Rhododendron maximum impacts seed bank composition and richness following Tsuga canadensis loss in riparian forests

Southern Appalachian riparian forests have undergone changes in composition and function from invasive pathogens and pests. Castanea dentata mortality in the 1930s from chestnut blight (Cryphonectria parasitica) and Tsuga canadensis mortality in the 2000s from the hemlock woolly adelgid (Adelges tsugae) have led to the expansion and increased growth of Rhododendron maximum, an evergreen subcanopy shrub. A better understanding of seed bank characteristics and the various abiotic and biotic factors that affect the seed bank may be useful in determining the restoration potential of forest communities following invasion-related disturbances. We compared the seed bank of two deciduous forest types: hardwood forests with a dense R. maximum subcanopy (hereafter, RR) and hardwood forests without R. maximum (hereafter, HWD). We evaluated numerous microenvironmental variables through principal component analysis (PCA) and correlated the derived PCA axes scores to seed bank density and richness across forest types. We found that seed bank density was comparable between the forests types; however, seed bank richness was much lower in RR than HWD and the species composition was dissimilar between forest types. Twenty-eight of 64 (44%) species in the seed bank of HWD were not found in the seed bank of RR. Species that were represented in both forest types were often found in contrasting densities. Most notably, seed bank densities of several woody species were considerably higher in RR (85%) than HWD (45%), while herbaceous seed bank density was lower in RR (11%) than HWD (50%). Mineral soil pH, soil nutrient availability, and soil moisture were lower, and organic soil (Oi + Oe + Oa) depth and mass were greater in the RR than HWD forest type. PCA correlations revealed that PCA4 (represented by understory density and Oe + Oa phosphorus and carbon/nitrogen ratio) was negatively correlated with total seed bank density. PCA1 (represented by Oe + Oa cations and phosphorus, understory richness, ground-layer cover, and mineral soil pH) and PCA4 were positively correlated with total seed bank richness. These results suggest that the soil seed bank will not be the primary mode of recruitment to establish a diverse and herbaceousrich community if a RR is present.


