Skip to main content

Calculations for: Detector dog work assessing probability of detection for Yellow crazy ant

Cite this dataset

Hoffmann, Benjamin (2022). Calculations for: Detector dog work assessing probability of detection for Yellow crazy ant [Dataset]. Dryad.


The use of detector dogs within environmental programs has increased greatly over the past few decades, yet their search methods are not standardised, and variation in dog performance remains not well quantified or understood. There is much science to be done to improve the general utility of detector dogs, especially for invertebrate surveys.

We report research for detector dog work conducted as part of yellow crazy ant eradication. One dog was first used to quantify probability of detection (POD) within a strictly controlled trial. We then investigated the search patterns of two dogs when worked through sites using different transect spacings. Specifically we quantified their presence within set distances of all locations in each assessment area, as well as the time they took to assess each area. In a GIS we then calculated the relative percentage of the entire search area within six distance categories, and combined this information with the POD values to obtain a site-level POD.

The calculated relationship between distance and POD was extremely strong (R2 = 0.998), with POD being 86% at 2 m and 28% at 25 m. For site-level assessments conducted by the two dogs, both dogs achieved highest site-level POD when operated on the lowest transect spacing (15 m), with POD decreasing significantly as transect spacing increased. Both dogs had strong linear relationships between area assessed and time, with the area assessed being greater when the transects had greater spacing. The working style of the two dogs also resulted in significantly different assessment outcomes. In one hour one dog could assess approximately 9.2 ha with transects spaced 20m apart, and 6.8ha with transects spaced 15 m apart, whereas the second dog could only assess approximately 6.9 ha with transects spaced 20 m apart, and 4.9 ha with transects spaced 15 m apart.

Our study provides insight into the ability of dogs to detect yellow crazy ants, and sets the basis for further science and protocol development for ant detection. With the lessons learnt from this work we then detail protocols for using detector dogs for ant eradication assessments.


Study area

The study was conducted between July and August of 2020 and June and July of 2021, within northeast Arnhem Land near the town of Nhulunbuy (12°11’S, 136°46’E) in Australia’s Northern Territory. The regional climate is tropical monsoonal with high temperatures (1733 °C) throughout the year and an annual rainfall of approximately 1200 mm falling predominantly during the summer wet season. The landscape is primarily savanna woodland dominated by Eucalyptus tetrodonta (height and canopy cover approximately 15 m and 20%, respectively), with an understorey up to three meters of mainly Acacias and grasses (Williams et al., 1996) (Figure 1). The weather throughout the sample periods was predominantly dry and sunny (working temperature range 18.5–29.5 °C), with relatively high humidity (50–92%) and low to moderate winds (0–6.5 knots). For this work, vegetation with a dense understorey was avoided, because it was difficult for dogs and the handler to move uninterrupted.

Target species

Yellow crazy ant is one of the world’s most invasive ant species (Holway et al. 2002; Wetterer 2005). It is a medium sized (4 mm) species, which in this environment nests at ground level within leaf litter, hollowed wood, or underground holes (Hoffmann 2015). This species is naturally supercolonial, whereby nests are not discrete entities; instead, workers of all nests display no aggression to each other and will move among nests, thereby forming large, interconnected and contiguous populations that often cover hundreds of hectares (Haines and Haines 1978; Rao et al. 1991; Rao and Veeresh 1991; Hoffmann 2014). The study region contains many spatially discrete yellow crazy ant populations (Hoffmann and Saul 2010), many of which have been subject to eradication attempts since 2004 using many baits, treatment timings, and treatment regimes (Hoffmann 2010), but the success or failure of many of the treatments to achieve eradication remains unknown. It is these treated populations that were used for area-wide searches to determine if any nests were persisting.

Detector dogs

Jet is a male working-line Springer Spaniel and is predominantly used for detecting Koala scats in eastern Australia (Figure 1). In 2018, when Jet was four years old, he was also trained on the scent of yellow crazy ant. Jet was subsequently used sporadically around Lismore and Terrania in eastern Australia and Nhulunbuy in central north Australia to detect yellow crazy ants. Frankie is a two-year-old female working-line Springer Spaniel also predominantly used for detecting Koala scats in eastern Australia, and was similarly trained to detect yellow crazy ants in early 2021.

Training the dogs to detect yellow crazy ants was conducted using the following protocols. Unscented cotton pads were first placed in yellow crazy ant nests for periods of between 12 and 36 hours to absorb the ant’s odour. The pads were then removed using stainless steel tongs and placed into double zip locked bags. Other pads had ants crushed onto them and were then also placed into double zip locked bags. These pads with ant odour were then placed randomly in an open grassy area so that they protruded partially from the ground. Some non-scented pads were also placed throughout the area. Individually, the dogs were then commanded to search the area and were rewarded (toy reward) when they showed interest in an ant-scented pad, regardless of whether the odour was sourced from crushed ants or nests. The dogs were not rewarded if they showed interest in an unscented pad. This process was repeated at least twice daily for approximately four weeks until the dogs were actively indicating on scented pads only. The second stage of training involved placing 10-50 live ants in small plastic or stainless-steel containers with a mesh cover. These containers were placed in the field, in a both open and more vegetated areas, along with similar containers that did not contain ants. Again, the dogs were commanded to search the area and were only rewarded when they indicated on vessels containing ants. This process was repeated daily for approximately four weeks in conjunction with the scent pad training. After the dogs were deemed to be only indicating on yellow crazy ant odour, they were moved to an area where yellow crazy ants were inhabiting and commanded to search the area. When the dogs indicated, whether it be individual ants, numerous ants, or a nest, they were rewarded. This process was repeated at least daily for at least three months before they were officially used to detect yellow crazy ants, and then sporadically thereafter in between times they were worked for detecting the ants.

