The biodiversity benefit of native forests and mixed‐species plantations over monoculture plantations

China's Grain for Green Program (GFGP) is the largest reforestation programme in the world and has been operating since 1999. The GFGP has promoted the establishment of tree plantations over the restoration of diverse native forests. In a previous study, we showed that native forests support a higher species richness and abundance of birds and bees than do GFGP plantations and that mixed‐species GFGP plantations support a higher level of bird (but not bee) diversity than do any individual GFGP monocultures (although still below that of native forests). Here, we use metabarcoding of arthropod diversity to test the generality of these results.


| INTRODUC TI ON
An important challenge for conservation science is to quantify the biodiversity impacts of major policy initiatives, especially in regions undergoing large shifts in land use change. Nowhere is this more true than in China, which combines a high level of native biodiversity (Tao, Huang, Jin, & Guo, 2010) with a large human population that is increasing its ecological footprint (Liu & Diamond, 2005;Pyne, 2013;Sayer & Sun, 2003;Xie et al., 2012). Moreover, for decades, China has had the managerial, political and financial capacity to implement the largest land-sustainability programmes ever seen, from nature-reserve protection to reforestation to de-desertification (Bryan et al., 2018;Liu et al., 2003;Wu et al., 2019;Xu, Wang, & Xue, 1999). These programmes have caused major land use changes and successfully slowed land degradation caused by economic activities (Liu et al., 2008;Ouyang et al., 2016;Ren et al., 2015). For example, China established its first nature reserve in 1956 and reached 2,740 reserves at the end of 2015 (Ma, Shen, Grumbine, & Corlett, 2017). Nearly two-thirds of the area of those nature reserves have national-level status, meaning that they receive the highest level of protection and funding, and analysis of Landsat imagery has shown that national-level reserves successfully deter deforestation (Ren et al., 2015).
Two other major land-sustainability programmes are the Natural Forest Protection Program (NFPP, also known as Natural Forest Conservation Program) and the Grain for Green Program (GFGP, also known as the Sloping Land Conservation Program and the Farm to Forest Program), which were implemented after widespread flooding in 1998 (Liu et al., 2008;Xu, Yin, Li, & Liu, 2006;Yin, Yin, & Li, 2009). The NFPP aims to reduce soil erosion and flooding by protecting native forests in the upstream watersheds of the Yangtze and Yellow rivers (Liu et al., 2008;Ren et al., 2015). The GFGP complements the NFPP by controlling soil erosion on sloping land. The government pays cash and grain to farmers in exchange for tree planting on sloping farmland (Delang & Yuan, 2015;Liu et al., 2008;Ma et al., 2017;Xu et al., 2006;Zhai, Xu, Dai, Cannon, & Grumbine, 2014). Having reforested 9.06 million ha of cropland over 16 years (~2014) since its inception in 1999, the GFGP is the world's largest reforestation programme.
Guided by the goal of soil erosion control, and operating under the implicit assumption that any type of tree cover should achieve this goal, the GFGP has predominantly established tree plantations ("plantations" hereafter) on retired croplands, rather than restoring native forests (Hua et al., 2018(Hua et al., , 2016Zhai et al., 2014). However, compared with native-forest ecosystems, plantations are known to support lower levels of biodiversity across the world's forest biomes and across taxa (Barlow, Overal, Araujo, Gardner, & Peres, 2007;Bremer & Farley, 2010;Brockerhoff, Jactel, Parrotta, Quine, & Sayer, 2008;Gardner, Hernandez, Barlow, & Peres, 2008;Lindenmayer & Hobbs, 2004), although certain management regimes, such as maintaining understorey structure and mixed cropping, can somewhat increase biodiversity (Hartley, 2002). On the other hand, compared a higher arthropod biodiversity value of mixed-species plantations than previously indicated by bees alone.

Main conclusion:
In our previous study, we recommended that GFGP should prioritize the conservation and restoration of native forests. Also, where plantations are to be used, we recommended that the GFGP should promote mixed-species arrangements over monocultures. Both these recommendations should result in more effective protection of terrestrial biodiversity, which is an important objective of China's landsustainability spending. The results of this study strengthen these recommendations because our policy prescriptions are now also based on a dataset that includes over 500 species-resolution taxa, ranging across the Arthropoda.

