Alien invasive species are a global concern and a threat to biodiversity (Pimentel et al. 2000, Van Wilgen et al. 2001). They also negatively impact agricultural and forestry sectors with substantial economic costs associated with their direct impacts, eradication, control and restoration efforts (Pimentel et al. 2000, 2001). Like many invasive species, Cardiospermum species have been introduced for their economic value prior to becoming problematic (Pimentel et al. 2000, Van Wilgen et al. 2001). Cardiospermum species have been extensively moved around the world for both their medicinal (Venkatesh Babu and Krishnakumari 2006, Subramanyam et al. 2007) and ornamental (Carroll et al. 2005a) values.

The ornamental attraction of Cardiospermum species are their inflated balloon shaped fruit (Fig. 2). Coincidently this trait also contributes to their colonisation success, since these balloons can float in seawater and stay viable for long periods of time, facilitating long distance dispersal, even between landmasses (Carroll et al. 2005a, Simelane et al. 2011). For example, Cardiospermum grandiflorum was introduced to the Cook Islands as a result of a hurricane (Meyer 2004), whilst increased spread of balloon vines in Australia was associated with a major cyclone and subsequent flooding (Carroll et al. 2005a). We floated Cardiospermum grandiflorum fruit structures in seawater and found some of them capable of floating more than 25 weeks with seed remaining viable. (E. Gildenhuys et al., unpubl. data). Upon dehiscence, each seed is attached to a circular blade that permits further transport by wind.

Invasive Cardiospermum species are considered “transformer weeds” (Mc Kay et al. 2010), as they often extensively cover native vegetation, depriving it of sunlight and thus photosynthesis (Mc Kay et al. 2010, Simelane et al. 2011). Cardiospermum invasions also have substantial economic impacts on sugarcane and soybean production (Johnston et al. 1979, Jolley et al. 1983, Voll et al. 2004, Subramanyam et al. 2007, Murty and Venkaiah 2011). For example, in Brazil Cardiospermum halicacabum reduces soybean crop yields by up to 26% (Dempsey et al. 2011, Brighenti et al. 2003). The problem with controlling Cardiospermum infestations in soybean crops is the difficulty of mechanically excluding their seeds, which are similar in size and shape to those of soy (Brighenti et al. 2003).

Currently two Cardiospermum species are globally considered important invaders. Cardiospermum grandiflorum is classified as an invasive species in Australia, southern Africa, Cook Islands and many other Pacific islands (Mc Kay et al. 2010) while Cardiospermum halicacabum is considered a weed in Australia with its status (native or introduced) undetermined in most other parts of its range (Henderson 2001, Harris et al. 2007). In Australia, Cardiospermum grandiflorum is considered amongst the “most destructive life forms of rainforests” (Werren 2002), while in South Africa Cardiospermum grandiflorum is classified as a Category 1 weed which means it’s cultivation is prohibited and control is mandatory (Henderson 2001).

South Africa’s Working for Water program launched a research initiative in 2003 to find biological control agents against Cardiospermum grandiflorum (Simelane et al. 2011). Eight insects and two fungal agents have been identified and are currently undergoing host-specificity testing in South Africa (Simelane et al. 2011). Most are capable of feeding and developing on other Cardiospermum spp. in South Africa, in particular Cardiospermum halicacabum and Cardiospermum corindum (Mc Kay et al. 2010). Three promising agents were identified, a seed-feeding weevil (Curculionidae: Cissoanthonomus tuberculipennis), a fruit-galling midge (Cecidomyiidae: Contarinia spp.) and the rust fungus Puccinia arechavaletae (Simelane et al. 2011). Concerns about potential non-target impacts of candidate control agents on Cardiospermum corindum and Cardiospermum halicacabum, as well as the debated native status of these congeners in southern Africa (Table 1), have so far prevented the release of these agents.

The ornamental trade of Cardiospermum halicacabum and Cardiospermum grandiflorum spans more than 100 years. For example, in Australia the first herbarium records of Cardiospermum grandiflorum date back to 1923, collected around Sydney, New South Wales (Carroll et al. 2005a). Currently invasive populations are found throughout the east coast of Australia between Sydney and Cairns although less abundantly to the north of Brisbane (E. Gildenhuys, pers. obs.). More recently the species has spread inland to forest areas such as Toowoomba (Queensland) and the Blue Mountains (New South Wales) (Carroll et al. 2005a, E. Gildenhuys, pers. obs.). Cardiospermum halicacabum is more abundant in the northern parts of Australia such as Darwin and Cairns, and is seldom found along the east coast south of Rockhampton, Queensland (E. Gildenhuys, pers. obs.). It is speculated that Cardiospermum halicacabum was introduced during James Cook’s second voyage in the 1770’s long before the introduction of Cardiospermum grandiflorum (Bean 2007, Harris et al. 2007).

