To characterize green roof soil microbial assemblages, soil samples were collected from 32 green roof sites across New York City (Figure 1). New York City has a mean annual temperature of 16.4°C and average annual precipitation is ~1,110 mm (Graphical Climatology 2018). All samples were collected between July 10 and 21, 2017. Green roof buildings ranged from 1.5 to 12 stories and vegetated roofs ranged from ~96–6,070.28 m² in area (Table 1). The average area covered by soil core sampling is 88.73 m² to account for species-area relationships and the areas surveyed on green roofs ranged from 15 to 100 m². Two categories, Sedum and mixed-vegetation, were used to broadly characterize the plant communities on each green roof. Sedum green roofs denote rooftops planted exclusively with Sedum species while the mixed-vegetation category captures roofs planted with any or all of the following: grasses, wildflowers, and succulents (Figure 2).

To survey soil bacterial and fungal communities on each roof, we collected nine randomized 10-cm deep soil samples using 2.5-cm diameter soil corers. Since host plant species may influence root-associated microbial assemblages (Smalla et al., 2001; Kowalchuk et al., 2002; Singh et al., 2008), we collected soil samples adjacent to species of Sedum from the Sedum green roofs and adjacent to species of Asteraceae on the mixed-vegetation roofs. We chose this method of targeted sampling to account for potential variation in soil microbial community composition due to differences in plant species. Asteraceae plants were selected for targeted sampling due to frequent use of Asteraceae plant species on green roofs (McGuire et al., 2013; Aloisio et al., 2016). Due to variation in green roof design, samples were arranged according to the layout of vegetation. On green roofs with continuous patches of Sedum or Asteraceae plants, soil cores were obtained at three points along three randomly arranged transects. The second method was unique to mixed-vegetation roofs, where all patches of Asteraceae species were numbered and subsequently selected for sampling using a random number generator. We also collected additional roof information from either the green roof database managed by the Green Roof Researchers Alliance, when possible, or contacted green roof managers directly (Table 1).

To assess the effects of green roof soil microbial communities on plant physiological response to heat and drought, we conducted a replicated, fully factorial experiment in the Arthur Ross Greenhouse at Barnard College, Columbia University (New York, NY, USA). We germinated 120 Panicum virgatum (Poaceae) and 120 Solidago nemoralis (Asteraceae) individuals from seed (Greenbelt Native Plant Center, Staten Island, NY, USA) and propagated 120 Sedum tetractinum (Crassulaceae) plants from cuttings (NYC Department of Parks and Recreation Citywide Nursery, Bronx, NY, USA). Soil microbial communities were established in greenhouse pots via inoculation with field-collected soil from eight conventional green roofs planted with Sedum mats and eight green roofs planted with mixed-vegetation (grasses and wildflowers), and autoclaved green roof media (rooflite Semi-intensive green roof mix) as a control. Hereafter, the soil treatments are referred to as Sedum soil, mixed-vegetation soil, and autoclaved control soil, respectively. Experimental pots were arranged in nine total treatments (three focal plant species × three soil types) (Figure S1).

In the greenhouse, plants underwent four treatments: drought stress, heat stress, heat + drought stress, and a well-watered, non-heated control. Following a 90-day period to allow soil inocula to fully establish in each treatment, drought was induced on 90 of the plants by withholding water for 10 days, simulating summer drought conditions in NYC. Germination heat mats (VIVOSUN) were placed beneath pots receiving heat and heat + drought treatments for 10 days, increasing substrate temperature from an average of 27.2–29.7°C. Pots that were not drought treated received water every 3 days, which was estimated before the stress treatments began by measuring soil moisture and pot weight each day (soil moisture declined from an average 22.3% moisture on the day of watering to ~2% after 3 days post-watering).

We used plant survivorship, biomass, and g_s to evaluate individual plant responses to abiotic stressors. Twice per week, plants were recorded as either alive or dead. Final fresh and dry biomass was collected for S. tetractinum by first removing all soil from the plant root system, and separating above-ground and below-ground parts of the plant. Prior to oven-drying the plants, all roots and shoots were weighed to obtain fresh mass. Dry mass was estimated by weighing the roots and shoots after they were oven-dried for 48 h at 60 degrees C.

For each plant-soil-treatment combination, one leaf on four individuals was selected for repeated g_s. For all plants, g_s, leaf temperature, and pot weight measurements were collected from 10:00 to 13:00 using an SC-1 leaf porometer (Decagon Devices). Measurements were obtained on alternate days for control and drought, and heat and heat and drought treatments, following watering schedules, totaling five samples for each of the four replicates in each plant-soil-treatment combination. All measurements were collected 2 days after watering for the control and heat treatments. Immediately after the treatments were completed, three soil samples from each plant-soil-treatment combination were collected to investigate root-associated fungal and bacterial communities (108 samples total).

