Chasing the unknown: predicting seed dispersal mechanisms from plant traits


  • Fiona J. Thomson,

    Corresponding author
    1. Australian Wetlands and Rivers Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
    2. Evolution & Ecology Research Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
      Correspondence author. E-mail:
    Search for more papers by this author
  • Angela T. Moles,

    1. Evolution & Ecology Research Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
    Search for more papers by this author
  • Tony D. Auld,

    1. Department of Environment, Climate Change and Water, Sydney, NSW 2220, Australia
    Search for more papers by this author
  • Daniel Ramp,

    1. Australian Wetlands and Rivers Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
    Search for more papers by this author
  • Shiquan Ren,

    1. Australian Wetlands and Rivers Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
    Search for more papers by this author
  • Richard T. Kingsford

    1. Australian Wetlands and Rivers Centre, School of Biological, Earth and Environmental Sciences, University of New South Wales, Sydney, NSW 2052, Australia
    Search for more papers by this author

Correspondence author. E-mail:


1. The dispersal capabilities of most plant species remain unknown. However, gaining basic dispersal information is a critical step for understanding species’ geographical distributions and for predicting the likely impacts of future climate change. Dispersal mechanisms can indicate short- or long-distance dispersers, and highlight important biological interactions.

2. To predict dispersal mechanisms for species where information is limited, we used generalized linear mixed models with basic life-history and ecological traits. Sets of models were created (using Australian species) for six dispersal categories: wind, unassisted, water, ant, vertebrate-ingestion and vertebrate-attachment dispersal mechanisms. We validated our models on the dispersal mechanisms of 50 Australian, 30 Californian, 30 Swiss plant species and a global compilation of 70 species.

3. Growth form, seed mass and vegetation type were the main predictor variables. Our models predicted dispersal mechanisms for Australian and Californian plant species equally well (c. 70% correct) and to a lesser extent for the Swiss flora (c. 50% correct). Our models predicted observed dispersal mechanisms (c. 50% correct) equally well to inferred dispersal mechanisms (based on seed morphology).

4.Synthesis. Our approach of using easily obtainable traits for predicting dispersal mechanisms of species allows dispersal information to be predicted for species where little is known. From here, the application of realistic dispersal curves to the predicted dispersal mechanisms will further understanding on the dispersal capabilities of species.


Dispersal is a key stage in the life cycle of plants, allowing offspring to move away from the parent plant and establish in new areas. Local and long-distance dispersal are necessary for the persistence and migration of species, and are linked to maintenance of plant diversity (Ozinga et al. 2009) and the diversification of species (Ricklefs & Renner 1994; Dodd, Silvertown & Chase 1999; Lengyel et al. 2009). Although studies on dispersal have increased, it is still unknown how the majority of species disperse.

Predicting the impacts of climate change on species distributions has become an important issue in conservation biology (Pearson & Dawson 2003; Thuiller et al. 2008). Predictive species distribution models need inclusion of dispersal capabilities (Davis et al. 1998; Pearson & Dawson 2004; Midgley et al. 2006; Brooker et al. 2007; Thuiller et al. 2008). However, the lack of dispersal information has forced modellers to include unrealistic assumptions of unlimited dispersal (no restrictions on the dispersal capabilities of the species) and/or no dispersal (where species are incapable of dispersal) (Hampe 2004; Guisan & Thuiller 2005; Pearson et al. 2006). This omission of realistic dispersal scenarios can lead to huge variation in the predicted ranges where species will move and persist under climate change (Engler & Guisan 2009). The addition of realistic dispersal scenarios, even approximations, will help reduce uncertainty of predicted species distributions (Engler & Guisan 2009).

Ideally, we would use quantitative data on the likelihood of seeds travelling particular distances (dispersal curves). However, such data are difficult to gather, and are not currently available for enough species. Models have been developed using plant traits and environmental factors that produce realistic dispersal curves for individual species, although primarily for wind-dispersed species (Nathan, Safriel & Noy-Meir 2001; Tackenberg, Poschlod & Bonn 2003; for review, see Kuparinen 2006). In the meantime, dispersal mechanisms provide useful proxies for dispersal curves, which are traditionally used to classify the dispersal of plant species. Such a classification system is coarse, but nevertheless dispersal mechanisms are often linked to dispersal distances (Willson 1993; Gómez & Espadaler 1998). Short-distance dispersal mechanisms are ant, ballistic and unassisted dispersal (although seeds with no particular dispersal structures can be wind dispersed), whereas long-distance mechanisms include wind, water and vertebrate dispersal (Willson 1993; Gómez & Espadaler 1998; Vittoz & Engler 2007).