The introduction of Cardiospermum grandiflorum into South Africa occurred approximately 100 years ago (Simelane et al. 2011). Today it is classified as a major weed, and is present and considered invasive in five provinces, of which Kwazulu-Natal and the Eastern Cape are the most affected (Henderson 2001, Simelane et al. 2011). The first records of Cardiospermum halicacabum in South Africa dates back to 1917, 1919 in Namibia and 1930 in Botswana (Global Biodiversity Information Facility: GBIF, http://data.gbif.org/welcome.htm). It is classified as a minor weed in southern Africa, though its native status is debated, with slight impacts compared to Cardiospermum grandiflorum (Henderson 2001).

Cardiospermum halicacabum and Cardiospermum grandiflorum are also present in North America (Carroll and Loye 2012). Cardiospermum halicacabum is more widespread than Cardiospermum grandiflorum, the latter apparently restricted to a small area in suburban Los Angeles (S. Carroll, pers. obs.). Due to the evident ability of some Cardiospermum species to disperse over long distances (Carroll et al. 2005a, Simelane et al. 2011), it is possible that the presence of Cardiospermum halicacabum in North America is due to natural dispersal from South and Central America, rendering a native status. On the other hand, if seeds escaped horticultural and agricultural environments, they should be awarded non-native status (Subramanyam et al. 2007). Cardiospermum halicacabum was reported in the Spontaneous Illinois Vascular Flora before 1922 and was described as abundant in Oklahoma in the 1820’s (James 1825); thus, if not native, Cardiospermum halicacabum was introduced more than 180 years ago.

Cardiospermum halicacabum is also present in China and India. In China it is described as a common weed in forest margins, shrublands, grasslands, cultivated areas and wastelands of the east, south and southwest (Flora of China, www.eFloras.org) – though considered native by some – [Pacific Island Ecosystems at Risk (PIER)]. In India it is widespread and considered non-native (Raju et al. 2011). The history of Cardiospermum halicacabum in these countries is unknown, but it is widely used for medicinal purposes (Subramanyam et al. 2007).

A comprehensive understanding of the biology and ecology of Cardiospermum halicacabum and Cardiospermum grandiflorum is important because of the invasive potential and biogeographic uncertainties which characterise these two taxa. Such information will also contribute to making informed decisions on their conservation (if native) or control (if invasive). This is especially true since the extent to which thesespecies are invasive is essentially unknown and the uncertainties of their classification in most areas suggest the possibility of a cosmopolitan native distribution.

The morphology of these two species is similar, with both being adapted for tropical and subtropical climates. Cardiospermum grandiflorum is a large, semi-woody perennial, whereas Cardiospermum halicacabum is smaller, less woody and commonly annual. Cardiospermum grandiflorum has elongated fruit (4.5–6.5 cm in length) compared to the more compact fruit of Cardiospermum halicacabum (2.5–3.0 cm in length) (Fig. 2A and B). Fruit structures consist of three dorsally keeled membranous capsules each consisting of three internal blades (Weckerle and Rutishauser 2005). The fruit are septifragal with the capsules breaking away from each other when fruit are ripe, changing colour from green to brown (Weckerle and Rutishauser 2005). Seeds of the two species differ, with a kidney shaped hilum on Cardiospermum halicacabum seeds and a round hilum on Cardiospermum grandiflorum seeds. Both species normally produce three seeds per fruit (Weckerle and Rutishauser 2005), are climbers with tendrils and have large flat biternate leaves. The leaves and stems of Cardiospermum grandiflorum have small reddish hairs that are absent in Cardiospermum halicacabum (Henderson 2001). Flowers are white and yellow with Cardiospermum halicacabum flowers smaller (2–3 mm) compared to those of Cardiospermum grandiflorum (7–11 mm) (Henderson 2001). The average length of Cardiospermum halicacabum is 1–3 m, while Cardiospermum grandiflorum is slightly taller with an average of 2–5 m, though both are capable of greatly exceeding these lengths (Henderson 2001).