INTRODUCTION
Human activities over the past century have strongly impacted forest disturbance regimes in many parts of the world. Repeated outbreaks of native insect pests and exotic invasive species have caused widespread tree mortality in North America (Orwig et al. 2002, Poland and McCullough 2006, Kurz et al. 2008 and Europe (Lausch et al. 2013, Maclean et al. 2018a, where introduced species continue to invade new territory despite efforts to contain their spread. Examples of non-native forest insects and pathogens that have resulted in tree mortality include gypsy moth (Lymantria dispar L.), emerald ash borer (Agrilus planipennis Fairmaire), chestnut blight (Cryphonectria parasitica (Murr.) Barr), Dutch elm disease (Ophiostoma novo-ulmi), and hemlock woolly adelgid (Adelges tsugae Annand), and these disturbances have inflicted ecological and economic damage (Lovett et al. 2016). Invasive species introductions are expected to continue, and even intensify, with potentially far-reaching consequences for forest communities and their associated ecosystem services (e.g., carbon storage, nutrient cycling; Ellison et al. 2005, Lovett et al. 2016, Dukes et al. 2009).
Changes in forest disturbance regimes may in some cases result in increased abundance of one or more understory plant species (Mallik 2003, Royo andCarson 2006). In the southern Appalachian Mountains, Castanea dentata (Marsh.) Borkh. tree mortality in the mid-1930s due to the chestnut blight resulted in expansion of Rhododendron maximum L. shrubs , and the more recent Tsuga canadensis (L.) Carri ere tree mortality from infestation by the hemlock woolly adelgid has led to further increased growth of R. maximum . Rhododendron maximum is an ericaceous evergreen shrub that, at high densities, inhibits tree seedling recruitment and limits overstory regeneration (Clinton 1995, Lei et al. 2002. Rhododendron maximum inhibits seedling recruitment by substantially altering resource availability through physical and chemical pathways, suggesting active management of this species in post-Tsuga canadensis forests may be required for restoration. Where thickets are present, incident light might be reduced to <2% full sun during the growing season due to R. maximum's thick canopy (Clinton 1995, Nilsen et al. 2001, Lei et al. 2006). Under such conditions, incident light is often below the compensation point for many deciduous species and net photosynthetic carbon gain may be insufficient to support seedling development (Lei et al. 2006). The impact of low light levels on seedling recruitment under R. maximum thickets is compounded by competition for water and nutrients (Horton et al. 2009). Within R. maximum thickets, available water, soil-extractable cation concentration, nitrogen mineralization rates, and pH can all be markedly lower compared to open understories (Nilsen et al. 2001). Differences in belowground resources may be explained by the poor quality (relatively low nutrient concentrations, high lignin concentration, and low decomposition rate) of R. maximum leaf litter and the development of a thick recalcitrant soil organic layer Hendrick 2007, Horton et al. 2009).
Similar to other ericaceous species, R. maximum has low nutrient requirements, yet a high nutrient retention capacity. Although R. maximum leaves represent <20% of their average total biomass, leaf longevity (6-8 yr) makes them an important nutrient pool (Monk et al. 1985). Most of these resources are resorbed following senescence, and only a small fraction remain in leaves prior to leaf fall. In addition, both the leaves and roots of R. maximum are rich in phenolic compounds capable of forming recalcitrant polyphenolic-organic complexes. These complexes reduce nitrogen mineralization, which in turn decreases inorganic nitrogen availability in soil beneath R. maximum (Wurzburger and Hendrick 2007, Horton et al. 2009). Because R. maximum frequently occurs in close association with T. canadensis trees, continued growth and expansion of R. maximum shrubs is expected as they replace declining T. canadensis stands (Roberts et al. 2009. Thus, at least the partial removal of R. maximum may be needed to promote recovery of ecosystem structure and function. In addition, successful restoration may require the replacement of plant communities that have been locally extirpated or are severely depressed. Under these conditions, the seed bank may represent a potential source of propagules for recruitment of some target species (Saatkamp et al. 2014) and therefore should be considered when planning and implementing restoration activities. For example, Hille Ris Lambers et al. (2005) found that more than 40% of the seeds of Liriodendron tulipifera and Betula spp. available for germination come from a persistent seed bank. If other plant life-forms (e.g., forbs, graminoids, and shrubs) are also found in the seed bank, this could provide propagules for a diverse plant community (Schuler et al. 2010, Small and McCarthy 2010, Saatkamp et al. 2014. A better understanding of seed bank characteristics and the various abiotic and biotic factors that affect seed bank density and composition may be useful in determining the restoration potential of forest communities following invasion-related disturbances (see review, Gioria and Py sek 2016). To investigate the contribution of the soil seed bank as a recruitment source, we compared the seed bank of two deciduous forest types: hardwood forests with a dense R. maximum subcanopy (hereafter, RR) and hardwood forests without R. maximum (hereafter, HWD). We hypothesized that (1) seed bank density would be lower in RR than HWD; (2) if a viable seed bank was present, then seed bank richness would be lower in RR than HWD; (3) light transmittance, soil moisture, mineral soil pH, and nutrient availability would be lower, and organic soil depth and mass would be greater in RR than HWD; and 4) variation in the seed bank characteristics would relate to variation in microenvironment and edaphic conditions across the two forest types. We also considered whether the seed bank could contribute to the restoration of plant communities in southern Appalachian riparian forests affected by adelgid-induced T. canadensis mortality.

Study sites
We conducted our study at two sites in the southern Appalachian Mountains of western North Carolina, USA: the Coweeta Basin (CWT; 35°03 0 N, 83°25 0 W) and White Oak Creek Basin (WOC; 35°15 0 N, 83°35 0 W). Mean annual temperature is 12.6°C at CWT and 10.8°C at WOC Knoepp 2005, Laseter et al. 2012). Mean annual precipitation is~200 cm at both sites, with little seasonal variation; however, dry years are increasingly common. Soils are deep sandy loams underlain by folded schist and gneiss (Thomas 1996). Vegetation is characterized as southern mixed deciduous forest with overstory codominance by Quercus, Acer, Carya, and Liriodendron species and understory dominance by Rhododendron maximum (Day et al. 1988). Forests at both sites were 70-90 yr old and were infested with hemlock woolly adelgid between 2003 and 2005. Substantial Tsuga canadensis decline (80% crown loss, 33% mortality) was documented by 2007 (Elliott and Vose 2011), and complete hemlock mortality was observed by 2012. At the time of this study (2014), overstory T. canadensis trees had been dead for at least two years.
We established 32 plots at CWT and 24 plots at WOC in mesic riparian areas with similar elevation range, slope, topography, tree species composition, and abundances of dead T. canadensis (>40% basal area prior to mortality). Plots were located at low-to-moderate slopes (<30%), and elevation ranges from 760 to 1060 m at CWT and 1160 to 1390 m at WOC. First, we installed and delineated the 20 9 20 m RR forest type plots with a dense RR, and then, we located nearby 10 9 10 m HWD forest type plots without a R. maximum subcanopy. We established HWD plots close to RR plots (within 100 m of at least one RR plot), so that we had equal numbers of RR and HWD plots and covered the same geographic and physiographic distribution of RR plots at each location. To avoid edge effects and any influence of R. maximum subcanopies on HWD plots, all plots were placed well within their respective forest type with at least a 15-m buffer on all sides. This sampling design resulted in a total of 56 plots (28 RR, 28 HWD). In the RR forest type, R. maximum mean density was 10,000 AE 918 stems/ha, and ranged from 2,300 to 21,000 stems/ha; and R. maximum mean basal area was 6.43 AE 0.50 m 2 /ha, and ranged from 2.84 to 11.93 m 2 /ha. Even though RR and HWD plots sizes were dissimilar, the sampling was conducted similarly across plots (i.e., soil seed bank, ground-layer vegetation, microclimate, and soils). We adjusted for plot size in the overstory layer calculations of density, basal area, and leaf area index (LAI), and understory density (see Vegetation sampling below).