K E Y W O R D S
Arthropoda, biodiversity, China, forest management, Grain for Green Program, metabarcoding, reforestation with croplands, plantations are known to support different species assemblages, with potentially higher levels of biodiversity, although there are indications that croplands in low-intensity agricultural systems-which the croplands retired under GFGP tend to be (Hu, Fu, Chen, & Gulinck, 2006)-may support considerable biodiversity which potentially exceeds that associated with plantations (Allan, Harrison, Navarro, Wilgen, & Thompson, 1997;Buscardo et al., 2008;Elsen, Ramesh, & Wilcove, 2018). Together, these insights suggest that plantations should have been expected to support low levels of biodiversity and that the GFGP could support more biodiversity if it restored native forests.
Indeed, this is what Hua et al. (2016) found. They surveyed bird and bee communities in GFGP-related tree covers in south-central Sichuan, comparing native-forest remnants to GFGP-financed tree-cover types, which include monoculture stands of bamboo, Eucalyptus and Japanese cedar, as well as "mixed plantations", which are mostly patchworks (checkerboards) of two to five different monocultures and, to a lesser extent, bona fide tree-level mixtures (Hua et al., 2018). Most importantly, this study documented that bird and bee species diversities were higher in native forests than in any of the monocultures. In addition, they found that in mixed plantations, bird diversity for non-breeding species was higher than in any of the individual monocultures, albeit lower than in native forests. In contrast, bee diversity was equally low in mixed plantations and monocultures. The lack of a boost to bee diversity in mixed plantations was not surprising, because as with monocultures, the understorey vegetation in mixed plantations was notably lacking in flowering plants (Hua et al., 2016).
The above findings, however, raise the question of why bird diversity was increased just by planting monocultures of different tree species next to each other. One possibility that could not be investigated in Hua et al. (2016) is that general arthropod diversity might also have been boosted in the mixed plantations, as, unlike bees, other arthropods can exploit a range of food resources available even in plantations, via direct consumption of plants and fungi, and via decomposition, parasitism and predation of other animals, including other arthropods (Jactel & Brockerhoff, 2007). Increased arthropod diversity might in turn support more bird diversity. In addition, as a large component of biodiversity, how arthropods themselves (and subgroups thereof) are affected by the GFGP is an important part of understanding the GFGP's biodiversity effects.
For instance,  compared primary forest and Eucalyptus plantations in Brazil and found that birds achieve highest diversity in primary forest, while bees have similar levels of species richness in primary forest and Eucalyptus plantations. They also found that butterflies and dung beetles achieve low diversity but that fruit flies and moths achieve high diversity in Eucalyptus plantations.
The purpose of this study is to test the generality of Hua et al.'s (2016) results by interrogating the "rest of the biodiversity" that was captured in the same sites analysed by Hua et al. (2016). We employ the technique of metabarcoding, which combines traditional DNA barcoding with high-throughput DNA sequencing to characterize the biodiversity of mixed samples of eukaryotes (Cristescu, 2014;Deiner et al., 2017;Yu et al., 2012), and which has been shown to be a reliable and efficient method for biodiversity characterization (Ji et al., 2013). Through metabarcoding the non-bee arthropods caught in the same pan traps previously used to trap bees in Hua et al. (2016), we hope to answer the following questions: (a) Do native forests support higher levels of arthropod species richness and diversity than all four GFGP plantations? (b) For all GFGP plantations, do mixed plantations support higher levels of arthropod species richness and diversity than do the three individual monocultures? (c) How does community composition compare among these tree covers and what underlies the potential differences?

| Study location
The study region and locations are as in Hua et al. (2016). In short, our study region was a 7,949-km 2 area in south-central Sichuan In Hua et al. (2016), we used the term "mixed forests", but in Hua et al. (2018), we switched to "mixed plantations".
The two other surveyed land covers were croplands (CL) and native forests (NF). Croplands mostly consist of low-intensity plantings of rice, corn and vegetables and are the land cover type that has been reforested by GFGP. Native forests are broadleaf, subtropical, evergreen forests that have been subject to decades of selective logging and other forms of extraction. Because this region of China has been inhabited for millennia, there are no undisturbed native forests. Croplands are typically located on flatter land than are the tree covers, as GFGP reforestation targeted sloped land, and the native forests are concentrated towards the more hilly, southern end of the study region. For sampling, we chose larger expanses (>60 ha) of these six land cover types: BB, EC, JC, MP, NF and CL.