Probability of detection

Experimental design

This work aimed to quantify POD at varying distances under strict time, behavioural and environmental conditions. Time limitations resulted in this portion of the work only being conducted by one dog, Jet. Locations were selected each day based on the wind direction, with transects being positioned perpendicular to the wind as much as possible. Locations used were roadside edges beside open bushland with very little midstorey layer (for ease of access), to the best of our knowledge were not populated with A. gracilipes, and were at least 500m from known A. gracilipes populations. It is known that wind variability affects how scent travels through air (Syrotuck 1972, Snovak 2004) and such wind variability affects dog detection abilities (Shivik 2002). Our preliminary work found that wind behaviour greatly influenced results, so we created two wind categories: ideal and non-ideal. Ideal was when the wind at dog level was moving constantly and in a single direction (when overhead winds were around 20 knots). Non-ideal conditions ranged from calm conditions through to gusty winds up to around 16 knots whereby the wind would also swirl and constantly change direction at dog level. All work was conducted during the cooler parts of the day (from 6:30–9:30 am and 3:30–6 pm) when conditions were conducive for yellow crazy ant foraging (Hoffmann 2015) and also for minimising panting, which decreases a dog’s sniffing rate and scent detection (Gazit and Terkel 2003).

Stainless steel cappuccino shakers (70 mm diameter, 98 mm long) with 50 mm diameter gauzed lids were used as cannisters to hold live ants. The ants were collected from a nearby population the day prior, stored overnight in a 20 L plastic bucket with a moist tissue to prevent dehydration, and placed in the cannisters just prior to going to the field assessment locations. Unlike for other utilities like vertebrate scat detection (Mackay et al. 2008), we knew of no prior publications about detection thresholds for dog training or assessments using ants, so we were unsure how many ants would be best to put in the cannisters. Preliminary work attempted to use two low ant quantities, 10 and 20 ants, but we observed that Jet continually found the control cannisters (no ants), even at great distances, potentially because he was familiar with the cannisters from prior work with Koala scats. To help overcome this behaviour, we used 50 ants so that the ant scent would be prominent. We also considered that this number could quite likely emulate the odour of hundreds of ants in a nest, which in this region typically has a 20 mm diameter entrance (Hoffmann 2015).

Cannisters were spaced at least 30m apart along the transect, and placed at four distances away from the transect: 2, 10, 20, and 30 m (Figure 2). Preliminary work also found that Jet was very capable of following the path of the person setting the cannisters. To prevent this issue the cannisters were set from the opposite side of the assessment area. Separate transects were used for upwind and downwind trials. In all cases the view of the cannister was obscured so that Jet could not simply visually find it. This involved placing it within perennial grass clumps, behind stems, or with a light covering of leaf litter (but not covering the gauze). Both the dog and the handler were not in the area when cannisters were laid to prevent bias, nor was the handler told whether cannisters were treatments or controls until they were detected by the dog.


Just prior to Jet commencing walking along each transect, environmental conditions were recorded from both the live meteorological bureau data and using a hand-held Kestrel weather station at approximately 50 cm high (dog height). Data recorded were wind direction and speed, temperature, and humidity.

Jet was tightly controlled to minimise random or false searches and ensure that detections were at the appropriate distance. He was walked on a short leash along a clear path (edge of windrow/vegetation) and was only allowed to walk away from this path greater than the distance of his lead if he pulled strongly. He was walked at a pace of approximately half a metre per second. When given the freedom to pursue a scent, he needed to display an expected zig-zag search path, and/or be seen to only be pursuing a cannister scent (i.e. not the trail of the person setting the cannisters). If Jet was seemingly walking randomly, not following a scent, following the scent of the setter, or pursuing the likes of a prior cannister, he was ordered to return to the handler. Also, if he was indeed on the scent of a cannister but lost the scent or could not find it, he was promptly called to return.

When Jet found a cannister of ants (and signalled by sitting), the trial was paused, and he was given a reward, being time to play with his ball. When he found a control cannister, the trial did not stop and he was not rewarded, but we noted whether he signalled that he had found yellow crazy ants or not. If he found any cannister after being instructed to return to the handler (e.g. he smelled and found a cannister farther along the transect rather than one he initially searched for), or if he clearly violated any of the other rules determined above, the found cannister was voided from the data. Found canisters were collected immediately. Jet was walked along the entire length of a transect in one direction, and was given a second opportunity in the opposite direction to find any cannisters he did not find on the first run.