Both taxa produce flavone aglycones and cyanogenic compounds that likely protect them against predators such as soapberry bugs (Subramanyam et al. 2007). Soapberry bugs (genera Leptocoris, Jadera and Boisea from the family Rhopalidae) feed exclusively on seeds of Sapindaceae and are predators of Cardiospermum (Carroll et al. 2005b, Carroll 2007). An example of the impact of invasive Cardiospermum populations includes an evolved increase in beak length of the native Leptocoris tagalicus soapberry bug feeding on invasive Cardiospermum grandiflorum in Australia (Carroll et al. 2005b). Soapberry bugs co-occur with the widespread distribution of Cardiospermum and thus may be a factor in Cardiospermum reproduction globally. A treatment of soapberry bugs that feed on Cardiospermum halicacabum and Cardiospermum grandiflorum can be found in Carroll and Loye (2012).

The germination and growth success of Cardiospermum halicacabum is well studied because of its medicinal value, as well as its impact on soybean plantations and on natural riparian areas (Dempsey 2011). In contrast, no studies exist addressing these topics for Cardiospermum grandiflorum, despite the need for additional biological information about this environmental weed. Optimum germination of Cardiospermum halicacabum takes place at 35°C, with high oxygen concentrations increasing germination success (Johnston et al. 1979, Jolley et al. 1983, Dempsey 2011). Therefore, in natural habitats, establishment may be more likely in conditions with warm, well-oxygenated soils. Seeds and young plants are able to survive flooded, saturated and dry conditions while performing best in intermediate conditions (Dempsey 2011).

Despite morphological similarity, these two species differ markedly. They occasionally occur sympatrically but mostly prefer different habitats with Cardiospermum halicacabum dominating tropical and Cardiospermum grandiflorum subtropical areas (Henderson 2001). Although both species invade forest margins and watercourses, Cardiospermum grandiflorum also thrives in disturbed urban open areas while Cardiospermum halicacabum predominantly invades wood- and grasslands which highlights its threat to plantations (Henderson 2001).

To date, managing and reducing impacts of Cardiospermum invasions has mostly involved manual removal or burning (Subramanyam et al. 2007). Manual removal involves cutting plants at the base enabling the top part to die off after which roots are dug out which is thus labour intensive (Mc Kay et al. 2010). Chemical control of larger plants includes treatment with paraquat, glufosinate-ammonium, lactofen, carfentrazone-ethyl, sulfentrazone, glyphosate or 2, 4-dichloraphenoxy acetic acid (Subramanyam et al. 2007). However, the use of chemical control could potentially be problematic for two reasons, firstly because of non-target impact on underlying vegetation and secondly the typical proximity of invasions to waterways makes environmental contamination a threat (Simelane et al. 2011). Another key problem in the management of Cardiospermum invasions is the persistent seed bank. If the weedy canopy is cleared it opens the door for long-lived seeds to sprout (FloraBase 2012).

In collaboration with South Africa’s Working for Water program, a biological control programme was initiated against Cardiospermum grandiflorum in 2003. However due to the taxonomic uncertainty surrounding Cardiospermum halicacabum and Cardiospermum corindum (discussed earlier, Table 1), biocontrol agents cannot be released, hampering effective management in South Africa. The importance of clarifying the geographic native ranges of all Cardiospermum species currently found in South Africa for the successful biological control of Cardiospermum grandiflorum is therefore evident. If Cardiospermum corindum and Cardiospermum halicacabum are indeed native to southern Africa, only agents that are specific on Cardiospermum grandiflorum can qualify for release in South Africa, and thus far, these agents have proved particularly difficult to rear and test under quarantine conditions (D. Simelane, pers. comm.). On the other hand, if Cardiospermum corindum and Cardiospermum halicacabum are not native to southern Africa, all suitable agents against Cardiospermum grandiflorum qualify for release in South Africa.

