Introduction

Habitat loss and degradation of natural ecosystems are major drivers of the global biodiversity crisis1,2, with more than half of the terrestrial land surface converted for anthropogenic uses3. Croplands and pastures have become one of the larger terrestrial land cover types on the planet4, with the net increase in tropical regions exceeding 100 million ha/decade, during the 1980s and 1990s5,6. Across the tropics, between 1980 and 2014, oil palm production increased 15-fold7, contributing to land-use change and intensification and impacting global biodiversity hotspots. Specifically, oil palm plantations occupy 21 million hectares, mostly in Indonesia and Malaysia8. In the face of this biodiversity crisis, there is currently an unprecedented political will to restore degraded ecosystems and landscapes globally9. Therefore, it is fundamental to bring a complementary perspective to the United Nations (UN) decade on Ecosystem Restoration by expanding the restoration scope from degraded and abandoned lands to include the enrichment of biodiversity in monoculture-dominated landscapes.

Embedding small patches of native trees (“tree islands”) in degraded landscapes is a promising strategy to enhance biodiversity and facilitate landscape restoration10. By actively planting trees or through natural regeneration, integrating natural habitats into monoculture-dominated landscapes can positively affect environmental heterogeneity10,11,12, where heterogeneous habitats can be associated with higher species diversity across taxa and spatial scales13,14. However, it remains uncertain to what extent environmental heterogeneity at the landscape scale (i.e., landscape heterogeneity) can be leveraged to enhance the diversity of multiple taxonomic groups (i.e., multi-taxa diversity) in monoculture-dominated landscapes.

To inform landscape restoration practices, it is essential to integrate insights from community assembly mechanisms; for example, through beta diversity, defined as “the extent of change in community composition”15. The assembly of ecological communities is determined by biotic and abiotic filtering, environmental drift, and dispersal16,17. For instance, through direct and indirect species interactions, biotic filtering may play an important role in shaping biodiversity18,19,20 and the spatial organisation of (meta)communities21,22,23,24, explaining the growing interest in understanding the role of biotic interactions on community assembly in restoration contexts25,26,27. In particular, interactions between above- and below-ground components of ecosystems, e.g., between plants and soil organisms, can drive ecological processes at the community and ecosystem level28. Yet, our understanding of assembly mechanisms of multi-taxa communities in human-modified landscapes, particularly in the tropics, remains limited29,30.

Here, we assessed if multi-taxa diversity can be enhanced in large monoculture-dominated landscapes by embedding environmentally dissimilar tree islands. Furthermore, we investigated to what extent biotic associations are central to defining the spatial distribution of multi-taxa communities (i.e., multi-taxa beta diversity). To this end, we used comprehensive data from a tropical biodiversity enrichment experiment (EFForTS-BEE [Ecological and socio-economic functions of tropical lowland rainforest transformation systems: biodiversity enrichment experiment]31), located in Sumatra, Indonesia, a global hotspot of biodiversity loss32 and recent tropical deforestation33. Embedded within a 140-ha oil palm plantation, 52 experimental tree islands were established, varying in island size (from 25 to 1600 m2) and planted tree diversity (from zero to six native tree species planted). In our study, we defined a landscape as a geographical area distinguished by natural and human-induced landscape elements34, with tree islands as the landscape elements (and no other surrounding land-use patches). This landscape-scale perspective with tree islands makes EFForTS-BEE unique among the largest network of tree diversity experiments worldwide (TreeDivNet35). We analysed multi-taxa diversity sampled three to five years after establishment, when the tree islands substantially differed in vegetation structural complexity as a result of varying planted diversity and island size36. We calculated beta diversity partitioned into its turnover and nestedness components, the main patterns driving differentiation among communities37. Turnover takes place when some species are replaced by others in different sites38, and nestedness occurs when from rich sites, small subsets are formed39,40 (i.e., species losses and gains). We used community data of understorey arthropods, soil biota (fungi, bacteria, and fauna), herbaceous plants, and trees (excluding planted trees). We hypothesised that tree islands, varying in vegetation structural complexity and soil conditions, will increase total landscape diversity (i.e., gamma diversity) by tree islands fostering unique species resulting in higher species turnover across islands (Fig. 1).