Soil seed bank sampling
In late July 2014, we divided each plot into four equal quadrants and extracted a soil core from the center of each quadrant using a PVC collar (7.62 cm inside diameter 9 10 cm depth). To prevent premature germination and minimize disturbance, soil cores were left in collars and capped in order to safely transport to the climate-controlled greenhouse at University of Texas at San Antonio. We assessed the seed bank using the seedling emergence approach (Thompson and Grime 1979, Thompson 1987, Keyser et al. 2012, Maclean et al. 2018b. Soil samples were composited by plot and cold-wet stratified at 4°C for two months without light to break seed dormancy (Milberg andAndersson 1998, Baskin andBaskin 2001) and to assure the largest number of species break dormancy and germinate. Following stratification, we sorted samples by hand to remove pieces of root, rocks, and coarse woody debris and then spread each sample evenly over a 3-cm layer of potting soil and vermiculate (1:1) in a 52.7 9 32.7 9 8.0 cm germination tray. Trays were arranged randomly throughout the greenhouse, and 15 trays containing only potting soil and vermiculate were interspersed between sample trays to detect contamination from outside seed sources. All trays were monitored daily, watered as needed, and rearranged monthly for 11 months. Newly emerged seedlings were identified to species and counted, except for Rubus and Betula species, which were identified to genus. Following identification, we clipped seedlings at their base to prevent competition with seeds that had not yet emerged. Unidentified specimens were transferred to separate pots and grown until identification was possible. When no new germination was observed, we removed all remaining seedlings, mixed each soil sample by hand, and monitored emergence for one subsequent month. Other studies have monitored seed bank trays for a much shorter period (e.g., Keyser et al. 2012, Maclean et al. 2018b, and this extra step, mixing and observing for one or more months, is often not taken (e.g., Augusto et al. 2001, Small and McCarthy 2010, Keyser et al. 2012. In our study, seed bank samples from both forest types (RR, HWD) across sites (CWT, WOC) experienced the same greenhouse emergence method described above.

Vegetation sampling
To characterize the vegetation composition of HWD and RR forest types, we sampled the existing vegetation in each plot by layer: overstory, understory, and ground-layer. In the overstory layer, we measured all trees and shrubs ≥2.5 cm at diameter at breast height (DBH, 1.37 m above ground) to the nearest 0.1 cm. Leaf area index (m 2 projected leaf/m 2 ground area) was estimated using DBH and allometric equations developed for woody species in the southern Appalachians (Boring and Swank 1986, Martin et al. 1998, Elliott et al. 2002. In the understory layer, we counted all trees and shrubs (<2.5 cm DBH and ≥0.5 m height) in a 4.0 m wide belt nested within each plot. The ground-layer was sampled in mid-to-late June 2014, the time of peak biomass accumulation; however, some spring ephemerals could have been missed. We placed two 1.0 9 1.0 m quadrats in opposite plot corners and recorded the percent cover of all plants (woody stems <0.5 m height and all herbaceous plants) using a scale that emphasizes intermediate accuracy (Gauch 1982): in 1% intervals from 1% to 5%, in 5% intervals from 5% to 20%, and in 10% intervals above 20%. All species nomenclature follows Gleason and Cronquist (1991).