We used BIOMOD version 1.1.5 (Thuiller et al. 2009) implemented in R version 2.15.1 (R Development Core Team 2012) to predict potentially suitable climate habitats for Cardiospermum halicacabum, Cardiospermum grandiflorum and Cardiospermum corindum. Locality records were sourced from public databases [GBIF; Henderson 2007] and personal observations. We discarded records with spatial uncertainty (e.g., points in the ocean) and those from botanical gardens or with missing or duplicate values. Since no absence data is available for Cardiospermum species, but is needed for modelling, 10, 000 pseudo-absence background points were created per species, by random sampling of the Köppen-Geiger climate classification. We employed generalized boosted regression models (GBM), a method uniting regression trees with boosting (for a more comprehensive description see Elith et al. 2008). For all analyses, seven climatic variables were sourced from BioClim (Hijmans et al. 2005), based on their importance for species survival and low co-linearity (Table 2). Importance, and thus the contribution of each variable to the model was assessed using Pearson rank correlation between standard predictions and those based on random permutations for each variable separately (Thuiller et al. 2009). If correlations between these two predictions were high, the specific variable was regarded as less important. Co-linearity between different variables was limited to <0.70 using Spearman rank correlation coefficients. Consequently, precipitation of the wettest quarter was dropped for modelling of Cardiospermum halicacabum due to a high correlation with precipitation of the warmest quarter. A raster of 6 arc min was used to extract variables since a more coarse resolution is realistic for global scale prediction, while also accounting for sampling error. Models were calibrated with 70% of the data and evaluated with the remaining 30%. A cut-off value was determined with BIOMOD’s default setting, representing the best probability threshold which maximizes the percentage of presence and absences correctly predicted for the evaluation data (Thuiller et al. 2009). Area under the receiver-operator-curve (AUC, Hanley and McNeil 1982) and the true skill statistic (TSS, Allouche et al. 2006) were used for model evaluation. AUC scores between 0.95 and 1 indicate an excellent, 0.9 and 0.95 a good and 0.6 and 0.8 a fair model (Thuiller et al. 2005). TSS values of 0.8–1 are excellent, 0.6–0.8 good and 0.0–0.6 fair for predicting accuracy (Allouche et al. 2006).

The accuracy of species distribution modelling is influenced by false positives and negatives (Thuiller et al. 2005, Fawcett 2006). Therefore a second aim of our species distribution modelling approach was to evaluate the accuracy with which this technique can predict potential invasive regions using models calibrated with native range data only. South and Central America were used as the native range for all three species since native status is debated in all other regions. A model calibrated using these records were then used to project suitable climate regions globally as described above. Known global occurrence records were then used as independent data to evaluate modelling accuracy.

Australia: Global data models for all three species performed well, with AUC values above 0.9 and TSS values above 0.65 (Table 3). Bioclimatic predictions show that a large proportion of Australia is climatically suitable for Cardiospermum corindum, a species currently absent in this country. Both Cardiospermum halicacabum and Cardiospermum grandiflorum have been introduced to Australia and are classified as invasive weeds. The suitable climate range for Cardiospermum corindum in Australia is much larger than predicted for both Cardiospermum grandiflorum and Cardiospermum halicacabum and as such ornamental or medicinal introductions of Cardiospermum corindum into Australia should be prevented (Fig. 4A, B, C). Modelling also predicted that the east coast of Australia is climatically highly suitable for Cardiospermum halicacabum, such that any risks from its establishment in this area should be assessed. Cardiospermum grandiflorum appears to be a more rapid colonizer than Cardiospermum halicacabum in Australia and it is already present in most predicted areas. It is however likely to become locally more abundant in areas where it is already found (Fig. 1B and Fig. 4B).

Europe and Asia: Our modelling approach identified Europe as mostly climatically unsuitable for Cardiospermum (Fig. 4A, B, C). Areas of suitable climate are present for all three species in certain parts of Asia including India (where Cardiospermum halicacabum and Cardiospermum corindum are present), Thailand and Pakistan, with Cardiospermum grandiflorum potentially being the most restricted taxon (Fig. 4B). Cardiospermum corindum has high climatic suitability in southern Yemen, southern India, Thailand, Myanmar and southern China (Fig. 4A). The southernmost tip of Yemen seems climatically suitable for Cardiospermum halicacabum, with India, Thailand, Cambodia, Vietnam, Myanmar, Japan, Taiwan and parts of China highly suitable (Fig. 4C). Many of these regions are already occupied by Cardiospermum halicacabum. Climatically suitable habitat for Cardiospermum grandiflorum in Asia only appears to be present in southern India, Sri Lanka and parts of Vietnam (Fig. 4B).