| Sampling design
Each land cover type was represented by at least two locations set ≥ 15 km apart. All tree-cover stands sampled had closed canopy. For each land cover type, we sampled with at least 10 one-ha quadrats, within each of which we operated 40 fluorescent pan traps for 24 hr (Bartholomew & Prowell, 2005; Figure S1). In total, we sampled 74 quadrats (BB: 10, EC: 10, JC: 12, MP: 10, NF: 16, CL: 16). Different quadrats were separated by ≥300 m if placed in the same tree-cover stand. Samples were stored in 100% ethanol at ambient temperature until shipment to the laboratory, where they were stored at −20°C before DNA extraction. The original reason for using pan traps had been to trap bees, which we individually DNA-barcoded in Hua et al. (2016). Here, we analyse the bycatch.

| Amplicon preparation
For each of the 74 quadrats, we pooled all 40 pan traps into a single sample. Three quadrats had very few individuals, and we pooled them with their nearest neighbour of the same land cover type (EC01 + EC02 + EC03; NF02 + NF03), leaving us with 71 samples.
Storage ethanol was removed by air-drying on single-use filter papers. Our samples were dominated by Diptera and Hymenoptera, as expected. We equalized input DNA across species by using one leg of every individual larger than a mosquito (~5 mm long) and the whole body if smaller (e.g. midges). This was to reduce the effect of largebiomass individuals out-competing small-biomass individuals during PCR, which improves taxon detection (Elbrecht, Peinert, & Leese, 2017). DNA extraction followed the protocols of Qiagen DNeasy Blood & Tissue Kits, followed by quantification via NanoDrop 2000 (Thermo Fisher Scientific).
We amplified a 319-bp fragment of COI using forward primer LCO1490 (5′-GGTCAACAAATCATAAAGATATTGG-3′) and reverse primer mlCOIintR (5′-GGNGGRTANANNGTYCANCCNGYNCC-3′) (Leray et al., 2013). All samples were carried out with two rounds of PCR. In the first round, both forward and reverse primers were tailed with tags (12-17 bp) for sample identification. In the second round, we added Illumina adapters to the amplicons from the first PCR, thus avoiding the tag jumping that can arise during library preparation of amplicon mixtures (Schnell, Bohmann, & Gilbert, 2015). A table of tags and primers is in Supplementary Information (Table S1). All PCRs were performed on a Mastercycler Pro (Eppendorf) in 20-µl reaction volumes, each containing 2 µl 10× buffer (Mg 2+ plus), 0.2 mM dNTPs, 0.4 µM of each primer, 1 µl DMSO, 0.4 µl BSA (bovine serum albumin) (TaKaRa Biotechnology Co. Ltd), 0.6 U Ex Taq DNA polymerase (TaKaRa Biotechnology) and approximately 60 ng genomic DNA. Both F I G U R E 1 Study area in south-central Sichuan Province, subdivided into counties and shaded by elevation. Each cross represents a pan-trap sampling location, colour-coded by land cover type: BB = bamboo monoculture, blue; EC = Eucalyptus monoculture, light green; CL = croplands, orange; JC = Japanese cedar monoculture, red; MP = mixed plantations, purple; NF = native forests, dark green rounds of PCR started with an initial denaturation at 94°C for 4 min, followed by 35 cycles of 94°C for 45 s, 45°C for 45 s and 72°C for 90 s, and finishing at 72°C for 10 min. PCR products were gel-purified with QIAquick PCR Purification Kit (Qiagen). One sample failed to amplify.
We pooled the 70 PCR products into two libraries and sequenced on the Illumina MiSeq (Reagent Kit V3, 300PE) at the Southwest Biodiversity Institute Regional Instrument Center in Kunming. The total number of paired-end reads returned was 13,601,908.

| Data analyses
The bioinformatic script, including parameters, for the analyses below is in Supplementary Information and will be archived in datadryad.org, along with sequence data and metadata. The R scripts and data tables are on https ://github.com/dougw yu/Sichu an2014.
Below, R packages are indicated with single quotes, and other software is italicized.