Fig. 1: Tropical biodiversity enrichment experiment (EFForTS-BEE) and conceptual figures.
figure 1

(a) 52 experimental tree islands were embedded within a 140-ha oil palm plantation. Tree islands varying in tree native planted diversity and island size; (b) example of a tree island using a drone image; (c) Conceptual example of a heterogeneous landscape with four tree islands varying in size and diversity level; (d, e, f) examples of beta diversity (partitioned into turnover and nestedness components) for understorey arthropods, herbaceous plants, and soil bacteria. If multi-taxa beta diversity is driven by habitat differentiation, higher landscape heterogeneity (resulting from islands differing in their vegetation structural complexity) is expected to be associated with the beta diversity of multiple taxa. In contrast, if multi-taxa beta diversity is driven mostly by stochastic processes such as dispersal, landscape heterogeneity may not be associated with changes in beta diversity. Beta diversity patterns may be driven by species turnover, with higher turnover resulting in higher gamma diversity, or by nestedness (i.e., gain and species losses in light grey in the bars). Positive associations between landscape heterogeneity and beta diversity of multiple taxa translate into greater dissimilarity in vegetation structural complexity between islands being associated with dissimilar multi-taxa communities e.g., (c, d) landscape heterogeneity and understorey arthropods, and (c, e) landscape heterogeneity and herbaceous plants. A positive association between the beta diversity of two taxa (e.g., herbaceous plants and soil bacteria) (e, f) implies that tree islands that differ in herbaceous plant composition also differ in soil bacteria composition.

Full size image

To reveal the factors shaping the spatial distribution of multi-taxa communities (beta diversity, turnover, and nestedness) across tree islands, we used partial correlation networks, which quantify associations among landscape heterogeneity and beta diversities (or its underlying components) across the six taxa. Our landscape heterogeneity variables included (i) the dissimilarity in three-dimensional vegetation structure (vegetation structural complexity) using mean fractal dimension (MeanFRAC) from terrestrial laser scans and (ii) the dissimilarity in soil conditions using soil phosphorus concentration between islands. We selected MeanFRAC because it is associated with enriched tree island conditions, i.e., planted tree composition, richness, and tree island size36 and soil phosphorus because lowland tropical forests are associated with strongly-weathered soils poor in rock-derived nutrients such as phosphorus41. Further, both, MeanFRAC and soil phosphorus showed the highest network connectivity compared with other vegetation structural complexity metrics and soil measurements (see methods). Partial correlations can provide insights about associations shaping the spatial organisation of communities across taxa, e.g., similar niche requirements, dispersal limitations, and potential biotic interactions due to co-occurrences; this approach is particularly helpful in hyperdiverse regions such as the tropics, where biotic interactions are predicted to strongly structure community assembly21 but assessing interactions is extremely challenging42,43. In the network, nodes represent landscape heterogeneity and beta diversity (or one of its two components) for each taxon. The links in the network represent associations between the nodes. For example, positive associations between landscape heterogeneity and beta diversity of multiple taxa translate into greater dissimilarity in vegetation structural complexity, soil conditions, or both between islands being associated with dissimilar multi-taxa communities. A positive association between the beta diversity of two taxa (e.g., herbaceous plants and soil bacteria) implies that tree islands that differ in herbaceous plant composition also differ in soil bacteria composition. Similarly, a positive association between turnover (or nestedness) between herbaceous plants or soil bacteria implies that tree islands that foster unique species (or are driven by species losses and gains) for herbaceous plants also show the same pattern(s) for soil bacteria (Fig. 1).

