Supporting information for:
Consumer preferences for seeds and seedlings of rare species impacts tree diversity at multiple scales
Hillary S. Young, Douglas J. McCauley, Roger Guevara, Rodolfo Dirzo

This file includes:

SI 1.Seed and seedling densities

SI 2.Identity of seed and seedling predators

SI 3.Landscape-based clustering index

SI 4.Tree dispersal and seedling mortality

SI 5.Duration of simulation runs

SI References

SI Table 1: Seed dimensions

SI Figure 1:Patchy distributions of forest types across Palmyra

SI Figure 2:Temporal patterns of diversity

SI Figure 3: Images of natural abundance seedling densities

SI Figure 4: Results from longer runs of simulations

Supplementary Information 1:

Seed and seedling densities

The densities of seeds and seedlings used in this study (seeds - 2.5/m2P. fischerianus, 3/m2C. nucifera, 5/m2 for other species; seedlings – no greater than 0.01/m2) are within natural ranges of seed and seedling density for these species in this system.Individual P. fischerianus fruits can have up to 200 “seeds” (truly phalanges with 1-15 seeds per phalange)); when the fruit ripens and falls, the seeds are generally scattered all at once resulting in highly localized seed patches.In vegetation transects surveyed in 2008-2010 (ranging from 100 m2 to 300 m2) which counted only fresh seeds and live seedlings weobserved P. fischerianus seed densities upwards of 6/m2 and seedling densities upwards of 2/m2(Young et al 2011, Young et al unpublished data). For C. nucifera, an individual palm produces 50-80 fruits per year (Chan & Elevitch 2006), all of which fall directly beneath the canopy; in our vegetation transects this resulted in observed seed densities (mature, intact, unsprouted seeds only) exceeding 4/m2 and seedling densities exceeding 0.5/m2.Presumably due to the very high observed seed predation rates described in this study, there is virtually no standing seedbank of seeds of small seeded species (T. argentea,S. sericea, and P. grandis); however, all of these species can produce large, and pulsed amounts of seeds (easily exceeding 1000 seeds/tree at given time periods, Burger 2005, Young et al unpublished data), such that seed densities of these three species used in this study also fall within normal ranges. Seedling density of these species is also typically fairly low (ranging up to 0.03/m2 in an average year, which still exceeds densities used in this experiment), but seedling carpets of these species do form occasionally in gaps or after masting events (SI Fig 3).

Supplementary Information 2:

Identity of seed and seedling predators

In addition to the five common species of land crabs, non-native rats (Rattus rattus) are also abundant at Palmyra, and are known seed and seedling predators (Wegmann 2009). To determine which seed and seedling predators were dominant in this system we used time triggered and heat sensing cameras (Reconyx Inc.). Cameras were trained on seeds of four tree species (C. nucifera, P. grandis, S. sericea, T. argentea) at eight sites (four each in Cocos forests and dicot forests) and seedlings of two species (C. nuciferaand P. grandis, n = 15 of each) in each surveyed forest type. Cameras were run for 24 hours in the seed experiment and 5 days in the seedling experiment. Additional data on predation of the larger C. nucifera seeds were gathered by examination of the diagnostic damage rendered to these seeds.Only Birguslatroand R. rattus can open intact (unhusked) nuts, and both have been observed to do so on site by investigators.

Identity of seed and seedling predators from camera monitoring showed that land crabs were particularly important predators on the smaller seeded dicot species (e.g. P. grandis, S. sericea, and T. argentea). Of 200 seed predation events on various dicot species, we observed 192 predation events by crab Coenobitaspp., 5 by crab Cardisomaspp., and 3 by Rattus rattus. Of 28 observed predation events of P. grandisseedlings, 24 were by Coenobitaspp., 3 by Cardisomaspp., and 1 by R. rattus.

Based on scars on the husk remains, the coconut crab B. latro appeared to be the dominant seed predator on mature C. nucifera seeds, accounting for 100% (n = 26) of the observed seed predation events for this species (as evidenced in distinct husking pattern of residual nuts). B. latroalso caused all 6 predation events on C. nuciferaseedlings. Rats appeared to cause the vast majority of non-lethal herbivory on C. nuciferaseedlings; however this was not quantified as it was often not possible to conclusively identify the herbivore for this type of herbivory.

All species of crabs and R. rattus were present at all sites.However, we did not quantify changes in consumer community composition or consumer abundance across sites. While changes in abundance of all seed and seedling predators among forest types should not result in patterns observed (as that would likely affect all species equally), it is possible that changes in species composition of seed and seedling predators across forest types (rather than selective predation of rare seeds and seedlings), could account for some of variation in predation pressure across sites. Nevertheless the selective increase of animals in sites where favored food sources are absent, seems counter-intuitive and unlikely.By contrast, a preference for rare food items within a species is both physiologically logical (as a strategy to gain rare nutrients) and consistent with other observations (Allen and Anderson 1984, Smith 1987, Thacker 1996). Moreover, the net ecological effect of the herbivore community on seed and seedling survival (selection for rare species) remains the same, regardless of whether changing predation rates are due to within species preferences or changes in species assemblage.

