Automated Surveillance of Lepidopteran Pests with Smart Optoelectronic Sensor Traps

Abstract

Several lepidopterans are pests in horticulture and pose biosecurity risks to trading countries worldwide. Efficient species-specific semiochemical lures are available for some of these pests, facilitating the implementation of surveillance programmes via trapping networks. These networks have a long history of success in detecting incursions of invasive species; however, their reliance on manual trap inspections makes these surveillance programmes expensive to run. Novel smart traps integrating sensor technology are being developed to detect insects automatically but are so far limited to expensive camera-based sensors or optoelectronic sensors for fast-moving insects. Here, we present the development of an optoelectronic sensor adapted to a delta-type trap to record the low wing-beat frequencies of Lepidoptera, and remotely send real-time digital detection via wireless communication. These new smart traps, combined with machine-learning algorithms, can further facilitate diagnostics via species identification through biometrics. Our laboratory and field trials have shown that moths flying in/out of the trap can be detected automatically before visual trap catch, thus improving early detection. The deployment of smart sensor traps for biosecurity will significantly reduce the cost of labour by directing trap visits to the locations of insect detection, thereby supporting a sustainable and low-carbon surveillance system.

1. Introduction

Invasive species pose an ever-growing threat to the stability of native and agricultural ecosystems worldwide [1,2]. Biosecurity programmes are key to preventing invasive species incursions brought on by our ever-more-connected global economy [3]. While the front-line effort against the spread of invasive species often focuses on pathway risk management and import policy [4], for species that pose a particularly high risk, targeted post-border surveillance programmes can be worth the investment [5]. Highly species-specific sex pheromone lures (i.e., semiochemicals) are commercially available for many lepidopteran pest moth species [6,7]. These semiochemical lures can be used in traps to attract the targeted insect, greatly improving the probability of detection in the early stage of an incursion [8]. While these semiochemical trapping networks offer great efficiency gains over passive surveillance methods, labour input can be further reduced through the use of more advanced trapping technology.

Currently, trappers are required to manually inspect each trap, regardless of whether the trap contains any insects of interest. The trap visitation frequency and network size of a surveillance programme are generally limited by the labour costs associated with these manual inspections, representing about 90% of the total cost of the surveillance programme [5,9]. In New Zealand, for instance, there are over 9000 semiochemical traps deployed across the country for the targeted surveillance of Tephritidae fruit flies and gypsy moth (Lymantria dispar), which are checked once every two weeks during summer, costing about NZ$1M/year. For the 2018–2019 season, only 7107 out of 177,430 (4%) surveillance trap inspections found suspect insects [10,11]. Thus, more than 95% of the time, traps are empty. By replacing manual traps with “smart” traps that are able to autonomously report when insects are detected and thus direct trappers’ inspection, we could greatly reduce the labour cost and carbon footprint associated with running an insect surveillance network. These smart traps also have the potential to further reduce the detection lag in the event of an incursion from weeks to days, if not minutes, greatly reducing the risk of spread and thus increasing the chance of successfully eradicating the target species at a lower cost [12].

With the rapidly decreasing cost and power consumption of computing and the development of low-power networking, efforts have been made to bring trapping technology into the 21st century. Some research teams have focused on the use of cameras as a tool for remote visual inspection of traps for caught insects [13]. While these camera-based smart traps are effective at automating trap inspections, the high computational complexity and power consumption associated with transmitting and processing images have limited the scale at which they can be deployed effectively. This issue only compounds as image resolution and upload frequency increase, resulting in the need for solar panels, large batteries and associated infrastructure to accompany each trap [14]. Other research teams have focused on optoelectronic beam break type sensors as a simple, low-power, low-cost sensing solution [15]. These sensors are especially suited to biosecurity applications, where trap catch is often very low and thus reports are only generated when an insect has been detected, significantly reducing the network overhead expenses and power consumption. The optoelectronic sensors have proven to be effective not only in detecting insects, but also in identifying the species caught, such as fruit fly or mosquito, based on disruption of the electric signal patterns caused by insect wing beat frequency while passing through the sensor field [16,17]. This concept of optically recording the wing beats of insects into an electric signal has a long history, with early work beginning in the late 1970s [18]. These optoelectronic sensors were able to effectively record the wing beat patterns of flying insects in the laboratory, however, due to technological limits at the time, they were less practical for use in the field [19]. Inspired by this early work, researchers have developed field stable versions of these sensors and integrated them into traps for use in the automatic monitoring of insects. These efforts have produced effective prototypes for monitoring fruit flies in McPhail-type traps [20] and crawling insects in pitfall traps [21].