OTU filtration and taxonomic assignment
From the resulting sample X OTU table, we used "lulu" 0.1.0 (Frøslev et al., 2017) to combine OTUs that were likely from the same species but which had failed to be clustered by CROP. "lulu" identifies such "parent-child" sets by calculating pairwise similarities of all OTUs (using vsearch) to identify sets of high-similarity OTUs and then combining OTUs within such sets that show nested sample distributions.
For example, four OTUs might be highly similar, and within this set of four, one OTU contains the most reads and is observed in ten samples. This OTU is the parent, and daughters are inferred if they are present in a subset of the parent's samples. We ended with 1,506 OTUs.
A common filtering step is to remove OTUs made up of few reads (e.g. 1-read OTUs), as these are more likely to be artefactual (e.g. Yu et al., 2012;Zepeda-Mendoza, Bohmann, Carmona Baez, & Gilbert, 2016). For instance, PCR errors can generate clusters of sequences that are sufficiently different from the parent that they cannot be identified as daughters. Such OTUs are more likely to be small because novel haplotypes typically arise in a later PCR cycle. However, the definition of small is subjective and differs with the size of the sequence dataset. We therefore used "phyloseq" 1.19.1 (McMurdie & Holmes, 2013) to plot the number of OTUs that would be filtered out at different minimum OTU sizes (see http://evomi cs.org/wp-conte nt/uploa ds/2016/01/phylo seq-Lab-01-Answe rs.html, accessed 19 July 2018), and we chose a minimum OTU size of 44 reads, which was roughly the graph's inflection point and thus filtered out the most OTUs for the lowest minimum size. We ended with 594 OTUs.
We then used PyNAST 1.2.2 to align the 594 OTU sequences to a reference alignment of Arthropoda COI sequences (Yu et al., 2012) at a minimum similarity of 60%; one sequence failed to align and was deleted. The remaining sequences were translated to amino acids using the invertebrate mitochondrial codon table, and we removed 32 OTUs with sequences that contained stop codons. We carried out taxonomic assignment of the OTUs using a Naïve Bayesian Classifier (Wang, Garrity, Tiedje, & Cole, 2007) trained on the Midori UNIQUE COI dataset (Machida, Leray, Ho, & Knowlton, 2017). Sixteen OTUs assigned to non-Arthropoda taxa and two OTUs assigned to Collembola were removed. We ended with 543 OTUs.
Finally, we inspected the OTU table and set to zero those cells that had <5 reads representing that OTU in that sample, as these were more likely to be the result of sequencing error (Yu et al., 2012).
In addition, we removed two samples (rows) that contained ≤100 reads total (i.e. samples with little data) and removed seven samples (rows) with <5 OTUs because these samples were potentially overly influential in analyses of species richness. These seven samples included two from native forests and five from monocultures (three BB, one EC and one JC), meaning that we disproportionately removed monocultures, making our species diversity analyses below more conservative. After these sample removals, seven OTUs were removed because they were left with few (<20) reads. Because we do not consider OTU size to be a reliable measure of biomass or abundance (Nichols et al., 2018;Piñol, Mir, Gomez-Polo, & Agustí, 2015;Yu et al., 2012), we converted the OTU table into a presence/absence (0/1) dataset. Throughout, our bias was to remove false-positive detections even at the expense of losing true-positive detections, thereby resulting in a dataset with less, but more reliable (and thus more replicable), data. We ended with 536 OTUs and 61 samples.

OTU phylogenetic diversities
Because we used a combination of CROP+"lulu" and "phyloseq" to combine and remove small OTUs that were likely to be artefactual, the remaining OTUs were more likely to represent true presences.
Nonetheless, it remained possible that we had oversplit some biological species into multiple OTUs, as there is no single correct similarity threshold for species delimitation, and this oversplitting might have occurred more often for some taxa in some land cover types, leading to artefactual differences in species richness. However, oversplit OTUs should cluster together in a phylogenetic tree and thus contribute less to estimates of phylogenetic diversity than would OTUs from different biological species. Phylogenetic diversity should thus be a robust estimator of alpha diversity (Yu et al., 2012). To estimate sample phylogenetic diversities, we used "iNextPD" 0.3.2 (Hsieh & Chao, 2017). We built a maximum-likelihood (ML) tree in The algorithm used a rapid bootstrap analysis and searched for the best-scoring ML tree (-f a), with -N 1,000 times bootstrap and -p 12,345 as the parsimony random seed. Three OTU sequences produced very long branches in the ML tree, which would skew estimates of phylogenetic diversity, and we removed them. Two of these OTUs were found in all land cover types (and thus would not have been informative), and one was only found in some cropland samples (and thus would not have informed analyses of the tree-cover sites).

