The Sentinel Bait Station: an automated, intelligent design pest animal baiting system

Introduction

Vertebrate pests such as feral cats (Felis catus), European red foxes (Vulpes vulpes), feral pigs (Sus scrofa) and wild dogs (Canis familiaris) cause major negative impacts to agriculture and the environment (Braysher 1993; Fleming et al. 2017). Controlling introduced pests requires the integration of methods such as aerial and ground baiting to distribute toxic baits (Newsome 1990; Fleming et al. 2014; Ballard et al. 2020), trapping (Meek et al. 1995; Fleming et al. 1998; Meek et al. 2022) and the use of other devices (e.g. Canid Pest Ejectors (CPE’s); Busana et al. 1998; Marks et al. 1999; Marks and Wilson 2005) to achieve reductions in abundance that exceed the pest population’s ability to increase (Hone 1994). Ground poison-baiting can mitigate negative impacts but can be costly, and sustaining the level of control necessary to overcome population increase can be operationally demanding. Consequently, ongoing improvement of existing methods via technological innovation (Meek et al. 2020) is a useful objective.

A study on the effectiveness of controlling the reinvasion of a peninsula by foxes (Vulpes vulpes) highlighted the challenges associated with maintaining predators at low levels (Dexter et al. 2007). Invasion by foxes across a narrow isthmus provided an opportunity to increase the encounter rate with toxic baits. However, traditional bait stations can hold only single baits, so maximising availability to targets required daily replacement of baits (Fleming 1996, 1997). Time spent replacing baits in situations such as this comes at the cost of other operational imperatives, and demonstrates the need for a baiting system that can deliver toxin to successive targets with a reduced reliance on people to frequently replenish baits.

Non-target species add to the demand on bait replenishment by consuming, moving and/or degrading baits. Behavioural means of discriminating between targets and non-targets is already utilised in some control technologies (e.g. CPEs), but arguably reduces the likelihood of targets consuming toxin. Advances in image recognition, especially by our research group (Falzon et al. 2012; Falzon et al. 2020; Shepley et al. 2021a, 2021b), on incorporating such technology into a baiting device, show potential for creating target specific baiting that both protects vulnerable non-target species and reduces the removal of baits intended for pests by unaffected non-target vertebrates and invertebrates.

With the technological advances in computer science and engineering, designing automated toxin delivery devices specific to target species is now achievable (Meek et al. 2020; Moseby et al. 2020; Ross et al. 2020; Corva et al. 2022). We aimed to provide a reliable means of automatically replenishing baits while minimising undesirable bait loss (e.g. removal by non-targets) to increase the number of a target species encountering baits per unit of human effort. We designed, built and tested a proof-of-concept prototype to evaluate the potential to automatically and repeatedly deliver a bait to target pest species (such as wild dogs and foxes) using image recognition technology.

Methods

Development

Development involved designing and integrating two main components: the smart camera trap (Fig. 1a) and the bait-dispensing unit (Fig. 1b). The fundamental concept was to transform a standard camera trap into an intelligent cyber–physical system that could reliably and automatically dispense baits to animals of target species by using recent advances in machine learning. The design specifications were to build a system that could be placed in the field unattended for extended periods of time (e.g. multiple weeks or until all baits are taken).

Fig. 1.

Sentinel Bait Station (SBS) camera trap: (a) smart camera trap (220 × 120 × 100 mm) and (b) bait-dispensing unit (180ø × 150 mm).

The intelligent camera consisted of a single-board computer, a combined passive infrared and microwave heat-in-motion sensing module and camera module (consisting of camera, dynamic near infrared filter and flash module outside of the visual range). This setup enabled the camera to automatically detect animal movement (day or night) at a range up to 15 m from the camera, photograph the animal and process the image to determine whether it was a target or non-target.

The SBS software system is an event-driven, finite state machine design equipped with machine vision, logic and wireless communication modules to provide advanced functionality. When the system is awaiting the presence of a target, the single board computer (Raspberry Pi 3B+, Raspberry Pi Foundation, Cambridge, UK) remains in idle mode until receiving an interrupt from the heat-in-motion sensing module. When heat-in-motion is sensed, the camera captures images that are processed by the species recognition algorithm. The species recognition algorithm was trained according to the location invariance methodology presented in Shepley et al. (2021b). During inference, the algorithm assigns a classification confidence score to detected objects (0–100%), with a higher score denoting a higher internal confidence by the algorithm that the target species is present in the image. A confidence threshold can be adjusted to the specific use case, as the main tuneable variable for determining whether the model favours false positives or false negatives. For this study, a confidence value of 5% was chosen (i.e. any image where the algorithm confidence value was more than 5% confident of a target species being present resulted in a positive detection). This value was chosen during development and early testing of the system to maintain a higher sensitivity to target species. The image detection model utilised within this design is MobileNet v1 model, which is a deep convolutional neural network (CNN) optimised for embedded system computer vision applications (Howard et al. 2017). The dataset used to train the model comprised of both publicly available images of target and non-target species sourced from FlickR and iNaturalist, along with an infusion dataset of camera trap images featuring strong similarities to those used in this study. The model was trained using transfer learning in accordance with the location invariance method to maximise robustness to unfamiliar camera trap settings. Importantly, none of these images were part of the simulation dataset mentioned below.