We hypothesised that landscape heterogeneity (i.e., dissimilarity in vegetation structural complexity and dissimilarity in soil conditions), will be highly connected in the networks (i.e., have a high strength) because it influences multi-taxa beta diversity via habitat provision and environmental filtering. Specifically, dissimilarity in vegetation structural complexity might influence the spatial composition of multi-taxa by providing a higher number of niches and habitats, not only above- (e.g., structurally complex and open habitats)13,44 but likely also below-ground45. In addition, dissimilarity in soil conditions may influence the spatial composition of herbaceous plants and trees as well as soil biota via bottom-up processes46. While dissimilarity in vegetation structural complexity is expected to be mainly associated with beta diversity of above-ground taxa, associations between vegetation structural complexity and below-ground taxa might be driven by plant-soil feedbacks28,46, e.g., soil fauna in tropical plantations relies on vegetation via plant litter and root-derived resources47. Moreover, we hypothesised plants, either through vegetation structural complexity or composition of herbaceous plants or trees, to be highly connected in the networks, by connecting above- and below-ground taxa via non-trophic and trophic interactions. For example, vegetation may connect above- and below-ground biota via biotic interactions with pathogens, mycorrhizal fungi, and decomposers28,48, as well as with understory herbivores and pollinators49,50.

Results and discussion

Gamma and beta diversity across tree islands embedded in an oil palm plantation

Across the 52 tree islands, we recorded 958 morphospecies of understorey arthropods, 8159 operational taxonomic units (OTUs) of soil fungi, 47,856 OTUs of soil bacteria, 27 taxonomic groups of soil fauna (Supplementary Table S4), 75 herbaceous plant species, and 50 trees species—excluding planted trees, (gamma diversity; all classifications are referred to as “species” below). Overall, across the 52 tree islands, beta diversity (calculated as Jaccard pairwise dissimilarity) varied among taxa, ranging from 0.31 for soil fauna to 0.77 for understorey arthropods. Beta diversity was mainly driven by species turnover, while nestedness, except for trees and soil fauna, played a minor role (Fig. 2). Specifically, the highest species turnover was found for soil fungi, understorey arthropods, and soil bacteria, accounting for ~94% of the total beta diversity. Herbaceous plant turnover made up 78% of total beta diversity. Species turnover was lower for trees (52%) and soil fauna (59%). We did not find major differences in the results when calculating beta diversity using Sørensen pairwise dissimilarity (Supplementary Figs. S2 and S5). Hence, our results consistently indicate that beta diversity is primarily associated with the uniqueness of species assemblages rather than smaller assemblages being a subset of larger ones. These results align with studies in tropical regions where beta diversity patterns across different organisms have identified species turnover as the dominant component driving beta diversity in environmentally heterogeneous ecosystems51,52,53. Consequently, promoting the uniqueness of species assemblages with multiple tree islands appears as a promising strategy for enhancing biodiversity in monoculture-dominated landscapes, at least during the first years after tree island establishment.

Fig. 2: Turnover and nestedness components of beta diversity.
figure 2

Beta diversity for the six taxa were calculated with the Jaccard index. Similar results were found when beta diversity was calculated using Sørensen pairwise dissimilarity (Supplementary Fig. S2).

Full size image

The differences in beta diversity across taxa that our study revealed might be explained by ecological processes related to dispersal ability, body size, and life history. For instance, due to the long lifespan of trees, the influence of processes such as local extinction and colonisation may require more time than for other taxa. Furthermore, tree beta diversity patterns may be shaped mainly by seed sources in the surrounding landscape and by tree species with higher dispersal capacities54, explaining the unexpected high nestedness in human-modified ecosystems compared to tropical forests for trees51. While we expect overall patterns to hold, the influence of differences in sample coverage across taxa—particularly incomplete coverage for highly diverse taxon such as fungi—in terms of turnover and nestedness under- or over-estimations remains unknown. Finally, taxonomic resolution may impact our ecological understanding55, particularly for soil fauna that mainly was assessed at the level of orders (that often represent functional groups56). Contrasting resolutions reflect the challenge of biodiversity assessment in the species-rich tropics42. Despite that, soil fauna was a good indicator of overall multi-taxa community dissimilarity (see below). Therefore, we expect this crucial role to remain or be strengthened with higher resolution but increases in resolution will likely result in higher beta diversity due to higher turnover.