Strong correlations between plant traits, environmental factors and dispersal mechanisms, indicate that predicting dispersal mechanisms of plant species is possible (Hughes et al. 1994). The ability to predict dispersal mechanisms for different regions, where little local dispersal information is available, would be the ultimate goal. Plant traits associated with dispersal mechanisms include: growth form (Westoby, Rice & Howell 1990; Jurado, Westoby & Nelson 1991), seed mass (Moles et al. 2005; Will, Maussner & Tackenberg 2007), seedbank type (Hughes et al. 1994), life span (Jurado, Westoby & Nelson 1991) and plant height (Hughes et al. 1994). Some dispersal mechanisms are associated with vegetation type or structure (e.g. frugivory is common in rain forest communities) and soil fertility (e.g. myrmecochory is often associated with nutrient-poor soils) (Westoby et al. 1991; Forget et al. 2007). In addition, abiotic conditions can influence the presence and proportions of dispersal mechanisms found within plant communities. Examples include changes across moisture, light, nutrient, temperature and elevation gradients shifting the frequency of dispersal mechanisms (Willson, Rice & Westoby 1990; Ozinga et al. 2004; Butler et al. 2007; Almeida-Neto et al. 2008).

We aimed to create a framework for predicting plant species dispersal mechanisms when no dispersal information is known, using significant, yet simple, plant traits and ecological information. We addressed two specific questions: (i) can basic plant traits be used to create predictive models for dispersal mechanisms within a temperate floristic region and (ii) can predictive models based on a flora from one temperate region (Australia) predict dispersal mechanisms for other temperate regions (including Switzerland and California, USA)?

Materials and methods

We collected information on vascular plant species from the Greater Sydney Region in New South Wales (NSW), Australia, encompassing the Central Coast and Central Tablelands botanic subdivisions (Anderson 1961), and the Greater Blue Mountains World Heritage Area. This area has a diverse range of vegetation communities which include dry and wet sclerophyll forest, eucalypt woodlands, shrublands and heaths, while rain forest communities occur in wet, sheltered and/or nutrient-rich soils (Keith 2004). Elevation in the region ranges from 0 to 1360 m a.s.l., while mean annual rainfall ranges between 600 and 1500 mm and the annual average daily minimum and maximum temperatures are 6–15 and 15–24 °C, respectively. We investigated only species native to the study region, excluding introduced or exotic species from other bioregions in Australia or elsewhere in the world.

Dispersal database

We identified 3143 vascular plant species occurring within the Greater Sydney Region from a compilation of vegetation surveys from the NSW Department of Environment and Climate Change and Water, and the Sydney Royal Botanic Gardens data base (acquired March 2007). We gathered dispersal mechanism information for each species from published and unpublished data sets, derived from different methods of data collection: seed morphology analysis, field studies and field observations (Willson, Rice & Westoby 1990; French 1991; Benson and McDougall 1993–2002, 2005; NSW National Parks and Wildlife Service 2002; B. L. Rice, unpublished data). Seed morphology and dispersal mechanisms, although related, are not the same; seed morphology can only imply dispersal mechanisms for species. Detailed studies of species in the field or seed dispersal studies that record attachment potentials or terminal velocities give better indications of the dispersal mechanisms used by species. We acknowledge that deriving dispersal mechanisms based on morphology is imperfect, and we have added a test on a data set based on actual observations of the dispersal vector to determine the extent of this problem. We chose eight different dispersal categories (Fig. 1): wind (anemochory), unassisted (gravity or no particular dispersal morphology), water (hydrochory), ant (myrmecochory), endozoochory (ingestion), epizoochory (attachment), vertebrate (ingestion and attachment) and ballistic (ballochory). Seed dispersal through attachment and ingestion is often grouped as a generic ‘vertebrate’ dispersal category but we separated them to further investigate these mechanisms. Dispersal mechanisms are not mutually exclusive or preferential (Ozinga et al. 2004), hence species could be assigned more than one dispersal mechanism. Seeds of species in the unassisted category had no definite dispersal structure on the seeds and were typically small and probably wind dispersed (Willson, Rice & Westoby 1990). We removed subspecies records, except where they were the only records for a species.