Once a target is detected, the computer communicates wirelessly (Fig. 2) to the bait-dispensing unit to offer a bait to the target animal. Two-way wireless communication between the single-board computer within the camera and the microprocessor (ESP32, Espressif Systems, Shanghai, China) within the bait-dispensing unit used the Message Queuing Telemetry Transport (MQTT) protocol over a wireless local area network (WLAN).

Fig. 2.

The system component diagram representing the ‘Smart Camera Trap’ equipped with a Raspberry Pi single-board computer coupled to heat-in-motion and camera units detecting the presence of a target predator and activating the bait dispenser via wireless communication to an ESP32 microprocessor.

The rotating carousel within the bait-dispensing unit presents either an empty or baited compartment (Fig. 3) as required. During development, the carousel was designed to house five standard-sized manufactured dog baits (60 g each), but capacity could be increased. A stepper motor operates the carousel and a locking pin is used to maintain the desired position when the motor is inactive. An infrared light beam sensor (outside the visible range for canids) is used to determine if a bait is present or absent from the currently available compartment; this allows the SBS to ensure baits are presented to target animals, and if not consumed, enclosed for safe-keeping until the next detection occurs.

Fig. 3.

Computer-aided design (CAD) image of the bait dispenser showing the internal carousal design (the square structures within each triangle segment are the bait chambers and the empty segment is the starting point).

The field-tested design enabled the following operating scenario: the SBS camera is deployed with the bait dispenser approximately 5 m away, within the field of detection (Fig. 4). When the camera detects heat-in-motion it takes a pre-programmed number of photographs (e.g. for this study, 10 images at 1-s intervals). The on-board computer then rapidly performs image recognition to identify whether the animal is a target species or not. When a target is recognised, the computer instructs the bait-dispensing unit to offer a bait by turning the carousel to an open position. If the bait is not removed after 120 s, the carousel closes and awaits a new target detection. Alternatively, if the bait is removed, the device recognises this and will not expose another bait until 120 s has passed and another target is detected.

Fig. 4.

Field deployment involved positioning of the SBS camera unit at 1 m height and 4–5 m distant from the ground dispenser unit.

Results

Field testing

Although both wild dogs and foxes were potential targets, no foxes were detected during field testing. Camera traps used to monitor the SBS system and immediate area recorded target and non-target species displaying interest in the SBS bait dispenser (Fig. 5). The SBS machine vision recognition module was able to detect the target species under a wide range of conditions; however, some false detections still occurred (Fig. 6d, f). During field testing, the accuracy of the system was 0.90 (day: 0.96; night: 0.69; Table 2). The balanced accuracy was 0.82 (day: 0.85; night: 0.73; Table 2), indicating a drop in performance due to class imbalance but the classifier still maintained average–good per event classification performance. The precision value of 0.83 indicates that the system offered a bait when there was a target present in 83% of events (day: 0.71; night; 0.88), with four false positive events occurring throughout the trial. The sensitivity value of 0.68 indicates that a bait was offered in 68% of events where a target was present (day: 0.71; night: 0.61), and the specificity value of 0.96 indicates that bait was withheld in 96% of events when a target was not present (day: 0.98; night: 0.85).

Fig. 5.

The SBS bait dispenser attracted attention from both (a) target and (b) non-target species.

Fig. 6.

Detection events as recorded by the SBS camera module. Successful detections occurred under a range of challenging conditions including: (a) multiple target species present; (b) motion blur due to fast moving target species; (c) both target species and non-target (human) within the same image; and (d) camera trap illumination obscuring target species (note shadow of target species within image). In some circumstances, such as (e), both a true positive and false positive detection occurred simultaneously, and in other situations, such as (f), a false positive detection (the bait dispenser was incorrectly detected as the target species) occurred.

Table 2.Detection data from the Cooper Basin field trials of the sentinel bait station.