Microenvironment measurements
We measured soil water content (%), soil temperature (°C), organic soil (defined below) depth (cm), and photosynthetically active photon flux density (PPFD; lmol photonsÁm À2 Ás À1 ) at four equidistant points along a diagonal transect within each plot. Soil water content was measured at 20 cm depth using a handheld HydroSense Soil Water Measurement System (Campbell Scientific, Logan, Utah, USA), and soil temperature was measured at 10 cm depth with a Type T Thermocouple (Barnant Instruments, Barrington, Illinois, USA). Organic soil depth was measured to the nearest 0.1 cm. PPFD incident was measured with a portable light meter (Sunfleck Ceptometer, Decagon Devices, Pullman, Washington, USA) at 1.0 m above the forest floor and within AE 2 h of solar noon under clear, sunny conditions. Additional PPFD incident measurements were taken over each ground-layer quadrat to capture the light transmittance across the plot (n = 6 per plot). PPFD open was measured in open conditions within 30 min of the PPFD incident measurements. Light transmittance was calculated as PPFD incident Ä PPFD open and expressed as percent. Microenvironment measurements were taken three times in 2014 over the summer months (June, July, and August), and monthly values were averaged as a growing season estimate per plot.

Organic and mineral soil sampling and nutrient concentrations
We sampled organic and mineral soil because plant nutrients and viable seeds are found in both, and fine roots and mycorrhizae can acquire nutrients directly from organic layers, particularly the Oa (humus layer). We collected two organic soil samples from each plot using a 0.09m 2 sampling frame. Samples were separated into two organic soil layers: Oi (litter, where senesced leaves and twigs are deposited in the fall) and Oe + Oa (Oe = fermentation, where leaves have fractured and are partially decomposed; Oa = humus, dark, and decomposed, no longer recognizable as leaves or twigs). Each layer was placed in a paper bag, oven-dried at 60°C to a constant weight, and weighed. Samples were then composited by plot and layer, ground to <1 mm, and analyzed for total C, N, P, Ca, Mg, K, and Al concentrations. We determined total C and N by combustion on a Flash EA 1112 NC Elemental Analyzer (Thermo Scientific, Waltham, Massachusetts, USA) and total Ca, Mg, K, P, and Al by dry-ashing a subsample at 480°C, digesting it in HNO 3 acid, and analyzing it on an inductively coupled plasma spectrophotometer (Horiba, Edison, New Jersey, USA; Brown et al. 2015).
We collected mineral soil samples in each plot to a depth of 10 cm using an Oakfield soil probe. Each sample was a composite of 15-20 individual samples distributed systematically across the plot to provide a representative plot sample. All soil samples were air-dried and sieved to <2 mm before analysis. We determined total soil C and N by combustion as above, and soil pH in a 1:1 soil to 0.01 mol/L CaCl 2 slurry using an Orion portable pH meter (model 250A) with a Thermo Scientific Orion pH probe (Brown et al. 2015).

Statistical analyses
We conducted all statistical analyses using SAS computer software (v9.4; SAS Institute, Cary, North Carolina, USA). We used mixed linear models (PROC MIXED) to evaluate the main effects of forest type (RR, HWD) and site (CWT, WOC) and forest type 9 site interaction on seed bank and environmental variables. If overall Ftests for the interaction effect were significant (P ≤ 0.05), we used least square means (LS means, Tukey-Kramer-adjusted t-statistic) tests to evaluate significance. Degrees of freedom were approximated using Satterthwaite's formula (Littell et al. 2004).
We used principal component analysis (PCA, PROC FACTOR) to reduce the dimensionality of environmental variables across forest types (RR, HWD) and sites (CWT, WOC) into a set of uncorrelated principal components (Graham 2003) and to examine the combined influence of multiple environmental variables on seed bank density and richness. All environmental variables (vegetation, microenvironment, and soils) were evaluated in the PCA. Principal components with eigenvalues >1.0 were retained for further analysis (Kaiser 1960), and environmental variables with loadings >|0.50| were considered significant (Tabachnick and Fidell 2001) and used to define components. Extracted principal component axes scores were then related to seed bank variables using Spearman's rank correlation analysis (Wagner 2013, Chatterjee and Hadi 2015).

