Introduction

Following establishment, the ability to predict spread is integral to managing biological invasions (Liebhold and Tobin 2008). The establishment of non-native species is expected to continue (Seebens et al. 2021), despite increased efforts to strengthen biosecurity measures (Haack et al. 2014; Poland and Rassati 2019). Increases in international shipping and resulting use of solid wood packaging material have driven increased arrival rates of bark and ambrosia beetles (Aukema et al. 2010; Meurisse et al. 2019; Lantschner et al. 2020), and multiple biosecurity measures (e.g., International standard for phytosanitary measures No. 15 (ISPM 15)) have been developed to prevent invasions by bark and wood-boring insects. Species of ambrosia beetles have continued to invade (Gugliuzzo et al. 2021), however, including in countries requiring ISPM-15 compliance such as the United states (US) (Coleman et al. 2019), United Kingdom (Inward 2020), and New Zealand (Ministry for Primary Industries 2019), among others.

The spread of invading insects is driven by short-distance dispersal via natural means (e.g., flight, crawling) and long-distance dispersal typically via human activities. Regarding domestic invasion pathways, the movement of firewood is considered the most frequent mode of long-distance dispersal for bark and woodboring insects (Meurisse et al. 2019; Solano et al. 2021). There is substantial variation in spread rates within and among invasions by woodboring insects, as spread of Agrilus planipennis (Coleoptera: Buprestidae) varied through time from 13.8 to 55.7 km/yr (Ward et al. 2020) and spread of Sirex noctilio (Hymenoptera: Siricidae) varied among regions of the southern hemisphere from 11.7 to 78.0 km/yr (Lantschner et al. 2014). Variation in rates of spread might be driven by differences in life history traits and/or habitat features (Lantschner et al. 2014; Fahrner and Aukema 2018; Nunez‐Mir et al. 2022), but the patterns and drivers of spread for invading ambrosia beetles—some of which cause severe economic and environmental damage by vectoring pathogenic fungi (Hulcr and Dunn 2011; Ranger et al. 2015; Coleman et al. 2019; Gugliuzzo et al. 2021; Olatinwo et al. 2021)—are not well understood.

Knowledge of the factors that govern spread of ambrosia beetle species could guide monitoring and trapping efforts and allow management agencies to budget and prepare for invader arrival (Sharov et al. 2002; Sadof et al. 2017). Both bark and ambrosia beetles, along with their fungal symbionts, can be vectored in firewood (Jacobi et al. 2012). However, despite having similar life histories to bark beetles, ambrosia beetles exhibit different invasion patterns, appearing more likely to undergo contiguous spread with fewer long-distance jumps (Rassati et al. 2016). Ambrosia beetles also have obligate mutualisms with fungi that they vector and feed on exclusively, and some ambrosia beetle species are parthenogenic and/or inbreeding (Andersen et al. 2012; Kirkendall et al. 2015). These life history traits might mediate their ability to invade, for example, depending on the diet breadth of mutualistic fungi or presence of parthenogenesis that might facilitate host procurement or reduce Allee thresholds, respectively (Hughes et al. 2017).

Laurel wilt is an insect-pathogen complex involving the primary vector, Xyleborus glabratus Eichhoff (Coleoptera: Curculionidae; redbay ambrosia beetle), and the damaging fungus, Harringtonia lauricola T.C. Harr., Fraedrich & Aghayeva (Ophiostomatales: Ophiostomataceae) (Kendra et al. 2013; Olatinwo et al. 2021; De Beer et al. 2022). This complex, native to Asia (Rabaglia et al. 2006; Hulcr and Lou 2013; Wuest et al. 2017), is one of the most virulent non-native disturbance agents in forests of the United States (US) (Hughes et al. 2017; Fei et al. 2019). Redbay ambrosia beetle was initially detected in the US near Port Wentworth, GA in 2002 (Rabaglia et al. 2006), with redbay mortality first attributed to laurel wilt in 2003 (Fraedrich et al. 2008; Harrington et al. 2008). Multiple species of Lauraceae are highly susceptible to laurel wilt, including avocado trees (Persea americana), a major concern for the commercial avocado industry (Mayfield et al. 2008; Kendra et al. 2013) that in 2013 was valued at US$14 million and US$415 million in Florida and California, respectively (Kendra et al. 2013). Despite the ongoing threats posed by laurel wilt to both forests and agroecosystems, little is known about the relative importance of factors, such as propagule pressure versus habitat invasibility, contributing to spread at regional scales.