Time of day	True positive (count)	Baits taken (count)	True negative (count)	False positive (count)	False negative (count)	Accuracy (ratio)	Balanced accuracy (ratio)	Precision (ratio)	Sensitivity (ratio)	Specificity (ratio)
Day	5	3	90	2	2	0.96	0.85	0.71	0.71	0.98
Night	14	4	11	2	9	0.69	0.73	0.88	0.61	0.85
Overall	19	7	101	4	11	0.90	0.82	0.83	0.68	0.96

From 19 bait offerings made to targets, seven were removed from the device by a target species. When using the Doggone® baits, 2 of 8 baits that were offered were taken, but after switching to fresh beef pieces, 5 of 13 baits that were offered were taken. There were four false positive events, where the image recognition falsely detected non-targets as a target. The misclassified objects were the bait-dispensing unit (n = 2), vegetation (n = 1) and a person (n = 1). There were 11 false negative events (day: n = 2; night: n = 9) where a target was present, but a bait was not offered due to misclassification.

Discussion

To the authors’ knowledge, the SBS is a world-first solution comprising a camera trap with in-built species recognition capacity and wireless communication to a multiple bait delivery unit. Laboratory and field testing of this prototype device has yielded proof-of-concept evidence that the SBS can detect a target and present a bait (i.e. the system can detect heat-in-motion, rapidly capture and process a series of images to determine the presence of the target species and transmit a message to the carousel to offer a bait).

The response time of the system was similar for images taken during the day or night. The images were able to be processed in less than 2 s (i.e. the time taken from a motion trigger to the single-board computer analysing the picture) with the system completing the bait dispensing in approximately 10 s (i.e. the time taken from a motion trigger to the bait being offered). Therefore, on average, an 8-s delay was observed in offering the bait due to both the time taken for the actuators to move the bait into the open position and power saving methods with the wireless interface required to reduce the power consumption of the ground unit. This is likely adequate time for bait presentation prior to a target leaving the immediate area; additional testing could assess this.

It is highly likely the software can be optimised to further improve system performance. During the field testing, just over a third of bait offerings resulted in a bait being taken by the target. This was despite targets at the site having a ready supply of fresh food, possibly reducing their interested in baits when compared with targets under greater food stress or elsewhere. It was also possible that the manufactured baits were not attractive to dogs at the study site. As outlined in the methods, during the field trial the bait was changed over to beef pieces to increase the palatability and attractiveness of the bait to the target species. Although the change in bait increased the proportion of baits being taken when offered, the purpose of this study was to review the ability of the target species (in this example, wild dogs) to take baits from the bait-dispensing device rather than the baits’ palatability. Therefore, the authors assert that the change in bait type has little bearing on the results or findings.

Based on the simulation data, the sensitivity of the system was slightly better at night when using the infrared illumination to take photos in low-light scenarios. Specifically, the sensitivity (0.20) for night-time images indicates that for every five images of a target species taken at night-time, one of these images will result in a true positive event and is therefore a bait being offered to the target species. Conversely, the sensitivity for images taken during the day (0.09) suggests that 10 images would be required for one true positive event to occur and a bait to be offered to the target species. During field testing, the difference between day and night events was not evident, and this is largely attributed to how an event was considered. Specifically, within the field testing, 10 images were taken in succession after motion was detected, which increased the likelihood of a correct image recognition during each event and reduced the limitation of low sensitivity shown in the simulation. This system design decision was based on practical observations, where the target species investigated the bait dispenser, permitting multiple images to be collected in succession. This set of images can be rapidly processed by the machine vision module, with a high likelihood that at least some of the images within the batch will register detection of the target species. By utilising this algorithm design, it was feasible to incorporate a detection algorithm of lower computational complexity and hence more rapid processing times and lower power usage (at a cost of slightly reduced detection accuracy per image). The overall sensitivity during field testing (0.68) would indicate that approximately four out of five times (that a target entered the field of view of the camera) a bait was offered.

The low sensitivity observed in the simulations, which resulted in a low accuracy value, was due to the high number of false negative events. Such low sensitivity, and therefore accuracy, was not as prevalent during the field testing due to taking multiple images in succession when heat-in-motion was detected. The field testing showed that when a target was present near the system, a bait was offered in 68% of circumstances. This was slightly higher during the day (71%) than during the night (61%). Laboratory testing also involved a wider range of environments than field testing, which provides a challenging scenario for machine vision systems (Shepley et al. 2021b). Evaluations on a per image basis across a broad range of environments could have contributed substantially to the number of false negatives. Optimising model performance across diverse environments needs further investigation. Finally, due to the low number of false positive events compared with true negative events, the system demonstrated a very high specificity value for both the laboratory simulations (0.99) and field testing (0.96).

Overall, the system was better at detecting when a target species was not present within an image as opposed to identifying that there was a target species within the image. This aspect was a component of the machine vision algorithm design: detection thresholds were set to be conservative to minimise the number of bait offerings to non-target species. From a practical perspective for the specific use case field tested, a low false positive rate is more favourable because it is preferable to withhold a bait from a target species than to offer a bait to a non-target species, as currently occurs in most baiting situations. Additional development of the identification algorithms could be undertaken to further improve the accurate identification of target and non-target species (e.g. more images from the camera and environments used for deployment within the training dataset would likely improve the overall specifies recognition performance). To incorporate the SBS in other applications, it could be more effective to tune the algorithm to improve the false negative rate at the expense of the false positive rate, depending on the risk to other species if they were to access a bait.