Soil seed bank characteristics
Total seed bank density was not different between forest types or sites; however, the composition of the seed bank differed (Table 1). Herbaceous and graminoid seed densities were consistently lower in the RR than HWD forest type. In contrast, tree seed density was consistently greater in RR than HWD at both sites, and shrub seed densities were significantly greater in RR than HWD, but only at WOC (Tables 1, 2). A significant interaction effect (Table 1) revealed that shrub seed density was similar at CWT between forest types (t 1,35.2 = À0.62, P = 0.924); however, at WOC, shrub seed density was much greater in RR than HWD (t 1,35.2 = À3.51, P = 0.007; Table 2). Rhododendron maximum was a common associated species in the RR seed bank, accounting for 37%, but was absent in the HWD seed bank (Appendix S1: Table S1). Rubus sp. was the most abundant shrub species across sites, accounting for 71% and 89% of the shrub seed bank density in the RR and HWD, respectively (Appendix S1: Table S1). In the RR forest type, total seed bank density and plant life-form were not significantly related to R. maximum density (Appendix S1: Fig. S1). Only tree seed density was positively related to R. maximum ❖ www.esajournals.org , basal area, and density; understory density and richness; and ground-layer cover and richness); microenvironment (light transmittance, soil temperature, and soil water content); organic soil (Oi + Oe + Oa depth and mass, Oi N, C, K, Ca, Mg, P, and Al concentrations, and Oe + Oa N, C, K, Ca, Mg, P, and Al concentrations); and mineral soil (N and C concentrations and pH). Values in bold type indicate a significant forest type, site, or forest type 9 site interaction effect. Numerator degrees of freedom = 1 and denominator degrees of freedom (df) are provided in the table. , basal area, and density; understory density and richness; ground-layer cover and richness); microenvironment (light transmittance, soil temperature, and soil water content); organic soil (Oi + Oe + Oa depth and mass, and Oi N, C, K, Ca, Mg, P, and Al concentrations; Oe + Oa N, C, K, Ca, Mg, P, and Al concentrations); and mineral soil (N and C concentrations, and pH) in two deciduous forest types (RR, hardwood forest with a dense Rhododendron maximum subcanopy; and HWD, hardwood forest without R. maximum) within the Coweeta Basin (CWT) and White Oak Creek (WOC) sites. basal area (R = 0.44, P = 0.019, n = 28), whereas total seed bank and other plant life-forms were not related to R. maximum basal area (Appendix S1: Fig. S2).
Total seed bank richness was lower under RR than HWD at both sites (Tables 1, 2). The seed bank contained a much smaller proportion of herbaceous species (forbs and graminoids) and a greater proportion of woody species (trees and shrubs) in RR than HWD (Fig. 1). Herbaceous seed bank richness was significantly lower in the RR than HWD, but there was no difference between sites (Tables 1, 2). There was a greater number of graminoids at WOC than CWT, but no difference between forest types (Tables 1, 2). Common herbaceous species found in the HWD seed bank included Lobelia inflata, Viola blanda, Oxalis stricta, and Ageratina altissima; and Cardamine hirsuta and Oxalis stricta in the RR seed bank (Appendix S1: Table S1). Common graminoid species in both the RR and HWD seed banks included Juncus tenuis, Danthonia compressa, and Cyperus strigosus (Appendix S1: Table S1). We found a total of seven tree species in the seed bank (Appendix S1: Table S1). There was no significant difference in tree seed richness between RR and HWD (Table 1); however, tree seed richness in HWD was greater at CWT than WOC (t 1,39.2 = 3.34, P = 0.010), whereas there was no difference in tree seed richness in RR between sites (t 1,39.2 = 1.44, P = 0.481; Table 2). Betula spp. was the most abundant tree species in the soil seed bank across forest types and sites; however, its seed numbers were much greater in RR than HWD. Associated tree species included Liriodendron tulipifera, Oxydendrum arboreum, and Robinia pseudoacacia, which were found in both forest types (Appendix S1: Table S1).

