Citizen science can complement professional invasive plant surveys and improve estimates of suitable habitat

Citizen science is a cost‐effective potential source of invasive species occurrence data. However, data quality issues due to unstructured sampling approaches may discourage the use of these observations by science and conservation professionals. This study explored the utility of low‐structure iNaturalist citizen science data in invasive plant monitoring. We first examined the prevalence of invasive taxa in iNaturalist plant observations and sampling biases associated with these data. Using four invasive species as examples, we then compared iNaturalist and professional agency observations and used the two datasets to model suitable habitat for each species.


| INTRODUC TI ON
Monitoring is essential to early detection, risk analysis and effective management of non-native species, especially environmentally or economically harmful invasive species . However, systematic surveys by trained professionals are typically resourceintensive and limited in geographic scale (Hochachka et al., 2012), restricting the utility of observations in broad-scale applications, such as estimating species distributions . Although the collection of high-quality, comprehensive species occurrence data is a major challenge in conservation biogeography in general, nonnative species have been especially neglected (Foxcroft et al., 2017;Reaser et al., 2020), and the databases that do exist remain difficult to access, share and integrate (Wallace et al., 2020).
Biodiversity citizen science, in which volunteers participate in species data collection, offers a cost-effective means of addressing these data limitations, as the observations are often greater in spatial and temporal extent as well as more readily accessible (McKinley et al., 2017;Theobald et al., 2015). The availability and relative accessibility of citizen science data has great potential to expand current information systems for non-native and invasive species (Johnson et al., 2020).
Citizen science projects differ in scale and structure in order to balance participant recruitment and experience with data credibility (Freitag et al., 2016). This can result in varying levels of identification accuracy as well as geographic and taxonomic completeness, especially when observations are collected on an incidental basis (Dickinson et al., 2010). Nonrandom variation in observer ability, preferences or search effort can produce sampling bias in the observations due to the mismatch between participants' sampling patterns and actual species richness or abundance across space (van Strien et al., 2013). Sampling effort by citizen scientists, for example, is likely to be motivated by accessibility (i.e. roads or other human infrastructure) or interest in species and areas of conservation concern (Botts et al., 2011;Steger et al., 2017;Stolar & Nielsen, 2015;Tulloch et al., 2013). Thus, a trade-off between the quality and quantity of citizen science data is often assumed (Robinson et al., 2020).
However, sampling biases are not unique to citizen science (Theobald et al., 2015). Sampling by both professional and citizen scientists is taxonomically biased towards vertebrate species (Theobald et al., 2015). Herbarium data exhibit biases towards roadsides and more accessible, lower elevation areas, as well as seasonal bias towards spring and summer (Daru et al., 2018). In peer-reviewed literature, distribution data for some ground-feeding birds are biased towards threatened species and protected areas (Boakes et al., 2010). Because sampling biases in different datasets can be complementary, combining citizen science observations with professional data has increased the spatial coverage of monitoring for shorebirds (Robinson et al., 2020), insects (Hochmair et al., 2020;Wilson et al., 2020;Zapponi et al., 2017), large mammals (Farhadinia et al., 2018) and easily recognizable non-native aquatic species (Lehtiniemi et al., 2020). Following brief training, citizen scientists have produced similar field estimates of species cover and occurrence as professional scientists (Crall et al., 2011;Danielsen et al., 2014). Crall et al. (2015) demonstrated that citizen science can expand invasive plant monitoring in Wisconsin and lead to more realistic estimates of habitat through local and regional programs that involve identification training or collaboration with botanists. Nevertheless, the vast majority of citizen science plant data are incidental observations from low-structure programs (Di Cecco et al., 2021). Evaluation of the differences between low-structure citizen science data and professional plant surveys is needed to address assumed disparities in quality, which may limit the use of citizen science observations in conservation and scientific applications, such as habitat suitability modelling (Lewandowski & Specht, 2015;Riesch & Potter, 2014;Theobald et al., 2015).
Habitat suitability models (HSMs) use the relationship between species occurrence records and the environmental conditions at those locations to predict the species' distribution across sampled and unsampled space. In invasive species management, HSMs have several applications, including predicting potential spread, disease risk or range shifts under climate change (Guisan & Thuiller, 2005;Newbold, 2010;Srivastava et al., 2019). Model predictions can be used to identify areas vulnerable to invasion and to guide survey and monitoring efforts that are critical to early detection (Guisan et al., 2013). Although the use of citizen science observations in habitat suitability modelling (HSM) has increased steadily over the last decade, plant data specifically are currently under-used (Feldman et al., 2021). The quality of model training data is critical to model accuracy, and ideally, species records would be representative of the entire modelled environment (Kramer-Schadt et al., 2013). Predictions based on biased data may more closely reflect survey effort rather than the true distribution of suitable habitat (Phillips et al., 2009) and consequently lead to erroneous forecasts of invasive species ranges (Dimson et al., 2019;Katz & Zellmer, 2018). While data about the observation process can be used to account for uneven sampling effort (Johnston et al., 2021), the majority of citizen science observations come from unstructured projects that do not collect this information (Di Cecco et al., 2021). Data filtering treatments, including spatial thinning or subsampling of records or culling by survey effort or observer expertise, can be used to address sampling bias to a certain extent (filtering will not address a complete lack of occurrence data in a region). The effect of these treatments can vary depending on the species modelled, but several studies have shown that filtered eBird data can successfully produce more accurate models that are K E Y W O R D S biodiversity monitoring, citizen science, habitat suitability model, iNaturalist, invasive species, sampling bias comparable to those based on more structured survey data (Robinson et al., 2018;Steen et al., 2019Steen et al., , 2021. Additionally, eBird-based studies have demonstrated that citizen science and professional observations can contain unique biases (Coxen et al., 2017;Robinson et al., 2020;Tanner et al., 2020). Combining multiple datasets to potentially mitigate their respective biases has thus become more common; integration methods range from data pooling to more formal techniques that can account for sampling issues (Fletcher et al., 2019).
This research focused on the Hawaiian Islands, a biodiversity hotspot with both exceptionally high levels of endemism and ongoing habitat loss (Myers et al., 2000). Native plant species have become outnumbered by non-native species, some of which are highly invasive and pose significant threats to native ecosystems (Cuddihy & Stone, 1990). Native Hawaiian plants, having evolved in isolation, without herbivores, and largely without broad-scale disturbances like wildfires, are highly vulnerable to competitive displacement by invasive species (Carlquist, 1974;Gillespie et al., 2008;Richardson & Pyšek, 2006).
Using the example of iNaturalist in Hawai'i, this study has three main components that explore the utility of low-structure citizen science as a source of invasive plant monitoring and habitat suitability modelling data. iNaturalist is a global, multi-taxa citizen science platform that we selected for its potential to support monitoring of a diversity of regions and taxonomic groups (www.inatu ralist.org).
(1) First, we investigated potential biases in sampling between native and non-native species by conducting a sampling bias analysis of all iNaturalist vascular plant observations. As previous studies have reported citizen science bias towards rare or threatened species (Matteson et al., 2012;Tulloch et al., 2013), we hypothesized that native plant species would be sampled at a higher rate than non-native species, but that both native and non-native iNaturalist observations would be spatially biased towards more accessible and disturbed sites.
(2) Second, we assessed potential differences in spatial sampling bias between citizen science and professional management agency data through a species-specific comparison of four example invasive species that have been well-sampled by both groups. We hypothesized that iNaturalist and professional data would be similarly biased towards open spaces and areas accessible by road or trail, but that professional observations would be biased towards sites dominated by native vegetation, as the efforts of management agencies in Hawai'i often focus on remote native areas containing species of conservation concern. (3) Finally, to understand the implications of the biases identified above to non-native HSMs and their associated projections, we used iNaturalist and professional observations to train a series of HSMs for the four study species. We hypothesized that single-source model predictions (i.e. models based on iNaturalist or professional data only) would be distinct from one another and less comprehensive compared to predictions based on combined datasets, and that using filtered subsets of iNaturalist data would produce model predictions in which oversampled environments were relatively de-emphasized.  (Thomas & Rock, 2015).

