Assessing the ecological niche and invasion potential of the Asian giant hornet

Environmental Factors Affecting V. mandarinia.

We assessed environmental factors mediating V. mandarinia occurrence using records gathered via the “spocc” package in R (10). We obtained 343 records from V. mandarinia’s native range; 119 were filtered out by enforcing a distance of 5 km between observations, resulting in 224 unique records (Fig. 1A). To determine climatic variables that constrain V. mandarinia, we considered temperature (annual mean, range, maximum of warmest and minimum of coldest months), precipitation (annual and wettest and driest months), and annual mean radiation. We relied on environmental variables because little is known about how other biological factors, such as species interactions or evolution, might impact V. mandarinia. For example, V. mandarinia predators are virtually unknown, and no congeners that might compete against V. mandarinia for resources occur in western North America (9).

Realized Niche Modeling.

We assessed realized niches occupied by native and introduced V. mandarinia populations using minimum ellipsoid volumes, which depict niche breadth in three dimensions. We generated three environmental dimensions that summarized 90% of overall variation in the eight bioclimatic variables using principle component analysis in NicheA (http://nichea.sourceforge.net/).

Ecological Niche Modeling.

We used an ensemble model to assess habitat suitability for V. mandarinia (11), which averaged predictions across five commonly used ecological niche models: 1) generalized additive, 2) generalized linear, 3) general boosted, 4) random forest, and 5) maximum entropy. Our approach included statistical models (1 and 2) that infer relationships between variables, and machine-learning models (3 to 5) that seek to obtain a general understanding of the data to make predictions. A common limitation of statistical models is that they require assumptions about the distribution of variables, whereas machine-learning models often suffer from overfitting and provide limited information about biological mechanisms affecting species distributions. However, while these differing approaches each have specific assumptions and limitations, by averaging predictions across models using a consensus method, ensemble models limit biases of any particular approach (11). Fifty percent of records were used for model training and 50% for validation. We used a “random” method in biomod2 to select 10,000 pseudoabsences to improve model fit, from <400 m or >1,400 m elevations in “accessible” areas of V. mandarinia that were delimited by buffering minimum convex polygons of observed points at 400 km. We used area under the curve of receiver operating characteristic plots and true skill statistic to assess model fit.

Habitat modification has also been linked to invasiveness in some Vespidae, and areas with suitable climate and high human activity may be more susceptible to invasion (4, 7). We mapped human footprint, an indicator of human-mediated disturbances that combines population density and infrastructure (12) with climate suitability, using a bivariate approach (Fig. 1D).

Dispersal Simulation.

V. mandarinia dispersal is mediated by queens. Dispersal behavior for most Vespa species are unknown, but invasive populations of the congener V. velutina have expanded by 78 km/y in France (13) and 18 km/y in Italy (14). To simulate V. mandarinia spread, we used the “MigClim” package in R (15). MigClim uses a dispersal time step, which we set as 1 y because queens form colonies once per year (1). We ran simulations for 20 y using habitat suitability from ensemble models and two dispersal scenarios: 1) short-distance only and 2) both short- and long-distance dispersal. Datasets to simulate dispersal had a resolution of 5 arcmin per pixel (∼5.5 km). We defined short-distance dispersal as less than ∼49.5 km (9 pixels), the average spread for V. velutina from two European studies (13, 14); short-distance dispersal was modeled with a dispersal kernel that assumed an exponential decline in the probability of movement at greater distances. We assumed that long-distance dispersal can occur up to ∼110 km (20 pixels), which reflects the maximum rate of spread observed for V. velutina in Europe (13). Within MigClim, long-distance dispersal events are generated with a defined probability and within a defined distance range. We set the dispersal probability as 1 to reflect that the assumption that 100% of source cells produce propagules in any given year; we also assumed long-distance dispersal occurred at distances greater than 9 pixels (49.5 km) but less than 20 pixels (110 km). While longer-distance dispersal via human activity is possible, we chose conservative values that likely capture the vast majority of human-mediated dispersal events (15). We assumed there are no geographic or environmental barriers to dispersal.