Small insects like mosquitoes or fruit flies generally beat their wings rapidly over hundreds of hertz [22]. Therefore, these high wing-beat frequency signals can easily be separated from direct current (DC) electric signals and low-frequency interference caused by the sun via the use of band pass filtering techniques to exclude the low-frequency region. Recordings from these sensors are often presented in an audio-like format with a low-end cut-off around 20 Hz, which may exclude insects passing through the sensor field without beating their wings (e.g., walking or gliding). In contrast, the wing-beat activity of Lepidoptera (moths and butterflies) is much slower, often in the low tens of hertz, thus the band pass filtering approach would exclude much of the signal produced by these insects, even when beating their wings. An alternative filtering approach based on modulating the emitters and demodulating the received signal was developed by Potamitis and Rigakis [23] to remedy this issue, allowing the complete frequency spectrum to be recorded down to 0 Hz, while also excluding DC interference from the sun. In addition, by capturing the full frequency spectrum of an insect passing through the sensor, we can now explore the full range of behaviours exhibited by insects entering a trap to improve the accuracy of species identification.

In this research paper, we present our results from the development of an optoelectronic trap that adapts simple and low-cost optoelectronic sensors to fit delta-type traps, commonly used for monitoring lepidopterans, and employing a filtering approach to produce interference-free recordings of low-frequency wing beat patterns from lepidopterans. Furthermore, we integrated wireless communication to facilitate remote detection and simplified the hardware requirements to reduce both the power consumption and cost of these smart devices. Finally, we were interested in exploring the digital signals to better understand how insects behave as they enter the smart sensor trap as well as identify species-specific characteristics that can be diagnostic of particular organisms for future automatic species identification. These improvements should offer a novel sustainable solution for automatic monitoring of invasive Lepidoptera for biosecurity.

2. Materials & Methods

2.1. Smart Trap Design

2.1.1. Optoelectronic Sensor Improvement

Simple optoelectronic beam break circuits monitor the disruption of light traveling between a series of infrared emitters and photodiode receivers. This disruption causes a drop in the photocurrent generated by the receivers proportional to the degree of disruption. This circuit offers an effective method for detecting objects crossing the emitter beam and is often used in automated production lines. The drop in photocurrent generated by an insect’s disruption of the emitter beam is converted to a readable voltage through a current to voltage converter, generally in the form of a trans-impedance amplifier. In this trap-based application, the circuit must be able to function in a high ambient light outdoor environment. For this reason, a method was required for filtering low-frequency interference from the sun to prevent baseline drift potentially creating false positive detections caused by fluctuations in ambient light, while still recording low-frequency activity produced by Lepidoptera passing through the sensor field. In this trap-based application, the circuit must function in a high ambient light outdoor environment. For this reason, a method was required for filtering low-frequency interference from the sun to prevent baseline drift potentially creating false positive detections caused by fluctuations in ambient light, while still recording low-frequency activity produced by Lepidoptera passing through the sensor field. An analogue front end was designed with an effort to use as few active components as possible to minimise the device cost, complexity and power consumption. Through this optimisation process, we discovered that with some modifications to the amplifier configurations, a differentiator amplifier based around a single opamp could act as an effective AC coupled trans-impedance amplifier, rejecting the DC photocurrent generated by the sun while still converting pulsed photocurrent from the emitters. A more detailed explanation of this circuit can be found in Welsh [24]. This differentiator-based analogue front end was paired with an STM32 microcontroller. A pulse width modulation (PWM) output was used to pulse the emitters while band-pass sampling of the resulting signal demodulated the sensor output. This band-pass sampling was achieved by timing the sampling of the on-board analogue-to-digital converter (ADC) with each emitter pulse so that only the peak of each corresponding receiver pulse was recorded. This system achieved a similar output to that described by Potamitis [23] while using only one active component in the analogue front end. A flow diagram of this system is presented in Figure 1.

The samples from the ADC were constantly written via direct memory access to a circular buffer. An on-board analogue input watchdog was used to trigger the retrieval of data from this circular buffer only when a threshold crossing event occurred. The threshold value which triggered the analogue watchdog was set at a fixed ratio of the baseline, which was updated at the end of each buffer using a moving average when no threshold crossing events were detected to account for any slow drifts in the baseline due to dust or debris build-up on the sensor. This also allowed the threshold to be modified depending on the size of the target insect to help reduce false positives caused by smaller by-catch. By utilising the peripherals of the microcontroller in this way, the main processor remained in sleep mode until an insect was detected, significantly reducing the device’s power consumption.

