Effectiveness of canine-assisted surveillance and human searches for early detection of invasive spotted lanternfly

Abstract

Prevention and early detection of invasive species are championed as the most cost-effective and efficient strategies for reducing or preventing negative impacts on ecosystems. Spotted lanternfly (SLF), Lycorma delicatula, is a recently introduced invasive insect whose range in the United States has been expanding rapidly since it was first discovered in Pennsylvania in 2014. Feeding by this planthopper can cause severe impacts on agricultural production, particularly grapes (Vitis spp.). Human visual surveys are the most common search method employed for detection but can be ineffective due to the insect's cryptic egg masses and low density during early stages of infestation. Therefore, finding alternative early detection methods has become a priority for agencies tasked with addressing SLF management. This study experimentally tested whether trained detector dogs could improve the probability of detecting SLF in both agricultural and forest settings. We surveyed transects in 20 vineyards and their adjacent wooded areas in Pennsylvania and New Jersey, USA, and used a multiscale occupancy model to estimate detection probability achieved by human observers and detection dogs as a function of SLF infestation level, weather, and habitat covariates. We modeled transect-level occupancy of SLF as a function of infestation level, habitat type, topographic position index, and distance to forests. Occupancy probability of SLF was higher on vines within vineyards than in forests, and occupancy declined with increasing distance from forests, which is informative for future search efforts. Detection probability of SLF was lower at forested sites but was higher at high infestation sites. Detection dogs had a lower detection probability than humans in the vineyards, but the detection probability of dogs was >3× greater than that of humans in forested sites. Our study suggests that detection dogs are more effective than human visual searches as an early detection method for SLF in forested areas, and utilizing detector dogs could strengthen SLF early detection efforts. This study demonstrates the potential applicability of using canine-assisted search strategies combined with occupancy models to enhance the surveillance and prevention of other difficult-to-detect invasive species.

INTRODUCTION

Accelerating rates of biological invasions pose a major threat to the sustainability of agriculture and natural resources, ecosystem services, and biodiversity in the United States and worldwide (Chornesky et al., 2005; Vanbergen et al., 2018). Invasive species have the potential to decrease the productivity of crop systems, forests, fisheries, and waterways, resulting in negative ecological, economic, and societal impacts. They cost the US economy billions of dollars annually from control efforts and losses to agriculture, timber, tourism, and other industries (Pimentel et al., 2005). Further, invasive species can threaten human health through direct impacts (Baker, 2017) or by serving as vectors for diseases (Kauffman & Kramer, 2017).

Preventing the introduction of invasive species is championed as the most cost-effective and efficient strategy for avoiding negative impacts on systems (Finnoff et al., 2007). However, once a species has been introduced, the next best option is to detect the population and eliminate it before it becomes established and spreads more widely across the landscape. This strategy, referred to as “Early Detection and Rapid Response” (EDRR), is used when an invasive species is recently introduced into an ecosystem, and a control method is available to remove it (Reaser et al., 2020). EDRR has the benefit of lower overall costs than other strategies since eliminating invasions results in no additional control costs or further damage (Alvarez & Solís, 2018). However, the challenge with this strategy is that small, isolated populations are difficult to locate, and detection usually occurs after the population has spread and is therefore more easily detected. To improve the likelihood of detecting these early introductions, extensive efforts have been made to develop systems, tools, and technologies to aid in locating species in early detection scenarios (Martinez et al., 2020).

Spotted lanternfly (SLF), Lycorma delicatula, is an invasive planthopper native to Asia that has increased exponentially and has caused severe ecological and economic impacts on agricultural production since its first discovery in southeastern Pennsylvania in 2014 (Dara et al., 2015; Urban, 2020). This phloem-feeding insect feeds on at least 103 species of plants (Urban & Leach, 2023), but primary interest in the United States has been on the serious damage that can be caused by feeding on grapes (Vitis sp.), fruit trees, and hops (Humulus spp.). Vineyards are especially vulnerable, but nurseries, tree farms, and other fruit farms are also at risk of damage. These planthoppers expel sticky “honeydew” when feeding, and sooty mold then grows on the excrement and blocks photosynthesis, leading to a reduction in carbohydrate resources, weakened plants, and reduced fruit yields when populations are large (Urban, 2020). In vineyards, SLF infestations can cause 80%–100% vine mortality within a single growing season and have resulted in a threefold increase in the number of insecticide applications, at a cost increase of 170% (Pfeiffer et al., 2019). In addition to threatening agricultural interests, SLF can be a nuisance pest in areas where it feeds on tree of heaven (Ailanthus altissima), another invasive species that is well established in the United States. Tree of heaven thrives in edge habitats and disturbed areas with full sunlight. These environments can include human travel corridors like highways and railways, which can facilitate the spread of SLF (Kasson et al., 2013).

Despite quarantine measures in infested areas, SLF has continued to spread and is especially of concern in states with robust grape and wine industries. For example, in New York State, which shares a border with Pennsylvania, there were no confirmed infestations at the time this study began. The grape and wine industry in New York supports over 71,000 jobs, resulting in US $6.6 billion in economic output in 2019 (Dunham, 2019). Estimated losses associated with SLF in Pennsylvania range between US $42.6 and $99.1 M to agricultural industries and between US $152.6 and $263.3 M to forest industries (Harper et al., 2019). Recent modeling efforts predict SLF has a high probability of establishing in California's multi-billion-dollar grape-producing counties (which produce 82% of the US grape crop) by 2033 if no treatments to control the populations are applied (Jones et al., 2022). Spotted lanternflies are spreading quickly, both on their own and with human-mediated transport of egg masses, and infestations are now present in 18 states (NYSIPM, 2023); therefore, early detection and efforts to monitor spread are of major importance (Urban & Leach, 2023).

There exist several challenges associated with the early detection of SLF. These planthoppers have one generation per year, with overwintering eggs that are laid on varied surfaces from September to November, hatching in May to June (Dechaine et al., 2021), depending on temperatures. The overwintering population can be used to provide information about the abundance and distribution of insects during the subsequent growing season (Leather et al., 1993). During the growing season, adults and nymphs are very mobile, making population estimation challenging. During the overwintering stage, SLF eggs are laid in one-layer masses, which are often quite cryptic when covered by a brown-gray waxy layer; masses are variable in size but approximately 2.5 cm long, with an average 27–35 eggs/mass (Liu, 2022).

Human visual surveys are the most common search method employed for SLF, but survey efficacy is questionable due to the cryptic nature of some SLF egg masses and their low density during early stages of infestation. Additionally, managing large established populations of SLF is difficult to impossible, making it critical to detect small incipient populations. Thus, alternative and innovative early detection methods are being used to detect SLF presence including drones and environmental DNA (eDNA) (Valentin et al., 2020). While each of these methods has the potential to be effective in specific contexts, detection dogs may offer the greatest detection potential in agricultural and forest settings. Detection dogs—dogs trained to detect the presence of specific targets via scent and then indicate the location of the target to their human handler (Hurt & Smith, 2009)—have demonstrated cost-effectiveness (Orkin et al., 2016) and increased detection rates compared with human visual searches for species such as birds, desert tortoises (Gopherus agassizii), and kit foxes (Vulpes macrotis) (Homan et al., 2001; Nussear et al., 2008; Smith et al., 2001).

The earliest recorded study of using dogs (Canis lupus familiaris) to detect invasive insects was in 1974, when dogs were trained to detect spongy moth (Lymantria dispar) egg masses (Wallner & Ellis, 1976). Dogs have been used to detect brown marmorated stink bugs (Halyomorpha halys) (Lee et al., 2014), Asian longhorned beetles (Anoplophora glabripennis) (Hoyer-Tomiczek et al., 2016) and citrus longhorned beetles (Anoplophora chinensis) (Arnesen & Rosell, 2021), red palm weevil (Rhynchophorus ferrugineus) (Suma et al., 2014), hermit beetle (Osmoderma eremita) (Mosconi et al., 2017), and emerald ash borer (Agrilus planipennis) (Hoyer-Tomiczek & Hoch, 2020). Detection dogs provide a means for L. delicatula detection as they can be trained to identify egg masses, whether the egg masses are alive or dead (Essler et al., 2021).

While some studies have reported positive proof-of-concept abilities in naturally occurring insects (Mosconi et al., 2017; Rutter et al., 2021), more commonly, studies have evaluated detection dogs searching for insect scent in controlled, indoor lineup settings to ascertain basic detection ability (Pfiester et al., 2008), discrimination between potentially similar odors (Aviles-Rosa, Nita, et al., 2023; Essler et al., 2021; Lin et al., 2011), or to determine detection thresholds (Aviles-Rosa, Kane, et al., 2023). Additionally, dogs have been used to detect experimentally placed targets in a known environment (Hoffman et al., 2022; Hoyer-Tomiczek et al., 2016; Hoyer-Tomiczek & Hoch, 2020) which may differ from actual detection rates in natural environments. In practice, being able to detect an insect in a controlled setting is merely a first step toward understanding whether dogs will be helpful detection tools in realistic complex environments, under variable weather conditions, and at a scale that will be meaningful.

Our study objectives were to evaluate the utility of detection dogs as an early detection method for SLF. Specifically, we (1) compared human observers and detection dogs based on their probability of detection of SLF egg masses and their search efficiency, (2) identified environmental factors that influence a dog's ability to detect SLF, (3) modeled the probability of occurrence of SLF as a function of site-level landscape characteristics, and (4) informed optimal search strategies for SLF based on our findings.