Supplementary Information 3:

Landscape-based clustering index

Given the limitations of traditional metrics of landscape diversity (i.e. beta diversity) in extremely low diversity landscapes, we also examined coefficients of variation of SDI as a metric of landscape diversity. However this metric actually measures species turnover and at extremes (complete monodominance and perfectly overdispersed distributions) these values would converge. Thus to complement these metrics we also developed a new measurement of clustering: a "landscape based clustering index". In this metric we evaluated the degree of clustering across the landscape by comparing the average number of species empirically observed in a small plot as compared to the expected number observed if all species were distributed randomly across a landscape, given the actual observed abundance of each species. The clustering metric was calculated based on the following equation:

NP = Total number of plots

Nspi = Number of species in plot i

N = Total number of species in the landscape

Pij = Proportion of species j in plot i

Pj = Proportion of species j in the landscape relative only to the species in plot i

Based on this metric, landscapes could range from -1 (maximally overdispersed) to 1 (maximally clustered), with 0 being equal to a truly random distribution. This metric was calculated both based on the field data on forest distribution and for the computer simulation results (under both simulation scenarios).

Supplementary Information 4

Tree dispersal and seedling mortality

To provide a rough approximation of initial dispersal ability of the dominant trees to the system for the simulation, initial abundance of each tree species was set to be inversely proportional to fresh seed mass (measured in the field; Table SI 1). Thus the initial probabilities of arrival on the landscape in a given unit for each species were as follows: C. nucifera,0.000001; P. fischerianus,0.00001;P. grandisand T. argentea, 0.435; and S. sericea 0.130. To calculate mortality probabilities of seedlings that established in the simulation, we used simple regression models based on relative abundance of conspecifics among the nearest 100 square units. To simulate preference for rare species by consumers, we had one negative regression line (of mortality to relative dominance in the canopy) that corresponded to mortality of the locally dominant species and three positive regression lines, one for each of the other three species in the simulation. Three points were defined in these regressions for each species. The intercept of each regression line was the overall average rate of seed and seedling predation of each species in all forest types excluding monodominant stands (i.e., when relative dominance equals 1). Then we estimated the mortality penalties for each species at local dominance of 0.5 as the overall mean predation across all forest types including pure stands. Finally mortalities at local dominance of one species (pure stand) were those estimated from seed and seedling predation experiments. Field mortality probabilities were calculated as: 1 - (seed survivorship * seedling survivorship).

Based on preliminary seedling predation trials (Wegmann 2009), we assume seedling mortality for P. fischerianusto be very low; we set it equal to lowest observed seedling mortality (15%) in conspecific forests. To be conservative, we assumed only a 5% increase in mortality in forests where it is rare (this increase in mortality by forest type is less than half of the rates seen for other species). Results from model simulations which suggest a somewhat more important role for P. fischerianusthan actually observed in the landscape are likely due to the conservative nature of these estimates. Based on preliminary data of seedling predation at this site, seedling mortality of S. sericeawas assumed to be intermediate to seedling mortality of the other dicots, T. argenteaand P. grandis (Wegmann 2009). Since seedling mortality rates for locally rare species were not directly assessed in Pandanus forest, species-specific mortality rates were assumed to be the same here as in other sites where that species was locally rare.

The species utilized in this study likely utilize multiple and various dispersal strategies (e.g. Burger 2005, Chan andElevitch 2006) and average dispersal distances in this system is not known.The simulation data presented in this paper is run using the simplistic assumption that dispersal ability is inversely proportional to seed mass (Greene and Johnson 1993).However, we also ran the simulations with two alternative parameterizations on dispersal ability.We first examined results on diversity assuming equal dispersal ability among all tree species.Second, we also adjusted the total dispersal distance per species from very low (strong dispersal limitation) to large (little dispersal limitation), while still keeping the difference among species proportional to seed mass.In all these simulations a similar patchy landscape structure (with low local species diversity and high species turnover) emerged, similar to that reported in this paper, however when dispersal limitation is weak, the patchy structure takes longer to appear but ultimately remains more stable over time.In strongly dispersal limited situations, some species quickly become rare or even disappear from the landscape.These alternative simulations suggest that the overarching patterns of effects of negative density dependent predation across spatial scales on diversity are not likely to be changed by variation in dispersal limitation, but the speed at which such patterns appear and the time period at which they persist across the landscape may vary considerably.