| Sampling bias
To evaluate the potential of iNaturalist as a source of invasive species observations, we compared the number of native and non-native vascular plant species observed in iNaturalist to the species composition of Hawai'i overall. Species were divided into three native and three non-native classes. Native classes included endemic (occurring only in Hawai'i), endemic-listed (endangered or threatened at the State or Federal level) and indigenous (non-endemic native) species. Nonnative classes included invasive (known to cause significant economic or environmental harm), invasive-potential (likely to become invasive) and naturalized (established and not currently considered invasive) species. Plant species occurring on Hawai'i's four largest islands were identified from botanical checklists produced by the Bishop Museum's Hawai'i Biological Survey, which maintains an inventory of native, naturalized and non-native taxa in Hawai'i (Imada, 2012(Imada, , 2019. Because a current regulatory list of invasive species in Hawai'i is not available, invasive and potentially invasive species were categorized using the Hawai'i-Pacific Weed Risk Assessment (Daehler et al., 2004) and We investigated three environmental biases that are likely to influence citizen science sampling, including accessibility bias (clustering near roads and trails), status bias (preference for sites in or near areas of conservation interest) and disturbance bias A 250-× 250-m grid (i.e. 250-m resolution) was used to partition the study area into four classes for each bias type ( Table 1) for the four study species within these classes was quantified using a bias index as in Kadmon et al. (2004), which compares the observed distribution of records in space to the expected distribution as: where n d is the number of species records per class d, p d is the probability that a record is located in class d given its area and N is the total