Environmental variables: Vegetation, microenvironment, and soils
Overstory basal area was similar across forest types and sites; however, overstory LAI was lower under RR than HWD (Tables 1, 2). Overstory density was greater in RR than HWD, due to the high numbers of R. maximum stems >2.5 cm dbh (Appendix S1: Table S2), and it was greater at WOC than CWT (Tables 1, 2). Acer rubrum was the dominant overstory species in HWD at WOC and was codominant with L. tulipifera and Betula lenta in HWD at CWT. Associated tree species at both sites were Fagus grandifolia, Quercus rubra, Quercus montana, and members of the genus Carya. Many of the tree species present in HWD occurred in RR, albeit in lower densities (Appendix S1: Table S2).
Understory density and richness were consistently lower under RR than HWD (Tables 1, 2). The understory layer in the HWD forest type at both sites contained high densities of Gaylussacia ursina and the woody vine, Smilax rotundifolia. Other common understory species included Tsuga canadensis, A. rubrum, and Quercus coccinea at CWT; and F. grandifolia, Hamamelis virginiana, Pyrularia pubera, Halesia carolina, and Rubus sp. at WOC (Appendix S1: Table S3). Rhododendron maximum was the most abundant understory species in the RR forest type, accounting for approximately 80% of the total stem density at both sites. All other understory species were in low numbers in the RR forest type (Appendix S1: Table S3). In the RR forest type, density of understory species other than R. maximum was not significantly related to R. maximum density (R = 0.012, P = 0.952, n = 28) or R. maximum basal area (R = 0.089, P = 0.651, n = 28).
Both ground-layer cover and richness were consistently lower under RR than HWD (Tables 1, 2). The most abundant species in the ground-layer in RR were Galax aphylla, G. ursina, and R. maximum; abundant species in HWD included A. rubrum, Aster divaricatus, and Thelypteris noveboracensis (Appendix S1: Table S4). In the RR forest type, similar to the understory layer, R. maximum basal area was not related to ground-layer cover (R = À0.110, P = 0.577, n = 28) or richness (R = À0.067, P = 0.734, n = 28), likely because we chose RR plots with a dense R. maximum subcanopy.
Soil moisture, pH, and nutrient availability were lower, and organic soil depth and mass were greater in the RR than HWD forest type. Light transmittance below the canopy was low and similar under RR and HWD forest types (RR, 3.26 AE 0.65%; HWD, 4.12 AE 0.94%). Soil water content and soil temperature were lower in RR than HWD across sites. There was no difference in soil water content between sites, but soil temperature was significantly higher at CWT than WOC (Tables 1,  2). Organic soil mass and nutrient concentrations varied considerably by forest type and between sites (Table 2). Organic soil depth was greater in the RR than HWD forest type at both sites; and organic soil depth in RR was greater at WOC than CWT (t 1,42.2 = 4.35, P < 0.001). Total organic soil mass (Oi + Oe + Oa layer) was also consistently greater in RR than HWD (Tables 1, 2).
Oi N did not differ between forest types, whereas Oi C was consistently greater in RR than HWD (Tables 1, 2). Oi K was consistently lower in RR than HWD, and it was greater at CWT than WOC. Oi Ca and Oi Mg were not different between forest types; however, Oi Mg was greater at CWT than WOC in HWD (t 1,35.9 = 4.80, P = 0.003), while there was no difference between sites in RR (t 1,35.9 = À0.890, P = 0.999). Oi P was similar between forest types and sites; and Oi Al was greater in RR than HWD (Tables 1, 2).
Mineral soil pH was consistently lower in RR than HWD (Tables 1, 2). In HWD, soil pH was higher at WOC than CWT (t 1,12.1 = 5.86, P < 0.001); in RR, soil pH was similar between sites (t 1,10.8 = À1.92, P = 0.270). Mineral soil N and C were higher at WOC than CWT (Table 2), but there were no differences between forest types (Table 1).

Relationships among environmental variables and seed bank characteristics
To combine environmental variables and reduce dimensionality, we used PCA and then correlated the PCA axes scores with seed bank density and richness. The first four PCA axes explained a large proportion (64%) of the variation in environmental variables across forest types and sites (Table 3), compared to others using PCA and environmental data (e.g., Eide et al. 2017Eide et al. , L evesque et al. 2017. PCA4 was negatively correlated with total seed bank density (Fig. 2). PCA1 and PCA4 were positively correlated with total seed bank richness (Fig. 2). PCA1 was positively correlated with herbaceous seed density and richness, and graminoid seed density and richness; and negatively correlated with tree seed bank density (Appendix S1: Table S5). PCA4 was negatively correlated with herbaceous seed bank density and richness (Appendix S1: Table S5). Positive loadings for PCA1 were Oe + Oa cations (K + Ca + Mg), Oi cations, Oi P, understory richness, groundlayer cover and richness, and soil pH; and negative loadings were organic soil depth and mineral soil C:N ratio. Positive loadings for PCA4 were understory density and Oe + Oa P, and a negative loading was Oe + Oa C:N ratio (Table 3).