Insights of multi-taxa beta diversity through landscape heterogeneity and biotic associations

Beta diversity patterns across multiple taxa were correlated, with the network for beta diversity comprising 17 edges (Fig. 3a, Supplementary Table S6). The most connected taxa were soil fauna and bacteria (strength, i.e., the sum of absolute edge weights, = 0.82 and 0.71, with five and four edges with other nodes, respectively; Fig. 4a). By contrast, trees were the least connected (strength = 0.46, with four edges). The highest correlation coefficient was observed between soil fungi and bacteria beta diversity (+0.25). Turnover patterns for multi-taxa diversity were also correlated, with the network for turnover comprising eight edges (Fig. 3b, Supplementary Table S7). Turnover of soil fauna and understorey arthropods were disconnected from the network. In other words, neither the turnover patterns of soil fauna nor understorey arthropods follow dis(similar) turnover patterns of other taxa nor landscape heterogeneity. Finally, nestedness patterns for multi-taxa diversity were correlated except for trees (Fig. 3c, Supplementary Table S8), with the network retaining six edges. Yet, the nestedness network had low stability. Together, these results suggest that direct and indirect associations shape the spatial organisation of communities across taxa in tropical human-modified landscapes, supporting previous studies in temperate ecosystems23,24.

Fig. 3: The role of landscape heterogeneity and biotic associations shaping multi-taxa beta diversity.
figure 3

Landscape heterogeneity refers to the dissimilarity in vegetation structural complexity (measured through MeanFRAC) and soil conditions (measured through soil P) together. Nodes represent (a) total beta diversity, (b) turnover, and (c) nestedness of multiple taxa and dissimilarity in vegetation structural complexity and soil conditions. Edges thicknesses, i.e., line thickness, are proportional to partial correlation coefficients, with grey and red edges representing positive (i.e., greater dissimilarity in vegetation structural complexity between islands being associated with dissimilar multi-taxa communities or tree islands that differ in composition for a taxon also differ in composition for another taxon) and negative (i.e., greater dissimilarity in vegetation structural complexity between islands being associated with similar multi-taxa communities or tree islands that differ in community compositions for a taxon have similar community compositions for another taxon) correlations, respectively. Edge length is not meaningful. Nodes with partial correlation coefficients equal to or near zero are not included in the corresponding networks.

Full size image
Fig. 4: Importance of the individual taxa and landscape heterogeneity in shaping multi-taxa beta diversity.
figure 4

Landscape heterogeneity refers to the dissimilarity in vegetation structural complexity (measured through MeanFRAC) and soil conditions (measured through soil P) together. The centrality value (x-axis) for each node (y-axis) is presented. Nodes represent (a) the total beta diversity and (b) the turnover of multiple taxa and dissimilarity in vegetation structural complexity and soil conditions. The centrality value is quantified by the strength (i.e., the sum of absolute edge weights) in the undirected partial correlation networks and shown as standardised z-scores. Negative values indicate low centrality, whereas positive values indicate high centrality. Correlation stability coefficients of strength for beta diversity and turnover were 0.36 and 0.44, respectively. For nestedness, the correlation stability coefficient was lower than 0.25, suggesting lower stability of this network that is therefore not presented in this figure (see Supplementary Fig. S3). Other centrality measures, i.e., betweenness and closeness, are shown in Supplementary Fig. S3. Observed and non-parametric bootstrap mean and 95% CI estimated are shown in Supplementary Fig. S8 and S9.