| Model series and filtering treatments
To examine the effect of data source and data filtering on estimates of suitable habitat, we produced seven models for each of the four invasive study species. These included iNaturalist unfiltered and professional unfiltered models that used all available records from their respective sources (hereafter 'iNaturalist model' and 'professional model'), a combined unfiltered model that used all available records from both sources ('combined model'), and four models that used filtered subsets of iNaturalist observations: the iNaturalist thinned model, which targeted clustering in geographic space, and the iNaturalist accessibility-stratified, iNaturalist status-stratified and iNaturalist disturbance-stratified models, which targeted the environmental biases described previously.
Subsampling for the iNaturalist thinned HSM reduced spatial clustering by selecting iNaturalist records at a coarser resolution than that of the predictor layers (one record per 1-km cell). For stratified models, we created environmentally stratified subsamples proportionate to the area of each site class. Stratified treatments aimed to remove potentially redundant records in oversampled site classes (Varela et al., 2014).

| Model parameters
Maxent is a common correlative modelling method that contrasts true species presences with pseudo-absences generated from background data (Phillips et al., 2006). Target Nineteen bioclimatic variables were calculated from 250-m resolution mean monthly temperature and rainfall grids (Giambelluca et al., 2013(Giambelluca et al., , 2014 (Muscarella et al., 2014).
Models were also compared using threshold-dependent measures and predictor contributions. The maximum sum of sensitivity and specificity threshold, recommended because it minimizes omission and commission errors and its selection is less affected by the use of pseudo-absences (Liu et al., 2005(Liu et al., , 2013(Liu et al., , 2016, was used to calculate total suitable area predicted by each HSM series (as a percentage of study area). In order to observe whether bias in the species records would lead to similar biases in model predictions, we also calculated the distribution of suitable cells among site classes. Finally, Maxent tracks the contribution of each environmental predictor to model gain and reports its relative contribution as a percentage (Phillips, 2017). We used the predictors' per cent contribution to each HSM as another indicator of how independent sets of observations influenced model training. The maximum sum of sensitivity and specificity threshold was then used to identify the range of suitable values for top contributing predictors.

| Native and non-native observations
We obtained 13,186 iNaturalist research-grade records for 253 vascular plant species that were collected by 1506 unique users.
Metrics for models of a given species were generally similar.
AUC TEST for the professional L. leucocephala model was significantly higher but was similar to other models in AUC DIFF and omission rate.
Filtered iNaturalist models tended to have lower discrimination ability and higher overfitting, particularly for treatments that omitted a large number of records.
Overlap was lowest between iNaturalist and professional models (i.e. single-source HSMs) (D = 0.42-0.74), with relatively lower D scores within site classes depending on the species (see Appendix 2).
For H. gardnerianum, L. camara and P. cattleianum, overlap between the single-source HSMs was lowest in the bare ground class. For L. camara and L. leucocephala, differences were greater between the single-source HSMs within the native-dominated class and in sites with low to zero road/trail density.
Similarity between single-source and combined models also varied by species. The L. leucocephala combined HSM had higher F I G U R E 3 Distribution of iNaturalist records by (a) accessibility (class I = no roads), (b) status (I = within designated open space) and (c) disturbance (see Table 1 for additional class descriptions). Grey bars indicate % area represented by each class. ± indicates statistically significant over/underrepresentation.