We hypothesized that (1) occupancy rates vary across sites as a function of the overall SLF infestation level since sites with higher infestation levels should have a greater number of egg masses present at the subunit level (individual vines or 1-m transect segments at forest sites); (2) human and dog detection probability should be higher at sites with higher infestation levels because there should be more lanternfly eggs available for detection; (3) human and dog detection should be similar in vineyards, but dogs should be better at forest sites due to the ability of dogs to use olfactory senses where visual detection is more challenging; and (4) occupancy of SLF should be higher on vineyard transects closest to the forest due to the species association with tree of heaven at forested sites and the limited dispersal ability of SLF upon emergence.

METHODS

Training detection dogs and human observers

A Labrador Retriever and a Belgian Malinois were trained to detect adult SLF in August 2019 by the New York-New Jersey Trail Conference and Working Dogs for Conservation professional trainers using live SLF adults in Chester County, Pennsylvania. Dogs were trained using positive reinforcement (Braun, 2013) whereby they were trained to sit when they located a SLF adult and this behavior resulted in the dog receiving a reward (i.e., a ball toss). Training continued in New Jersey from August 2019 through January 2020 for 51 days and 398 repetitions per dog using training samples periodically collected from Pennsylvania and New Jersey. Dogs spontaneously generalized their training on adults to egg masses. Training took place in situ with live egg masses, and egg masses were collected to be used as training samples starting on 24 October 2019. Samples ranged from whole, covered egg masses attached to substrate, to partially covered or uncovered egg masses. The teams were first deployed to detect SLF adults on 25 October 2019 and surveyed for SLF egg masses starting on 12 December 2019. A total of 17 days were spent in training at field sites, surveying for egg masses between October 2019 and September 2020 at a variety of locations in New York and New Jersey including trailheads, parking lots, vineyards, and railways.

Before fieldwork commenced, human surveyors were trained to identify SLF egg masses and distinguish them from other egg masses, such as mantid oothecae, using photographs and field examples. Training included a test run at a flagged site where surveyors inspected transects of 20 vines or tree trunks to ensure consistent detection of SLF egg masses.

Site selection

Sites that were considered for selection in Pennsylvania and New Jersey were within the SLF infestation zone defined by Cornell Cooperative Extension (https://lookerstudio.google.com/u/0/reporting/b0bae43d-c65f-4f88-bc9a-323f3189cd35/page/QUCkC). We contacted 51 potential vineyards within this zone, and they were asked to approximate the level of SLF infestation that they believed to have (none, low, medium, high). We chose approximately equal numbers of vineyards that self-assessed as having low, medium, and high infestations. Of note is that all sites selected had a known, visible infestation and had a surrounding forested area adjacent to the vines. Vineyard owners of selected sites agreed to not manage or scrape egg masses until we had completed our surveys. Our final selection resulted in 20 vineyards (Figure 1).

FIGURE 1

Location of 20 vineyards surveyed for spotted lanternfly (Lycorma delicatula) egg masses in Pennsylvania and New Jersey, USA, during winter/early spring 2020–2021. Locations are denoted by black dots.

Surveys

We surveyed 20 sites between 11 December 2020 and 11 May 2021, when SLF are not active and only egg masses are present. Egg masses provide an excellent life stage for detection as they are sessile when overwintering and are present for many months (Keller et al., 2020); therefore, throughout this period, the population was considered closed to immigration or emigration. At each site, we recorded the detection or non-detection of SLF egg masses at survey units (transects) and survey subunits (each vine, pole, or 1-m segment of forest transect). We surveyed 12 transects per vineyard and 12 transects in the adjacent forest. Each vineyard transect included searching 20 vines, and forest transects were searched for 27 m. The spatial resolution of forest subunits was chosen to roughly match the vine spacing. The beginning and end of each vineyard and forest transect was marked with flagging. We conducted two repeat surveys using detection dogs and two repeat surveys using human observers. Each survey type (i.e., humans or dogs) was performed on the same day with two different observers/dogs at each site, but was timed so that observers/dogs were not surveying the same transect at the same time. The dog and human teams surveyed the site on different days, with only one visit by a dog/human per day to minimize the possibility of dogs keying in on scents left by another survey team. Surveys were completed using a total of three human observers and two dogs. To keep visits close to each other in time, we split the 20 sites into three groups of 6–7 vineyards per group and completed both the human and dog surveys for a group of sites before moving on to the next group. For each site, we recorded start time and end time for searching each transect for each observer. Human searchers thoroughly searched grapevine trunks, branches, canes, and the ground immediately surrounding each vine. Each pole or post in the vineyard transect was also searched thoroughly. For each vine and pole, we recorded whether an egg mass was detected and recorded the spatial location. Each dog wore a GPS unit, and the dog track was recorded. Both humans and dogs were allowed unlimited search time. A dog team consisted of the detection dog and a handler. Dogs searched on-leash and were directed to search vineyard rows and forest transects (Figure 2). If a dog signaled, the handler confirmed the dog detection and the dog was rewarded by a ball thrown for them upon positive confirmation of an egg mass. Forest transects were surveyed by searching for any tree, shrub, or searchable object (e.g., boulders, stumps) within a 2-m-wide belt transect. Any assessable tree branches that were overhanging the transect were searched.

FIGURE 2

(a) Detection dog and handler and co-author J. Beese searching a vineyard transect for spotted lanternfly (Lycorma delicatula) egg masses, showing vines and metal support poles. (b) Spotted lanternfly egg mass on the underside of a grape vine. Photo credits: Angela Fuller.

Modeling approach

We employed an occupancy modeling framework, where occupancy is defined as the probability that SLF occupies a sample unit during a specified period of time during which the occupancy state is assumed to be static. Occupancy models account for imperfect detection of the target species by repeat sampling of a site (MacKenzie et al., 2017). Occupancy models estimate ψ or the probability that a unit is occupied by SLF, and p, or the probability of detecting SLF at an occupied unit. Multiscale occupancy models (Nichols et al., 2008) extend occupancy models to allow for hierarchically structured sample units and account for site dependence at the lower levels, which we expected for both individual vines and 1-m forest transect segments in the same transect. We use multiscale occupancy models to estimate occupancy at the transect level, which we relate to the probability of invasion from a source population and occupancy at the subunit level (individual vines and poles in vineyards and 1-m forest segments), and then relate to the intensity of infestation. By using both detection dogs and human observers, we estimate detection probability for both survey methods and ultimately gain an insight into which method is the most effective in both vineyards and forested areas.

Data structure

We organized the human and dog detection data from the 20 sites into three data structures: vines, poles, and forest We defined survey units to be each transect and survey subunits (hereafter “subunits”) to be each vine or pole in vineyard sites or 1-m transect segments in forested sites. We assigned human detection data to occasions 1 and 2 and dog detection data to occasions 3 and 4. The resulting detection data for vines, poles, and forest at site $$$ j $$$ and occasion $$$ k $$$ are $$$ {y}_{jk}^{\mathrm{vine}} $$$, $$$ {y}_{jk}^{\mathrm{pole}} $$$, and $$$ {y}_{jk}^{\mathrm{forest}} $$$, respectively, for $$$ j=1,\dots, J=240 $$$ and $$$ k=1,\dots, K=4 $$$. Note that the $$$ j $$$ indices for vine and pole data correspond to the vineyard transects and the $$$ j $$$ indices for the forest data correspond to the forest transects. Each data element records the number of subunits with detections for transect $$$ j $$$ and occasion $$$ k $$$. The total numbers of subunits for transect $$$ j $$$ are $$$ {N}_j^{\mathrm{vine}} $$$, $$$ {N}_j^{\mathrm{pole}} $$$, and $$$ {N}_j^{\mathrm{forest}} $$$ for vines, poles, and forest, respectively. Finally, we mapped transect $$$ j $$$ to its respective site (vineyard) using the site indicator site_j.

State model

We used a multiscale occupancy model, nesting the subunits within transects. In this model, transects may or may not be occupied and each subunit may only be occupied if their associated transect is occupied. This nested structure accounts for the component of the correlation in occupancy probability between subunits in the same transect stemming from the transect occupancy status itself (i.e., if a transect is unoccupied, each subunit must be unoccupied). We defined the transect-level occupancy states for the vineyard and forest as $$$ {z}_j^{\mathrm{vine}} $$$ and $$$ {z}_j^{\mathrm{for}} $$$, respectively, for $$$ j=1,\dots, J $$$. We then defined the subunit-level occupancy states for vines, poles, and forest as $$$ {w}_j^{\mathrm{vine}} $$$, $$$ {w}_j^{\mathrm{pole}} $$$, and $$$ {w}_j^{\mathrm{for}} $$$, also for $$$ j=1,\dots, J $$$. The transect-level occupancy states are binary, taking values 0 or 1, and the survey unit-level occupancy states are binomial, taking any value between 0 and $$$ {N}^{\mathrm{type}} $$$ for each data type (type = vine, pole, or forest), reflecting the total numbers of vines, poles, or forest transect meters in each transect. At the transect level, we assumed $$$ {z}_j^{\mathrm{vine}}\sim \mathrm{Bern}\left({\uppsi}_j^{z_{\mathrm{vine}}}\right) $$$ and $$$ {z}_j^{\mathrm{for}}\sim \mathrm{Bern}\left({\uppsi}_j^{z_{\mathrm{for}}}\right) $$$ for vineyard and forest transects, respectively. Both $$$ {\uppsi}_j^{z_{\mathrm{vine}}} $$$ and $$$ {\uppsi}_j^{z_{\mathrm{forest}}} $$$ were modeled as functions of landscape covariates, described below. Then at the subunit level, we conditioned the vine and pole $$$ w $$$ states on the vineyard $$$ z $$$ state and the forest $$$ w $$$ state on the forest $$$ z $$$ state as follows:

$$$$ \left[{w}_j^{\mathrm{vine}}|{z}_j^{\mathrm{vine}}\right]\sim \mathrm{Binomial}\left({\uppsi}_j^{w_{\mathrm{vine}}}\times {z}_j^{\mathrm{vine}},{N}_j^{\mathrm{vine}}\right). $$$$

$$$$ \left[{w}_j^{\mathrm{pole}}|{z}_j^{\mathrm{vine}}\right]\sim \mathrm{Binomial}\left({\uppsi}_j^{w_{\mathrm{pole}}}\times {z}_j^{\mathrm{vine}},{N}_j^{\mathrm{pole}}\right). $$$$

$$$$ \left[{w}_j^{\mathrm{for}}|{z}_j^{\mathrm{for}}\right]\sim \mathrm{Binomial}\left({\uppsi}_j^{w_{\mathrm{for}}}\times {z}_j^{\mathrm{for}},{N}_j^{\mathrm{for}}\right). $$$$