Full size image

Our results point toward the key role of below-ground organisms in structuring multi-taxa beta diversity patterns. Soil biota (soil fauna, bacteria, and fungi) were central to the overall ecological community. In other words, beta diversity of soil biota was associated with beta diversity patterns of other taxa and with landscape heterogeneity variables (i.e., dissimilarity in vegetation structural complexity and soil conditions) (for different centrality indices, see Fig. 3 and Supplementary Fig. S3). Soil biota may act as an indicator of current conditions (i.e., tree islands), the result of legacy effects from previous land-uses (e.g., oil palm plantation or tropical forest), or both57. For example, soil fauna composition can be associated with differences in specific organic materials (reflecting the heterogeneity before the land-use conversion) and time delays because of the limited dispersion of soil fauna58. Similar beta diversity patterns between soil fauna and soil fungi may be underlain by species interactions (e.g., soil fungi as an important food source in soil food webs59), similar niche requirements, and/or dispersal limitations influencing soil biota (symbiotroph, pathotroph, and saprotroph, Supplementary Figs. S4–S7; Supplementary Tables S9–S11). Associations between the above- and below-ground systems, e.g., plant and soil biota, can result from positive or negative plant-soil feedbacks that influence community and ecosystem-level processes28,48. Soil fauna potentially influences vegetation dynamics and above-ground biodiversity60. For instance, soil biota has been shown to affect understorey arthropods (particularly pollinators, Supplementary Figs. S4–S6) when soil biota indirectly affects floral traits (e.g., bacteria, root herbivores, and mycorrhizal fungi), influencing pollination attractions and plant fitness61. While plant-soil feedback experiments would be required to disentangle the mechanisms of above- and below-ground associations shaping multi-taxa dynamics, here we provide further evidence highlighting the importance of integrating the below-ground compartment towards elucidating network structure and associations in monoculture-dominated landscapes.

Landscape heterogeneity (i.e., dissimilarity in vegetation structural complexity or soil conditions) played a crucial role in all three networks (Fig. 3). For instance, dissimilarity in vegetation structural complexity was the most connected node (strength = 0.84 with four positive and two negative edges to other nodes) in the beta diversity network. Besides, soil P was the most connected node (strength = 0.49 with four edges, Fig. 4b) in the species turnover network. The highest and lowest correlation of soil P was found with soil bacteria and fungi beta diversity, respectively (+0.18 and +0.11). This suggests that landscape heterogeneity can promote beta diversity by fostering different species compositions, reinforcing the role of enriched tree islands in influencing community assemblages and biodiversity at the landscape scale (i.e., beta and gamma diversity). Further, it implies that dissimilarity in abiotic conditions can directly or indirectly impact multiple taxa. The influence of vegetation structural complexity on multi-taxa diversity may act via altering light and microclimatic conditions62 and other characteristics associated with variation in local planted tree species diversity and identity, with both shaping vegetation structural complexity36. Furthermore, the influence of tree islands on multi-taxa diversity might reflect the removal of environmental filtering associated with conventional management, such as liming and fertilisation, which is responsible for biotic homogeneity in monoculture-dominated landscapes. Further possible mechanisms include enhanced nutrient cycling and plant litter decomposition28,63, particularly in ecosystems under transition (e.g., primary or secondary succession)64.

Conclusions

We conclude that enriching monocultures with tree islands varying in vegetation structural complexity (as a result of variation in planted tree diversity and/or island size)36 can foster unique ecological communities above- and below-ground and thereby promote multi-taxa diversity at the landscape scale (beta and gamma diversity). Additionally, we suggest distributing tree islands across the monoculture-dominated landscape to enhance multi-taxa diversity by capturing contrasting biotic and abiotic conditions. Landscape restoration strategies aiming to enhance multi-taxa diversity should consider not only key abiotic conditions but also the extent to which biotic associations play an important role in shaping ecological communities at the landscape scale. By enhancing biodiversity at the landscape level in monoculture-dominated tropical landscapes, we bring a complementary perspective to the UN Decade on Ecosystem Restoration and provide experimental evidence urgently needed for enriching biodiversity in productive agricultural landscapes.

