative abundance, making it the most neutral index for assessing ‘true species diversity’ and is equivalent to the exponential of the Shannon-Wiener index. For this study, we employ these commonly used Hill diversity indices, which we refer to as richness (i.e., Hill-richness; q = 0) and Shannon diversity (i.e., Hill-Shannon diversity; q = 1) for easy communication.

Data quality checks

The exclosures were in place for approximately six months before the collection of environmental and microbial data, which could have introduced treatment-related effects (exclosure vs. reference) in the data. To address this issue, we performed paired permutation tests (9999 permutations; Holm correction for multiple comparisons) for each individual environmental factor (i.e., aboveground biomass, litter mass, bulk density, grass/herb cover, moss/lichen cover, open sand cover, shrub cover, total vegetation cover, pH, SOM, vegetation richness and vegetation Shannon diversity) and microbial diversity index (bacterial and fungal richness and Shannon diversity). Our analysis revealed no significant treatment effect for any of the tested variables.

To examine the differences in the environmental factors included in our analyses across the sampled habitat types and the exhaustiveness of environmental gradient sampling, we conducted permutational ANOVAs (9999 permutations). If a significant difference was detected, we employed permutational post-hoc t-tests (9999 permutations; Holm correction for multiple comparisons) to determine pairwise differences across habitat types.

In order to account for potential differences and related problems arising from unequal read depths of the microbial sequence data, we used a sample-size- and sampling-coverage-based rarefaction/extrapolation approach where the Hill diversity index values were extrapolated to the asymptote to estimate the diversity at complete sampling coverage^94,95. These extrapolated Hill diversity index values were utilized in subsequent statistical analyses.

Procedure for structural equation model selection and multimodel inference

To identify the main drivers of soil microbial ASV α-diversity, we employed an information-theoretical approach to structural equation model (SEM) selection and multimodel inference^23,24,26. In observational field studies, such as the present one, variables often exhibit strong correlations, posing challenges for traditional stepwise selection methods in model selection and inference^23,27. Moreover, traditional methods do not consider model uncertainty and neglect other equally plausible candidate models²⁵. These obstacles can be overcome by ranking models in a candidate set by an information criterion, balancing fit and precision while also taking model complexity into account. Subsequently, this approach allows for an inference drawn from the relative likelihood of the models in the candidate set that best align with the observed data. Below, we describe the multiple steps of this information-theoretical approach in more detail:

Parameter minimization

The first step involves minimizing the number of SEM parameters to a suitable level considering the available data points and model complexity (i.e., approximately 7 parameters in our case) to address potential issues related to overfitting and collinearity. We therefore calculated Pearson’s pairwise correlation coefficients (r) following a permutation procedure (9,999 permutations; Holm correction for multiple comparisons) for each combination of our initial 12 environmental factors (Supplementary Fig. 1). A bivariate cut-off r-value of 0.7 was used to identify highly correlated factors⁹⁶, resulting in the exclusion of EC, bulk density, litter mass, open sand cover, shrub cover, and total vegetation cover. This left us with a set of 7 environmental factors: pH, SOM, aboveground biomass, moss/lichen cover, grass/herb cover, vegetation richness, and vegetation Shannon diversity.

Global structural equation model specification

In the second step we specified global SEMs that enabled us to partition the direct and indirect effects, and the relative importance, of the biotic and abiotic factors on the two soil microbial α-diversity indices (i.e., ASV richness and Shannon diversity for both bacteria and fungi). Based on prior knowledge, with a focus on coastal sand dune ecosystems, these global models comprised all hypothesized causal pathways amongst the environmental factors and the direct causal links between these factors and the diversity index (see Supplementary Table 2 for pathway justifications). Pathways were further based on “healthy” systems and experimental studies, allowing for the comparison of these hypothesized pathways to our final results and to the aforementioned pastoral studies and controlled experiments on the drivers of microbial α-diversity in intensively grazed systems. We used the same global SEM for each soil microbial α-diversity index to facilitate comparison across indices and microbial groups (Fig. 1). The global SEMs were fitted with ‘piecewise’ SEM⁹⁷, which allows for the incorporation of linear mixed-effect regression models (LMER models) with sampling location as a first-order random effect to control for possible location-based differences²⁴. This is made possible by the construction of individual LMER models for each parameter, prioritizing local over global parameter estimation. These individual models are then integrated into a unified directed acyclic graph, providing a well-suited framework for managing hierarchical nested data structures, accommodating non-normal error distributions, and addressing small sample sizes⁹⁸.