State model—Infestation level

We hypothesized that both (1) transect and subunit-level occupancy rates vary across sites as a function of the overall lanternfly infestation level at a site and (2) human and dog detection probability should be higher at sites with higher infestation levels because there should be more lanternfly eggs per subunit available for detection. We were unable to quantify infestation levels independently of the detection data, so we used a finite mixture model (Pledger, 2000) to classify the sites into two types based jointly on their estimated occupancy and detection probabilities that are described further below. We used a two-component mixture to estimate the site-level mixture components assuming $$$ {\upzeta}_j\sim \mathrm{Categorical}\left(\boldsymbol{\uppi} \right) $$$ where $$$ \boldsymbol{\uppi} $$$ is of length two, corresponding to the probabilities sites are level 1 (moderate infestation) or level 2 (high infestation). A grouping of sites with higher occupancy probabilities having higher detection probabilities would be consistent with our infestation level hypotheses.

State model—Transects

We modeled the transect-level expected occupancy probabilities as varying between major habitat type (vineyard vs. forest), as a function of the site-level infestation level, as a function of the interaction between habitat type and infestation level, and as a function of two habitat covariates, topographic position index (TPI) and distance to forest (included for the vineyard only). The TPI value was calculated by comparing the elevation of a site to its surrounding neighbors, with positive values indicating higher elevation than the surrounding terrain (i.e., higher relief). Areas with higher relief could serve as landing sites for dispersing adult SLF. Habitat covariates were measured at the transect midpoints. We hypothesized that lanternfly occupancy would be highest closer to the forest within the vineyard. The vineyard transect-level expected occupancy probability for transect $$$ j $$$, $$$ {\uppsi}_j^{z_{\mathrm{vine}}} $$$, is as follows:

$$$$ \mathrm{logit}\left({\uppsi}_j^{z_{\mathrm{vine}}}\right)={\upbeta}_0^z+{\upbeta}_1^z\left({\upzeta}_j-1\right)+{\upbeta}_2^z\mathrm{Dist}2\mathrm{Fo}{\mathrm{r}}_j+{\upbeta}_3^z\mathrm{TP}{\mathrm{I}}_j+{\upbeta}_4^z\mathrm{TP}{\mathrm{I}}_j^2, $$$$

where $$$ \mathrm{Dist}2\mathrm{Fo}{\mathrm{r}}_j $$$ is the distance to forest of transect $$$ j $$$, $$$ \mathrm{TP}{\mathrm{I}}_j $$$ is the topographic position index of transect $$$ j $$$, and $$$ {\upzeta}_j $$$ modulates the site-level high infestation offset, $$$ {\upbeta}_1^z $$$, zeroing it out for $$$ {\upzeta}_j=1 $$$ and including it for $$$ {\upzeta}_{\mathrm{j}}=2 $$$. Similarly, the forest transect-level occupancy probability for transect $$$ j $$$, $$$ {\uppsi}_j^{z_{\mathrm{for}}} $$$, is as follows:

$$$$ \mathrm{logit}\left({\uppsi}_j^{z_{\mathrm{for}}}\right)={\upbeta}_0^z+{\upbeta}_5^z+\left({\upbeta}_1^z+{\upbeta}_6^z\right)\left({\upzeta}_j-1\right)+{\upbeta}_3^z\mathrm{TP}{\mathrm{I}}_j+{\upbeta}_4^z\mathrm{TP}{\mathrm{I}}_j^2, $$$$

where the intercept $$$ {\upbeta}_0^z $$$ is shared with the vineyard transects, $$$ {\upbeta}_5^z $$$ is the offset for the forest habitat type relative to the vineyard, and $$$ {\upbeta}_6^z $$$ is the interaction between high infestation and the forest habitat. TPI effects are shared with the vineyard transects.

State model—Subunits

We included the same effects for the subunit occupancy probabilities, except we further differentiate subunit detection within the vineyard by substrate type, including vines (20 vines per transect), metal poles (180 transects), wood poles (60 transects), and “other” poles (28 transects), which were a mixture of wood, metal, and/or concrete poles. The subunit-level occupancy probability for vines in transect $$$ j $$$, $$$ {\uppsi}_j^{w_{\mathrm{vine}}} $$$, is as follows:

$$$$ \mathrm{logit}\left({\uppsi}_j^{w_{\mathrm{vine}}}\right)={\upbeta}_0^w+{\upbeta}_1^w\left({\upzeta}_j-1\right)+{\upbeta}_2^w\mathrm{Dist}2\mathrm{Fo}{\mathrm{r}}_j+{\upbeta}_3^w\mathrm{TP}{\mathrm{I}}_j+{\upbeta}_4^w\mathrm{TP}{\mathrm{I}}_j^2, $$$$

with the covariates and parameters sharing the same meaning as they do in the transect occupancy process above. Then, the subunit-level occupancy probability for poles in transect $$$ j $$$, $$$ {\uppsi}_j^{w_{\mathrm{pole}}} $$$, is as follows:

$$$$ \mathrm{logit}\left({\uppsi}_j^{w_{\mathrm{pole}}}\right)={\upbeta}_0^w+{\upbeta}_1^w\left({\upzeta}_j-1\right)+{\upbeta}_2^w\mathrm{Dist}2\mathrm{Fo}{\mathrm{r}}_j+{\upbeta}_3^w\mathrm{TP}{\mathrm{I}}_j+{\upbeta}_4^w\mathrm{TP}{\mathrm{I}}_j^2+{\upbeta}_7^w+{\upbeta}_8^w{I}_j^{\mathrm{metal}}+{\upbeta}_9^w{I}_j^{\mathrm{other}}, $$$$

where $$$ {\upbeta}_7^w $$$ is the offset for poles from vines, $$$ {\upbeta}_8^w $$$ is the offset for metal poles relative to wood, $$$ {I}_j^{\mathrm{metal}} $$$ is a transect-level indicator for metal poles, $$$ {\upbeta}_9^w $$$ is the offset for “other” poles relative to wood, and $$$ {I}_j^{\mathrm{other}} $$$ is a transect-level indicator for “other” poles. Finally, the subunit-level expected occupancy probability for forest transect $$$ j $$$, $$$ {\uppsi}_j^{w_{\mathrm{for}}} $$$, is as follows:

$$$$ \mathrm{logit}\left({\uppsi}_j^{w_{\mathrm{for}}}\right)={\upbeta}_0^w+{\upbeta}_5^w+\left({\upbeta}_1^w+{\upbeta}_6^w\right)\left({\upzeta}_j-1\right)+{\upbeta}_3^w\mathrm{TP}{\mathrm{I}}_j+{\upbeta}_4^w\mathrm{TP}{\mathrm{I}}_j^2, $$$$

where $$$ {\upbeta}_5^w $$$ is the offset for the forest habitat type relative to the vines and $$$ {\upbeta}_6^w $$$ is the interaction between high infestation and the forest habitat.

Detection model

We assumed detection conditional on subunit occupancy varies by subunit type (vine, pole, forest), between observer type (human or dog), between observer ID (three humans, two dogs), and as a function of infestation level and weather covariates. The weather covariates we considered were (1) snow (presence/absence), (2) precipitation when the site was surveyed (presence/absence), and (3) wind (only for dog observers). We considered an interaction between observer type and habitat type to test the hypothesis that humans will compare more favorably with dogs in the vineyard because we expected the vineyard rows are easier for humans to search effectively than the forest. We also considered an interaction between observer type and precipitation and snow because they may affect visual detection for humans and olfactory detection for dogs differently. The expected detection probability for vine, pole, and forest subunits in transect $$$ j $$$ on occasion $$$ k $$$ for human occasions (1 and 2) is as follows:

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{vine}}\right)={\upbeta}_0^p+{\upbeta}_2^p\left({\upzeta}_j-1\right)+{\upbeta}_6^p{I}_{jk}^{H2}+{\upbeta}_7^p{I}_{jk}^{H3}+{\upbeta}_9^p{I}_{jk}^{\mathrm{precip}}+{\upbeta}_{11}^p{I}_{jk}^{\mathrm{snow}}, $$$$

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{pole}}\right)={\upbeta}_0^p+{\upbeta}_2^p\left({\upzeta}_j-1\right)+{\upbeta}_4^p+{\upbeta}_6^p{I}_{jk}^{H2}+{\upbeta}_7^p{I}_{jk}^{H3}+{\upbeta}_9^p{I}_{jk}^{\mathrm{precip}}+{\upbeta}_{11}^p{I}_{jk}^{\mathrm{snow}}, $$$$