Figure 1.

 The number of occurences for each dispersal mechanism type: wind, unassisted, water, ant, vertebrate ingested, vertebrate attached, vertebrate (ingested + attached) and ballistic. Black bars are for all species with known dispersal mechanisms within the Greater Sydney Region (= 1176), some species had > 1 dispersal mechanisms. Grey bars represent the subset of species used in the modelling process (= 687), species without all variables were excluded from model building.

We collated information on dispersal mechanisms for 1176 vascular plant species (37% of the flora of the region). This is one of the most comprehensive compilation of dispersal data within Australia. Thirty per cent of families (60 families, 170 species) had no species with known dispersal mechanisms and were generally represented by only a few species (mean = 2.8 ± 0.6 SE species per family). The other 138 families had at least one species with a known dispersal mechanism, but knowledge of dispersal mechanisms was generally poor: 77 families (2111 species) had more than half their species with no known dispersal mechanism. The most species-rich families within Australia, Myrtaceae, Poaceae and Fabaceae (Crisp, West & Linder 1999), varied in their proportions of species with known dispersal mechanisms (55%, 15% and 38%, respectively, Table 1). The Proteaceae had the highest proportion (80%) of species with known dispersal mechanisms (Table 1) and are among the ten most diverse families in the Greater Sydney Region. Families with more than two species also tended to have more than one type of dispersal mechanism (2.1 ± 1.2 SE) although often there was a dominant mechanism. For example, Fabaceae were predominantly ant dispersed (Table 1). Dispersal of seeds by wind was the most common dispersal mechanism in the Sydney region (12% of all species in the Greater Sydney Region), followed by dispersal by ants (10%), unassisted dispersal (8%) and vertebrate dispersal through ingestion (6%) (Fig. 1). Ballistic, attachment and water dispersal mechanisms were relatively rare with 79, 52 and 34 occurrences, respectively (Fig. 1).

Table 1.   Summary of dispersal mechanisms (DM) for the 10 largest families found in the Greater Sydney Region for the large data set (= 1176), with number of species with known DM, total number of species in the family, the percentage of species with known dispersal mechanism, count of DM in each family and the number of occurrences (or species) for each dispersal mechanism type (wind, unassisted, water, ant, vertebrate ingestion, vertebrate attachment, vertebrate both and ballistic)
Familyspp. with DMTotal spp.% spp. DM# DM*Dispersal mechanism types
WindUnassistedWaterAntVertebrate ingestionVertebrate attachmentVertebrate both†Ballistic
  1. *Excluding vertebrate both.

  2. †Vertebrate both ingestion and attachment.


Life history and ecological attributes

For species with known dispersal mechanisms, we collected information on seven life-history and ecological traits from the literature and available data bases: vegetation type, growth form, height, seed-bank type, seed mass, perenniality and taxonomic class (i.e. dicotyledons) (Table 2). We placed species into four growth form categories: tree, shrub, herb and climber. Species could have two growth forms; for example Acacia floribunda occurs as either a small tree or a shrub. We collected maximum height data (or scape height when maximum height was unknown) and placed species into four height categories: < 0.5, 0.5–2.0, 2.0–5.0 and > 5.0 m. These reflected the spread of the data and reflected height maxima and minima for certain dispersal mechanisms. For example, ant dispersal is common for species > 2 m (Westoby, Rice & Howell 1990; Hughes et al. 1994).