| Environmental predictors
The relative contributions of the environmental predictors varied among models of the same species (see Appendix 4, Figure S4.13).
Climate variables (rainfall of the warmest quarter and temperature annual range) made higher contributions to L. camara and P. cattleianum professional and combined models, and were less important in iNaturalist-based HSMs for these species. Models of H. gardnerianum and L. leucocephala shared some top predictors (elevation and soil great group), but the per cent contribution of these variables still had differences.
Predictor values classified as suitable varied among models of the same species (see Appendix 4, Figures S4.14-S4.16). For F I G U R E 4 (a) Accessibility bias (class I = no roads or trails), (b) status bias (I = within designated open space) and (c) disturbance bias in iNaturalist versus professional agency records for four invasive species (see Table 1 for additional class descriptions). ± indicates statistically significant over/underrepresentation.  I  II  III IV  I  II  III IV  I  II  III IV  I  II  III IV   0 I  II  III IV  I  II  III IV  I  II  III IV  I  II  III I  II  III IV  I  II  III IV  I  II  III IV  I  II  III IV   0

| iNaturalist bias towards non-native species
Observer preference for rare species has been observed in studies of other taxonomic groups. Citizen scientists tend to prefer areas where threatened bird species are known to occur (Tulloch et al., 2013), and to underreport more common butterfly species (Matteson et al., 2012). However, we found that non-native taxa represented the majority of iNaturalist plant species and observations, while endemic Hawaiian plants, especially those that are threatened or endangered, were underrepresented. This is perhaps expected in Hawai'i, where non-native species currently dominate native Hawaiian flora both in terms of species richness and land area (except on Hawai'i Island) (Hughes et al., 2017). Frequent encounters with rare, threatened or endangered species would be unexpected given the more restricted ranges and smaller population sizes of native island species in general (Paulay, 1994). Furthermore, specieslevel plant identification can be difficult (Roman et al., 2017), and citizen scientists may be more likely to record larger, more familiar or more widely distributed plant species that are easier to identify (Boakes et al., 2016). In Hawai'i, the most-recorded plants on iNaturalist were woody species that would be well-known to local observers: ʻōhiʻa (Scaevola taccada), a common coastal native also used in commercial and residential plantings; and noni (Morinda citrifolia), a Polynesian introduction with numerous cultural uses.
Species considered invasive or potentially invasive were recorded at a higher rate than other non-native plants. They represented 12.3% of all plant species in Hawai'i, but 26.9% of the species recorded by iNaturalists and 29.8% of the observations. This could indicate that invasive species are more common than native species in areas surveyed by iNaturalists, but a fuller analysis of native versus non-native species bias would need to account for prevalence, which this study did not do. iNaturalist observers are possibly more motivated or able to record these taxa. While invasive species have received a fair amount of media coverage in Hawai'i, a 2003 survey showed that the general public is relatively unconcerned about invasive plants (Daehler, 2008).
Residents may also have positive associations with certain species.
Kukui (Aleurites moluccanus), for example, is the Hawaiian state tree and a Polynesian introduction with invasive potential (Daehler et al., 2004).
Other invasive trees have gained emotional or cultural value as well, including P. cattleianum (Warner & Kinslow, 2013) and the early 20th century introduction Falcataria moluccana (Niemiec et al., 2017). Whatever the motivation, the relative abundance of non-native plant records on iNaturalist is encouraging for invasive species management purposes.