Materials and Methods

Study area

This study was conducted in the Biodiversity Enrichment Experiment (EFForTS-BEE) located in Jambi province, Sumatra, Indonesia. The main aim of EFForTS-BEE is to evaluate the potential of establishing tree islands10 within an industrial oil palm plantation as a restoration measure to enhance biodiversity and ecosystem functioning while maintaining financial benefits31,65. The area is characterised by a humid tropical climate with two peak rainy seasons (March and December) and a dryer period extending from July to August31. The mean temperature is 26.7 ± 1.0 °C, and the mean annual precipitation is 2235 ± 385 mm (1991–2011). The predominant soil type in the region is loamy Acrisol66. EFForTS-BEE was established in December 2013 and consists of 52 experimental plots, i.e., tree islands, varying in island sizes of 25 m2, 100 m2, 400 m2, and 1600 m2, and planted tree diversity level 0, equal to no tree planted, 1, 2, 3, and 6 tree species planted in a plot. Tree islands were embedded in a 140-ha oil palm plantation (01.95° S and 103.25° E, 47 ± 11 m a.s.l.), with the planting of the oil palm starting in 200131. Fertilisation, herbicide, and pesticide application stopped after the establishment of tree islands31. Regular management, including manual weeding of the understorey, stopped 2 years after tree planting, allowing for natural regeneration in the tree islands. The experiment follows a random partition design aiming to disentangle the linear effects of tree diversity and plot size and the non-linear effects of tree species composition31. For details of the experimental design, see ref. 31. The planted species represent native, multi-purpose trees used for the production of fruits (Parkia speciosa Hassk, Archidendron jiringa (Jack) I.C.Nielsen, and Durio zibethinus L.), timber (Peronema canescens Jack, and Shorea leprosula Miq.), and natural latex (Dyera polyphylla (Miq.) Steenis)67.

Data collection

We quantified vegetation structural complexity by terrestrial laser scanning between September and October 201636. We calculated Effective Number of Layer (ENL), which describes the vertical structure of forest stands and is influenced by the stand height and the vegetation distribution across vertical layers68. In addition, we calculated Mean Fractal Dimension (MeanFRAC), defined as the arithmetic mean of fractal dimensions of the polygons formed by cross-sections of the 3D point cloud, describing the geometric complexity of the stand69. MeanFRAC is associated with enriched tree island conditions, i.e., planted tree composition, richness, and tree island size36. Finally, we calculated the stand structural complexity index (SSCI) by combining ENL and MeanFRAC in a single indicator that is a holistic measure of stand structural complexity69,70. By construction, the indicators do not scale with the area. All the indicators were calculated based on one single scan in the centre of each plot, thereby capturing the potential influence of edge effects associated with differences in island sizes.

Soil nutrient variables, including total carbon (C) and nitrogen (N) concentration (g mg−1), C-to-N ratio, and plant-available P concentration (mg g−1), were quantified using the same soil samples as for soil fungi collected in December 2016 (see below). Total C and N were determined via the combustion method in a C/N analyser57. Plant-available P was quantified following Bray & Kurtz71. The soil samples were mixed with Bray-I Extraction solution, shaken for 60 min, and filtered with phosphate-free filters. P concentration of filtrates was measured using inductively coupled plasma mass spectrometry57.

The diversity data for this study were collected between October 2016 and May 2018. We sampled above-ground and below-ground taxa, including understorey arthropods, soil biota (soil fungi, soil bacteria, and soil fauna), herbaceous plants, trees, vegetation structural complexity measures, and soil conditions, with all measurements within the 52 tree islands, i.e., plots. Arthropods sampled at the height of the understorey vegetation (referred to as “understorey arthropods”) were sampled three times with six pan traps (2 × 3 pan traps) equally distributed within each plot, for 45 h from October 2016 to January 2017. The traps were made of white plastic bowls coloured with yellow UV paint72 and filled with water and a drop of detergent. All individuals were preserved in 70% Ethanol, sorted by morphospecies, and subsequently identified into the higher taxonomic classification possible (i.e., 14 groups/families) and their corresponding functional groups (e.g., Table S5).