Table 2.   Summary of data used for modelling dispersal mechanisms, wind, unassisted, water, ant, vertebrate (ingestion and attachment), ingestion, attachment and ballistic dispersal, its description and source from the literature, available data bases or derived data
VariableData descriptionSource
Vegetation typeIndependent variable, vegetation types where the species occurs, five categories – heath, shrubland, forest, rain forest and grassland or swampBenson & McDougall (1993–2002, 2005), Botanic Gardens Trust (2008)
Growth formIndependent variable, four categories – herb, shrub, tree, climberBenson & McDougall (1993–2002, 2005), Botanic Gardens Trust (2008)
HeightIndependent variable, maximum plant height of species, four categories – < 0.5, 0.5–2, 2–5, > 5 mRobinson (2003), Costermans (2005), Benson & McDougall (1993–2002, 2005), Botanic Gardens Trust (2008)
SeedbankIndependent variable, one category; canopy seedbank or ‘other’ (combining transient or soil seedbanks)NSW National Parks and Wildlife Service (2002)
Seed massIndependent variable, three categories – < 0.1, 0.1–100, > 100 mgSeed Information Data base Kew Botanic Gardens (Liu et al. 2008); Moles et al. (2005)
PerennialityIndependent variable, either annual or > 1 year life spanNSW National Parks and Wildlife Service (2002), Benson & McDougall (1993–2002, 2005), Botanic Gardens Trust (2008)
GroupIndependent variable, three categories – dicotyledon, monocotyledon or ‘other’ (includes cycads)Botanic Gardens Trust (2008)

Seed mass was categorized: < 0.1, 0.1–100 and > 100 mg (following Westoby, Leishman & Lord 1996). Seed bank was either a canopy seed bank or not (Hughes et al. 1994). Other seed-bank types, persistent soil or transient seed banks, may influence dispersal mechanisms; however, these data were seldom available so were not included in the analysis. Species were defined as either annual or perennial. Plant longevity or life span data would have been more informative but were not available for sufficient species. Finally, we defined species by their taxonomic class: dicotyledon, monocotyledon or other types (e.g. cycads). Twenty-five vegetation types were reduced to five broad vegetation types based on the ecological structure, using a hierarchical cluster analyses (SPSS 15.0, 2006; SPSS, Chicago, IL, USA): rain forest, forest (woodland, forest, wet and dry sclerophyll forest), shrub, heath and ‘grassland or swamp’ (included marsh) (see Appendix S1 in Supporting Information). All species were assigned a binary classification of either a 0 (no) or 1 (yes) for each variable or trait. Limited data availability for certain traits restricted the final number of species for the modelling, but we identified dispersal mechanisms and all associated life-history and ecological trait information for 687 species from 93 families (Fig. 1). This reduced data set was the model-building data set (= 687) (Appendix S2).

Statistical Analysis

We expected dispersal mechanisms to be highly correlated within families, hence we fitted generalized linear mixed-effects models (GLMM; Pinheiro & Bates 2000) using the lme4 package in R (R Development Core Team 2010), using family as a random factor. Mixed models were built for each dispersal mechanism, to produce sub-models. Univariate analyses with GLMM were fitted to evaluate relationships of life-history and ecological trait variables on the dispersal mechanisms. Significant variables from the univariate analyses were then incorporated into a multivariate GLMM for each dispersal mechanism. Model selection was based on the selection of independent variables that contributed significantly to each final GLMM sub-model (Pinheiro & Bates 2000; Leung et al. 2008). All GLMM were constructed with a binomial response variable and a logit-link function. Alpha was set at 0.05 for all analyses and all tests were two-sided tests. There are different valid methods for model selection for GLMM inferences, each with advantages and disadvantages: Wald tests, likelihood ratio tests and information criterion analyses (Bolker et al. 2009). We used a null hypothesis approach (Wald tests) for our analyses rather than Akaike’s Information Criterion (AIC) because: (i) we focused on significant independent variables that explained each response variable, (ii) likelihood tests and information criteria such as AIC may sometimes not be valid asymptotically (Pinheiro & Bates 2000) and (iii) AIC does not deal appropriately with the taxonomic clustering inherent in our data (Vaida & Blanchard 2005).

Predictions for species’ dispersal mechanisms were produced by calculating the probability for each dispersal mechanism from the sub-models. The sub-model that produced the highest probability value was then deemed the predicted dispersal mechanism. For example, the model assigned Daviesia mimosoides (Fabaceae) a 6% probability of using dispersal through ingestion, 2% for wind, 2% for unassisted, 1% for attachment and < 1% probability for water dispersal. The highest probability for this species is for ant dispersal (84%), so it was (correctly) assigned as an ant-dispersed species. Species could theoretically have high probabilities for more than one dispersal mechanism. The data we used for model creation typically only assigned singular dispersal mechanisms to species, therefore we did not test the accuracy of our models for predicting diplochory or polychory in species. For our predictive model, species had to be in one of the families used in model building and all variables had to be available for each species (Appendix S3 for model).