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{for}}\right)={\upbeta}_0^p+{\upbeta}_1^p+\left({\upbeta}_2^p+{\upbeta}_3^p\right)\left({\upzeta}_j-1\right)+{\upbeta}_6^p{I}_{jk}^{H2}+{\upbeta}_7^p{I}_{jk}^{H3}+{\upbeta}_9^p{I}_{jk}^{\mathrm{precip}}+{\upbeta}_{12}^p{I}_{jk}^{\mathrm{snow}}, $$$$

where $$$ {I}_{jk}^{H2} $$$ and $$$ {I}_{jk}^{H3} $$$ are the site-by-occasion-level indicators taking value 1 if “human 2” or “human 3” surveyed site $$$ j $$$ on occasion $$$ k $$$, respectively, and 0 otherwise. Similarly, $$$ {I}_{jk}^{\mathrm{precip}} $$$ and $$$ {I}_{jk}^{\mathrm{snow}} $$$ are the site-by-occasion indicators taking value 1 if there was precipitation or snow when site $$$ j $$$ was surveyed on occasion $$$ k $$$, respectively, and 0 otherwise. As for the occupancy processes, $$$ {\upzeta}_j $$$ modulates the site-level high infestation offset, $$$ {\upbeta}_1^p $$$, zeroing it out for $$$ {\upzeta}_j=1 $$$ and including it for $$$ {\upzeta}_j=2 $$$. Finally, $$$ {\upbeta}_1^p $$$ is the offset for forest, $$$ {\upbeta}_4^p $$$ is the offset for pole (not differentiated by pole type), and $$$ {\upbeta}_3^p $$$ is the interaction between forest and high infestation. The expected detection probability for vine, pole, and forest subunits in transect $$$ j $$$ on occasion $$$ k $$$ for human occasions (3 and 4) is as follows:

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{vine}}\right)={\upbeta}_0^p+{\upbeta}_2^p\left({\upzeta}_j-1\right)+{\upbeta}_5^p+{\upbeta}_8^p{I}_{jk}^{D2}+\left({\upbeta}_9^p+{\upbeta}_{10}^p\right){I}_{jk}^{\mathrm{precip}}+\left({\upbeta}_{11}^p+{\upbeta}_{12}^p\right){I}_{jk}^{\mathrm{snow}}+{\upbeta}_{13}^p\mathrm{win}{\mathrm{d}}_{jk}, $$$$

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{pole}}\right)={\upbeta}_0^p+{\upbeta}_2^p\left({\upzeta}_j-1\right)+{\upbeta}_4^p+{\upbeta}_5^p+{\upbeta}_8^p{I}_{jk}^{D2}+\left({\upbeta}_9^p+{\upbeta}_{10}^p\right){I}_{jk}^{\mathrm{precip}}+\left({\upbeta}_{11}^p+{\upbeta}_{12}^p\right){I}_{jk}^{\mathrm{snow}}+{\upbeta}_{13}^p\mathrm{win}{\mathrm{d}}_{jk}, $$$$

$$$$ \mathrm{logit}\left({p}_{jk}^{\mathrm{for}}\right)={\upbeta}_0^p+{\upbeta}_1^p+\left({\upbeta}_2^p+{\upbeta}_3^p\right)\left({\upzeta}_j-1\right)+{\upbeta}_5^p+{\upbeta}_8^p{I}_{jk}^{D2}+\left({\upbeta}_9^p+{\upbeta}_{10}^p\right){I}_{jk}^{\mathrm{precip}}+\left({\upbeta}_{11}^p+{\upbeta}_{12}^p\right){I}_{jk}^{\mathrm{snow}}+{\upbeta}_{13}^p\mathrm{win}{\mathrm{d}}_{jk}, $$$$

where $$$ \mathrm{win}{\mathrm{d}}_{jk} $$$ is the wind speed recorded for transect $$$ j $$$ on occasion $$$ k $$$ during dog surveys and $$$ {I}_{jk}^{D2} $$$ is a site-by occasion indicator taking value 1 if dog observer 2 surveyed transect $$$ j $$$ on occasion $$$ k $$$ and 0 otherwise. $$$ {\upbeta}_5^p $$$ is the offset for dogs, $$$ {\upbeta}_{10}^p $$$ is the interaction between dogs and precipitation, and $$$ {\upbeta}_{12}^p $$$ is the interaction between dogs and snow. Because the wind speed was not recorded for 9/480 and 55/480 transect-visits in the vineyard and forest areas, respectively, we treated the missing data as latent variables to be estimated along with the other model parameters (Bonner & Schwarz, 2006), assuming $$$ \mathrm{win}{\mathrm{d}}_{jk}^{\mathrm{vine}}\sim \mathrm{Lognormal}\left({\upmu}^{\mathrm{vine}},{\upsigma}^{\mathrm{vine}}\right) $$$ and $$$ \mathrm{win}{\mathrm{d}}_{jk}^{\mathrm{for}}\sim \mathrm{Lognormal}\ \left({\upmu}^{\mathrm{for}},{\upsigma}^{\mathrm{for}}\right) $$$ for vine and forest wind speeds, respectively.

Search time

We compared the search time and efficiency between humans and dogs as a function of major habitat type (vineyard vs. forest) and infestation level. Both human and dog teams recorded the time it took to search 12 vineyard and 12 forest transects at each site. To produce observer by habitat by infestation level estimates of mean search time, we fit a linear model with the linear predictor $$$ {\upmu}_i^t={\upbeta}_0^t+{\upbeta}_1^t{I}_i^{\mathrm{Dog}}+{\upbeta}_2^t{I}_i^{\mathrm{Forest}}{\upbeta}_3^t{I}_i^{\mathrm{HighInfest}}+{\upbeta}_4^t{I}^{\mathrm{Dog}}{I}_i^{\mathrm{Forest}}+{\upbeta}_4^t{I}_i^{\mathrm{Forest}}{I}_i^{\mathrm{HighInfest}}+{\upbeta}_4^t{I}_i^{\mathrm{Dog}}{I}_i^{\mathrm{HighInfest}} $$$, where $$$ {I}_i^{\mathrm{Dog}} $$$ is an indicator for “dog,” $$$ {I}_i^{\mathrm{Forest}} $$$ is an indicator for “forest,” and $$$ {I}_i^{\mathrm{HighInfest}} $$$ is an indicator for “high infestation.” Then, we assumed the search time for transect i was $$$ {y}_i^t\sim \mathrm{Normal}\left({\upmu}_i^t,{\upsigma}^t\right) $$$.

Search efficiency

To compare how efficient humans and dogs were at detecting lanternflies, we estimated their search efficiency as a function of infestation level and major habitat type (vineyard or forest). We defined search efficiency to be the expected number of detections per hour. Because search time was recorded per 12 transects, either in the vineyard or in forest, search efficiency can be computed by dividing an estimate of the expected number of detections across 12 transects by the estimated search time. To obtain the expected number of detections across 12 transects, we first produced a posterior for the unconditional occupancy probabilities. Our multiscale occupancy model calculated the subunit occupancy probabilities conditional on the transect-level occupancy process; however, we require unconditional occupancy probabilities for the subunits, which are averaged over the transect-level occupancy states (Nichols et al., 2008). Because we use the Bayesian estimation, we multiplied the transect and subunit-level occupancy probability posteriors to produce a posterior for the expected unconditional occupancy probabilities. For example, to compute the unconditional occupancy probability of a vine at a moderate intensity site for MCMC iteration $$$ i $$$, we compute $$$ {\uppsi}_{\mathrm{vine}\mathrm{y}\mathrm{ard},\mathrm{moderateintensit}\mathrm{y}{\mathrm{y}}_i}^u={\uppsi}_{\mathrm{vine}\mathrm{y}\mathrm{ard},\mathrm{moderateintensit}{\mathrm{y}}_i}^z{\uppsi}_{\mathrm{vine},\mathrm{moderateintensit}{\mathrm{y}}_i}^w $$$. We can then make posterior inference from the unconditional occupancy posteriors and also produce the search efficiency posteriors. The next step is to produce posteriors for the expected number of detections per subunit. To do this, we multiplied the unconditional occupancy posterior by the posterior for the detection probability. For example, to compute the expected number of detections per subunit or vines at moderate intensity sites for MCMC iteration $$$ i $$$, we computed $$$ {\upmu}_{\mathrm{vine}\mathrm{yard},\mathrm{moderateintensity}{\mathrm{y}}_i}^d={\uppsi}_{\mathrm{vine}\mathrm{yard},\mathrm{moderateintensity}{\mathrm{y}}_i}^u{p}_{\mathrm{vine},\mathrm{moderateintensity}{y}_i} $$$. Then, the expected number of detections across 12 transects in the vineyard or forest for iteration $$$ i $$$ was calculated by multiplying $$$ {\upmu}_i^d $$$ by the mean number of subunits in a vineyard or forest transect. Finally, the search efficiency posterior (expected detection per hour) for iteration $$$ i $$$ is obtained by dividing the ith iteration of the posterior for the expected number of detections across 12 transects by the ith iteration of the posterior for the search time.

Model fitting and parameter estimation

We fit the multiscale occupancy model in the program Nimble (de Valpine et al., 2017). The code and results can be found on ScienceBase (https://doi.org/10.5066/P1744ZNW). We ran three chains for 200,000 iterations, enough to ensure the Gelman–Rubin statistic (Gelman & Rubin, 1992) upper 95% CI was 1.01 or lower for all parameters and discarded 10,000 samples as burn in. For the search time analysis, we fit the model using the MCMCglmm R package (Hadfield, 2010) following the same protocol. Posterior modes were used as point estimates and 95% highest posterior density (HPD) intervals were used for interval estimates.

RESULTS

We surveyed 12 transects in each vineyard ($$$ \overline{x} $$$ = 20 m, range = 20–21 m) and 12 transects in the forest adjacent to the vineyard ($$$ \overline{x} $$$= 26.6 m, range = 17–38 m) at each of the 20 sites. The finite mixture allocated eight sites to “high infestation” and 12 sites to “moderate infestation” with posterior probabilities of 0.995 or greater, indicating nearly no uncertainty in assigning sites to the two infestation levels.