2.1.2. Trap Refinement to Integrate the Sensor

Different geometric arrangements of the emitters and receivers were trialled to fit the delta-trap openings. While delta-type traps generally have a triangular opening, we found that triangular arrangements of the receivers resulted in large amounts of distortion in output amplitude across the sensor field due to the uneven amounts of received light. A rectangular arrangement with rows of eight emitters and eight receivers, each interspaced by 10 mm with 42 mm separating each emitter-receiver pair, was chosen for the final design to minimise the amplitude distortion. Theoretically, this allowed the sensors to record any object larger than 10 mm passing through the sensor field with 100% certainty. Detection of smaller objects was certainly possible; however, there was a possibility of them passing between the active areas of the sensor field, the likelihood of which increases as the size of the object decreases. Emitters with high radiant intensity and a small half-angle (Osram SFH 4542-BWCW) were used to further reduce amplitude distortion created by crosstalk between emitter-receiver pairs while also maximising the light intercepted by each receiver (Osram SFH 2500-FA).

A new smart trap was designed to enclose the sensor electronics while adhering as closely as possible to the conventions of a standard delta-type trap; see Figure 2A,B for a comparison of the standard delta-trap to our new smart sensor trap. In New Zealand, biosecurity traps can be placed on publicly accessible land. Therefore, there was a desire to design the traps to obscure the electronic components, reducing the perceived value of the device to decrease the risk of theft or vandalism when deployed in the field. The trap was 3D printed from polylactic acid (PLA) and the sensor was placed in one end cap, which was fixed into the trap body by screws (Figure 2C). The data logging and networking electronics are in a waterproof enclosure incorporated into the ceiling of the trap. The other end cap is easily removable via a tab which can be released by applying pressure at the base of the cap tabs, allowing the trap to be inspected and serviced without disruption of the sensor (Figure 2D). This end of the trap also includes an opening for airflow; however, it is covered with mesh to prevent insects from entering without passing through the sensor at the other end. This modification of the standard delta trap (i.e., one entrance for the smart trap versus two entrances for the standard trap) can impact the trap catch if the opened entrance fitting the sensor is not oriented in the downwind direction so that the pheromone lure can attract insect upwind. Producing a trap with sensors at both ends would eliminate this issue; however, this would increase the power consumption of the device and make servicing the trap without accidentally triggering the sensor difficult. A better compromise may be to add a wing or a wind vane to the trap and allow it to pivot in the wind, ensuring that the pheromone plume always attracts insects to the open end [25,26].

2.1.3. Data Communication

For the laboratory experiments, detections of events were logged as timestamped CSV files to an on-board micro-SD card. Later for the field trial, wireless logging of insect detection was implemented via the use of a Particle Boron© cellular networking module. Data from the sensor were received by Particle Boron through the physical connection on the circuit, facilitated by UART (Universal Asynchronous Receiver/Transmitter). The data were then temporarily retained in the module’s flash memory while waiting for a network connection then transmitted to an InfluxDB database via a HTTP post request. This removed the need for an SD card on the device and greatly simplified the data collection process, see Figure 1.

The field trial environment resembled real deployment conditions including limited availability of network medium. In such an environment, the cellular network was the only network media fit for the purpose. The Particle Boron module offered seamless hardware integration and simplified the rigorous process of registration and connection to a cellular network.

2.2. Smart Sensor Trials

Our smart traps were tested in the laboratory in a controlled wind-tunnel environment with three species of laboratory-reared moths: Helicoverpa armigera (n = 60♂, 40♀), Spodoptera litura (n = 35♂, 60♀) and Epiphyas postvittana (light brown apple moth (LBAM)); (n = 80♂, 40♀). These three moth species have wingspans of 30–40 mm, 30–38 mm and 16–25 mm respectively. They are pests in New Zealand and biosecurity threats to some other countries. The traps were also tested in an apple orchard field with wild E. postvittana.