The distribution of laurel wilt in the US is non-contiguous, seemingly achieved by diffusive spread and long-distance jumps (Liebhold and Tobin 2008). Investigations during the early stages of the invasion (2004–2006) indicated that spread was proceeding at approximately 55 km/yr and that invasion speed was likely to be influenced by host density and suitable climate (Koch and Smith 2008). Now that the invasion has progressed over another decade, several aspects of spread and invasion risk remain unknown. Our specific objectives were to quantify (i) any changes in rates of spread through time, (ii) the frequency and distance of long-distance jumps, and (iii) the effects of spatial contagion, forest composition, climate, and human activities on invasion risk. We hope this research will help identify factors that mediate county-level invasion risk and provide insight into the invasion dynamics of laurel wilt and, potentially, other economically and ecologically damaging insect-pathogen complexes.

Materials and methods

Invasion data

County-level data on the occurrence (presence/absence) and year of invasion (2004–2021) for laurel wilt were obtained in February 2022 from the Georgia Forestry Commission (Fig. 1). The database has been updated annually based on reports from state, federal, and university forest health experts from across the southeastern US. These reporting efforts are passive and not achieved through targeted aerial surveys aimed at detecting new infestations. Although laurel wilt is a highly virulent disease that can kill naïve North American host trees such as redbay and sassafras within weeks of initial infection (Fraedrich et al. 2008; Olatinwo et al. 2021), there is potentially a lag between establishment and detection in a county (Riggins et al. 2010) for which we were unable to account. For example, laurel wilt could have been in a county for multiple years prior to someone (e.g., a forest health specialist or landowner) noticing the disease symptoms, with detection lags potentially owing to limitations of resources for surveys and/or reporting lags by non-specialists. Moreover, even if an infestation was limited to a single stand of trees, the entire county would be recorded as invaded. Our analyses of the invasion were thus conducted on a coarse scale (county-level), potentially resulting in overestimated rates of spread. Limitations on survey and reporting efforts, as well as reporting methodology that documents invasion at the county-level, should be considered when interpreting our findings.

Fig. 1
figure 1

County-level invasion by laurel wilt in the southeastern US (2004–2021)

Full size image

We conducted several analyses of the county-level invasion data to provide insight into patterns and drivers of laurel wilt spread. For all analyses, the first discovery point was set at 32.1480°N and 81.1629°W near Port Wentworth, GA and distances between counties were estimated using county centroids. Spatial data were projected with the Albers equal-area conic projection (NAD 1983). Spatial and statistical analyses were conducted using R (R Core Team 2021) and the following packages: geosphere (Hijmans 2019), ggpubr (Kassambara 2020), sd (Pebesma 2018), sp and spdep (Bivand et al. 2013), terra (Hijmans 2021), tidyverse (Wickham et al. 2019).

Historical spread

Spread rates were estimated by calculating the effective range radius (Shigesada et al. 1995), equivalent to the approach used by Koch and Smith (2008) to estimate annual rates of spread for this disease in the first three years following detection of tree death. This method entails summing the area of counties invaded per year (A), multiplying that value by two to reflect the semicircular distribution of the invasion (2 × A), dividing 2 × A by π, taking the square root (\(\sqrt{(2\times A)/\pi }=r\)), and, lastly, regressing r on year. To verify our approach was equivalent to Koch and Smith (2008) and quantify potential changes in spread rates since their analyses, three time intervals were evaluated: (i) 2004–2006 (i.e., the time interval evaluated by Koch and Smith (2008)), (ii) 2006–2021, and (iii) 2004–2021 (i.e., the entire time series). Additionally, a piece-wise regression analysis was conducted using the segmented package in R (Muggeo 2008), enabling us to quantitatively detect potential shifts in radial spread rates through time.