Transect-level occupancy

At the transect level, the occupancy probability was lower, on average, in the forest (0.85) than in vineyard transects (0.94) and higher at high infestation sites than at moderate infestation sites, with no support for an interaction between forest and high infestation (Table 1). We found a negative relationship between occupancy probability and distance to forest (only modeled in the vines; Figure 3). Occupancy probability of vineyard transects was >0.82 (95% CI: 0.72–0.92) if the transect was <75 m from the forest edge (Figure 3). There was a positive relationship between occupancy and TPI, with a quadratic term that was consistent with 0.

TABLE 1. Spotted lanternfly (Lycorma delicatula) occupancy process parameter estimates at the transect and subunit levels on the logit scale.

Parameter	Transect			Subunit
	Est	Lower	Upper	Est	Lower	Upper
Intercept	$$$ 2.02 $$$	$$$ 1.33 $$$	$$$ 2.95 $$$	$$$ 0.93 $$$	$$$ 1.06 $$$	$$$ 0.79 $$$
Forest	$$$ 1.48 $$$	$$$ 2.28 $$$	$$$ 0.48 $$$	$$$ 1.82 $$$	$$$ 2.20 $$$	$$$ 1.22 $$$
High infest	$$$ 2.87 $$$	$$$ 1.83 $$$	$$$ 4.49 $$$	$$$ 1.40 $$$	$$$ 1.25 $$$	$$$ 1.56 $$$
Forest × high infest	$$$ 0.42 $$$	$$$ 1.99 $$$	$$$ 1.47 $$$	$$$ 0.02 $$$	$$$ 0.56 $$$	$$$ 0.48 $$$
Dist to forest	$$$ 0.70 $$$	$$$ 1.14 $$$	$$$ 0.29 $$$	$$$ 0.08 $$$	$$$ 0.00 $$$	$$$ 0.16 $$$
TPI	$$$ 1.49 $$$	$$$ 0.76 $$$	$$$ 5.77 $$$	$$$ 0.74 $$$	$$$ 0.61 $$$	$$$ 0.86 $$$
TPI^2	$$$ 0.24 $$$	$$$ 0.54 $$$	$$$ 3.62 $$$	$$$ 0.66 $$$	$$$ 0.78 $$$	$$$ 0.54 $$$
Pole	…	…	…	$$$ 0.37 $$$	$$$ 0.01 $$$	$$$ 0.67 $$$
Pole—wood	…	…	…	$$$ 1.77 $$$	$$$ 2.17 $$$	$$$ 1.41 $$$
Pole—other	…	…	…	$$$ 2.10 $$$	$$$ 2.74 $$$	$$$ 1.54 $$$

• Note: “Est” is the posterior mode, and “Lower” and “Upper” are the lower and upper 95% highest posterior density (HPD) intervals. Factor intercept levels are “vine” and “moderate infestation.” “Forest” is the offset for the forest area, “High Infest” is the offset for high infestation sites, and “Forest × High Infest” is the interaction between forest and high infestation sites. “Dist to Forest” is the effect of distance to forest (only included for the vineyard area), and “TPI” and “TPI²” are the linear and quadratic effects of the topographic position index. For subunit-level occupancy in the vineyard area, “Pole” is the offset from “vine” and “Pole—Wood” and “Pole—Other” are the offsets for wood and other pole types from “Pole.” No poles were in the forested area.

FIGURE 3

Probably of occupancy, p(occupy), of spotted lanternfly (Lycorma delicatula) egg masses on vineyard transects as a function of distance to the forest edge.

Subunit-level occupancy

At the subunit level, we found the same directional relationships for forest, high infestation, and their interaction as we did at the transect level. Average subunit occupancy was 0.47 in vineyard transects and 0.13 in forest transects. In contrast with the transect level, we found the effect of distance to forest to be negative at the subunit level, although the lower 95% CI bound covered 0, and the TPI relationship was more strongly quadratic (Table 1). Metal poles had a higher occupancy probability than vines, but the occupancy probability for wood and “other” poles was consistent with that for vines. Finally, the unconditional occupancy probabilities for substrate type by infestation level are depicted in Figure 4, with selected pairwise comparisons in Table 2. We found that vines were more likely to be occupied than forest, and metal poles were more likely to be occupied than vines at both infestation levels.

FIGURE 4

Unconditional occupancy probability posteriors, p(subunit occupied), for vines, metal poles, and forest subunits at moderate and high infestation sites with topographic position index (TPI) set to the mean value and distance to forest for vines set to the mean value. Overlapping posteriors do not necessarily indicate differences consistent with 0—see Table 4 for pairwise comparisons.

TABLE 2. Pairwise comparisons for spotted lanternfly (Lycorma delicatula) transect-level, subunit-level, and unconditional occupancy parameters with distance to forest and topographic position index set to the mean values.

Type	Comparison	Est	Upper	Lower
Transect	Vine vs. forest|moderate infest	$$$ 0.21 $$$	$$$ 0.01 $$$	$$$ 0.36 $$$
Transect	Vine vs. forest|high infest	$$$ 0.01 $$$	$$$ 0.01 $$$	$$$ 0.06 $$$
Transect	High vs. moderate infest|vine	$$$ 0.16 $$$	$$$ 0.07 $$$	$$$ 0.25 $$$
Transect	High vs. moderate infest|forest	$$$ 0.37 $$$	$$$ 0.09 $$$	$$$ 0.52 $$$
Subunit	Vine vs. forest|moderate infest	$$$ 0.23 $$$	$$$ 0.19 $$$	$$$ 0.27 $$$
Subunit	Vine vs. forest|high infest	$$$ 0.41 $$$	$$$ 0.37 $$$	$$$ 0.45 $$$
Subunit	Vine vs. metal pole|moderate infest	$$$ 0.07 $$$	$$$ 0.15 $$$	$$$ 0.00 $$$
Subunit	Vine vs. metal pole|high infest	$$$ 0.08 $$$	$$$ 0.14 $$$	$$$ 0.01 $$$
Subunit	High vs. moderate infest|vine	$$$ 0.34 $$$	$$$ 0.30 $$$	$$$ 0.37 $$$
Subunit	High vs. moderate infest|forest	$$$ 0.15 $$$	$$$ 0.11 $$$	$$$ 0.19 $$$
Unconditional	Vine vs. forest|moderate infest	$$$ 0.20 $$$	$$$ 0.17 $$$	$$$ 0.24 $$$
Unconditional	Vine vs. forest|high infest	$$$ 0.41 $$$	$$$ 0.37 $$$	$$$ 0.45 $$$
Unconditional	Vine vs. metal pole|moderate infest	$$$ 0.06 $$$	$$$ 0.13 $$$	$$$ 0.00 $$$
Unconditional	Vine vs. metal pole|high infest	$$$ 0.08 $$$	$$$ 0.14 $$$	$$$ 0.01 $$$
Unconditional	High vs. moderate infest|vine	$$$ 0.38 $$$	$$$ 0.34 $$$	$$$ 0.42 $$$
Unconditional	High vs. moderate infest|forest	$$$ 0.17 $$$	$$$ 0.14 $$$	$$$ 0.20 $$$

• Note: “Est.” is the posterior mode in probability units, and “Lower” and “Upper” are the 95% HPD interval lower and upper bounds.
• Abbreviation: HPD, highest posterior density.

Detection

Detection probability was lower in the forest and higher at high infestation sites with a greater difference in the vines than in the forest (negative forest × high infest interaction; Table 3). Dogs had a lower detection probability (0.45) than humans (0.85) in the vines, but their detection was more than three times higher than that of humans in the forest (dog forest = 0.51, human forest = 0.15) (Figure 5). We found small but supported differences in detection probabilities between the three human observers, but no difference between the two dog observers (Table 3, Figure 6). We also found that detection probability was higher for vines than for poles.

TABLE 3. Spotted lanternfly (Lycorma delicatula) detection process parameter estimates on the logit scale.

Parameter	Est	Lower	Upper
Intercept	$$$ 0.85 $$$	$$$ 0.64 $$$	$$$ 1.10 $$$
Forest	$$$ -2.11 $$$	$$$ -2.62 $$$	$$$ -1.71 $$$
High infest	$$$ 0.90 $$$	$$$ 0.65 $$$	$$$ 1.17 $$$
Forest × high infest	$$$ -0.63 $$$	$$$ -1.11 $$$	$$$ -0.12 $$$
Dog	$$$ -1.52 $$$	$$$ -1.75 $$$	$$$ -1.29 $$$
Dog × forest	$$$ 2.51 $$$	$$$ 2.29 $$$	$$$ 2.73 $$$
Dog × high infest	$$$ 0.10 $$$	$$$ -0.13 $$$	$$$ 0.33 $$$
Dog—Obs 2	$$$ -0.17 $$$	$$$ -0.30 $$$	$$$ -0.05 $$$
Human—Obs 2	$$$ -0.17 $$$	$$$ -0.37 $$$	$$$ 0.04 $$$
Human—Obs 3	$$$ 0.26 $$$	$$$ 0.09 $$$	$$$ 0.42 $$$
Pole	$$$ -0.43 $$$	$$$ -0.66 $$$	$$$ -0.27 $$$
Precip	$$$ -0.06 $$$	$$$ -0.30 $$$	$$$ 0.17 $$$
Precip × dog	$$$ -0.17 $$$	$$$ -0.97 $$$	$$$ 0.66 $$$
Snow	$$$ 0.20 $$$	$$$ -0.08 $$$	$$$ 0.44 $$$
Snow × dog	$$$ -0.41 $$$	$$$ -0.73 $$$	$$$ -0.11 $$$
Wind (dog only)	$$$ 0.09 $$$	$$$ 0.05 $$$	$$$ 0.12 $$$

• Note: “Est” is the posterior mode, and “Lower” and “Upper” are the lower and upper 95% HPD intervals. Factor intercept levels are “Vine,” “Moderate Infestation,” “Human,” “No Snow,” and “No Precipitation.” “Forest” is the offset from vines, “High Infest” is the offset from moderate infestation, and “Dog” is the offset from Humans. “Forest × High Infest” is the interaction between forest and high infestation, “Dog × Forest” is the interaction between dogs and forest, and “Dog × High Infest” is the interaction between dogs and high infestation. There is one offset for the second dog observer and two offsets for the second and third human observers. “Precip” and “Snow” are the offsets from “No Precipitation” and “No Snow,” respectively, and “Precip × Dog” and “Snow × Dog” are the interactions between dogs and precipitation and snow, respectively. “Wind” is the effect of wind speed on dog detection only.
• Abbreviation: HPD, highest posterior density.