To additionally prevent Jet from targeting smells from the cannisters instead of the ants, individual cannisters were only used once per morning or afternoon session, and were later soaked in boiling water for at least five minutes to remove volatiles.

After all voided trials were removed from the data, what we considered minimum sufficient replication (minimum 13 assessments per treatment) was only achieved for non-ideal conditions, because ideal conditions did not occur with sufficient frequency/duration during the assessment period between 11 July to 25 August 2020 (Table 1). Results for the ideal conditions were still analysed, but must be interpreted with caution. Notably, what we have classed as non-ideal conditions predominate in this region, so this reflects the on-ground reality.

Coverage and site-level POD

Experimental design

This work investigated the search patterns of both dogs (Jet and Frankie) by quantifying their presence within set distances of all locations in an assessment area (coverage) when worked through a site using different transect spacings. Multiple areas that had in years prior received toxic treatments to eradicate yellow crazy ant populations (Hoffmann 2010) were selected for real post-treatment assessments to determine either persisting presence or absence (eradication) of the ants. Sites were not standardised for size or vegetation, but areas with consistently open ground-layer vegetation were selected as much as possible to allow free movement of the dogs and the handler. Parallel transects crossing the areas were delineated in the iPhone application iGIS, at distances of 15, 20, or 25 m. Preliminary work found that using an application, whereby the handler could constantly locate themselves relative to the transect line, greatly improved efficiency and accuracy of the work as opposed to just visually estimating the transects. These specific distances were selected to accommodate the dogs’ zig-zag search style, which in preliminary observations typically extended about 15 m either side of a transect (influenced by some verbal control from the handler). Note that these transect spacings are the analogue of “effective sweep width” used in search theory (Glen and Veltman 2018). The direction of the transects was not random, but was 45o to the wind to maximise the distance the dog would travel through a scent plume.

The dogs were worked individually for assessments, never together, and always in different areas. Just prior to commencing assessments, the dog that was to conduct the assessment was fitted with a Garmin GPS collar that recorded point locations every 2.5 seconds. The handler then walked along the transects, using iGIS as a guide, while the dog worked to detect yellow crazy ants. The handler gave no instructions of where to look, only keeping the dog moving along the transect and not too far perpendicularly away from the transect. The handler moved along the transect at the pace of a slow walk, approximately 1m every four seconds. Assessments were conducted between 11 July to 25 August 2020 and 23 June to 18 July 2021.

Because these were real assessments there were occasionally instances when persisting ants were detected. In these instances the work sessions were not used for the coverage assessments because the presence of the ants influenced the dog’s search paths.

Area-time relationships

For each of the coverage assessments, the duration of each assessment and distance the dog walked (from the GPS data) were also recorded. The area that the dog assessed was then calculated by multiplying the distance the dog walked by the transect spacing (e.g. 1000 m with 15 m transect spacing = 1.5 ha). Note that this area calculation is what the dog could potentially assess, but does not necessarily reflect the actual area covered (i.e. a dog could walk varying distances within any sized area).


Probability of detection

The percentage of times that cannisters containing ants were detected at the four distances, upwind and downwind, and within ideal and non-ideal conditions was calculated. Following consideration of the replication of all trials, coupled with the later protocol determination that assessments would be conducted from downwind to upwind within sites, the average of the results for the two upwind situations (ideal and non-ideal conditions) was used to determine the polynomial of the relationship. This equation was used to calculate POD at exact distances.

Coverage and site-level POD

At the end of each area assessment session, the GPS tracks of the dog were downloaded, and uploaded into the GIS program ArcGIS 10.6. The paths were then buffered using the following distance categories from the dogs: 0–2 m, 2.01–5 m; 5.01–10 m; 10.01–15 m; 15.01–20 m; and 20.01–25 m. These distance categories were determined subjectively based on a visual assessment of the shape of the POD graph and opinion of how to maximise site-level POD calculations. From the buffers, we then quantified the relative percentage of the entire search area within each distance category These relative percentages were then combined with the POD values from the POD graph to calculate a single site-level POD value for each search, and in some cases where searches were combined, across an entire area of a treated ant population. The exact POD values used for the categories were: 86.0978; 73.796; 56.433; 42.995; 33.482; and 27.894 respectively. So for example, if all six categories each comprised 20% of the search area, these POD values were each multiplied by 0.2 and summed to give an overall POD of 64.14%. One-way ANOVAs and Tukey’s post-hoc tests were used to test for differences in POD between the dog and transect spacing combinations. The assumption of data homogeneity was confirmed using Cochran’s tests.

Area-time relationships

The area and time data were plotted separately for both dogs and for transect separations of 15 and 20 m only. Uniformity of the four linear area-time relationships was statistically tested using an Analysis of Covariance and Tukey’s post-hoc test. The assumption of data homogeneity was confirmed using a Cochran’s test


Rio Tinto (Canada)