We then categorized each invaded county as either contiguous or non-contiguous. Contiguous counties were those with at least one invaded neighboring county at the time of their invasion, whereas non-contiguous counties had no known previously invaded neighboring counties. We use long-distance or discrete jumps here to refer to invasion of non-contiguous counties, and we assume such events were achieved with assistance from humans or—although we do not consider it a key driver—aeolian dispersal. We also caution that within county spread likely includes discrete jumps over shorter distances, but the county-level resolution of the data—the highest available for laurel wilt at regional scales—preclude accounting for such dispersal events. The number of counties invaded per year was quantified for each of the three county designations (all counties, contiguous counties, and non-contiguous counties) and length of discrete jumps (mean ± SE, median) into non-contiguous counties was estimated.

County-level drivers of spread

Cox proportional hazards (CPH) models were developed using the survival package in R (Therneau and Grambsch 2000; Therneau 2021) to identify county-level correlates of invasion risk. Annual county-level invasion data can be treated as “time-to-event” data that are amenable to survival model techniques such as CPH models (Hastings et al. 2005; Ward et al. 2020). Cox proportional hazards models can be particularly useful for modeling invasions because, in addition to time-independent predictors (e.g., mean annual temperature or human population density), they can incorporate time-varying predictors (Thomas and Reyes 2014; Therneau et al. 2022): as the area invaded grows the invasion risk to different counties changes. For example, when laurel wilt makes a long-distance jump to a given county i, the counties neighboring county i become at higher risk. Cox proportional hazards models can account for this phenomenon via incorporation of a time-varying predictor that adjusts the annual risk accordingly (Thomas and Reyes 2014; Ward et al. 2020). These models require events to be reported in intervals of time, such as a detection of laurel wilt reported between the start of year i to year i + 1 (i.e., 2004–2005, 2005–2006, …, 2021–2022). Each county was assigned a “0” until invasion occurred, at which point a”1” – indicating the occurrence of an event (invasion)—was assigned. Structuring the data for this analysis meant that each county appeared in the data set “invasion year—2004″ times, as we assumed that counties could not become uninvaded (no successful eradications of laurel wilt have been reported). For example, if a county was invaded in 2004, it was assigned a “1″ for the 2004–2005 interval and would not be associated with future intervals (e.g., 2005–2006, 2006–2007, etc.); counties that were not invaded through 2021 thus appeared 18 times each in the analysis. The response variable was “time to county-level invasion” and the unit of observation was individual county.

All candidate models were developed on invasion data from 2005 to 2018, with 2019–2021 (79 invaded counties, ~ 30% of total invaded counties) withheld for evaluating model predictions (see “Future spread” below). We excluded three county detections in 2004 from the response side of the model (i.e., counties invaded in 2004 still contributed to propagule pressure estimates, described below), given that no counties were reported as invaded prior to 2004 (meaning that using the previously invaded range to estimate propagule pressure was infeasible for those instances). We developed a predictor to account for the spatial contagion s for each county i estimated on an annual basis using an inverse-distance weighted metric:

$${s}_{i}=\sum_{j=1}^{J} \frac{1}{{d}_{ij}^{2}}$$
(1)

in which d is the distance between county i and each previously invaded county j (Mally et al. 2021). The spatial contagion variable was estimated for each county in each year to account for annual changes in the invaded range of laurel wilt. Spatial contagion was the only time-varying predictor such that each county in each time interval had a unique estimate of spatially-derived propagule pressure based on its distance to all previously invaded counties. In addition to the term for spatial contagion, predictors representing habitat invasibility (e.g., host and non-host biomass per county (Fig. 2), bioclimatic predictors) and propagule pressure (e.g., human population density) were considered. We note that some time-independent predictors in our model, such as host biomass or human population density, are not truly static through time but we have assumed that temporal changes in these predictors did not occur systematically across the study area and/or would not be of a magnitude that would influence our conclusions. A detailed description of variables is provided in Table 1.