We initially validated our model using species not included in the original model building for three different regions: study region (50 species), California (30 species) and Switzerland (30 species). Dispersal mechanisms for these species were provided by D. Benson and L. von Richter from the Sydney Botanic Gardens (Australia, unpublished data; = 15) and [Kanowski et al. (2008); = 9]. Additional seed mass data from NSW Seedbank staff at the Botanic Gardens Trust were obtained for 26 species which were initially excluded from the model-building data set (but had known dispersal mechanisms). For 30 species from California, we obtained data for dispersal mechanisms (Keeley 1991) and trait data [CalFlora (; United States Department of Agriculture PLANTS Database (; Liu et al. 2008]. For Swiss species, we obtained data for dispersal mechanisms (Vittoz & Engler 2007) and trait information [LEDA Traitbase (Knevel et al. 2003); Liu et al. 2008; Interactive flora of NW Europe (, edited by C. Stace, R. van der Meijden and I. de Kort, based upon Stace (1997)]. We created a secondary validation data set, containing 70 plant species from around the world, using studies that recorded dispersal events by particular dispersal agents (F. J. Thomson, unpublished data). This data set allowed us to validate our model against observed or ‘realized’ dispersal mechanisms, while the remainder of our analyses are based on dispersal mechanisms inferred from seed morphology. We used separate chi-square tests to test whether there were differences between modelled and real dispersal mechanisms for the three regions, and between realized and inferred dispersal mechanisms.


Model selection

All dispersal mechanism sub-models included at least one of the following three variables: growth form, seed mass and vegetation type, with seed-bank type included in the wind sub-model (Table 3). The wind dispersal sub-model had three significant variables, relating to growth form and canopy seed bank (Table 3). Trees and species with a canopy seed bank were more likely to be wind dispersed, whereas shrubs were unlikely to have wind dispersal strategies. The major wind-dispersed taxa in our model-building data set (= 687), were Myrtaceae (= 89 wind-dispersed species), Asteraceae (= 50 species) and Proteaceae (= 25 species), with a total of 24 families containing wind-dispersed species (Fig. 1). Based on our predictive sub-model, unassisted dispersal was common among herbaceous or small-seeded (< 0.1 mg) species, and species that occur in grasslands or swamps. Tree species do not usually have unassisted dispersal. In our model-building data set 43 families had species with unassisted dispersal, most commonly Myrtaceae (= 58 unassisted species), Proteaceae (= 27) and Orchidaceae (= 21).

Table 3.   The final sub-models for the generalized linear mixed-effect analysis for wind, unassisted, water, ant, vertebrate ingestion, vertebrate attachment and vertebrate both (ingestion and attachment) dispersal mechanisms (= 687). Terms shown are the deviance (DEV), the variable coefficients (β) ± standard errors (SE), P-value (P)
Sub-modelVariableDEVβ ± SEP
WindShrub399−1.322 ± 0.378<0.001
Tree3991.521 ± 0.394<0.001
Canopy seedbank3992.084 ± 0.451<0.001
UnassistedHerb5682.806 ± 0.524<0.001
Tree568−0.970 ± 0.3630.008
Grass or swamp5680.756 ± 0.3630.038
Seed mass <0.1 mg5681.441 ± 0.4700.002
WaterHerb1111.411 ± 0.5950.018
Forest111−2.076 ± 0.595<0.001
Seed mass >100 mg1111.542 ± 0.7800.048
AntClimber441−1.892 ± 0.7180.008
Herb441−1.560 ± 0.5540.005
Tree441−1.496 ± 0.4300.001
Grass or swamp441−1.957 ± 0.549<0.001
Rain forest441−1.515 ± 0.4440.001
Vertebrate ingestionClimber2552.959 ± 1.1830.012
Herb255−4.187 ± 1.197<0.001
Tree2551.820 ± 0.6540.005
Rain forest2553.847 ± 0.675<0.001
Seed mass <0.1 mg255−3.636 ± 1.4870.015
Seed mass >100 mg2554.390 ± 0.932<0.001
Vertebrate attachmentGrass or swamp1531.796 ± 0.6740.008
Vertebrate bothClimber3761.686 ± 0.7260.020
Tree3761.608 ± 0.4670.001
Rain forest3762.204 ± 0.399<0.001
Seed mass <0.1 mg376−2.795 ± 1.1490.015
Seed mass >100 mg3763.312 ± 0.731<0.001