To characterize bacterial and fungal communities from the green roof survey and fungal communities from the greenhouse experiment, DNA was extracted from ~0.25 g of each soil sample from the soil survey and greenhouse experiment using a DNeasy PowerSoil Kit (Qiagen). For fungal OTUs, PCR amplification of the internal transcribed spacer region 1 (ITS1F) was carried out using the fungal-specific primer pair, ITS1F and ITS2R, adapted for the Illumina platform (McGuire et al., 2013). For bacterial OTUs, PCR amplicons for the 16S rRNA gene and a corresponding Illumina dataset were generated.

For ITS data, reads were de-multiplexed, quality-filtered, and processed using QIIME (Caporaso et al., 2010). The UNITE fungal reference database was used to define operational taxonomic units (OTUs). High-quality sequences from each sample were rarified to 13,000. OTUs were assembled using the QIIME V1.9.1 1 pipeline.

For 16S data, sequences were de-multiplexed and all sequences with a quality score <20 were excluded from analysis. OTUs were selected using the UCLUST algorithm and taxonomy was assigned using the greengenes reference database (McDonald et al., 2012) with the RDP classifier specialized for 16S rRNA sequences (Wang et al., 2007). For both ITS and 16S data, data were transformed using the “phyloseq” and “dplyr” packages in R (McMurdie and Holmes, 2013; Wickham and François, 2015) to join metadata with the cleaned OTU tables prior to downstream analysis.

In R, we computed Shannon-Weiner and Simpson diversity indices (Oksanen et al., 2018) to compare bacterial and fungal alpha diversity measures between vegetation communities and tested statistical significance of diversity indices between vegetation communities using PermANOVA in R (Anderson, 2001). We assessed fungal and bacterial community similarity by using Bray-Curtis dissimilarity index to calculate pairwise distances among fungal and bacterial communities due to our method of sampling (maximum area of 100 m² on a green roof) and its use in similar studies evaluating community composition on green roofs, including microbial communities (McGuire et al., 2013; Parkins and Clark, 2015; Partridge and Clark, 2018). Using non-metric multidimensional scaling (NMDS) plots, we visualized fungal and bacterial community clustering. We tested statistical significance of observed clustering patterns by plant community using PermANOVA in R. A random forest model was used to test if vegetation type could be classified based on bacterial and fungal scaled relative abundance (Liaw and Wiener, 2012; Evans and Murphy, 2018). To explore which taxa were associated with mixed-vegetation compared to Sedum green roof systems, we used the R package indicspecies (De Caceres and Legendre, 2009) to run an indicator taxa analysis.

We assessed fungal and bacterial community similarity on a subset of samples from the NYC Parks Department Five Borough Building, using the same index and analysis that was employed for the city-wide assessment (Bray-Curtis dissimilarity index, PermANOVA).

For the greenhouse experiment, we used generalized linear models (GLMs) to explore if each variable (plant species, soil treatment, and environmental treatment) was predictive of plant response (survivorship, biomass, and g_s). GLMs were employed as the data were not normally distributed. We ran a GLM at each sampling time for g_s and used the final biomass and survivorship as response variables. Since final survivorship is a binary response variable, we specified the family as binomial in the GLM for plant survivorship.

Using the R package FUNGuild (Nguyen et al., 2015), we categorized all OTUs into the following functional groups: animal endosymbiont, animal pathogen, arbuscular mycorrhizal, ectomycorrhizal, endophyte, epiphyte, ericoid mycorrhizal, fungal parasite, lichen, plant pathogen, saprotroph, and unclassified. All data visualization and figures were created using “ggplot2” (Wickham and François, 2015).

After filtering the dataset for high-quality sequences, there were 13,332 fungal OTUs among the 266 samples with an average of 494 OTUs per sample. The most abundant phyla were Ascomycota (67.2%), Basidiomycota (19.1%), Glomeromycota (6.4%), and Mortierellomycota (3.5%) in both vegetation systems (Figure 3). Green roofs planted with mixed-vegetation communities had a greater number of fungal OTUs in each sample (540.85 ± 110.28) compared to roofs planted with Sedum species (411.90 ± 75.51). A total of 486,026 bacterial OTUs were classified and samples were rarefied to an even depth of 82,376 OTU per sample. Though there were no significant differences in alpha diversity among indices for bacteria, there was significantly higher observed (df = 1, F = 78.39, p < 0.001) and Shannon (df = 1, F = 7.024, p < 0.01) alpha soil fungal diversity on mixed-vegetation green roof systems.