Fig. 2
figure 2

County-level biomass (kg) of a redbay (Persea borbonia) and b sassafras (Sassafras albidum) across the US

Full size image
Table 1 Variables used in Cox proportional hazards models investigating drivers of county-level invasion by laurel wilt across the southeastern United States
Full size table

Predictors except bioclimatic variables were ln-transformed and then all predictors were standardized ((x-mean)/sd). We confirmed there was no substantial collinearity (|r|> 0.7) (Dormann et al. 2013) between standardized predictors before proceeding. The effects of this suite of candidate variables on invasion risk were quantified through three parallel analyses: evaluating (i) counties that contained hosts (redbay and/or sassafras) according to USDA Forest Service—Forest Inventory and Analysis data (Table 1), (ii) invaded counties as of 2020 plus a 500 km buffer to reflect jump dispersal (99% of jump events were < 500 km; see “Results”), and (iii) the entire contiguous US (Appendix S1). The purpose of conducting multiple analyses was to assess the robustness of results to changes in the spatial extent considered. That is, host abundance might be expected to drive invasion when evaluating the entire contiguous US, but such an effect at that spatial extent may be due to the presence—rather than abundance—of hosts in the southeast (Fig. 2).

For each of the three parallel analyses, we fit a full model and then used multi-model inference via the MuMIn package (Barton 2022) to summarize the effects of predictors across multiple candidate models and identify three best-fitting models, one at each spatial extent. The fit of each candidate model was evaluated using Akaike’s information criterion (AIC) and then the dredge() command was used to rank models using ΔAIC compared to the best-fitting model, indicated by the lowest AIC value. Within each parallel analysis, we selected all models with ΔAIC < 2 compared to the best-fitting model and used the model.avg() and confint() functions to obtain model averaged coefficients and 95% confidence intervals. Akaike weights (\({\mathrm{w}}_{i}\)) estimating the likelihood of each model relative to all models (Burnham and Anderson 2002; Wagenmakers and Farrell 2004) were calculated using:

$${\mathrm{w_{i}(AIC)}}=\frac{\mathrm{exp}(-0.5\times {\Delta_{i} \mathrm{(AIC)}})}{\sum_{k=1}^{K}\mathrm{exp}(-0.5\times {\Delta_{k} \mathrm{(AIC)}})}$$
(2)

where \({\mathrm{w}}_{i}\) can be interpreted as the probability that model i is the best model (Wagenmakers and Farrell 2004) at that spatial extent. A flow chart depicting the process of the parallel analyses is provided in Appendix S2.

Future spread

We evaluated the potential for future spread into redbay and sassafras by first estimating how much of the county-level area containing these hosts had been invaded. Separately for each of these two host tree species, we calculated the percent of (i) counties containing hosts that were invaded and (ii) total host biomass occurring in invaded vs. uninvaded counties, both taken as proxies for the amount of suitable habitat invaded. Additionally, given the importance of cold temperatures in invasion risk (see Results), we calculated the mean monthly minimum temperatures (Table 1) separately for invaded and uninvaded counties to determine the potential invasibility of remaining habitat. We caution, however, that other hosts, such as northern spicebush (Lindera benzoin), occur in more northern areas and were not considered.