| Spatial sampling bias
iNaturalist data showed significant sampling biases towards open space and cells with higher road/trail density, which is consistent with other citizen science datasets (Botts et al., 2011;Mair & Ruete, 2016;Tulloch et al., 2013), and towards areas with heavily disturbed and non-native-dominated vegetation. Non-native species observations were more strongly skewed towards these sites. These sampling patterns were not unexpected, and it is possible that they reflect actual species distributions or abundance rather than observation bias. Roads have been shown to have a positive effect on island plant species richness, due in part to the creation of novel habitats during road construction (Irl et al., 2014). For invasive species specifically, roads may serve as a dispersal pathway (Pauchard et al., 2009) and source of ignition for fire-prone species (Ellsworth et al., 2014). L. camara presence, for example, has been shown to be positively associated with roadside disturbance (August-Schmidt et al., 2015).
If that were the case, one might expect the independent records from professional agencies to corroborate the spatial patterns in the iNaturalist data. However, accessibility sampling patterns in the professional records were either less pronounced than (L. camara, L. leucocephala) or the inverse (H. gardnerianum, P. cattleianum) of those in iNaturalist. Both datasets were similarly skewed towards open space and adjacent sites but differed in the proportion of records observed in each status class, particularly for L. leucocephala and H. gardnerianum ( Figure 4b). iNaturalists made far fewer observations in nativedominated sites compared to the professional agencies, which in turn tended to undersample heavily disturbed sites (Figure 4c). This disagreement indicates some degree of bias in at least one of the two sources, and that each dataset represents distinct environments.
The impact of sampling bias on HSM accuracy depends on the gradient of relevant environmental conditions included in F I G U R E 6 Distribution of each HSM's suitable cells by site accessibility, status and disturbance (± indicates statistically significant over/underrepresentation; iNat_unfiltered = iNaturalist unfiltered HSM; iNat_disturb = iNaturalist disturbance-stratified HSM; iNat_ status = iNaturalist status-stratified HSM; iNat_access = iNaturalist accessibilitystratified HSM; iNat_thinned = iNaturalist thinned HSM).  Disturbance   I  I I  I II  IV  I  I I  I II  IV  I  I I  I II  IV  I  I I  I II  IV   0  well-sampled areas (Kadmon et al., 2004). Each of the disturbance bias classes appeared to contain a limited range of conditions, as oversampling in the heavily disturbed, non-native-dominated and native-dominated classes was associated with disagreement between single-source models in how suitable area was distributed.
Yet, we did not find a consistent relationship between accessibility and status sampling bias patterns and distribution of suitable habitat. Rather, models for H. gardnerianum, L. camara and L. leucocephala generally agreed with each other regardless of the sampling bias in the training records. This indicates that open space and road/trail networks in Hawai'i span a range of environmental gradients relevant to those study species, and thus sampling bias related to these site features does not necessarily restrict the utility of observations (Kadmon et al., 2004;McCarthy et al., 2012). However, investigating complementary datasets is important when sampling data are biased with respect to vegetation disturbance in this region.

| Effect of filtering treatments on iNaturalist HSMs
The thinning treatment produced HSMs that were highly similar to unfiltered iNaturalist HSMs, and had limited effects on predictions of suitable area and variable contributions. Other researchers have observed that similar thinning treatments can either improve model performance by reducing overfitting Fourcade et al., 2014) or decrease performance due to the random approach as well as loss of information (Steen et al., 2021;Varela et al., 2014).
In this study, thinning to a 1-km resolution may not have been coarse enough to impact the models beyond small increases in estimated suitable area. Relatively few records were removed from training at this resolution (see Appendix 1, Figure S1.1).
Targeting clusters in environmental space, versus thinning records geographically, is reported to have a more positive effect on model performance, but smaller datasets may also produce less consistent results (Varela et al., 2014). The effects of stratified filtering treatments in this study varied by bias type, with the most notable effect observed for disturbance bias. Disturbance-stratified Reduced sample sizes due to filtering were often associated with greater divergence from the unfiltered model, as well as lower performance metrics. While a smaller, evenly sampled dataset has been found to be more effective than a larger, biased one (Bean et al., 2012;Varela et al., 2014), sample size can have a substantial effect on HSM performance (Gábor et al., 2020), and thus small samples may be further negatively impacted by environmental filtering. We observed a strong negative correlation between per cent decrease in sample size and overlap with the unfiltered HSM (i.e. the more records were filtered out, the less similar the iNaturalist models became). Whether the departures from the unfiltered HSMs are due to targeted filtering or the loss of training data is difficult to determine. Controlling for sample size (e.g. as in Boria et al., 2014) in future work would permit a clearer distinction between the effects of filtering, data source and information loss.