FIGURE 5

Detection probability posteriors, p(detect), for humans and dogs at moderate and high infestation sites. Both human and dog posteriors are averaged across equally weighted observers. Overlapping posteriors do not necessarily indicate differences consistent with 0. See Table 4 for pairwise comparisons.

FIGURE 6

Detection probability posteriors, p(detect), for three human observers and two dog observers at moderate and high infestation sites for vines and forest.

For weather effects, we found that dogs had a lower detection probability in the presence of snow (adding the snow and dog by snow interaction yields a log odds of −0.23, 95% CI −0.43 to −0.06), and a higher detection probability at greater wind speeds, which were both higher and more variable in the vineyard area versus the forest (Table 4). The detection probability posteriors can be found in Figure 5, and the posterior probabilities for selected pairwise comparisons are included in Table 5.

TABLE 4. Other spotted lanternfly (Lycorma delicatula) parameter estimates.

Parameter	Est	Lower	Upper
Mixture—p(moderate infest)	$$$ 0.60 $$$	$$$ 0.39 $$$	$$$ 0.79 $$$
Wind—vineyard mean	$$$ 0.64 $$$	$$$ 0.58 $$$	$$$ 0.70 $$$
Wind—forest mean	$$$ 0.29 $$$	$$$ 0.23 $$$	$$$ 0.34 $$$
Wind—vineyard SD	$$$ 0.68 $$$	$$$ 0.64 $$$	$$$ 0.73 $$$
Wind—forest SD	$$$ 0.57 $$$	$$$ 0.53 $$$	$$$ 0.60 $$$

• Note: “Mixture” is the probability a site is a “Moderate Infestation” site on the probability scale. “Wind—Vineyard Mean” and “Wind—Forest Mean” are the mean wind speeds in the vineyard and forest during dog surveys, respectively. “Wind—Vineyard SD” and “Wind—Forest SD” are the SDs of the wind speeds in the vineyard and forest during dog surveys, respectively.

TABLE 5. Pairwise comparisons for the spotted lanternfly (Lycorma delicatula) detection parameters averaged across human and dog observers, with no precipitation or snow and “ind” set to the mean value for dogs.

Comparison	Est	Upper	Lower
Human vs. dog|moderate infest, vine	$$$ 0.35 $$$	$$$ 0.30 $$$	$$$ 0.39 $$$
Human vs. dog|high infest, vine	$$$ 0.25 $$$	$$$ 0.22 $$$	$$$ 0.28 $$$
Human vs. dog|moderate infest, forest	$$$ 0.22 $$$	$$$ 0.29 $$$	$$$ 0.16 $$$
Dog vs. human|high infest, forest	$$$ 0.27 $$$	$$$ 0.31 $$$	$$$ 0.22 $$$
Vine vs. forest|moderate infest, human	$$$ 0.48 $$$	$$$ 0.40 $$$	$$$ 0.56 $$$
Vine vs. forest|high infest, human	$$$ 0.58 $$$	$$$ 0.54 $$$	$$$ 0.62 $$$
Vine vs. forest|moderate infest, dog	$$$ 0.09 $$$	$$$ 0.20 $$$	$$$ 0.03 $$$
Vine vs. forest|high infest, dog	$$$ 0.06 $$$	$$$ 0.00 $$$	$$$ 0.13 $$$
High vs. moderate infest|vine, human	$$$ 0.15 $$$	$$$ 0.10 $$$	$$$ 0.20 $$$
High vs. moderate infest|forest, human	$$$ 0.06 $$$	$$$ 0.03 $$$	$$$ 0.13 $$$
High vs. moderate infest|vine, dog	$$$ 0.25 $$$	$$$ 0.20 $$$	$$$ 0.29 $$$
High vs. moderate infest|forest, dog	$$$ 0.09 $$$	$$$ 0.02 $$$	$$$ 0.22 $$$

• Note: “Est.” is the posterior mode in probability units, and “Lower” and “Upper” are the 95% HPD interval lower and upper bounds.
• Abbreviation: HPD, highest posterior density.

Infestation level

The high infestation sites were characterized by both higher occupancy probabilities and higher detection probabilities. There was a greater difference in detection probability as a function of infestation level in the vineyard area than in the forest. Combining the occupancy and detection results, at both infestation levels there were more detections per subunit by humans in the vineyard areas than by dogs, more by dogs in the forest, and more at high infestation sites than at moderate infestation sites (Figure 7).

FIGURE 7

The expected number of detections per subunit for humans and dogs stratified by substrate type (vine vs. 1-m forest segment) and infestation level (moderate or high).

Search time/search efficiency

The mean human search time for 12 transects was 2.44 h (95% CI = 2.09–2.80) in the vineyard area and 1.30 h (95% CI = 1.08–1.53) in the forest. For dogs, the mean search time was 1.87 h (95% CI = 1.65–2.10) in the vineyard area and 2.08 h (95% CI = 1.82–2.33) in the forest. Generally, dogs took less time than humans in the vines and more in the forest, and humans took less time in the forest than in the vineyards (Table 6). Humans did not take more time at high infestation sites (2.44 h) than at moderate infestation sites (2.45 h) in the vineyards, but dogs did, with 2.3 h of search time in high infestation sites and 1.59 h in moderate infestation sites (Table 6). Within forested sites, dogs took longer at high infestation sites (2.70 h) than at moderate infestation sites (1.69 h), whereas humans searched a similar amount of time at high infestation sites (1.48 h) and moderate infestation sites (1.19 h). Selected pairwise comparisons are found in Table 7, which shows that humans were slower than dogs at moderate infestation sites in the vineyard; humans were faster than dogs in the forest at both infestation levels; and all other pairwise comparisons were consistent with no difference.

TABLE 6. Spotted lanternfly (Lycorma delicatula) search time parameter estimates in hours to search all 12 vineyard or forest transects per vineyard.

Parameter	Est	Lower	Upper
Intercept	$$$ 2.46 $$$	$$$ 2.15 $$$	$$$ 2.75 $$$
Dog	$$$ 0.88 $$$	$$$ 1.25 $$$	$$$ 0.45 $$$
Forest	$$$ 1.25 $$$	$$$ 1.67 $$$	$$$ 0.87 $$$
High infest	$$$ 0.03 $$$	$$$ 0.44 $$$	$$$ 0.43 $$$
Dog × forest	$$$ 1.36 $$$	$$$ 0.86 $$$	$$$ 1.86 $$$
Forest × high infest	$$$ 0.36 $$$	$$$ 0.20 $$$	$$$ 0.81 $$$
Dog × high infest	$$$ 0.70 $$$	$$$ 0.21 $$$	$$$ 1.23 $$$

• Note: “Est” is the posterior mode, and “Lower” and “Upper” are the lower and upper 95% HPD intervals. Factor intercept levels are “Human,” “Vine,” and “Moderate Infestation.” “Forest” is the offset from vines, “High Infest” is the offset from Moderate Infestation, and “Dog” is the offset from Humans. “Forest × High Infest” is the interaction between forest and high infestation, “Dog × Forest” is the interaction between dogs and forest, and “Dog × High Infest” is the interaction between dogs and high infestation.
• Abbreviation: HPD, highest posterior density.

TABLE 7. Pairwise comparisons for spotted lanternfly (Lycorma delicatula) search time in hours to search 12 transects in the vineyard or forest area and search efficiency in detections per hour.