For the wind-tunnel experiments, moths were received as pupae and set up in plastic bags until adult emergence. The pupae were maintained in a wind-tunnel room at 18–22 °C and 40% relative humidity (RH) under a 16:8 h light phase with dusk starting artificially at 3 p.m. On emergence, for each species and sex separately, 10 to 15 adults were released in the wind tunnel (90 cm in height, 55 cm wide, and 190 cm long) downwind, and one smart trap was suspended from the ceiling of the wind tunnel upwind. A white LED was placed on the smart trap sticky base to act as a generic lure to attract moths from the different species into the trap and a wind speed of 0.5 m/s was used to replicate optimal environmental conditions for these moths to fly. A Raspberry Pi HD camera with the infrared filter removed was used to record some of the trap visits to better understand insect behaviour at the trap entrance. Time stamps were used to match insect detection by our sensors and video recordings. The moths were set up usually just before dusk (3 p.m.) and left to fly around in the wind tunnel for 23 h, after which, data from the sensors and camera were collected and compared to actual insect catch visually counted on the sticky base. These wind tunnel experiments were used to better understand insect behaviour while entering the trap and comparing signal detection and wing beat frequency (WBF) across the three different species.

For field experiments, five smart sensor traps along with five standard delta traps were placed in a small organic apple orchard managed by the Lincoln University Biological Husbandry Unit in New Zealand (43°38′57.58″ S, 172°27′26.47″ E). The traps were spaced randomly every 3 m across two rows and placed at shoulder height (ca. 1.50 m). The traps were baited with the sex pheromones of the female E. postvittana (DESIRE LBAM, UPL, Auckland, New Zealand) to attract males, and were fitted with a sticky base to catch the insects. Once a week, traps were inspected, and the number of insects caught on the sticky base was visually counted. Event signals from the smart traps were sent automatically on detection to an online platform from which digital data were downloaded regularly. Sticky bases were replaced weekly after inspection and the lure changed after 6 weeks as recommended by the provider (UPL, Auckland, New Zealand). The field trial ran for a total of 84 days from early February to late April.

2.3. Data Processing

The data were collected as a series of wind tunnel and field trial sessions. Sessions comprised several detection events stored in CSV files. To extract only event segments with enough information to be useful for identification, an event extraction process was created, see Figure 3. First, high-frequency noise was removed using a low-pass filter (LPF) of 500 Hz. Next, binary low-frequency event (LFE) and high-frequency event (HFE) signals were created following this process: focus on the low/high frequency components using a LPF (30 Hz) or high-pass filter (HPF, 15 Hz); create a 4 Hz-resolution spectrogram; calculate the low-/high-frequency intensity (LFI/HFI) signal, defined as the mean magnitude per time step of the spectrogram; threshold the magnitude signal into a binary event signal; remove events shorter than 0.05 s and merge events with less than a 0.1 s gap between them. Finally, the LFE and HFE signals were combined into a single event signal such that when either or both LFE or HFE were high, the event signal was high; otherwise, it was low.

Each discrete event was classified into one of three categories: “fast flight”, “identifiable”, and “walking”. Events shorter than 0.1 s were classified as “fast flight” and deemed unsuitable for species classification, since there was insufficient information in the event. Events not containing any HFE signal were classified as “walking” (i.e., non-identifiable). The remaining events were cropped to the longest HFE segment and classified as “identifiable”. This included cases where the insect walked in with their wings beating enough for potential species identification.

Confidence intervals were calculated for the proportion of event categories for each species and sex using the Wilson score interval in a one-vs-rest method [27], e.g., the proportion of “identifiable” H. armigera events out of all H. armigera events.

There was a high degree of variability of event signals within a given species. In order to visualise all identifiable events by species in a way that conveyed the general shape of events of a species, as well as the species variability, line-array plots of the event spectrums were used, where every event spectrum is plotted as a line with low opacity resulting in a blending of each individual event. Event spectrums were created by selecting the 0.5 s window containing the largest amount of HFI and calculating the log of the power spectral density (PSD).

One commonly cited characteristic of wingbeat is the fundamental frequency. A close proxy for this is the dominant frequency, which is the frequency with the greatest magnitude. In some cases, the dominant frequency is the second harmonic rather than the fundamental. Histograms overlaid with violin plots were used to visualise the distributions of dominant frequencies.

3. Results

3.1. Insect Detection

The number of digitally detected events was always higher than the number of insects caught visually detected on the sticky base in the field (Figure 4) and in the wind tunnel experiment, see

Table 1. Summary of the total number of moths caught on the smart trap sticky base, i.e., visually counted, versus the total number of moths digitally detected by the smart sensor in the field (i.e., wild) and wind-tunnel experiments.