Green roof soil fungal communities were significantly correlated with both biotic and abiotic factors. Mixed-vegetation and Sedum green roof plant communities had significantly different fungal communities (R = 0.27, p < 0.001; Figure 4A) and bacterial communities (R = 0.03, p < 0.01; Figure 4B). Soil fungal communities were compositionally distinct across plant genera (R = 0.38, p < 0.001), roof age (R = 0.13, p < 0.001), and by green roof system (R = 0.44, p < 0.001).

To account for our large sample size (n = 266), we focused on green roof systems sampled from the NYC Parks Department Five Borough Building, which is planted with a variety of green roofs containing different plant palates and soil types. This allowed a local test of differences in microbial communities associated with different plant communities, while keeping the location constant. We sampled four green roof systems, which are on the same roof and were built between 2009 and 2011. Among the four green roof systems on the Five Borough Building, there were significant differences in fungal community composition between vegetation type (R = 0.8237, p < 0.001) and across the green roof systems (R = 0.6292, p < 0.001; Figure 5A). There were significant differences in bacterial community composition by vegetation type (R = 0.69; p < 0.001; Figure 5B).

We ran random forest models to determine if vegetation type (mixed-vegetation or Sedum) could be classified based on bacterial and fungal scaled relative abundance. Both bacterial and fungal communities could classify vegetation type the majority of the time. Running 1,000 permutations, 99% of samples from mixed-vegetation roofs were classified using bacterial communities compared with 97% with fungal communities; and 90.7% of samples from Sedum green roofs were correctly classified with bacterial communities and 94% with fungal communities.

Saprotrophs were the most abundant fungal functional group on both mixed-vegetation (55.71%) and Sedum (40.83%) green roof systems (Figure 6). Plant pathogens comprised 14.51% of OTUs on Sedum green roofs, compared to 11.11% on mixed-vegetation roofs. The relative abundance of AM fungi was ~4 times higher on mixed-vegetation green roofs (1.37%) compared to Sedum green roofs (0.32%).

A total of 398 fungal OTUs were significantly associated with either the mixed-vegetation or the Sedum green roof sites. Twenty-nine of the 42 fungal indicator families associated with mixed-vegetation roofs aligned with Ascomycota and nine aligned with Basidiomycota. Two families, Archaeosporaceae and Claroideoglomeraceae, are in the phylum Glomeromycota (AM fungi). All indicator families for Sedum green roofs were within Ascomycota or Basidiomycota.

There were 48 bacterial indicator families associated with mixed-vegetation green roof systems, 22 of which aligned to the phylum Proteobacteria. Indicator taxa included genera recognized as plant growth-promoting bacteria, such as Bacillus (Glick, 2012) and Azospirillum (Saikia et al., 2014). There were 38 bacterial indicator families for Sedum green roofs, including Rhizobiaceae and Bradyrhizobiaceae, both of which contain plant-associated bacterial species.

Spearman's correlation between green roof abiotic and biotic factors revealed a strong association between the substrate depth and vegetation type (r = −1). A redundancy discriminant analysis revealed a much larger proportion of unconstrained variance (0.82) compared to constrained variance (0.18), though vegetation type explained the most variance in soil fungal community composition compared to all other variables analyzed.

The distinct soil microbial assemblages on Sedum compared to mixed-vegetation green roofs support our first hypothesis that green roof soil microbial community composition shifts with host plant species as observed in natural ecosystems (van der Heijden et al., 1998; Reynolds et al., 2003; Zak et al., 2003). Potential ecological mechanisms supporting these findings include microbial host-specificity (Santos-Gonzalez et al., 2007), host plant root structural differences (el Zahar Haichar et al., 2008), and functional group complementarity (Newsham et al., 1995). We found that within a single roof (the NYC Parks Five Borough Building roof), microbial community composition varied with plant composition and that 99% of the time, vegetation communities could be classified based on bacterial and fungal, strongly suggesting that soil microbial assemblages are not solely driven by differences in roof location and roof planting history, but also by the plant community assemblages themselves.

There is ample material supporting microbial community composition as drivers of plant community structure (Reynolds et al., 2003; van der Heijden et al., 2006; Bever et al., 2010) through direct interactions between plant roots and microbial functional groups, such as pathogens and mutualists, and from microbial nutrient cycling. As constructed ecosystems, green roofs are planted with a specific palette, enabling a directional perspective on the effect of plant communities on soil microbial composition. Previous work has highlighted the role of microbial composition on successional plant community establishment. Unmanaged green roofs are ideal sites to explore the role of microbial communities in plant community development on green roofs.