Next, we used the best-fitting model developed on data from the entire contiguous US to forecast invasion risk. We used this model and spatial extent because one key motivation of this work was to estimate invasion risk outside of the natural range of redbay and sassafras, as many members of the Lauraceae occur outside the natural range of these two hosts, such as in avocado growing regions of California. Models across all spatial extents provided similar forecasts, however, especially models evaluating the two larger spatial extents for which predictions were nearly identical (Appendix S3). We also note that because CPH models do not predict outside the window of time on which they were developed (data from 2005 to 2018 were used as training data), we used the most recent hazard function in our model (i.e., the 2018–2019 interval) to generate forecasts. However, values of the spatial contagion predictor were updated to reflect the distribution of laurel wilt in the preceding time step. For example, we used all counties invaded as of 2020 to estimate/update the spatial contagion predictor when making forecasts into 2021.

We predicted three years of invasion risk (2019–2021) for every remaining uninvaded county using the predict() command from the survival package with type = survival. This approach provides an estimate, p, of the probability of survival—or remaining uninvaded—and thus “1- p” was taken as the probability of invasion. We then paired the three years of invasion estimates with invasion status data (i.e., per year each county either remained uninvaded or became invaded) and used a receiver operating characteristic curve (ROC) and area under the curve (AUC) to compare estimated invasion risk vs. observed invasion. The ROC and AUC analyses were completed using the pROC package (Turck et al. 2011).

Results

Historical spread

Analyses of county-level data from 2004 to 2021 indicated that invasion speed averaged 28 ± 1 SE km/yr (Fig. 3a) but that spread decelerated ~ 5 years after the initial detection. Our analysis using the effective range radius, an approach equivalent to Koch and Smith (2008), indicated that after progressing at 55 ± 1 km/yr from 2004 to 2006, spread was 26 ± 1 km/yr from 2006 to 2021 (Fig. 3a). When analyzing the data using piecewise regression, the spread rate decreased from 40 ± 3 to 24 ± 0.4 km/yr around 2009 ± 0.5 SE (Fig. 3b).

Fig. 3
figure 3

Spread rates of laurel wilt across the southeastern US (2004–2021). Spread was estimated using the effective range radius and different time intervals were considered to a quantify spread rates across the entire invasion time series as well as changes in spread following initial analyses by Koch and Smith (2008) and b identify any breakpoints in spread as determined using piecewise regression. a Model statistics [solid black; dashed black; dashed gray] along with intervals of time evaluated: F1,16 = 1032.98, p < 0.0001 (2004–2021); F1,1 = 3210.36, p < 0.0001 (2004–2006); F1,9 = 2566.44, p < 0.0001 (2006–2021). b Model statistics [solid black; dashed black] along with intervals of time evaluated: F1,4 = 195.30, p = 0.0002 (2004–2009); F1,11 = 3227.38, p < 0.0001 (2009–2021). The breakpoint in spread on panel b was estimated to have occurred in 2009 ± 0.5 yr (SE)

Full size image

As of 2021, laurel wilt had been detected in 275 counties (Fig. 4a). The number of new detections from 2004 to 2021 averaged 15 ± 2 per year. The number of counties invaded per year varied markedly (3–33), with the smallest number of invaded counties per year occurring from 2004 to 2008 (i.e., < 10 counties were invaded per year in four out of those five years) and the two largest values, 30 and 33, occurring in 2019 and 2020, respectively (Fig. 4b). In 2021, 16 counties were invaded. Out of the invaded counties, 203 were into contiguous areas and 72 were into non-contiguous areas. The long-distance jumps averaged 164 ± 16 km to the nearest known previously invaded area (Fig. 4c), with the longest jump of 570 km being recorded into Jackson County, Mississippi in 2009.