Supplementary Information 5

Duration of simulation runs

All analysis of diversity from simulations reported in this paper was performed at that point when the relative abundance of the most dominant species in the density-dependent model approximated that actually observed in the field (6000 generations).However, since species abundances were not stable at this point, we also examined the effects on local and landscape level diversity over a much longer simulation run (25000 generations).We found that increasing the length of simulation run revealed overall similar patterns of diversity as seen in shorter runs.(SI Fig 4)However, species that were, on average, more preferred by consumers (T. argenta, S. sericea, and P. grandis) became less common or actually dropped out of simulation (particularly for S. sericea) as less preferred species were able to encroach on stands of these more preferred species.This long-term loss of preferred species appears robust to variation in initial abundance of preferred species, although it changes the time period over which a species disappears.For example, even when we dramatically increased initial relative abundance of S. sericeato 90% of the initial abundance, it disappears after 20,000 generations.

Supplementary Information References

Wegmann AS (2009) Limitations to tree seedling recruitment at Palmyra Atoll.PhD dissertation University of Hawai'i, Honolulu

Greene DF , Johnson EA(1993) Seed mass and dispersal capacity in wind-dispersed diaspores.Oikos. 67: 69-74.

Chan E, Elevitch CR (2006) Cocos nucifera(coconut)ver 2.1.In Species profiles for Pacific Island agroforestry.Levitch, C.R. (Ed.). Permanent Agriculture Resources (PAR), Holualoa, Hawaii, USA.

Burger AE (2005) Dispersal and germination of seeds of Pisoniagrandis, an Indo-Pacific tropical tree associated with insular seabird colonies. J Trop Ecol21:263-271

Allen JA, AndersonKP(1984) Selection by passerine birds is anti-apostatic at high prey densities. Biol J Linn Soc 23:237-246

Smith TJ (1987) Seed predation in relation to tree dominance and distribution in mangrove forests. Ecology 68:266-273

Thacker RW (1996) Food choices of land hermit crabs (Coenobitacompressus H Milne Edwards) depend on past experience. J Exp Mar Biol Ecol199:179-191

Supplementary Table 1
Seed Dimensions

Seed length (longest dimension), mass (g), and volume for all species of plants used in this analyses.For each dimension seeds that share the same letter are not significantly different in size from one another based on ANOVA.

Species / n / seed length (cm) / seed mass (g) / seed volume (cm3)
C. nucifera / 100 / 26.2 ± 2.5(A) / 1637.8 ± 373.3 (A) / 2766.7 ± 538.7 (A)
P. grandis / 100 / 1.2 ± 0.1(B) / 0.03 ± 0.01 (B) / 0.2 ± 0.1(B)
S. sericea / 100 / 0.9 ± 0.1(BC) / 0.1 ± 0.01 (B) / 0.1 ± 0.2 (B)
T. argentea / 100 / 0.5± 0.1(C) / 0.03 ± 0.01 (B) / 0.2 ± 0.03 (B)
P. fischerianus / 246 / 6.5 ± 0.6 (D) / 105.2±30.8 (C) / 292.5 ± 40.0 (C)

Supplementary Figure 1

SI Fig 1: At Palmyra atoll consumer preference for rare species keeps tree diversity low at small-scales, but facilitates high landscape-level heterogeneity at large-scales. Panels (a) and (b) show smoothed histograms (shades of grey indicate integrated histograms from multiple bin widths) of relative dominance of common tree species, including C. nucifera(a) and dicots P. grandis, T. argentea, and S. sericea(b), recorded in 100 m2 transect surveys. The skewed dominance distributions of these species at this scale of sampling reveals low levels of diversity at small-scales, and are consistent with a patchy landscape (low local diversity, high landscape heterogeneity). Panel (c) shows an example from Palmyra of how these small-scale patches of low diversity forest assemble to form highly heterogeneous landscapes.

Supplementary Figure 2

SI Fig 2: Temporal patterns of diversity in the preference for rarity predation simulation and null model simulation. a) Snap shot views of the spatial arrangement of species distribution after 0, 6000 and 12000 iterations for both models. b) Changes in relative abundance over 12000 iterations for five most common species show very different patterns under the two model simulations.

Supplementary Figure 3

SI Fig 3: Seed and seedling densities used in this study were well within range of natural variability.Even for small seeded dicot species where seed and seedling densities are typically low, high densities of seeds and seedlingscan occur periodically.Below, a P. grandisseedling carpet observed in 2012 (a, photo courtesy of A. Miller-terKuile) and a typical group of C. nuciferasprouts (b).

Supplementary Figure 4:

SI Fig 4: Longer runs of simulations (25,000 generations) in both null model and preference for rarity simulations continue to show similar patterns of diversity (a), although species preferred by consumers (i.e. S. sericea) drop out entirely by the end of the run.The preference for rarity simulations continue to show much higher than expected landscape level diversity (as shown in coefficient of variation of Shannon diversity index; b) and much higher clustering levels than under null model (c).

1