Effects of forest type on seed bank characteristics
Despite having similar seed bank densities, RR and HWD shared few other seed bank characteristics. Those species only present in the seed bank of HWD were predominantly graminoids and perennial forbs that are dispersed over short distances, and these species contribute to a diverse understory and typically do not impede tree regeneration (Elliott et al. 2014(Elliott et al. , 2015. By contrast, species only found in the seed bank of the RR forest type shared few distinguishing Fig. 2. Spearman's rank-order correlation analysis of total seed bank density (seeds/m 2 ) and richness (species/plot) with principal component axes (PCA1, PCA4) generated from environmental variables across forest types (RR, deciduous forests with a dense Rhododendron maximum subcanopy; and HWD, deciduous forest without R. maximum) and sites (CWT, Coweeta Basin; and WOC, White Oak Creek). Significant loadings describing PCA1 were understory richness, ground-layer cover and richness, organic soil depth, Oi cations and P, Oe + Oa cations, and mineral soil C: N and pH. Significant loadings describing PCA4 were understory density and Oe + Oa P (see Table 3). PCA2 and PCA3 were not significantly related to total seed bank density and richness. Notes: LAI, leaf area index. RR is deciduous forest with a dense Rhododendron maximum subcanopy; and HWD is without R. maximum, and sites (CWT, Coweeta Basin; and WOC, White Oak Creek). Variables with significant loadings (≥| 0.50|) are set in bold type.
characteristics. Four of these species produce seeds that are self-dispersed, one produces winddispersed seeds and another produces animaldispersed seeds. None of these species were found in the standing vegetation of the RR forest type. Given their degree of dissimilarity in mode of dispersal, it is unclear why these species were found in the seed bank of RR but not HWD. The same cannot be said for Rhododendron maximum seeds, however, which were found in high densities in the seed bank of RR, but were entirely absent from the seed bank of HWD. Seeds from R. maximum were also reported in the seed banks of temperate forests described by Hille Ris Lambers et al. (2005), although no claim was made regarding the persistence of R. maximum seeds in the seed bank, given that seeds were too small to be collected in seed traps, and the relationship between yearly seed rain and seed bank density was unknown.
Seed bank densities of several woody species were considerably higher in RR than HWD, even of species that were represented in both the RR and HWD forest types. A comparison of these results with other studies of temperate forest seed banks in this region reveals that the seed densities of woody taxa in the RR forest type exceeded densities that had been previously reported for many species (Lei et al. 2002, Hille Ris Lambers et al. 2005, Keyser et al. 2012. For example, the seed bank density of Betula spp. in the RR forest type was nearly two times greater than in comparable temperate forests in the Coweeta Basin described by Hille Ris Lambers et al. (2005) and Lei et al. (2002). Likewise, seed bank densities of Rubus sp. and Liriodendron tulipifera in the RR forest type exceeded published values by threefold (Lei et al. 2002, Hille Ris Lambers et al. 2005. By contrast, seed bank density of tree species in the HWD forest type was within the range reported by Hille Ris Lambers et al. (2005) and Lei et al. (2002).

Seed bank relationships with environmental variables
Variation in seed bank richness and in plant life-form-specific seed bank density between forest types may be partially explained by differences in environmental conditions. For instance, mineral soil pH, soil nutrient availability, and soil moisture were lower, and organic soil depth and mass were greater in the RR than HWD forest type. These microclimate and edaphic conditions can affect the seed bank directly, by influencing seed losses from the seed bank through germination and decay, and indirectly, by influencing seed inputs to the seed bank from the standing vegetation.
Our research supports a general pattern of increasing seed bank richness with increasing soil fertility and decreasing soil acidity (Staaf et al. 1987, Thompson 1987, Leckie et al. 2000, Maclean et al. 2018b. Principal component analysis correlations revealed that multiple environmental variables influenced seed bank density and richness. PCA4 (represented by Oe + Oa phosphorus and carbon/nitrogen ratio, and understory richness) was negatively correlated with total seed bank density. PCA1 (represented by Oe + Oa cations and phosphorus, mineral soil pH, understory richness, and ground-layer cover) and PCA4 were positively correlated with total seed bank richness.
Lower soil nutrient availability and greater soil acidity, such as in RR, may restrict the reproductive capacity of some species in the standing vegetation, potentially reducing the quantity and richness of seed inputs from local seed rain. Lei et al. (2002) reported no significant effect of R. maximum subcanopies on the quantity of seed rain from several tree species (Acer rubrum, L. tulipifera, Betula lenta, and Quercus spp.), but because some of these species are winddispersed, it was unclear whether seeds collected in seed traps were produced locally by maternal trees within the R. maximum thickets or were introduced from nearby forest stands. Many tree species have long-distance dispersal (>100 m), particularly those that are wind dispersed (Clark et al. 1999), whereas most forest herbs have short-distance or local seed dispersal (Bakker et al. 1996). In our study, both the HWD and RR forest types had an intact overstory tree canopy that could provide a local seed source, and longdistance dispersal is possible for many tree species. We did not quantify seed rain and do not know the origin of seeds (i.e., local or longdistance dispersal); however, we did find greater seed numbers of L. tulipifera and Betula spp. in the seed bank of RR than HWD (Appendix S1: Table S1).
❖ www.esajournals.org Increasing soil nutrients and decreasing soil acidity, as shown in the PCA, were positively correlated with seed bank richness and negatively correlated with tree seed density (Appendix S1: Table S5). We also found that tree seed density was positively related to R. maximum basal area (Appendix S1: Fig. S2). While it is unclear whether soil fertility has a direct effect on seed viability and germination (Bekker et al. 1998), lower nutrient availability and greater soil acidity could inhibit microbial activity, slowing decomposition and increasing seed longevity for some species (Champness and Morris 1948, Leck et al. 1989). Hence, while edaphic factors may limit local seed production from some species in the standing vegetation, they could also preserve seed longevity for species that are capable of dispersing into the RR forest type. Taken together, these results may explain the scarcity of herbaceous seeds and the abundance of tree seeds in the seed bank of the RR forest type, as the latter are more likely to be dispersed over greater distances and seed longevity, particularly Betula spp., is much greater than the former.
Leaf litter input in the RR forest type may also influence the reproductive capacity of seeds in the seed bank. Rhododendron maximum can produce as much as 125 kg/ha of leaf litter per year (Monk et al. 1985) resulting in the formation of thick recalcitrant organic soil layer. Under these circumstances, the steady build-up of organic soil beneath R. maximum, as seen in our study, could bury seeds too deeply for germination.