Niche theory suggests that organisms exhibit specific niche breadths and optimums along environmental gradients. Consequently, the relationship between soil microbial ASV α-diversity and the environmental factors, and between the environmental factors considered in this study themselves, may not necessarily be linear. Therefore, we examined both linear and quadratic relationships between each environmental variable and response variable individually. Univariate LMERs were employed, with sampling location as a first-order random effect, and their AICc values (i.e., the Akaike Information Criteria corrected for small sample sizes)⁹⁹ were compared. An overview of the functional forms representing the relationships with the lowest AICc values (and that can biologically be substantiated) can be found in Supplementary Figs. 2-6. In cases where a clear unimodal relationship was detected, composite variables were created for the corresponding environmental factor. This composite was constructed by taking the coefficient estimates of linear and polynomial terms and multiplying these terms by their estimates. The resulting composite variables were then incorporated into the global SEMs.

We recognize that bidirectional relationships or feedback loops can exist between our used environmental factors. However, piecewise SEM is only able to fit recursive models. We therefore fitted alternate potential global SEMs with changing arrow directions where bidirectional relationships or feedback loops hypothetically exist and compared model-to-data fits. Models with the lowest AICc value were selected as final global models for further analyses steps. Overall goodness-of-fit for the final global SEMs was further evaluated using the χ²-distributed Fisher’s C statistic. For all microbial ASV α-diversity indices the overall fit was good (p > 0.05), and d-separation tests indicated that there were no missing pathways between environmental factors.

Structural equation candidate model set selection

In order to construct a set of candidate models, we created a list for each microbial ASV α-diversity index of SEM models with all possible additive combinations of direct effects. Relationships between environmental factors were not expected to differ between models and were therefore kept constant over each individual model. Subsequently, the models were fitted with piecewise SEM and ranked by their respective AICc values. A candidate model set was created by selecting only those models with a ΔAICc of less than 4.0 compared to the top model²⁴. We chose this cutoff to balance inclusivity with model parsimony, as a higher cutoff would risk incorporating spurious results, while a lower cutoff might exclude models that are still well-supported^22,24. Thus, this approach ensures that a sufficiently broad set of candidate models is retained without introducing unnecessary complexity. The relative likelihood of each model in the candidate set was quantified by its Akaike weight, which is normalized across the set of candidate models to summate to one, and is interpreted as probability¹⁰⁰.

Path coefficient averaging and causal effect decomposition

As a fourth and final step, we determined the weighted model-averaged and standardized path coefficients along with their 95% confidence intervals. We achieved this by bootstrapping (1,000 iterations) each path coefficient estimate and confidence interval across all models within the candidate set and weighing the outcomes by the Akaike weight of the respective model, hereby following the zero-method as proposed by Lukacs et al. ¹⁰¹. The zero-method adjusts for model-selection bias by shrinking the estimates towards zero, with the degree of shrinkage being inversely proportional to pathway importance. Path coefficients and their associated confidence intervals were initially standardized by scaling them based on their mean and standard deviation. Furthermore, to address multicollinearity concerns, each coefficient was divided by its respective variance inflation factor (VIF) value. This transformation resulted in standardized semipartial pathway coefficients, shedding light on the unique contribution of each pathway in explaining variation in a response variable, independent of other predictor pathways within the model¹⁰². This approach facilitates the prioritization of effects by their size and allows for comparisons across the outcomes of different models¹⁰³. Subsequently, these model-averaged path coefficients were used to decompose the total causal effect into its direct and indirect components, thereby evaluating the relative importance of each environmental factor on the various soil microbial ASV α-diversity indices.

All statistical analyses were conducted using R statistical software version 4.2.2 with the packages lmerTest¹⁰⁴, MuMIn¹⁰⁵, iNEXT⁹⁵, MeanRarity¹⁰⁶, piecewiseSEM⁹⁷, and semEff¹⁰⁷.

Author contributions

Conceptualization: D.T.P.K, A.M.K., J.G.B.O., K.E.; Formal analysis: D.T.P.K.; Methodology: D.T.P.K., A.M.K.; Investigation: D.T.P.K., A.M.K., J.G.B.O.; Visualization: D.T.P.K.; Supervision: J.G.B.O., A.M.K.; Writing—original draft: D.T.P.K.; Writing—review and editing: D.T.P.K., E.M., J.G.B.O., A.M.K., K.E.

Data availability

The data supporting this study's findings are not openly available due to their sensitivity. However, they are available from the corresponding author upon reasonable request. The data are stored in controlled-access data storage at the Institute for Biodiversity and Ecosystem Dynamics.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44185-025-00081-x.