Only 2% of species in the modelling data set used water dispersal. It is unclear whether this is due to the natural rarity of water-dispersed species or to the lack of studies on these species in the Sydney region. Species with water dispersal mechanisms were predominantly herbs and species with large seed weights (> 100 mg), excluding forest species. We only had one water-dispersed species to test the accuracy of our model for the Sydney region, but the fact that this species was falsely predicted to be dispersed by ingestion suggests that the water section of our model needs refining. There were 14 families with species that were water-dispersed, but each had less than three species. There were negative associations between species dispersed by ants and the growth forms climber, herb and tree, and the vegetation types ‘grassland and swamp’ and rain forest (Table 3). In our model-building data set, the families with the highest number of ant-dispersed taxa were Fabaceae (n = 76 species) and then Rutaceae (n = 7 species).

Although many families (= 45) had species whose seeds were dispersed by ingestion, this dispersal type was rare within families; Ericaceae had the largest number of species (= 12) using ingestion in our model-building data set. Dispersal through ingestion was positively correlated with the rain forest vegetation type, the climber and tree growth forms, and species that had large seeds (> 100 mg; Table 3). Small-seeded species (< 0.1 mg) and herb growth forms were negatively associated with dispersal by ingestion. Species using attachment for dispersal tend to be found in grasslands or swamps; this vegetation type was the only variable in the attachment sub-model (Table 3). From our model-building data base, 11 families used attachment, with Poaceae (n = 7 species) and Asteraceae (n = 5 species) containing the largest number of taxa utilizing attachment. The vertebrate sub-model with attachment and ingestion dispersal mechanisms had similar variables to the ingestion sub-model, except herb growth form was not included (Table 3). This was almost certainly due to the small sample size of species using attachment (= 23), making a small contribution to the vertebrate sub-model.

Ballistic dispersal (= 45) was not significantly related to any of the life-history or ecological trait variables in the univariate analysis, with no variables meeting the threshold for inclusion in a multivariate GLMM. Most species reliant on ballistic dispersal in the model-building data base were from the families Rutaceae (= 35 species) and Euphorbiaceae (= 6 species), with the four remaining species from Oxalidaceae, Fabaceae and Rubiaceae. Ballistic plant species were sometimes diplochorous, using both ant and ballistic dispersal.

Model validation

We tested our model across three regions (the study region in Australia; California, and Switzerland). Forty-two out of 50 species from Australia, 23 out of 30 species from California and 19 out of 30 Swiss species had predicted dispersal mechanisms with probability values > 0.5 (i.e. the model suggested that these species had at least a 50% chance of having a particular dispersal mechanism); species that had prediction probabilities over 50% had their dispersal mechanism correctly predicted by the model 71% of the time in Australia, 65% of the time in California and 52% of the time in Switzerland.

Next, we assessed the accuracy of predictions across all species (including those whose strongest prediction had a probability lower than 0.5), to estimate the overall accuracy of the model for predicting dispersal mechanisms. Our model correctly predicted the dispersal mechanism for 66% of the 50 Australian species, 67% of the 30 Californian species and 47% of the 30 Swiss species.

The model predicted dispersal mechanisms equally well for Australian species and Californian species (χ2 = 0.015, d.f = 1, = 0.90). Predictive power was lower for the Swiss species, but the difference in accuracy between Australia and Switzerland was not significant (χ2 = 2.89, d.f. = 1, = 0.08). This lack of significance might be due to the relatively low sample size. For the Swiss flora, species utilizing ants or attachment for dispersal were often predicted as having unassisted dispersal.

For the Australian flora accuracy rose to 70% with the exclusion of the three ballistic species which could not be predicted as there was no model for ballistic dispersal. The three ballistic species in the Australian validation data set were all predicted as ant dispersed, probably reflecting the diplochorous species in our data. No ballistic species were included in the 30 Californian species. The accuracy of the predictions for the Swiss flora was 48% with the exclusion of ballistic species. Although all ballistic species were predicted incorrectly as ant dispersed in Australia or unassisted for Europe, ballistic, unassisted and ant-dispersal mechanisms are associated with relatively low dispersal distances, which should produce dispersal curves that are similar in distance. We found that separating the vertebrate sub-model into two separate ingestion and attachment sub-models was better because it increased the accuracy of the overall predictions of our models by 3% (validation data set).