Fig. 4
figure 4

County-level invasion by laurel wilt in the southeastern US. a Categorized by contiguous and non-contiguous invasions. b Annual number of contiguous and non-contiguous counties invaded. c Distance of long-distance jumps into non-contiguous counties

Full size image

County-level drivers of spread

The CPH models we developed at three different spatial extents differed markedly in some respects, but together indicated that proximity to invaded areas (spatial contagion), warmer minimum temperatures, and number of campgrounds increased risk of invasion (Fig. 5). At the smallest spatial extent (counties containing redbay and sassafras in rural forests), the best-fitting model and model-averaged coefficients, calculated using models with ΔAIC < 2 compared to the best-fitting model, suggested that the abundance of non-host trees had the largest—yet most variable—effect on invasion risk according to standardized slope coefficients (Fig. 5a). In terms of Z-value, however, the spatial contagion (Z = 16.49, p < 0.0001) was the strongest predictor, followed by minimum temperature (Z = 7.83, p <0.0001), abundance of non-hosts (Z = 4.28, p <0.0001), and density of campgrounds (Z = 2.60, p = 0.0092), which were all positive, statistically clear associations with invasion risk (Fig. 5a). Human population size was also positively associated with invasion risk in both the best-fitting model and according to model-averaged coefficients, but this was not a statistically clear association. Additional variables contributing to model-averaged coefficients but not appearing in the best-fitting model were income (negative effect), host abundance (positive), and mean annual precipitation (positive), but none of these effects were statistically clear.

Fig. 5
figure 5

Slope coefficients ± 95% confidence limits for Cox proportional hazards models predicting time-to-invasion, indicative of risk of invasion, by laurel wilt at the level of US county using three different spatial extents: a counties with host trees, b invaded area + 500 km buffer, and c the contiguous US (Appendix S1). Coefficients from model-averaging and a best-fitting model (lowest AIC) are provided. Descriptions of data and predictors are provided in Table 1

Full size image

In analyses at the two larger spatial extents, invaded area + 500 km buffer (Fig. 5b) and the contiguous US (Fig. 5c), the fits were nearly equivalent and, within each analysis, there were no models with ΔAIC < 2 compared to the best-fitting model. Thus, the model-averaged coefficients—all statistically clear associations—were exactly equivalent to the best-fitting models. At both extents, the spatial contagion was the most important driver of invasion risk: counties closer to previously invaded areas were at the highest risk of invasion (Fig. 5b, c). The next strongest predictor that was positively associated with invasion risk was minimum temperature, followed by precipitation, host abundance, number of campgrounds, and human population size. Income was negatively associated with invasion risk at both extents, whereas the only other predictor considered, abundance of non-host trees, did not appear to affect invasion risk when analyzing the larger spatial extents (Fig. 5b, c). We attribute the nearly equivalent results found when analyzing the largest spatial extents to the inclusion of several distantly located, uninvaded counties that have no redbay or sassafras. That is, switching extents from “the invaded area + 500 km buffer” to “the contiguous US” only contributed zeros (= uninvaded counties), and the absence of invasion events across these counties—mostly located far away from the invaded area—was likely accounted for statistically by the spatial contagion predictor.

Future spread

As of 2021, laurel wilt had invaded 68.2% of the counties that contained redbay (accounting for 60% of redbay biomass) but only 12.5% of counties that contained sassafras (accounting for 6% of sassafras biomass). However, invaded counties containing redbay and sassafras had mean monthly minimum temperatures of 4.0 and 0.57 °C, respectively, compared with 1.53 and  -4.54 °C, respectively, for noninvaded counties. Thus, the remaining uninvaded counties containing biomass of redbay and sassafras are generally located in areas with lower mean monthly minimum temperatures compared with the current invaded range.

County-level invasion risk in each from year 2019–2021 was forecasted using the best-fitting model developed using data from the contiguous US. Comparing predicted probability of invasion with observed invasions on an annual basis resulted in an AUC value of 0.92 (Fig. 6a). The median invasion probabilities forecasted for counties that actually became invaded in 2019, 2020, and 2021 were 0.04, 0.03, and 0.05, respectively, compared to predicted probabilities of <0.001 in each of those three years for counties that remained uninvaded (Fig. 6b).