Species	Sex	Total No. of Visual Detections	Total No. of Digital Detections
E. postvittana (Wild)	Male	194	575
E. postvittana	Male	32	42
E. postvittana	Female	10	19
H. armigera	Male	36	50
H. armigera	Female	35	49
S. litura	Male	26	68
S. litura	Female	21	58

.

3.2. Insect Behaviour

Visual inspection of the video recordings of insect activity near the trap entrance in the wind tunnel trials revealed that moths did not always get caught on a sticky base immediately upon entering the trap, but they often spent a large amount of time moving about in and out of the trap before becoming stuck on the base, sometimes exiting and re-entering the trap multiple times. This led to multiple detection events being recorded for a single insect caught, explaining the discrepancies previously observed between digital and visual detection.

In addition, we also observed a diverse array of entering behaviours from the video footage of the smart sensor trap. Some moths flew directly into the trap, others landed on the outside of the trap before flying into the trap slowly or walking into the trap. Of those that walked into the trap, some continued to beat their wings while others did not. This diverse array of behaviour was also observed in the digital event recordings, which were classified into three categories during the data processing: “fast flight”, “identifiable” and “walking”, see Figure 5.

There were many more “identifiable” events recorded for the wild E. postvittana than from the recordings done in the wind tunnel with the three species of lab-reared insects. About 70% of the digital detections from wild E. postvittana were “identifiable”, and thus could be further exploited for species identification. These recordings were made only from males attracted to the sex pheromone, compared with both sexes tested in the wind tunnel with an LED light as a generic attractant, which may explain the difference in event proportions. Among the digital recordings performed in the wind tunnel, considering the Wilson score intervals, there were no clear differences in event proportions between the species for each event category. Although, H. armigera had a significantly lower proportion of walking events than identifiable events.

3.3. Species Identification

The spectra of “identifiable” events showed visually distinctive patterns between the species, see Figure 6. There were no clear differences between sexes, and thus not presented. E. postvittana in the laboratory had generally lower-magnitude events with no clear dominant frequency. In the wild, spectra were more consistent with a range of dominant frequencies and a faint range of second harmonics. H. armigera and S. litura presented clear dominant and harmonic frequencies, even out to the fourth harmonic.

The distribution of dominant frequencies by species is shown in Figure 7. In general, the species were not well separated by this specific measurement. The wild E. postvittana had a range of dominant frequencies from 30–40 Hz, with 39 Hz occurring most frequently, while in the laboratory there was no prevalent dominant frequency. H. armigera and S. litura had most prevalent dominant frequencies in the range of 31–32 Hz, with widely overlapping distributions.

3.4. Power Consumption

In terms of battery longevity, the smart traps were able to last up to 21 days in the field trial. The power consumption of the device is driven mainly by two sources: the sensor and the network module (Particle Boron device), with the former the main user of power. The sensor is comprised of two main parts: the STM32F091RCT6 microcontroller and the beam breaker sensor, which is made up of eight LEDs and eight corresponding photodiodes. The STM32F091RCT6 microcontroller consumed 56 mW in sleep mode with peripherals on, and the beam breaker sensor consumed 660 µW. The Particle Boron device consumed 482 uW in its hibernation. During data processing, STM32F091RCT6 microcontroller power consumption increased to 88 mW and Particle Boron power consumption increased to 74 mW, according to the datasheet. Therefore, for every detection event, the device must wake up, reconnect to the network and transmit the data, during which it consumed 2.8 times more battery power. We found the time taken to reconnect to the cellular network varied from seconds to minutes depending on signal strength. Therefore, the longevity of the trap battery was influenced by the frequency of detection and cellular reception.

4. Discussion

Our novel smart sensor trap designed for lepidopteran surveillance was able to detect every moth entering a trap, whether it was flying or walking, and often overestimated the actual number caught on the sticky base of the trap, since moths were observed to fly in and out several times before getting caught. Therefore, our novel smart sensor trap provides a real-time alert system for automated biosecurity surveillance that enhances early detection.