We chose to target species in Asteraceae on the mixed-vegetation roofs, thus our results may reflect differences in root-associated microbial community composition among Sedum and Asteraceae plant species. Studies assessing root-associated microbial communities with more intensive sampling are crucial to further assess associations between plant and microbial assemblages on green roofs.

Our field survey did not control for green roof substrate type and initial inoculum, which is crucial to obtain a clearer picture of the relationship between plant and soil microbial community assemblages. Molineux et al. (2014) manipulated soil microbial communities using compost tea and AM fungi in two different substrate types, revealing the key role of substrate type and depth on green roof microbial ecology (measured by biomass). Nonetheless, the fact that we observed a difference in microbial communities associated with the different plant communities, irrespective of substrate type and initial inoculum, suggests that the effect of the plant community is a robust driver of microbial assembly in these systems.

McGuire et al. (2013) found distinct soil fungal communities between replicate green roofs and nearby city parks. Many bacteria and fungi are wind dispersed and AM fungi spores are more effectively transported by animals capable of reaching rooftops (Camargo-Ricalde, 2002), thus dispersal limitations could have been driving these differences. Initial inoculum in the starting green roof substrate or on plant roots growing in nursery soil may exhibit priority effects, which could impact the ability of other microbial taxa to establish after planting. Rumble et al. (2018) reported high levels of independent mycorrhizal colonization of Sedum species (S. album, S. spurium, S. reflexum) in a field experiment (92%), while John et al. (2014) reported no mycorrhizal colonization of Sedum acre, but colonization of all other green roof species, especially on Solidago bicolor (Asteraceae) and Danthonia spicata (Poaceae).

The most abundant microbial taxa we identified in green roof soils are also widespread in other temperate ecosystems. For example, species of Mortierella, the most abundant fungal genus in mixed-vegetation green roof soils, and has been confirmed in a variety of habitats including natural compost and forest soils (Wagner et al., 2013). On the Sedum roofs, the most abundant fungal genus on green roofs planted with Sedum was Corynespora, specifically the species Corynespora cassiicola, which is a plant pathogen. This fungal species is widespread in its geographic distribution and was recently described to cause “stem spot” an economically-important succulent plant in Europe (Madriz-Ordeñana et al., 2017). The pathogens identified in our survey may manifest as disease-causing agents for the green roof plants. Proteobacteria are also found in a variety of habitats and comprise a wide array of functions, ranging from pathogens to nitrogen fixing bacteria.

We found that AM fungi were abundant in all green roof samples, suggesting that plant-fungal mutualisms are sustained on green roof ecosystems, which corroborates with previous studies investigating microbial communities on green roofs (McGuire et al., 2013; John et al., 2014; Molineux et al., 2014; Rumble et al., 2018). Mixed-vegetation roofs had a greater abundance of AM fungi than the Sedum roofs, which had lower plant and microbial diversity. Mycorrhizal fungi colonize the root systems of 95% of plant families (Smith and Read, 1994), often benefiting the host by enhancing drought tolerance, increasing aboveground biomass, and promoting plant survivorship through enhanced nutrient uptake and resistance to pathogen attack (Augé, 2001). Due to environmental characteristics of green roofs, such as shallow growing media and drought and heat stress during the summer, grasses and forbs on green roofs likely highly depend on AM fungal symbionts for growth and survival. AM fungi can also mitigate the harmful effects of soil pathogens through reducing pathogen growth (Borowicz, 2001), eliciting plant defense mechanisms and competing with other rhizosphere biota associating with host plants (Azcón-Aguilar and Barea, 1997). Further research elucidating AM fungal community assembly and function in green roof systems is necessary to better understand the roles of these mutualists in stressful environments. We detected additional fungal functional groups, including ericoid mycorrhizal fungi and ectomycorrhizal fungi. A number of ectomycorrhizal host plants are temperate forest tree species common in New York City's canopy (Pregitzer et al., 2019), but these are not planted on green roofs; however, fungal spores in association with these tree species likely dispersed through the air and landed in the green roof substrate where we sampled.