Soil biota and herbaceous plants were surveyed in a subplot of 5 × 5 m area established within each plot31. The subplot was assigned randomly within each plot at a minimum distance of 1.5 m from the plot edge. Specifically, soil fungi were sampled and collected in December 2016 from three soil cores per plot (10 cm depth and 4 cm diameter) and identified through DNA extraction and next-generation sequencing57. OTUs were classified taxonomically using the BLAST algorithm (blastn, v2.7.1;73) and the UNITE v7.2 database (UNITE_public_01.12.2017.fasta;74). Soil bacteria were obtained for each subplot from three 10 cm cores of topsoil placed at 1 m far from the adjacent trees. The soil cores were mixed, homogenised and cleared from roots before DNA and RNA extraction and posterior classification75. In each plot, soil fauna communities were assessed in four soil samples of 16 × 16 cm using a spade down to a depth of 5 cm plus the entire overlying litter layer. The animals extracted from the soil samples by heat were counted and classified into taxonomic groups, corresponding to key functional soil invertebrate guilds (mainly groups/families, Supplementary Table S4)56,76,77. Herbaceous plants, described as all non-woody plants lower than 1.3 metres in height, were identified from February to March 2018. Trees refer to all free-standing woody plants with a minimum height of 1.3 m, inventoried in the total area of the experimental tree islands in August 2018, excluding the trees planted at the onset of the experiment.

Beta diversity and landscape heterogeneity

For each taxon, beta diversity was calculated using species incidence-based pairwise dissimilarity matrices (presence-absence data) with the function beta.pair from the package betapart version 1.5.478. We partitioned beta diversity into turnover and nestedness components37,78. The Jaccard pairwise dissimilarity (βjacc) among plots was computed as βjacc = βjtu + βjne, where βjtu accounted for the turnover fraction of Jaccard pairwise dissimilarity, and βjne accounted for the nestedness-resultant dissimilarity fraction, measured on a normalized scale from zero to one. We calculated beta diversity using community data (incl. operational taxonomic units, taxonomic groups, morphospecies, or species—referred to as species in the text). In addition, we calculated beta diversity using Sørensen pairwise dissimilarity, which incorporates turnover and richness differences as βsor = βsim + βsne. In this case, βsim accounted for turnover measured as Simpson pairwise dissimilarity, and βsne accounted for the patterns of beta diversity causing nestedness, measured as the nestedness-resultant dissimilarity fraction of Sørensen dissimilarity (Supplementary Figs. S2 and S5). While Jaccard considers the proportion of unique species in the entire pool, Sørensen considers the proportion of unique species per site79. For calculating landscape heterogeneity (dissimilarity in vegetation structural complexity and soil conditions), we calculated pairwise dissimilarity between all matrix rows, i.e., tree islands, using the function dist from the R stats package. We used the Euclidean distance method, calculated as a true straight-line distance between all matrix rows in Euclidean space.

Multivariate normality was tested with Mardia’s multivariate skewness and kurtosis coefficients using the function mvn from the R package MVN version 5.980. When the test did not state multivariate normality, a non-paranormal transformation to achieve Gaussian distribution was implemented using the function huge.npn and the setting shrinkage based on a shrunken Empirical Cumulative Distribution Function (ECDF) from the R package huge version 1.3.581.

Partial correlation networks

We applied partial correlation networks to study associations between landscape heterogeneity and beta diversity (turnover or nestedness) among multiple taxa. An association between taxa indicates the covariation of the spatial distribution of ecological communities among taxa. The advantages of partial correlation networks are threefold: first, they describe correlations between a set of conditionally independent variables82; second, they do not require a priori knowledge of the structure83; and finally, the correlations can be graphically represented and analysed to reveal key interdependencies and highly connected variables84. Partial correlation networks have been widely used to infer pairwise species interactions from observed presence-absence matrices83. A network is composed of nodes and edges, where the nodes represent the beta diversity (or turnover or nestedness) of the different taxa and the dissimilarity of vegetation structural complexity and soil conditions. The edges (i.e., links connecting pairs of nodes) represent correlations between nodes, in our case, undirected partial correlation coefficients23. Edges can be either positive or negative correlations (representing the covariation of the spatial distribution of ecological communities between taxa), and can be absent, indicating no or weak correlation between a set of variables85. When positive, the (dis)similarity in species composition between tree islands changes in the same direction for both taxa. When negative, the (dis)similarity in species composition for a taxon increases while it decreases for the other taxon.