For biosecurity surveillance, where insect detections are rare and bycatch is low, any insect activity at the trap is of interest and merits the investment of a targeted manual trap inspection. A surveillance system by which manual inspections can be directed based on trap activity alone offers significant efficiency improvements over the current standard practice of inspecting every trap bi-weekly, irrespective of catch. The detection of insects not retained in the trap should trigger an increased search effort around the traps to confirm detection. It may even be possible to use transient detections to estimate the populations through modelling without the need for absolute counts of caught insects [28]. This approach would include any insects that were detected but flew out of the trap without becoming stuck on the trap’s sticky base in population estimates, and as such could prove to be more accurate for population modelling and prediction than the current standard practice. Thus, using smart sensor traps for surveillance should improve early detection and more accurate population estimation leading to increased eradication success.

Our smart optoelectronic device can not only detect and count insects but also make detailed time series recordings of their behaviour when entering a delta-type trap. However, not all digital recordings contained exploitable information for species identification, such as when moths entered traps at high speed or without beating their wings (i.e., walking). It may be possible to increase the amount of behavioural information captured in recordings of fast-moving Lepidoptera by adding additional depth to the sensor field through the use of secondary optics as described by Potamitis et al. [29]. While this would add additional complexity and cost to the overall trap design, the potential of reducing the number of unidentifiable detected moths may make this approach worthwhile.

From our field trial, we found that more than 70% of the recordings could be “identifiable”. These behavioural recordings could be used in the future to identify the detected species based on their species-specific wing-beat patterns. Not only the dominant frequency can be useful to discriminate species, but many other parameters could be exploited. For instance, Hassall et al. [30] used over 50 bioacoustic parameters to reach 80% accuracy in discriminating species of weak-flying aphids and beetles with a similar optoelectronic sensor. Given the complexity of the spectrogram, deep learning algorithm may be appropriate to allow data-driven feature extraction and classification. Convolutional neural networks could be also used with the spectrogram as an image, or recurrent neural networks or deep learning transformers could be used directly on the time-series signal.

Our current trap design aimed to minimise the cost of production and power consumption. The design could be further improved, at additional cost, to increase insect detection and discrimination between insects flying in and out of the trap. One limitation of our current smart sensor trap design is the restriction of insects flying in via one entrance where the sensor is located, reducing the insect trapping rate if wind direction does not disperse the lure downwind. Easy improvements such as adding a wing or a wind vane could be made to the physical trap to increase aerodynamic and self-rotation with wind direction to maximise lure dispersal downwind to attract insects [25,26]. Another sensor could be implemented at the other end of the trap to match the standard trap, but this would require more power to run and may be less practical for manual handling. Furthermore, a system that uses two sensors per entrance could discriminate between insects flying in or out and thus calculate absolute counts by subtracting those exiting from those entering, but again this would consume more power and thus would be more expensive to implement.

The main drive for our design was to produce a low-powered sensor trap that could last a long time in the field. Ideally, the trap battery will last as long as the lure, therefore requiring no maintenance in case of no catch over the season, i.e., usually three months. In our experiment, we measured a longevity of three weeks (21 days) from our five smart traps deployed in an orchard, and this was negatively correlated with the number of event detections. In sleep mode, the sensor consumed around 57 mW, 2.8-fold less than in wake mode sending data (162 mW). Even so, for a typical event with less than 2 s of data, a conservative estimate of the time required to log, connect to the network and post the data is 1 min at most, assuming no issues with the connection. So, the extra power spent while in wake mode is almost negligible on the scale of days. The power consumption during sleep could be drastically improved by using a modern arm M0+ microcontroller such as the STM32G0 or STM32L0 series which are specifically designed for low power operation but have been impacted by the global chip shortage and thus not available for our experiments. Likewise, the power consumption during transmission and the time taken to reconnect could be reduced significantly by using LTE cat 1 cellular networking rather than the 3G network that we used for our field trial. In a real surveillance deployment where detection of the target species is anticipated to be much rarer, the trap will be able to last much longer.

In conclusion, remote real-time detection with our current smart sensor trap offers a significant upgrade over existing manual delta traps, especially for biosecurity surveillance applications. The ability to direct trappers to inspect traps that have reported activity, regardless of the species which caused it, will significantly reduce the cost and carbon emissions associated with running a semiochemical trap-based surveillance network. The significant reduction in cost for manual labour could allow the expansion of current surveillance programmes to cover a wider area for surveillance and/or the ability to expand to new species. By enabling incursions to be detected early with smart sensor traps, the probability of successfully eradicating insects before they become established should be increased. Finally, the future possibility to acquire large datasets with our current smart sensor trap will improve machine-learning models to increase accuracy in species identification. Such datasets could thus become a new tool for diagnostics in addition to surveillance.