Southern Africa: In South Africa bioclimatically suitable areas for Cardiospermum grandiflorum are in the Western Cape Province, while for Cardiospermum halicacabum they are in coastal areas in the Eastern Cape Province. Bioclimatically suitable areas in South Africa are the largest for Cardiospermum corindum, with the Western and Eastern Cape Provinces being highly suitable. Currently the species is limited to Limpopo, Mpumalanga and northern parts of Kwazulu Natal (SANBI). Spread and anthropogenic movements of Cardiospermum species in South Africa should therefore be closely monitored since a large part of South Africa appears climatically suitable for establishment. While Cardiospermum grandiflorum and Cardiospermum halicacabum are recorded as naturalised in parts of Namibia and Botswana, bioclimatic modelling did not predict either country as climatically suitable. Cardiospermum species are not widespread in these two countries and possibly only occur in areas with suitable microclimates. Such habitats typically differ significantly from surrounding environments and often result from human actions, and are therefore excluded in bioclimatic modelling based on more coarse data, such as this study (Kearney and Porter 2009).

Models calibrated with South and Central American native occurrence records performed fairly well when cross-validated using AUC and TSS, with values higher than 0.85 and 0.6 respectively. However this was not the case when these models were evaluated with independent data, thus known presence data not used in modelling. Cardiospermum halicacabum and Cardiospermum grandiflorum had low AUC and TSS values ranging between 0.60–0.80 and 0.30–0.45 respectively, only Cardiospermum corindum models performed fairly well (AUC > 0.85 and TSS > 0.55, Table 3).

These results indicate that models calibrated with native range occurrence records only, would not have accurately predicted the invasive spread of Cardiospermum grandiflorum in South Africa while underestimating its potential range in Australia. This lack of accuracy for identifying invasive regions using native data questions the suitability of using species distribution modelling alone when determining potential invasive regions.

Also contrary to what we expected, models calibrated using native range data predicted larger climatically suitable areas than models calibrated with global range data (Fig. 4; except for Cardiospermum halicacabum). We hypothesised that this is due to the more restricted climate zones created with the widespread pseudo-absence data of the global range, thus including more diverse habitats to exclude as suitable areas. We plotted the presence and absence points for both native and global range data for each variable against the probability of occurrence using the response plot function in R (Appendix, Fig. S1 A–F). In these figures it is clear that global data variables include a wider environmental range for pseudo-absences compared to the native range pseudo-absences, especially when considering the most significant variables based on variable importance (Table 2). To test if this is indeed the case we ran three additional models with the same settings as the previous models but using native range presence data and global pseudo-absences data. We used the same evaluation parameters as for the previous models (Appendix, Table S1, S2). This approach resulted in projections that more closely resembled global range model predictions or are even more restricted predictions (Appendix, Fig. S2). These results indicate that while native range data can be used to predict potential suitable areas, data are often over-fitted, thus over predict the extent of suitable habitats, due to less restricted absence data created from the native range.

While species distribution modelling is a popular tool for predicting potential invasive ranges its accuracy remains questionable (Araújo and Luoto 2007, Sinclair et al. 2010). Bioclimatic modelling did not accurately predict current invasive regions for the widely naturalized species Cardiospermum grandiflorum. Also native range data alone led to an over estimation of potential suitable habitats for Cardiospermum corindum and Cardiospermum grandiflorum. Our results comparing predictions based on native and global occurrence records are surprising and significant. We hypothesized that the reason for this observation is the more restricted climate zones created when using global pseudo-absences for model calibrations, an effect that can potentially be amplified for species characterised by incomplete range filling in their native ranges. A key assumption of species distribution modelling is pseudo-equilibrium, however this is probably unrealistic for most species and may therefore seriously impact model accuracy (Guisan and Thuiller 2005). On the other hand, bioclimatic predictions may be hampered if a species has undergone a niche shift in its invasive range (Broennimann et al. 2007). All the above-mentioned issues highlight how factors other than climate may play a crucial role in the accuracy of species distributions modelling. For example niche shift in the non-native range could be the result of release from natural enemies (Keane and Crawley 2002). Similarly, increased resource availability in the introduced range (Davis et al. 2000, Thompson et al. 2004) may increase habitat suitability while abiotic attributes of the new range may permit spread into novel habitats. In concert, dispersal limitations (Pulliam 2000), anthropogenic effects and unique historical factors (Jiménez-Valverde et al. 2008) may limit the distribution of species in their native ranges.

Thus, taking the contradicting results into account and also considering the many other factors that influences a species distributional range, lead us to conclude that while bioclimatic modelling is a useful approach, it should not be used as a stand-alone tool when making conservation decisions regarding the introduction of species into a novel range and caution should be exercised to ensure the quality of input data while also taking other factors into account as discussed above.