Understanding the flowering phenology of invasive alien plants is essential for predicting their potential impacts on invaded ecosystems and developing effective management strategies. However, achieving this on a global scale poses significant challenges, especially for widespread invasive species. Digital data offers an efficient and scalable solution to studying the flowering phenology of plants across diverse regions and environments. Here, we apply this approach to one of the most problematic groups of invasive plants in coastal areas worldwide, to some taxa in the genus Carpobrotus. We collected geotagged photographs from widely used online platforms (i.e., Google Maps, iNaturalist, and Instagram) at key tourist sites in six countries spanning native (South Africa) and non-native (Argentina, New Zealand, Portugal, Spain, and the United States of America) regions. These records were analysed to document the flowering phenology of Carpobrotus plants in different regions linked to their genetic lineages (clusters), focusing on the start, end, and peak flowering periods. Our results show that broad floristic region and sampling locality, rather than genetic lineage, are the primary determinants of flowering phenology in the Carpobrotus taxa studied. Non-native populations often displayed extended flowering periods compared to native populations, potentially enhancing their pollen availability and seed production, contributing thus to increased propagule pressure or seed bank. Apparent differences among genetic clusters in single-model analyses were not retained once a site‐level random effect was included, indicating that observed cluster contrasts reflect local environmental and sampling variation more than intrinsic genetic differences. This study highlights the use of digital data to address critical knowledge gaps in the flowering phenology of invasive plants across native and non-native ranges. By identifying extended flowering periods and their potential contribution to increased propagule pressure through prolonged seed production and subsequent accumulation in the soil seed bank, our findings provide valuable insights for developing targeted management strategies, such as optimizing intervention timing to coincide with flowering peaks.

1. Data collection

To study the flowering phenology of Carpobrotus taxa, we selected 29 sites (Table 1 in the associated manuscript) that represent populations from all three genetic clusters identified by Novoa et al. (2023) across both their native and non-native ranges. Overall, we selected locations from different world floristic regions (Liu et al. 2023): the Neotropic (Argentina: cluster Admixed), the Novozealandic (New Zealand: cluster A), the Holarctic (Portugal, Spain, and the United States: clusters A, B, and Admixed), and the African (South Africa: clusters A and C). Within the Holarctic realm, we further distinguished the Southern European (e.g., the Azores, the Portuguese volcanic islands in Macaronesia, and peninsular Spain) from the Californian subregion (western United States).

For data collection, we utilized a multi-platform approach including Instagram, iNaturalist, and Google Maps, with a focus on photographs uploaded between 2017 and 2022 (Fig. 1 in the associated manuscript). Records from Instagram were obtained by manually searching for images geotagged to the selected locations where Carpobrotus taxa were visible. To ensure accuracy in dating, we excluded images posted retroactively (e.g., not taken the day when the post was created), such as when users indicated in the text caption that this was in the past. For this reason, we also excluded commercial photography accounts since photoshoots are often posted at a later date. In regions where Instagram provided limited information, this was complemented by photographs obtained from iNaturalist and Google Maps, where we gathered user-submitted observations of Carpobrotus taxa. For each image, we recorded the location, date, presence or absence of flowers, colours of petaloid staminodes (which mimic petals), and density of flower presence, which were classified into five levels: "mass flowering (>50)", "many flowers (20–49)", "some flowers (3–19)", "few flowers (1–2)", and "no flowers" (Fig. 2 in the associated manuscript). The classification of flower density and other phenological parameters were performed by multiple authors. To try to keep the classification consistent, authors received guidance on how to evaluate the images, and the flower counts were binned into broad categories to reduce potential inaccuracies in interpretation. It is also important to note that the recorded density of flowers was influenced in part by whether the photographs were taken as close-ups or wide-view shots. For example, a record of 'few flowers' may indicate that only a small number of flowers were blooming, or it could reflect a close-up photograph of a plant during mass flowering. Either way, we were able to gather both landscape photographs and records of flower presence or absence, providing valuable data on flowering patterns.

Analyses

All statistical analyses were conducted in R 4.2.2 (R Core Team, 2022). To address our two research question: (i) whether flowering phenology differs among broad floristic regions and (ii) whether it differs among population-genetic clusters; we modelled the probability that a photograph contained flowers (0/1) with binomial generalised additive models (GAMs). Each model included a cyclic cubic spline for day-of-year, s(doy), to capture the seasonal flowering cycle. We fitted and compared four candidate models.

The baseline model (M0) contained only the seasonal spline. Model M1 added three fixed effects that address the first and second research questions simultaneously: observation year (fyear, 2017–2022) to capture inter-annual variability, floristic region (Africa/South Africa, Holarctic–California, Holarctic–Europe, Novozealandic) to test for geographic differences, and cluster (A, B, C or Admixed) to test the effect of genetic lineage. Model M2 retained the seasonal spline but accounted for uneven sampling effort by introducing a single random intercept for site. Lastly, the full model (M3) combined the fixed effects of M1 with the site random effect of M2, thereby testing regional and genetic predictors while simultaneously accounting for uneven sampling effort.

All GAMs were fitted with the 'mgcv' R package (v 1.9-1) using maximum-likelihood estimation of smoothing parameters, which allows Akaike's Information Criterion (AIC) to be applied directly to rank competing models (Wood 2011). Records from Argentina (Neotropic realm) were removed a-priori because the available sample (n=28) was below our minimum threshold of 30 observations per region. Diagnostic checks (i.e., residual-versus-fitted plots, QQ-plots, dispersion statistics and variance-inflation factors) indicated no over-dispersion, no influential residual structure and low collinearity among the fixed predictors. Reference categories were South Africa for region (native range), cluster A for genetic cluster and 2017 for observation year; all coefficients are interpreted relative to these baselines.

In separate analyses, we plotted petaloid-staminode colour and flower density as 100% stacked bar charts by genetic cluster (Fig. 5A in the associated manuscript). In the field, petaloid-staminode colour have been used to distinguish Carpobrotus taxa and could therefore provide an additional line of evidence for cluster identity and potential hybridization patterns.