Sixty one out of 70 species for which we had ‘realized’ dispersal mechanisms (based on observation rather than morphology) had probability values > 0.5, and the model correctly predicted the dispersal mechanism for 56% of these 61 species. Overall for the 70 species, the correct dispersal mechanism was predicted with 51% accuracy. This was not significantly different to correct predictions in data sets where dispersal mechanisms had been inferred from morphology (Australia: χ2 = 2.53, d.f. = 1, = 0.11; California: χ2 = 1.98, d.f. = 1, = 0.16; Switzerland: χ2 = 0.16, d.f. = 1, = 0.68). Thus, it seems that the accuracy of our model is not significantly affected by parameterization with morphologically derived dispersal mechanism data.


Dispersal mechanisms can be predicted through the use of simple plant traits. We have shown for the first time that models parameterized using one temperate flora (Australia) can be used to predict dispersal mechanisms for two other spatially discrete floras, without a significant reduction in the model accuracy (Switzerland, = 0.08 and California, = 0.90). We have also shown that our model can predict realized dispersal mechanisms equally well to inferred dispersal mechanisms (based on seed morphology). This represents a major advance in our ability to predict dispersal mechanisms.

The main significant predictors in our sub-models matched the literature. The common dispersal mechanisms wind, ingestion and ant had results related to growth form, seed size and vegetation type that reflected the literature (Westoby, Rice & Howell 1990; Hughes et al. 1994; Moles et al. 2005; Will, Maussner & Tackenberg 2007; Table 3) Our patterns were also consistent with the literature for the less common dispersal mechanisms, for which we had small sample sizes. For attachment to vertebrates, the vegetation type ‘grassland and swamps’ was the only important variable (Table 3). Species that use epizoochory or attachment for dispersal typically are low-growing species, common in grassland or high-light environments (Hughes et al. 1994; Ozinga et al. 2004). Characteristics of water-dispersed plants reflected empirical work with large seed mass associated with water dispersal (Moles et al. 2005) and herbaceous growth forms commonly found in riparian communities (Nilsson et al. 2002; Table 3).

There was a high consistency in the types of dispersal mechanisms used by species within several families, including Fabaceae, Orchidaceae and Asteraceae (Table 1). Other families, particularly Proteaceae, had a broad range of dispersal mechanisms (Table 1). The diversity of dispersal mechanisms may be due to the range of environments and vectors with which these taxa have evolved, or the evolutionarily constraints that families may have for evolving particular dispersal mechanisms. But dispersal traits may be evolutionary labile (Dodd, Silvertown & Chase 1999); for example, ant dispersal has independently evolved over 100 times and is a major driver of species diversification in angiosperms (Lengyel et al. 2009). The high species richness in the family Proteaceae could be because families with abiotic and biotic dispersal mechanisms have higher species richness compared to those with only biotic dispersal (Ricklefs & Renner 1994, 2000; Dodd, Silvertown & Chase 1999).

Our approach for predicting dispersal mechanisms of species worked at a regional scale and a larger global scale. This was surprising as Australia’s flora is often considered unique; however, this is limited to low taxonomic levels (Crisp, West & Linder 1999). At the family level, Australia is virtually identical to other areas (Crisp, West & Linder 1999). The accuracy of the dispersal mechanism predictions was not significantly different among regions, though it was lower for the Swiss flora. This lower accuracy for the European predictions was partly due to our ant-dispersal sub-model, which was negatively associated with herbaceous species (Table 3). While this is true in Australia, many ant-dispersed species in Europe and North America are herbs (Lengyel et al. 2009). Twenty per cent of misclassified species were classified as wind or unassisted dispersers, when they were the reverse. This can be attributed to the fact that unassisted dispersal can be part of the wind dispersal spectra, with lightweight seeds with no morphological adaptation to wind actually being reliant on wind dispersal (Willson, Rice & Westoby 1990). Although relevant for Australian and Californian species, the inclusion of the variable ‘canopy seed bank’ meant predicted probabilities for wind-dispersed species were low for those that did not occur in Mediterranean climates. The positive relationship between wind dispersal and trees and the negative relationship between unassisted dispersal and the tree growth form was the major separation between the two predictive sub-models wind and unassisted.