Beta diversity
To visualize changes in community composition across land cover types, we ran a Bayesian ordination with "boral" 1.6.1 (Hui, 2016), which is more statistically robust than non-metric multidimensional scaling (NMDS) analysis because "boral" is model-based and thus allows us to apply a suitable error distribution so that fitted-model residuals are properly distributed. We used a binomial error distribution and no row effect as we were using presence/absence data ( Figure S5). For the same reasons, we used "mvabund" 3.12.3 (Wang, Naumann, Wright, & Warton, 2012) to test the hypotheses that native forests and mixed plantations differ compositionally from each other and differ from the monocultures and croplands.
We also visualized changes in community composition with an "UpSetR" 1.3.3 intersection diagram, an alternative to Venn diagrams (Conway, Lex, & Gehlenborg, 2017), with a heatmap using the tabasco function in "vegan", and with a "betapart" 1.4-1 F I G U R E 2 Species richness estimates across land cover type. (a) Comparisons of Chao2 species richness estimates. Land cover types sharing the same superscript are not significantly different at the p = .05 level (Welch's t test) after table-wide correction for multiple tests (Bonferroni). (b) "iNEXT" estimates of species richness, Shannon diversity and "iNextPD" estimates of phylogenetic diversity by land cover type, using sample-based rarefaction and extrapolation. Native forests (NF) have the highest species richness and diversities, followed by croplands (CL) and mixed plantations (MP), followed by the three monoculture plantations (BB, EC and JC). Codes for land cover types as in Figure (Baselga & Orme, 2012) analysis, which partitions beta diversity into turnover and nestedness components using binary Jaccard dissimilarities, which we visualized with NMDS using the metaMDS function in "vegan". Finally, we used "metacoder" 0.2.0 (Foster, Sharpton, & Grunwald, 2017) to generate taxonomic "heat trees" to pairwise-compare the six land cover types and identify the taxa most strongly driving compositional differences.

| Alpha diversity
Species richness and diversity are highest in native forests and croplands, followed by mixed plantations, which are in turn richer and more diverse than the monoculture plantations, with the possible exception of bamboo.

| OTU richness and diversities
The Chao2 estimator indicates that native forests, mixed plantations and croplands have the highest estimated species richnesses and do not differ significantly from each other (Figure 2a). Importantly, all three monocultures (bamboo, Eucalyptus and Japanese cedar) exhibit less than half the species richness of native forests and around half the species richness of mixed plantations (Figure 2a). The pairwise differences between native forests and monocultures are all statistically significant (Table S2), and the pairwise differences between mixed plantations and the three monocultures are marginally or significantly different (Figure 2a; Table S2), all after table-wide correction.
The iNEXT analysis reveals even clearer contrasts: native forests have the highest estimated asymptotic species richnesses and Shannon diversities, followed by croplands and mixed plantations, F I G U R E 3 Phylogenetic distribution of OTUs by land cover type, created using "iNextPD". Terminal nodes are black and represent the OTUs. Internal nodes are white. Sizes of the squares on the right indicate each OTU's incidence frequency (number of samples in which the OTU is observed). Phylogenetic coverage is most complete in native forests (NF) and croplands (CL), followed by mixed plantations (MP), followed by the three monocultures (BB, EC and JC). Codes for land cover types as in Figure 1 BB CL EC JC MP NF 2.5 7.5 12.5 followed by the three monocultures (Figures 2b and S3). The iNEXTestimated richness and diversity of mixed plantations are significantly higher than all the monocultures, with the possible exception of bamboo, because the MP and BB confidence intervals touch.
Using "iNextPD" to visualize phylogenetic coverage by land cover type ( Figure 3) reveals that native forests and croplands exhibit almost complete coverage of the OTU tree, whereas mixed plantations and bamboo exhibit some coverage deficits, followed by larger coverage deficits in the other two monocultures.

| Beta diversity
Native forests are compositionally most similar to mixed plantations and most dissimilar to croplands. The differences in community composition are driven primarily by species turnover.