Fig. 6
figure 6

Diagnostic plots for the best-fitting Cox proportional hazards model developed on county-level data from across the contiguous US (Fig. 5c). a Receiver-operating characteristic curve with area under the curve (AUC) indicating the ability of the model to correctly predict invaded vs. uninvaded counties. b The distribution of estimated invasion probabilities split between counties that became invaded versus those that remained uninvaded

Full size image

Generally, predicted invasion was low, as 2689–2846 counties (86.5–91.6%) had a predicted invasion probability < 0.01 across the 3 years (Fig. 7). Counties with predicted invasion probabilities > 0.01 were, as expected, concentrated in the southeast and near the invasion front in each of the three years in our training dataset (Fig. 7).

Fig. 7
figure 7

Risk of invasion by laurel wilt estimated in three years (2019–2021) using the best-fitting Cox proportional hazards model trained on data from across the contiguous US (Fig. 5c). Counties outlined in black became invaded in each corresponding year. Spatial extent has been truncated to current invaded area plus 500 km buffer and only counties with an estimated probability of invasion > 0.01 are colored

Full size image

Discussion

Owing to its rapid spread and wide host range, the insect-pathogen symbiosis that results in laurel wilt has been characterized as one of the most damaging disturbance agents to invade North American forests (Hughes et al. 2017; Fei et al. 2019). We quantified landscape-level patterns of spread by laurel wilt, finding that the spatial contagion was the strongest driver of spread. Additionally, spread has been quickest into warmer regions with greater host biomass and higher human activity, although we caution that positive associations of invasion risk with human activity could in part reflect detection biases (i.e., invasions in rural areas are less likely to be noticed). Our analyses of future spread indicated that laurel wilt had invaded 68% of the counties that contain redbay and 13% of the counties that contain sassafras, but that several uninvaded counties containing sassafras are in northern, colder regions that are less likely to be invaded according to our forecasts (Fig. 7) and other climatic matching analyses (Koch and Smith 2008; Formby et al. 2018).

The invasion speed appeared to decrease approximately 6 years after the first detection of tree mortality (Fig. 3b), potentially owing to intense public awareness campaigns and/or survey efforts following initial detection. That is, initial survey efforts may have uncovered long-invaded counties that went undetected prior to increased awareness of the disease and its severity. We also provide evidence that the invasion slowed from 2020 to 2021, as the number of counties invaded dropped from 33 to 16. Whereas invasion theory suggests that spread can decrease as suitable habitat is exhausted (Shigesada et al. 1995; Shigesada and Kawasaki 1997), the recent decline in county invasion rates from 2020 to 2021 could be due to decreased travel, surveys, and/or reporting stemming from the coronavirus pandemic. However, this pattern may also indicate that laurel wilt has invaded most of the warmer counties with redbay across the southeastern US and that spread into sassafras containing regions at northern latitudes might be partially inhibited by the cold tolerance of the beetle vector (Formby et al. 2018). The pathogenic fungus has been detected on other scolytines, however, and alternative vectors that perform well in colder regions and, similar to redbay ambrosia beetle, attack healthy hosts– a somewhat unique characteristic among ambrosia beetle species—could serve to reaccelerate and/or maintain invasion speed. Additionally, there are still several uninvaded counties in the southeast that harbor sassafras (Figs. 1, 2b), which could be attributable to potentially low propagule pressure, dispersion of hosts at the sub-county level, and/or other invasibility factors not captured by our model(s).