The fact that AM fungi were most abundant in mixed-vegetation soil under drought conditions (Figure 10), compared to all other soil inoculum-treatment pairings in the experiment, suggests that AM fungi may play a role in maintaining plant cooling benefits under dry conditions on green roofs. Numerous field and greenhouse experiments have shown the effects of mycorrhizal fungi on stomatal conductance, the rate at which CO₂ enters and water exits the leaf, with higher stomatal conductance in drought vs. well-watered conditions (Augé, 2001; Augé et al., 2015). While adverse abiotic conditions reduce AM fungal diversity compared to non-disturbed soil, some AM fungi exhibit opportunistic life history strategies including investing energy into spore production (Declerck et al., 2001) and, in the case of Rhizophagus irregularis, rapidly colonizing plant roots after disturbance (Sýkorová et al., 2007). In the green roof survey samples, the three most abundant OTUs in Glomeromycota aligned to Rhizophagus irregularis, which occurs in high relative abundance with Asteraceae plants (Wehner et al., 2014), and have previously been observed in green roof fungal communities in New York City (McGuire et al., 2013). In this study, we collected and sequenced bulk soil communities from the rhizosphere of target plant species; this sampling strategy likely only detects AM fungal taxa extraradical mycelia and spores. Since plant biomass and survivorship during drought conditions has been tightly correlated with mycorrhizal colonization (Augé, 2001), quantifying percent root colonization and characterizing AM fungal taxa present in plant roots rather than just bulk soil is necessary to further elucidate the assembly patterns and functional roles of AM fungi in green roof systems.

The shifts in relative abundances of plant-growth promoting bacteria on mixed-vegetation compared to Sedum green roofs suggests the importance of integrating knowledge of soil bacteria into green roof strategies. Plant growth-promoting and endophytic bacteria promote plant vigor and survival under harsh environmental (Wang et al., 2016) and biotic conditions (van Loon et al., 1998), and thus likely play an important role in green roof ecosystems (McGuire et al., 2015; Fulthorpe et al., 2018). Xie et al. (2018) inoculated a variety of plant species with Bacillus amyloliquefaciens, a plant-growth promoting bacteria that associates with plant roots and confers benefits, such as pest resistance and nutrient absorption to the host. In addition to facilitating plant shoot growth, B. amyloliquefaciens demonstrated promotion of host colonization by AM fungus R. irregularis (Xie et al., 2018). The co-occurrence of plant growth-promoting bacteria and mutualist fungi associated with mixed-vegetation roofs suggest the role of root-associated biotic interactions in facilitating plant growth, survival, and subsequent ecosystem function on green roofs.

Obtaining direct measurements on plant g_s and biomass, and tracking survivorship of plants grown with different soil microbial inocula in controlled conditions enabled a more focused understanding of the soil microbial control on plant vitality and overall ecosystem services in green roof systems. The high survivorship of S. tetractinum aligns with results from previous green roof experiments, which have shown the resilience of Sedum species to abiotic stress on green roofs (Monterusso et al., 2005; Lundholm et al., 2010; Nagase and Dunnett, 2010). These results also demonstrate the negative impact of drought relative to heat on the survivorship of grasses and forbs.

The unique response of S. tetractinum compared with P. virgatum and S. nemoralis to abiotic stress treatment and soil inocula suggest potential functional complementarity between succulents, grasses, and forbs as they relate to cooling capacity. When subjected to drought, S. tetractinum had the greatest above and below-ground biomass, as Sedum species may be more adversely affected by overwatering than by under-watering (Sterphenson, 1994). We found that S. tetractinum plants had significantly greater root biomass when grown in autoclaved soil, which contrasts to previous work, in which plant-associated beneficial microbes increased shoot and root biomasss of Sedum alfredii (Ma et al., 2016). However, John et al. (2014) found no colonization of Sedum acre on green roofs. We speculate that S. tetractinum may have greater root biomass in autoclaved soil since it may have less percent root colonization by mycorrhizal fungi and as a result, needs to grow its root network to access nutrients in the soil.

Differences among plant species and stomatal conductance (g_s), which were especially pronounced during peak stress, highlighted the effects of plant life form and on potential green roof cooling benefits. Previous studies reported that tall forbs and grasses, including Solidago bicolor, experienced high water loss, compared to succulents (John et al., 2014) and that grass had the highest water loss across all watering treatments while forbs had equivalent water loss to grasses in well-watered and intermediate drought conditions (Wolf and Lundholm, 2008). Planting green roofs with grass and forb species may contribute to greater cooling benefits due to the relatively high transpiration rates in grasses and forbs compared to succulents, with the caveat that forbs lose cooling capacity under severe drought.

The elevated g_s of P. virgatum at peak heat and drought conditions indicate its potential usefulness in maintaining cooling benefits under abiotic stress conditions. Conversely, S. nemoralis had relatively low g_s at peak stress, but had elevated g_s when grown in mixed-vegetation soil during the heat and drought treatments. Solidago nemoralis had extremely low g_s under all treatments when grown in autoclaved soil. We speculate that AM fungi present in the field-collected inoculum enabled S. nemoralis to retain elevated g_s in a variety of conditions including well-watered and moderate heat stress.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.