We used the graphical lasso method (Least Absolute Shrinkage and Selection Operator) as implemented in the R package bootnet version 1.4.386 to build and analyse the networks. This method displays the unconditional association between two nodes once the influence of other variables is controlled (i.e., partial correlations82), reducing the risk of spurious relationships that can emerge from multicollinearity85. The Lasso method applies a regularisation penalty using a tuning parameter to reduce the number of parameters displayed. As a result, only a small number of partial correlations (i.e., the highest values) are used to explain the interconnections among variables82. We selected the tuning parameter with the Extended Bayesian Information Criterion EBIC87 using the function EBICglasso from the package qgraph version 1.6.988 (tuning parameter = 0.5). The partial correlations were represented graphically in networks with undirected weighted edges (i.e., there is an association, but the direction is not determined) using ggraph R package version 2.0.589. With the weighted networks, we consider the correlations among nodes and the weight of these correlations (partial correlation coefficients90).

We tested the influence of different landscape heterogeneity on network connectivity. To do so, we included various combinations of vegetation structural complexity metrics and soil conditions and measured the resulting number of edges in the network and the proportional changes. Variables included SSCI, ENL, and MeanFRAC as the vegetation structural complexity metrics and soil C, N, P, C-to-N ratio, as the soil condition variables. We found the highest network connectivity when MeanFRAC and soil P were included (Supplementary Tables S2 and S3). Other structural metrics or soil conditions did not increase network connectivity and were highly correlated with other environmental variables (Supplementary Table S1 and Fig. S1). Therefore, we only included MeanFRAC (named hereafter as vegetation structural complexity) and soil P in the final networks presented in this study.

We quantified the importance of specific nodes (i.e., certain taxon or a particular environmental variable) for structuring or maintaining the overall (i.e., multi-taxa) network by calculating three centrality measures commonly used in complex network approaches strength, betweenness, and closeness. Strength is the sum of absolute edge weights that a node has with the others82. The higher the strength value of a node, the higher the influence it has on influencing the composition and structure of the community24. Betweenness looks at the proportion of shortest paths between any pair of nodes that pass through a specific node. The shortest path is defined as the path with the minimum distance (calculated by adding the edges’ weights) needed to connect two nodes. Hence, a node with high betweenness lies “in-between” other nodes’ shortest paths in the network. High betweenness indicates that a node plays a crucial role in the connectivity and stability of the network, for example, implying a cascading effect with large consequences on the overall network when the node is lost91. Closeness describes the undirected connectance of a node to the other nodes in a network, calculated as the average distance of the shortest path from a specific node to all other nodes82. Because of its proximity to all other nodes, the node with the highest closeness centrality plays a crucial role in the overall network91 (Supplementary Figs. S3 and S7).

The accuracy of the parameters and measures estimated in a network depends greatly on sample size and variability90. Thus, we assessed the accuracy of the different networks (i.e., sensitivity to sampling variation) by estimating confidence intervals on the weight of the edges with a non-parametric bootstrapping of 1000 samples, with a confidence interval of 95%90, using the bootnet R package version 1.4.386. To assess the stability of centrality indices, we used a case-dropping subset bootstrap from the package bootnet. We calculated the correlation stability coefficient (CS-coefficient), which represents the maximum number of observations that can be dropped (in at least 95 % of the samples) so that the correlation between original centrality indices and the indices re-calculated with a subset of the data is 0.7 or higher82. The threshold considered stable for the CS-coefficient should be no <0.25 and desirably >0.5. Results of the sensitivity analysis are presented in Supplementary Figs. S8–S13.

Data were analysed with the software environment R, version 4.1.1 (R Development Core Team, 2021), using the packages ade492, betapart78, bootnet86, data.table93, ggplot294, ggraph89, glasso95, huge81, igraph96, MVN80, plyr97, qgraph88, reshape298, rlist99, tidyverse100, and vegan101. Our code is based on the R code provided by Ohlmann et al. (2018)23.