The spatial contagion was the main driver of invasion risk, but habitat characteristics appeared more important for driving this invasion than human activities (Fig. 5). Similar analyses of other forest invaders have identified human population density, often used as a proxy for several human activities that can lead to the movement of non-native species, as one of the most important drivers of invasions (Ward et al. 2020; Cook et al. 2021; Epanchin-Niell et al. 2022). For laurel wilt, it appears that warmer temperatures, precipitation, and host tree abundance were more strongly associated with invasion risk than human population density or number of campgrounds (Fig. 5). We used a composite metric of host abundance (redbay + sassafras) and did not attempt to disentangle the effects of warmer temperatures and abundance of redbay, which is mostly distributed in warmer regions. Nonetheless, both redbay and sassafras are highly suitable hosts, and the facilitative effect of host trees could indicate the increased propensity of beetle vectors finding a host following short and/or long-distance dispersal events. This conclusion—a stronger facilitative effect of host biomass compared with human population density—corroborates previous findings from a study on several non-native insects and pathogens (Ward et al. 2022), including laurel wilt, that used a different modeling framework (i.e., a logistic regression modeling the presence/absence of county-level occurrence, rather than time-to-invasion). One potential confounding factor is that sassafras, while found across a larger area than redbay, is more patchily distributed (Peters et al. 2020) and such spatial variation could inhibit the ability of beetle vectors to locate hosts. Additionally, the low Allee threshold for redbay ambrosia beetle—the initial invasion was likely established by a single individual or a small group of clones (Hughes et al. 2017)—potentially indicates that propagule pressure is less important for invasion success compared to habitat suitability.

Analyses across all spatial extents highlighted the importance of climate in driving range dynamics (Lehmann et al. 2020). The finding that warmer minimum temperatures increased invasion risk accords well with historical forecasts of suitable habitat for redbay ambrosia beetle (Koch and Smith 2008), which is native to southeast Asia. However, we caution that redbay ambrosia beetle appears capable of surviving in colder regions of the US than the current invaded range (Formby et al. 2013, 2018) and we did not incorporate northern spicebush, another suitable host for laurel wilt. Other aspects of temperature could be driving the positive association between invasion risk and warm temperatures, however, such as additional beetle generations in warmer areas leading to increased spread capacity (Fahrner and Aukema 2018). Nonetheless, the drivers of county-level spread and subcounty-level disease incidence differ, as disease incidence at the plot level was inversely related to maximum temperature (Choudhury et al. 2021). Studies at smaller spatial scales are critical for understanding tree- and stand-level progression of the disease; we caution that our findings pertain to county-level invasion dynamics and may not hold for sub-county dynamics for either rates of spread or responses of laurel wilt to environmental conditions.

We found that 72 long-distance jumps averaging 164 km occurred since the initial detection (Fig. 4c). Given that laurel wilt is vectored by an endophytic scolytine, it is likely to move via firewood (Meurisse et al. 2019). There were some instances of counties becoming invaded in clusters, such as when several contiguous counties were invaded within the same year (Figs. 4a, 7a). This pattern potentially highlights increased surveillance in counties following the discovery of a satellite infestation, rather than a simultaneously widespread invasion of a cluster of isolated counties within a single year. Similarly, the increase in invaded counties in 2019 and 2020 (Fig. 4) could reflect intensifying inspection efforts on sassafras as the invasion expanded into northern and western counties.

Despite advances in phytosanitary regulations, non-native ambrosia beetles and the fungi they vector have continued to invade and kill trees in several regions (Coleman et al. 2019; Ministry for Primary Industries 2019; Inward 2020; Gugliuzzo et al. 2021). Few analyses of the landscape-level dynamics of ambrosia beetles exist, but knowledge on drivers of spread is critical for delimiting newly detected populations, designing potential eradication programs, slowing spread, and, ultimately, mitigating environmental and economic impacts. The long distances between avocado-growing regions of California and Mexico and the invaded range of laurel wilt results in an extremely low estimated invasion risk, and thus a long-distance dispersal event over that distance is stochastic and exceedingly difficult to predict. Lastly, our results report on spread driven by natural and human-aided dispersal of redbay ambrosia beetle, but H. lauricola can be vectored by other scolytines (Ploetz et al. 2017) and the performance of such alternative vectors could vary with climate or time and result in invasion patterns deviating from those described here.