Type	Comparison	Est	Upper	Lower
Time	Human vs. dog|moderate infest, vine	$$$ 0.88 $$$	$$$ 0.45 $$$	$$$ 1.25 $$$
Time	Human vs. dog|high infest, vine	$$$ 0.14 $$$	$$$ 0.33 $$$	$$$ 0.59 $$$
Time	Human vs. dog|moderate infest, forest	$$$ 0.54 $$$	$$$ 0.91 $$$	$$$ 0.10 $$$
Time	Human vs. dog|high infest, forest	$$$ 1.26 $$$	$$$ 1.68 $$$	$$$ 0.74 $$$
Time	Vine vs. forest|moderate infest, human	$$$ 1.25 $$$	$$$ 0.87 $$$	$$$ 1.67 $$$
Time	Vine vs. forest|high infest, human	$$$ 0.94 $$$	$$$ 0.49 $$$	$$$ 1.41 $$$
Time	Vine vs. forest|moderate infest, dog	$$$ 0.07 $$$	$$$ 0.50 $$$	$$$ 0.31 $$$
Time	Vine vs. forest|high infest, dog	$$$ 0.42 $$$	$$$ 0.87 $$$	$$$ 0.06 $$$
Time	High vs. moderate infest|vine, human	$$$ 0.03 $$$	$$$ 0.44 $$$	$$$ 0.43 $$$
Time	High vs. moderate infest|forest, human	$$$ 0.30 $$$	$$$ 0.14 $$$	$$$ 0.73 $$$
Time	High vs. moderate infest|vine, dog	$$$ 0.72 $$$	$$$ 0.27 $$$	$$$ 1.13 $$$
Time	High vs. moderate infest|forest, dog	$$$ 1.03 $$$	$$$ 0.55 $$$	$$$ 1.45 $$$
Efficiency	Human vs. dog|moderate infest, vine	$$$ 3.55 $$$	$$$ 0.08 $$$	$$$ 7.04 $$$
Efficiency	Human vs. dog|high infest, vine	$$$ 12.98 $$$	$$$ 3.65 $$$	$$$ 22.59 $$$
Efficiency	Human vs. dog|moderate infest, forest	$$$ 0.95 $$$	$$$ 2.14 $$$	$$$ 0.10 $$$
Efficiency	Human vs. dog|high infest, forest	$$$ 1.02 $$$	$$$ 4.50 $$$	$$$ 3.05 $$$
Efficiency	Vine vs. forest|moderate infest, human	$$$ 14.15 $$$	$$$ 11.38 $$$	$$$ 17.55 $$$
Efficiency	Vine vs. forest|high infest, human	$$$ 39.73 $$$	$$$ 31.74 $$$	$$$ 48.75 $$$
Efficiency	Vine vs. forest|moderate infest, dog	$$$ 9.56 $$$	$$$ 6.68 $$$	$$$ 13.27 $$$
Efficiency	Vine vs. forest|high infest, dog	$$$ 25.77 $$$	$$$ 19.52 $$$	$$$ 32.97 $$$
Efficiency	High vs. moderate infest|vine, human	$$$ 35.27 $$$	$$$ 27.73 $$$	$$$ 44.52 $$$
Efficiency	High vs. moderate infest|forest, human	$$$ 9.80 $$$	$$$ 6.90 $$$	$$$ 14.14 $$$
Efficiency	High vs. moderate infest|vine, dog	$$$ 26.92 $$$	$$$ 19.55 $$$	$$$ 33.67 $$$
Efficiency	High vs. moderate infest|forest, dog	$$$ 10.12 $$$	$$$ 7.65 $$$	$$$ 12.69 $$$

• Note: “Est.” is the posterior mode, and “Lower” and “Upper” are the 95% HPD interval lower and upper bounds.
• Abbreviation: HPD, highest posterior density.

Combining the results for the number of detections per subunit and search time (i.e., search efficiency), there were more detections per hour in the vineyard area and at high infestation sites (Figure 7, Table 7). Search efficiency of humans (31.4 detections/h) was higher than that of dogs (24.0 detections/h) in the vineyard area, particularly at high infestation sites; humans (7.66 detections/h) and dogs (6.72 detections/h) were similarly efficient in the forest, with the greater number of dog detections due to the greater dog search time.

DISCUSSION

Detection

Detection dogs significantly outperformed human searchers in locating SLF egg masses in forested areas, with detection rates more than three times higher than humans. This finding highlights the value of leveraging dogs' olfactory abilities in dense, complex environments where visual detection by humans is hindered. Such environments, characterized by dense undergrowth and numerous potential substrates, present considerable challenges for human observers who rely solely on sight. In contrast, detection dogs can efficiently identify even the slightest traces of a target scent (Aviles-Rosa, Kane, et al., 2023; Hoyer-Tomiczek et al., 2016; Lin et al., 2011), making them an invaluable tool for early detection of invasive species in these settings. Dogs are less encumbered by their ability to locate egg masses in dense vegetation and can effectively search for egg masses along transects, making them ideally suited to locate cryptic egg masses in challenging environments. Similar to our observations, dogs detected trees newly infested by European spruce bark beetles (Ips typographus) at greater than 100 m (Johansson et al., 2019) and detected more infested trees than human observers, particularly in “random” locations where forest attributes differ from those typically described in infested areas (Vošvrdová et al., 2023).

In vineyards, however, the scenario was reversed: Human searchers exhibited higher detection probabilities than dogs. The structured and open nature of vineyards, with accessible vines and support poles, enables humans to conduct thorough visual inspections. Although SLF egg masses are cryptic, a slow and systematic search strategy is effective for humans to detect them when the search area is easy to see, well defined, and the number of places to search is limited. However, this effectiveness of humans is reduced in forest environments where the vegetation structure limits visual search, making these conditions more suitable for dog searches using olfaction.

Detection probabilities of both humans and dogs were found to be higher on vines than on poles (made of metal and wood) within the vineyards, despite the metal poles having a higher occupancy probability than the vines. The density of egg masses on vines and poles is unknown, but it is possible that egg masses were more abundant on vines than on supporting poles. This could explain the higher detection probabilities, as suggested by the higher detection probabilities of both humans and dogs at high infestation sites than at moderate infestation sites. This finding indicates that detection is likely to be higher when there are more egg masses available for detection. This relationship between detection and density of the target species was demonstrated in a study that found a positive relationship between spongy moth egg density and the number of egg masses detected by dogs (Wallner & Ellis, 1976). This relationship underscores a key challenge in early detection of invasive species: As infestations spread and densities increase, detection becomes easier, but at that point, management options may be more limited.

Based on our sample of two dogs and three human observers, the inter-observer variance in detection probability was low. Both dogs demonstrated consistent detection probabilities, with no significant difference between them. Among the human observers, there was slight variation in detection, whereby humans 1 and 2 were comparable, but human 3 exhibited a slightly higher detection probability than human 1 (Figure 6). These results highlight the potential for achieving low inter-observer variance in detection probabilities when both dogs and humans are consistently trained. The slight variation observed among the human observers could be attributed to differences in experience, attention to detail, or familiarity with the environment. However, the overall consistency suggests that rigorous and consistent training protocols could minimize inter-observer variance.

Our objective was to find SLF egg masses on trees and vines in areas already experiencing moderate to high infestations—amid whatever chemical signals the infested areas and individual trees and vines are already producing. It is possible that at earlier stages of infestation, dog detection by olfaction may have a benefit over visual inspection by humans not only because there are fewer visible signs but the scent of infestation may be a starker contrast to the ambient forest scent.

The dispersion of scent in the environment can be affected by wind currents, which in turn can impact how dogs detect that scent (Syrotuck, 1972). In vineyards, wind speed was found to increase the detection probabilities of dogs, with a linear relationship between wind speed and detection probability. Detection probability increased from approximately 0.40 with no wind to about 0.60 with 16 km/h winds. This conforms with Hoyer-Tomiczek and Hoch (2020) who found detection sensitivity for a wood-boring tree beetle to be better at ~8 km/h than at no to minimally detected wind speeds. In contrast, detections of a tree snake in a dense jungle environment decreased with increasing wind speeds, and the authors suggest that scent pooling in the absence of wind conferred an advantage for their small-area searches (Savidge et al., 2011). Wind speed or direction did not influence dog detection of desert tortoise (Nussear et al., 2008) or mammalian carnivore scats (Reed et al., 2011) when dogs searched off leash, and the authors suggest that wind might play a role when dogs are on-lead and have less freedom to search scents from all directions. Indeed, our dogs searched transects on-lead, so increased movement of scent from egg masses yielded higher detections during our winter/early spring study. At the opposite extreme, very high wind speeds can accelerate the dispersion of scent, resulting in a reduction in odor concentrations and limiting the distances over which scents can be detected (Conover, 2007), suggesting that searches for SLF egg masses by dogs are best when conducted when winds are light to moderate.

Search time/efficiency

Efficient allocation of resources is crucial for effective invasive species management, particularly in surveillance and early detection efforts. In the quest for optimizing resource allocation, cost-effectiveness of surveillance methods can be linked to the length of time it takes to search an area and the number of target species detected (efficiency). Thus, cost-effective methods for surveillance and early detection of invasive species are an important consideration. The fact that dogs were faster but less efficient than humans in vineyards, especially at high infestation levels, suggests that while dogs can rapidly cover large areas, they may not always outperform humans in terms of detections per unit time. This finding is particularly relevant for management strategies in agricultural settings like vineyards, where time constraints and labor costs are significant considerations. However, the situation changes in forested environments, where the complexity of the terrain and the lower visibility make human detection more challenging. Here, dogs were more efficient and much more effective at detecting egg masses than humans, particularly under low infestation levels. This highlights a critical advantage of using trained detection dogs in environments where visual searches are hindered by vegetation or other factors. Early detection in such contexts is vital because low-level infestations are easier to control and less costly to manage. Thus, integrating dogs into surveillance protocols in forested areas could improve the effectiveness of early detection efforts, reducing the lag time between introduction and detection of invasive species.

Moreover, while dogs required more time to search in high infestation areas due to the increased scent concentration, this behavior might be advantageous in pinpointing the exact location of the target, providing more precise information to managers. This precision could be beneficial for targeted control measures, such as the removal of egg masses or localized chemical treatments, which are more effective and less disruptive to the ecosystem than broad-scale interventions.

Our study also points to the need for balancing detection efficiency with the broader goals of invasive species management. While search efficiency (detections per hour) is an important metric, the ultimate goal is to detect infestations early enough to implement effective control measures. In this context, the ability of dogs to detect egg masses under lower infestation levels and in complex environments, where humans struggle, could be pivotal. These findings suggest that resource managers could consider not only the cost and time associated with different detection methods but also the ecological and economic benefits of detecting infestations at an early stage.