Even though the predictive accuracy of our model was not significantly different between observed dispersal mechanisms and those derived from seed morphology, the predictive capabilities were lower for the observed dispersal mechanism predictions (51% overall) compared to the Australian and Californian predictions (66% and 67%, respectively). Building predictive models based entirely on observed dispersal mechanisms and not those based on seed morphology would be preferable and would give new insights into the relationships between dispersal mechanisms and plant traits. As information availability increases, additional model refinement and improvement will be possible, but the key will be to improve predictive accuracy while still using simple and attainable variables within the model. The lower predictive accuracy for Swiss flora could be improved by extending our model-building data base with data from other regions, and using region as a random factor in the models, increasing the accuracy and breadth of predictions. For new areas, we recommend using a subset of species with known dispersal mechanisms (from the new study region) to assess the accuracy of the model predictions, using our model (Appendix S3). Future work could also investigate the ability to predict diplochorous or polychorous species, although information on species using multiple dispersal mechanisms remains limited (Vander Wall & Longland 2004).

The ability to predict species’ dispersal mechanisms is a considerable improvement on the present state of knowledge. Understanding of dispersal mechanisms could give insights into species at extinction risk. For instance, plant species that disperse by water or attachment to fur are more likely to decline than those using other mechanisms (wind or bird-assisted) in fragmented landscapes (Ozinga et al. 2009). However, dispersal mechanisms alone are a coarse measure of dispersal abilities. For some dispersal mechanisms (e.g. wind and some zoochory), detailed models are available allowing prediction of dispersal distance from variables such as seed morphology, seed terminal velocity and release height (Nathan, Safriel & Noy-Meir 2001; Tackenberg, Poschlod & Bonn 2003). Building such models for other dispersal mechanisms remains an important goal.

In the future, dispersal curves specifically for the region of interest or vectors present could be applied to the predicted dispersal mechanisms. This is important as the dispersal pattern, distance or spatial arrangement of seeds can be significantly different among vectors or regions (Jordano et al. 2007; Vittoz & Engler 2007). Separate attachment and ingestion sub-models were better than a combined vertebrate sub-model and this would also be a more informative classification of dispersal mechanisms when applying dispersal curves. Attachment and ingestion can produce different patterns and distances, with the type of animal vector causing the largest variation in dispersal patterns (Will & Tackenberg 2008).

Most valuably, our model allows relatively quick assessments of unknown dispersal mechanisms, capturing scarce dispersal knowledge using simple, readily available plant traits. Where a high level of accuracy is needed for species of ecological concern (rare species) or importance (keystone species), more intensive studies will still be required. Knowing dispersal mechanisms allows modellers to partition species into long-distance dispersers (e.g. wind-, water- or vertebrate-dispersed), and short-distance dispersers (e.g. ant and unassisted dispersal) (Willson 1993). Our modelling brings us a step closer to providing realistic species distribution models for regions of high biodiversity where dispersal information is scarce by providing information to establish sensible limits on species dispersal capabilities under future climate change.


Earth’s systems are experiencing rapid environmental change, with requirements to predict the outcomes for species, communities and ecosystem processes. Understanding species’ movement or dispersal is extremely important for predicting whether different species will be resilient to this change. For most species, there is no information on their dispersal ecology; in the Greater Sydney region alone, 1967 species (63%) had no dispersal data. We provided a modelling approach, based on easily attainable life-history traits, which predicted dispersal mechanisms for species with c. 70% accuracy. A natural extension of this work will be the application of realistic dispersal curves to the predicted dispersal mechanisms, to further predict dispersal capabilities of species where no dispersal information is currently available.


We thank Mark Westoby, Barbara Rice, and from the Sydney Royal Botanic Gardens Doug Benson, Lotte von Richter, Cathy Offord and Amelia Martyn for seed mass and dispersal data. Thanks to David Warton for statistical advice, Doug Benson for helpful discussion and two anonymous referees whose comments improved an earlier version. This work was part of the Managing for Ecosystem Change in the Greater Blue Mountains World Heritage Area project (LP0774833), funded by the Australian Research Council and the NSW Department of Environment, Climate Change and Water, Hawkesbury-Nepean Catchment Management Authority, Blue Mountains City Council, NSW Department of Primary Industry and the Blue Mountains World Heritage Institute.