| Differences in community compositions
Ordination with "boral" (Figure 4a) shows that the primary separation is between the tree-cover types and croplands, with a signifi- and 5). However, despite their overlap, "mvabund" analysis shows that the arthropod communities of mixed plantations and native forests are still significantly distinct from each other, and from the three monocultures and croplands (Table S3).

| Turnover versus nestedness
Consistent with the UpSetR result that the mode in each land cover type is unique species, we found that turnover, not nestedness, dominates compositional differences ( Figure 5; see Figure S7 for a heatmap visualization). In other words, the arthropod communities in the monocultures are not simply subsets of native forests or mixed plantations but contain distinct sets of species.

| Taxonomic compositions of and differences between land cover types
The 536 arthropod species in our metabarcoding dataset represent a wide range of arachnid and insect orders and, thus, represent a wide range of ecological functions (Figure 6), including generalist predators (Araneae, Formicidae) and more specialized parasites and parasitoids (Tachinidae, Phoridae, Braconidae) of other arthropods.
Although the "boral" ordination ( Figure 4a) reveals compositional similarity between mixed plantations and native forests, it does not reveal the taxa that are most responsible for this similarity, and for the differences with the other tree-cover types. With

Intersection Size
we observe the largest number of heat-tree-tip differences between any two land cover types. In other words, there are multiple specieslevel indicators of croplands (or in the case of the Mycetophila OTU, an indicator of Japanese cedar and Eucalyptus).

| Improving biodiversity conservation under the GFGP
Our study found that native forests support the highest levels of arthropod species richness, Shannon and Simpson diversities, and Faith's and phylogenetic diversity (Figures 2, 3, S3 and S4) and that most of those species are unique to native forests (Figure 4b), consistent with the patterns of bird diversity that were reported in Hua et al. (2016) and other biodiversity studies in plantations Gardner et al., 2008). In addition, our findings pertaining to the higher level of alpha diversity in mixed plantations over monocultures (Figures 2, S3 and S4), and their greater degree of compositional similarity to native forests relative to monocultures (Figures 4,5,6,S6,S7 and S8), corroborate those reported for birds (but not bees) in Hua et al. (2016) and are consistent with other studies of biodiversity in tree plantations. Butterfield and Malvido (1992) showed that mixtures of broadleaves and conifers resulted in a higher species richness of carabid beetles than in conifer monocultures, and Recher, Davis, and Holmes (1987) showed that some bird species are present when in Eucalyptus-pine mixtures but absent from pine monocultures. In short, mixed plantations not only support a higher diversity of non-breeding birds but also provide a small but detectable biodiversity boost for arthropods. Finally, we found that compositional differences among tree-cover types are almost entirely dominated by species turnover, not nestedness, meaning that some species were only detected in the monocultures. This result is consistent with the pattern of moth communities in primary, secondary and plantation forests studied by Hawes et al. (2009). In their findings, all three of their tree-cover types (primary and secondary forest, Eucalyptus plantation) contained large numbers of unique species in three moth families (Arctiidae, Saturniidae and Sphingidae).
Given the balance of evidence, we reaffirm our previous policy recommendations that the GFGP should prioritize the retention and restoration of native forests, and when restoring native forests is F I G U R E 5 NMDS (non-metric multidimensional scaling) ordination of beta diversity by land cover type (binary Jaccard dissimilarities), partitioned with "betapart". (a) Total beta diversity. (b) Beta diversity based on species turnover only. (c) Beta diversity based on species nestedness only. Turnover accounts for most the observed beta diversity across land cover types, which is visualized as greater distances between points in the turnover figure (b) and almost no distances between points in the nestedness figure (c). Codes for land cover types as in Figure 1  not possible, we secondarily encourage mixed-species plantings over extensive monocultures, at least in western China where we conducted this study. The foundation of these recommendations is now broadened to include 536 species-resolution taxa ranging across the Arthropoda. Given the growing understanding of the biodiversity implications of plantations compared with native forests in different forest biomes across the world (Bremer & Farley, 2010;Fierro, Grez, Vergara, Ramírez-Hernández, & Micó, 2017), these recommendations likely apply to other regions in China where GFGP is relevant, but their applicability will benefit from additional field studies and from anticipated technical advances in DNA-based biodiversity assessment. In the future, it will likely be insightful to carry out time series biodiversity surveys, as our dataset represents only a single time point, but the temporal turnover of forest arthropod communities is high (Barsoum et al., 2019). It is possible that the differences in biodiversity levels that we have detected are even stronger when integrated over time. Another important variable that we did not measure is sample biomass, given recent evidence that insect biomass has been dropping around the world (e.g. Hallmann et al., 2017). Because we observed high species richness and diversity in our cropland sampling sites (Figures 2, 3, S3 and S4), where agriculture is small-scale in nature, our a priori expectation is that biomass has probably not declined here as rapidly as elsewhere, but this clearly needs testing and should of course now be a standard metric in biodiversity surveys.
Greater levels of arthropod biodiversity in native forest is not a surprise, given their more diverse vegetation structures and species compositions, which are well known to be positively correlated with arthropod diversity (Castagneyrol & Jactel, 2012;Haddad et al., 2009;Stork, Mcbroom, Gely, & Hamilton, 2015;Zhang et al., 2016), but the greater diversity and similarity of mixed plantations to native forests is somewhat surprising, especially as they mostly just F I G U R E 6 Pairwise taxonomic comparisons of all land cover types. Upper right triangle: greener branches indicate taxa that are relatively more abundant (in numbers of OTUs) in the land cover types along the right column, and browner branches indicate taxa that are relatively more abundant in the land cover types along the top row. Lower left: taxonomic identities of the branches. Note that this is a taxonomic tree, not a phylogenetic tree. Legend: width indicates number of OTUs at a given taxonomic rank, and colour indicates relative differences in log 2 (number of OTUs). Codes for land cover types as in Figure 1. This mechanism is consistent with our finding that bamboo, which does not create closed canopies, exhibits the highest richness and diversity of the monocultures (Figures 2, S3 and S4). We note that 95% confidence interval overlap is considered an overly conservative test for statistical significance at the p = .05 level (MacGregor-Fors & Payton, 2013). A more diverse, and presumably higher-biomass, arthropod community in turn could also support a richer bird community, at least for the insectivorous subset of the community. Our results thus point to a plausible mechanism for why bird diversity is boosted in mixed plantations.
In this study, we report evidence for a biodiversity benefit of native forests over GFGP plantations, which we might think trades off against a greater value of timber sales from plantations. However, even excluding biodiversity, which they did not study, Cao, Zhang, and Su (2019) (Hua et al., 2016) to prioritize native-forest recovery and expansion over creating plantations.