| Complementary monitoring
Although often fewer in number, iNaturalist observations for the example invasive species covered a similarly broad range of conditions and were a valuable supplement to the relatively more structured, professional data in this study. iNaturalist observers and professional agencies also appeared to sample unique environmental conditions, as demonstrated by the suitable predictor values and moderate overlap between single-source HSMs. Neither source appeared to consistently provide more comprehensive information nor have greater similarity to the combined HSM, despite the higher number of professional records for three of the species (see Appendix 1, Figure S1.1).
When professional and iNaturalist sample sizes differed greatly, it is possible that the combined HSM was simply more similar to the source that contributed more training records. For example, P. cattleianum had four times as many agency observations as iNaturalist observations. However, iNaturalists sampled P. cattleianum more evenly across the islands and contributed to greater predictions of suitable area on Oʻahu and Kauaʻi in the combined HSM ( Figure 5).
Previous studies have noted that well-designed citizen science programs can collect data comparable to that of professional scientists (Chandler et al., 2017), and that combining structured survey data with eBird records improves model accuracy (Robinson et al., 2020). We found that opportunistic observations from the low-structure iNaturalist citizen science platform are useful in filling the gaps in professional invasive plant sampling. In Hawai'i, professional bias towards native-dominated sites and iNaturalist bias towards disturbed and non-native-dominated sites were reproduced in model predictions, which, in addition to highlighting the limitations of correlative HSMs in approximating species niches, has practical implications for HSMs as a management tool. Underestimates of invasive species distributions could limit land managers' ability to identify vulnerable areas and prioritize monitoring efforts accordingly. Citizen science data could be critical in monitoring areas that have not been officially surveyed and in producing more accurate forecasts of invasion risk across a broader range of environments.
Conversely, remote areas of the Hawai'i region were poorly sampled in iNaturalist. Kahoʻolawe, Lānaʻi, Molokaʻi and Niʻihau were excluded from this analysis, but the extremely low number of iNaturalist plant observations on these smaller, sparsely populated islands demonstrates an obvious limitation of low-structure citizen science data, which is that sampling is largely restricted to public access areas. Management agencies seeking to utilize citizen science as a complementary data source may use this information to guide their own survey efforts and, resources permitting, fill the known gaps in the landscape, which would enable the development of a more truly representative HSM. Additionally, when either professional or citizen science observations are the only monitoring data available for a region, care is warranted when interpreting the significance of unsampled space.
Our results showed that iNaturalist and professional observations, on their own, produced differing estimates of suitable habitat.
To take advantage of biased yet complementary survey efforts, we simply combined or 'pooled' all available data. However, more formal integrated modelling approaches may be able to better address sampling biases while more fully preserving the specific strengths of each dataset (Dorazio, 2014;Fletcher et al., 2019;Isaac et al., 2020;Miller et al., 2019;Pacifici et al., 2017). Utilizing additional attributes of the iNaturalist data could also improve the modelling process.
Although iNaturalist does not record absences, observer records of non-target species could be used to infer absence and select more meaningful pseudo-absences (Bradter et al., 2018;Milanesi et al., 2020). These data are readily accessible in iNaturalist.

| CON CLUS ION
iNaturalist observers have preferentially recorded and accumulated valuable occurrence data for non-native and invasive plant species in Hawai'i. A deeper understanding of this bias would support resource managers who wish to leverage smartphone-based citizen science apps like iNaturalist as an invasive plant reporting tool. Future research could explore why this bias exists or characterize the participants who tend to make these observations: is it common for first-time observers to identify invasive species, or are a few, experienced observers responsible for the majority of those observations?
The former could potentially indicate a general, active interest in non-native plants, which would facilitate the detection of incipient invasive species in addition to well-established ones, while the latter could reveal a need to broaden invasive species awareness.
iNaturalist observations were spatially biased towards open spaces and relatively more accessible, disturbed areas. However, we also showed that occurrence data collected by professional agencies exhibited similar or inverse sampling patterns for four selected invasive species, and that combining the two datasets produced more comprehensive HSM predictions. The informed integration of citizen science and professional data warrants consideration in similar habitat suitability modelling efforts, as opportunistic observations may expand data coverage in highly affected sites that are not typically included in professional surveys. While iNaturalist participation has grown exponentially in Hawai'i (Dimson & Gillespie, 2023), use of or access to citizen science apps may be limited in other regions.
Additional region-and taxa-specific studies could thus investigate the prevalence of this complementarity in order to encourage broader use of citizen science data in invasive species monitoring and management.

ACK N O WLE D G E M ENTS
We thank the iNaturalist citizen science community, Big Island Invasive Species Committee, Consortium of Pacific Herbarium, Kauaʻi Invasive Species Committee, Koʻolau Mountains Watershed Partnership, Oʻahu Army Natural Resource Program and Oʻahu Invasive Species Committee for collecting and sharing the species data that underpin this research. We also thank Franny Brewer and Tiffani Keanini for sharing helpful survey details and Helen Sofaer for providing useful suggestions that clarified the ideas we have presented here. Any use of trade, firm or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

FU N D I N G I N FO R M ATI O N
This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

CO N FLI C T O F I NTE R E S T S TATE M E NT
None.

PEER R E V I E W
The peer review history for this article is available at https:// www.webof scien ce.com/api/gatew ay/wos/peer-revie w/10.1111/ ddi.13749.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are openly available in the Dryad Digital Repository at https://doi.org/10.5068/D1769Q.