There were four images where a non-target was incorrectly recognised as a target. The bait-dispensing unit was not highly represented in the training data and this may have contributed to the bait-dispensing unit being incorrectly recognised as a target species on two occasions. To avoid such false positives in the future, it would be important to include the bait-dispensing unit within further training images. Interestingly, in the other two instances a bush and a person at the outer edge of the field of view were misidentified. All objects detected as false positives had a low confidence value compared with that of most true positive events. This supports further adjustments to the algorithm detection threshold, requiring multiple target detections per SBS unit trigger, iterative training revisions and improvements to the algorithm (e.g. the inclusion of more negative samples such as images with people close to the border would make the algorithm more robust to those variations, thus reducing false positives).

Overall, we sought to offer baits to targets, rather than non-targets, and automatically replace these baits, so that more targets could be baited per unit of human effort. With this considered, the most important metric for the performance of the system is the precision value (i.e. the number of baits offered to a target in comparison with all baiting events). During field testing, precision was relatively high (0.83) due to the low false positive event count. Compared with current practice, this would significantly reduce the number of baits available to non-targets. Further model improvements could enable the precision of the system to be improved through the modification of the confidence value (5% used in this trial). The precision could be increased by increasing the target confidence threshold (e.g. only dispense baits where the algorithm is 50% certain that a target is present). However, there is a trade-off between avoiding dispensing a bait to non-targets and dispensing a sufficiently high number of baits to targets. Further algorithm training would allow for a higher confidence value to be used within the model and therefore reduce the number of false positive events without significantly reducing the sensitivity of the system. In contrast to unconcealed baits (the current practice), the SBS only permitted bait availability for 2 min. Critically, the SBS permitted access to baits in circumstances where there was a high likelihood that the target species was present.

Future development of the SBS could be centred upon improving algorithm performance to detect target species across environmental conditions, which would translate to higher precision and sensitivity. Also, broader field tests are required to understand SBS performance under different contact rates with target and non-target species. Fortunately, the modular design of the machine vision component of the software permits model substitution, so the SBS could be trained to offer baits to a broader range of target species, such as rodents, feral pigs or feral cats.

The MobileNet v1 Convolutional Neural Network was state of the art at the time of development, integration and testing of the SBS. The fast pace of machine vision research has offered additional innovations and increase in performance that could be included in future algorithms. Options could include: extreme learning machine architectures that are optimised for drones and field robotics (Sadgrove et al. 2018); modifications and improvements to the MobileNet architecture (Ayi and El-Sharkawy 2020; Kavyashree and El-Sharkawy 2021); or modifications to improve the robustness of object detection algorithms (Shepley et al. 2020). The development of a more specific and extensive dataset for model training would also be very useful. Due to the prototype nature of the SBS, very limited image data were available to develop the machine vision model. Empirical evaluations of camera trap-specific CNN algorithm performance indicate that substantially more data could be required per class across environments to optimise performance (Shahinfar et al. 2020). Inclusion of negative classes that are known (from field trials) to trigger false positive detections (such as the bait dispenser) in further model training will permit further model refinement.

The engineering design of the bait-dispensing system could also be modified to increase its robustness and ability to securely store more baits. Further, to be able to use toxic baits during field trials and ultimately in practice, the system would be required to address legislative requirements for storing and distributing multiple baits in a single location. Although not observed within the field trials, it is the opinion of the authors that refinements to the mechanical design of the system to reduce noise associated with the bait-dispensing unit when opening and closing may help to limit the associated startle behaviours by the target species in response to unnatural noises (Meek et al. 2016).

Conclusion

Laboratory testing found the SBS was able to recognise target species (i.e. dogs and foxes) and dispense baits to them. Field testing was highly target specific, and target species were willing to take baits from the bait-dispensing unit when offered. The performance indicates that the SBS system could be developed for use in real-world control programs to reduce both non-target uptake and human effort required to provide baits to multiple target species.

Data availability

This proof-of-concept device is still the subject of further research applications, and is a potential commercial product for the Centre of Invasive Species and University of New England. As such, data will remain confidential until this process is complete, and will remain the subject of commercial confidentiality requirements.

Conflicts of interest

The authors declare no conflicts of interest.

Declaration of funding

Funding for this research was provided by the Department of Agriculture, Water and the Environment through the Centre for Invasive Species Solutions. Dr Ballard, Dr Fleming and Dr Meek were funded by the NSW Department of Primary Industries, and Dr Falzon and Mr Charlton were funded by the University of New England.