Occupancy

Understanding the spatial patterns of SLF occupancy is crucial for developing targeted management strategies, especially in agricultural landscapes such as vineyards where infestations can have significant economic impacts. Our results reveal that SLF occupancy is not only higher in vineyards than in adjacent forests but also exhibit distinct spatial patterns within these environments. The markedly higher infestation levels in vineyards suggest that these areas offer favorable conditions for SLF, possibly due to the availability of surfaces like metal poles that facilitate egg-laying. This aligns with existing knowledge on SLF behavior, where similar smooth-barked trees and artificial structures are preferred egg-laying sites (Dara et al., 2015). The concentration of SLF egg masses near vineyard edges underscores the importance of edge habitats as entry points for infestation. This edge effect, where occupancy probabilities were highest within 75 m of the forest boundary, suggests that forested areas adjacent to vineyards act as source habitats for SLF, enabling their dispersal into cultivated areas. This finding is consistent with other studies (Leach & Leach, 2020, 2020) and highlights the need for focused surveillance and control efforts at these ecotones to prevent the spread of SLF into the interiors of vineyards. By concentrating management efforts along these vulnerable edges, it may be possible to reduce overall infestation levels, thereby protecting the economic viability of vineyards. Our study also suggests that topography influences SLF occupancy, with higher infestation probabilities observed in vineyard transects situated at relatively higher elevations. This pattern may be driven by the dispersal behavior of SLF, which can use wind currents to move to higher elevations. Such microsites might serve as initial colonization points, especially for adult SLF, due to the availability of suitable landing substrates. While the dispersal capacity of SLF nymphs is not well understood (Urban, 2020), research has shown that when SLF hatch from egg masses in trees, they are known to either fall or blow out of the tree and reclimb into the canopy (Kim et al., 2011). Recognizing these topographical hotspots within vineyards could further refine detection and control strategies, allowing for a more efficient allocation of resources by prioritizing high-risk areas.

Dog and human detection considerations

While detection dogs have proven to be effective in locating invasive species like the SLF under various conditions, their use is subject to physiological and environmental constraints (Jenkins et al., 2018; Leigh & Dominick, 2015; Reed et al., 2011). In challenging environments, such as deep snow, extreme heat, or rugged terrain, dogs may face difficulties that reduce their efficacy. During our winter/early spring study, we observed that deep snow hindered the dogs' mobility and limited their search area, as snow-covered tree and vine bases obstructed access to egg masses that would otherwise be detectable. Although dogs have been successfully trained to locate large targets under snow (MacInnes, 1967; Smith & Stirling, 1975), this capability has not been extended to smaller targets like SLF egg masses, indicating a need for specialized training protocols for such conditions. SLF often lay eggs high in trees (Keller et al., 2020; Liu & Hartlieb, 2020), and for trees that are >6 m tall, studies have shown that a majority of SLF egg masses were concentrated in the upper canopy and branches; thus, surveyors could consider the use of binoculars to detect egg masses that are not immediately in view below 6 m (Keller et al., 2020; Lewis et al., 2023). While human observers could use binoculars for SLF detection, egg masses are cryptic, making such longer distance visual detection unreliable (Lewis et al., 2023). We observed dogs detecting egg masses on tree trunks 2.5 m aboveground, but this was rare, and most detections were between the ground and the height that could be reached by the dogs standing on their back legs.

Dogs have a limited number of work hours per day depending on environmental conditions and individual stamina. Paying attention to signs of fatigue is important for ethical standards as well as performance standards. Nussear et al. (2008) reported more detections per hour for dogs finding desert tortoise than human observers and that dogs finished searching their assigned areas with fewer hours of work (6 h for dogs, 8.5 h for humans). However, this faster search time did not result in additional area searched per day. It is important to plan field activities for what can be accomplished within a search day, rather than an hourly search rate, and to recognize that handlers are obligated to make conservative decisions on behalf of the dogs that cannot advocate for themselves.

The cost of detection dogs can be a deterrent to their use. Cost was a significant factor in 13% of instances where detection dogs did not outperform alternative methods (Grimm-Seyfarth et al., 2021). Training dog/handler teams may also require more initial investment than hiring human searchers, but other studies have shown that when the target is costly to find, detection dogs can be more cost-effective than human-only searchers (Cristescu et al., 2015). While more costly overall, detection dogs are more cost-effective given their increased number of detections (Long et al., 2007b). This suggests that the investment in training detection dog teams may yield substantial benefits for agencies and organizations over time. One advantage is that a single dog can be trained on multiple target scents or species (Long et al., 2007a; Vynne et al., 2011; Williams & Johnston, 2002). For example, our dogs were trained to detect several other target invasive species including Scotch broom (Cytisus scoparius), slender false brome (Brachypodium sylvaticum), sticky sage (Salvia glutinosa), oak wilt (Bretziella fagacearum), and kudzu (Pueraria spp.). Beyond their exceptional detection capabilities, detection dogs offer an additional benefit by enhancing public outreach efforts due to their charismatic nature and relatability to the general public.

While detection dogs are not without limitations, their specialized capabilities and versatility make them a valuable addition to invasive species detection and management programs. The successful integration of dogs into such efforts requires careful consideration of environmental conditions, operational constraints, and cost–benefit analyses to maximize their potential impact. By leveraging their strengths alongside human surveyors, we can develop more effective and comprehensive approaches to managing invasive species and protecting ecosystems from their impacts.

Early detection scenario considerations

The efficacy of early detection methods is important for managing invasive species, as it allows for intervention before the species becomes widespread and eradication efforts become impractical (Reaser et al., 2020). Eradication of the species is most feasible when there are only a small number of localized populations; it is under these conditions that early detection methods are most effective. Our study, focused on comparing the effectiveness of human and canine detection of SLF egg masses, was conducted in moderate to high infestation sites to allow for statistical comparisons. However, under lower infestation scenarios typical of early detection, the relative advantages of canine detection would likely be more pronounced. In these situations, where occupancy probabilities are lower, dogs' superior detection abilities could make them an invaluable tool for finding elusive egg masses that humans might miss. This enhanced capability could significantly improve the chances of early intervention and containment.

Applications

The utilization of canine scent detection dogs provides a promising strategy for enhancing the likelihood of early detection of the SLF and other invasive species. In forested areas, dogs proved to be both more effective and efficient than humans in detecting egg masses. Our results showed that in forested areas, dogs were significantly more effective and efficient than humans, detecting over three times as many egg masses. This suggests that deploying dogs in forested environments could substantially increase the likelihood of detecting nascent infestations, which are critical for timely management interventions. While our study utilized a systematic transect-based search approach, a more flexible search method—such as allowing dogs to freely search larger forest patches—might be more suitable for early detection, increasing both the area covered and the probability of locating low-density infestations. For vineyards, a targeted perimeter search along the vineyard-forest edge may be more efficient than a transect approach, given that SLF egg masses are often concentrated within 75 m of the forest edge.

Our study provides search strategies for detecting SLF egg masses, including searching for egg masses in vineyards located within approximately 75 m of the forest edge and in trees located at slightly higher elevations than their surroundings. While using dogs can also be effective in vineyards, it would only be cost-effective in early detection scenarios when it is unknown whether there is an existing infestation. When searching for egg masses on vines, it is important to search on the undersides of cordons, on renewal spurs, on trunks, in tight spaces between trunks and any type of support structure, on suckers, and at trunk bases, especially if there are weeds/foliage that provide cover. Moreover, egg masses were found to be frequent on the inner surfaces of metal poles in vineyards, providing an easy substrate for human searches. Interestingly, we found a higher occupancy of egg masses on metal poles than on wooden poles, suggesting that certain vineyard management practices, like choosing support materials, could influence SLF distribution. While this insight could inform future vineyard management decisions, more research is needed to understand whether switching to wooden poles could reduce overall SLF occupancy or simply shift the egg-laying preference to the vines themselves.

Our study utilized a multiscale occupancy model that provided estimates of occupancy at both the transect level and the subunit level, such as vines or poles, offering valuable information on differential occupancy based on substrate type. In addition, the model allowed us to estimate detection probability as a function of observer type (dog or human), infestation level, and weather conditions. This modeling approach, which accounts for the probability of detection of the target species, has broad applicability for dogs searching for various taxa, whether invasive, threatened, or endangered, across diverse habitats and geographic locations.

Ultimately, while dogs present an upfront cost in training and deployment, their superior detection capabilities in the forest and versatility—such as being trained for multiple target species—make them a valuable asset in early detection and management strategies. Beyond their practical utility, detection dogs can serve as effective ambassadors for public outreach, raising awareness and support for invasive species management efforts. By strategically integrating detection dogs into surveillance programs, we can improve our capacity for early intervention, helping to safeguard ecosystems from the impacts of invasive species.

AUTHOR CONTRIBUTIONS

Angela K. Fuller, Ben Augustine, Eric H. Clifton, Ann E. Hajek, Arden Blumenthal, Josh Beese, Aimee Hurt, and Carrie J. Brown-Lima designed the study. Eric H. Clifton, Ann E. Hajek, Arden Blumenthal, and Josh Beese collected field data. Ben Augustine and Angela K. Fuller conducted analyses/provided interpretation. Angela K. Fuller wrote the paper with contributions from authors.

ACKNOWLEDGMENTS

We would like to thank the vineyard owners for allowing us to survey their lands. We thank David Harris, James Liebherr, and Abby Bezrutczyk, who assisted as human searchers and Dia and Fagan as detection dogs. We thank D. Calvin for his thoughts regarding spotted lanternfly behavior and ecology. We thank Audrey Bowe, Sam Talbot, Linda Rohleder, Jessica Cancelliere, and Thom Allgaier for contributing to the project proposal and design. Funding was provided by the Cornell Atkinson Center for Sustainability. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the US government.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Open Research

DATA AVAILABILITY STATEMENT

Data (Fuller & Augustine, 2024) are available from the USGS ScienceBase-Catalog: https://doi.org/10.5066/P1744ZNW.