Implications for restoration
Restoration of southern Appalachian riparian forests affected by Tsuga canadensis mortality may involve at least the partial removal of R. maximum in order to promote recovery of ecosystem structure and function. Because there is often little to no herbaceous and tree seedling cover beneath R. maximum (Clinton andBoring 1994, Beckage et al. 2000), successful restoration will require the replacement of plant communities that have been locally extirpated or are severely depressed. Under these conditions, the seed bank may represent a potential source of propagules for recruitment of some target species, and therefore should be considered when planning and implementing restoration activities.
Our results indicate that seed bank communities under R. maximum are dominated by the tree and woody shrub life-forms, and by a low number of species. Although seed banks in the RR forest type contained a high density of potentially desirable tree species (e.g., A. rubrum, L. tulipifera, and Betula spp.), other common woody shrubs, such as Rubus sp. and R. maximum, in the seed bank could suppress the regeneration of some target species and reduce the vigor of others (Meilleur et al. 1994, Royo andCarson 2006). Competition from Rubus, however, may be short term due to its high-light requirements (Elliott et al. 2002, Elliott andKnoepp 2005). Rhododendron maximum seeds require high light for germination (Blazich et al. 1991) but are shade-tolerant and therefore will presumably germinate in mass if the canopy or subcanopy is removed as long as the substrate remains conducive for seed germination (e.g., moist soils, moss cover). Consequently, if the subcanopy of R. maximum is removed from riparian forest communities, a R. maximum seed bank could ensure its recolonization, particularly in the presence of ericoid mycorrhizae (Wurzburger and Hendrick 2009), without continued manipulation or active management (e.g., repeated fires). In contrast to the high numbers of tree and woody shrub seeds (85%) in the RR forest type, relatively fewer herbaceous species (14%) were found. Thus, the probability of a diverse herbaceous layer recruiting from the seed bank will likely be low, which could have long-term effects on nutrient cycling and overall forest plant diversity. For example, herbaceous foliage is substantially more nutrient dense than that in trees, can contribute~20% of the forest foliar litter and 70-90% of the forest plant species diversity (Muller 1978, Gilliam 2007, Welch et al. 2007).
Our results suggest that the soil seed bank may not be the primary mode of recruitment to establish a diverse herbaceous community even if R. maximum is removed from these forests, but it could likely facilitate forest canopy recruitment. The high proportion of aggressive ruderal species (i.e., Rubus sp.) in the seed bank could limit the growth and regeneration of some target tree species, however, if R. maximum is removed without sustained management. Lessons from studies comparing the effects of Rhododendron ponticum (evergreen similar to R. maximum), a non-native invasive in Scotland, on native community species may be relevant to seed bank implications here. Similar to our system, as R. ponticum density increases, native community species decline (Maclean et al. 2018b), and where R. ponticum had once been, even after clearing, the native community did not return even after 30 yr. Thus, after an initial R. maximum clearing, successive restoration efforts such as prescribed fire, herbicide, and cutting may be required to remove R. maximum; and soil amendments and seed introduction (Maclean et al. 2018a, b) may improve ecosystem function and promote diversity.

ACKNOWLEDGMENTS
We thank Patsy Clinton, Guillermo Silva, and Frank Llarena for assistance in field sampling. Drs. Beverly Collins and Michelle Baumflek and three anonymous reviewers provided helpful comments on the manuscript. This research was supported by Coweeta Hydrologic Laboratory, Southern Research Station, USDA Forest Service; the Coweeta LTER project funded by National Science Foundation grants DEB-1637522, DEB-1440485; and the College of Sciences, University of Texas at San Antonio. The use of trade or firm names in this publication is for reader information and does not imply endorsement by the U.S. Department of Agriculture of any product or service.