| Methodological comments on metabarcoding and studies of biodiversity patterns
Metabarcoding provides an efficient method for interrogating biodiversity samples, but because of its reliance on PCR, metabarcoding datasets tend to contain a non-trivial amount of noise. This noise manifests as a large number of false-positive OTUs, which are filtered out heuristically. Such false OTUs especially complicate efforts to estimate alpha diversity. Here, we applied several filtering steps to remove false OTUs, and we also used "iNextPD" to generate robust comparisons of alpha diversity by estimating phylogenetic diversity instead of species richness. This approach has been previously shown to be reliable (Yu et al., 2012). Another approach, which became available only after we had completed the wet-laboratory portion of our study, is to subject each sample to multiple, independently tagged PCRs (typically three) and to bioinformatically filter out sequences that fail to appear in at least two of the PCRs above some minimum number of reads; such sequences are more likely to be PCR or sequencing errors. This is implemented in the DAMe pro- With regard to studies of biodiversity patterns, we follow Magurran, Dornelas, Moyes, Gotelli, and McGill (2015; Magurran, 2016) in recommending that we should focus less on explaining change in species richness and more on explaining change in species composition as a function of natural and anthropogenic causes.
The argument is that anthropogenically disturbed communities can maintain species richness and even phylogenetic diversity, even as local, or worse still, endemic, species go extinct and are replaced by cosmopolitan species. In our study, croplands support an arthropod community similar in richness and diversity to that of mixed plantations and just below that of native forests ( Figures   2, 3, 4b, S3 and S4), but the species composition of croplands is distinct from those in native forests (Figures 4, 5, S6, S7 and S8).
Croplands therefore cannot compensate for the loss of the biodiversity dependent on native forests.

ACK N OWLED G EM ENTS
We