Towards reliable estimates of abundance trends using automated non‐lethal moth traps

Monitoring insect abundance or species richness at high spatial and temporal resolution is difficult due to personnel, maintenance, and post‐processing costs as well as ethical considerations. Non‐invasive automated insect monitoring systems could provide a solution to address these constraints. However, every new insect monitoring design needs to be evaluated with respect to reliability and bias based on comparisons with conventional methods. In this study, we evaluate the effectiveness of an automated moth trap (AMT), built from off‐the‐shelf‐hardware, in capturing declines in moth abundance, by comparing it to a conventional, lethal trap. Both trap types were operated five times on 16 plots from the beginning of July 2021 to the end of August 2021. On average AMTs recorded fewer individuals than conventional traps. However, both trap types depicted the same seasonal decline of approximately 3% per day, which corresponded to a total difference of ~85% over the sampling period. Given our sample size, both trap types had the same limitations in their reliability to detect smaller changes in abundance trends. This first proof of concept demonstrated that AMTs depict large magnitude events such as phenological patterns just as well as conventional, lethal traps. Therefore, AMTs are a promising tool for future autonomous and non‐lethal monitoring, which paves the way for high temporal coverage and resolution in insect monitoring. However, this initial quantitative field test revealed that its long‐term applicability must be preceded by several adjustments to the image quality, power supply and to data transfer.


INTRODUCTION
Recent reports on the decline in insect abundance and biomass (e.g. Hallmann et al., 2017Hallmann et al., , 2020Seibold et al., 2019) have raised considerable public concern together with calls for the immediate implementation of conservation strategies (Harvey et al., 2020). However, shortcomings in data availability have led to questions regarding the generalizability of this decline (Eggleton, 2020;Simmons et al., 2019).
Nonetheless, traditional survey methods are associated with various expenses, logistical challenges and a decreasing number of taxonomic experts, which limit the scalability, that is, the ability to increase the spatio-temporal coverage and resolution without delaying data processing. The development of automated monitoring systems that can be readily implemented and allow broad spatial and taxonomic coverage is therefore needed (Montgomery et al., 2021).
Visual surveys such as transect walks, net sweeping or hand catches are labour intensive. Conventional traps can reduce the demands of field work, but they are limited by their high maintenance and post-sampling efforts (Montgomery et al., 2021). Furthermore, caught specimens must be stored at least until their identification, but storage or freezer space may not be available or is limited. Lastly, conventional insect monitoring methods usually involve killing the insects for subsequent determination and biomass assessment, which has raised ethical concerns: Some traps can result in the slow death of the insects, which must be weighed against the knowledge and benefits for conservation gained from their use. The debate over consciousness of insects and their ability to feel pain should be recognised (Fischer & Larson, 2019). Even in the absence of detrimental effects imposed by the long-term monitoring of insect populations (Gezon et al., 2015), less destructive alternatives should be considered whenever possible (Drinkwater et al., 2019).
Given the above considerations, most monitoring campaigns focus on specific taxa or locations, with the latter often being chosen based on accessibility and a high density of the target taxon (Didham et al., 2020;Fournier et al., 2019). However, this could result in a bias due to non-random site selection and, in turn, to biased estimates of abundance or biomass and of the associated trends (missing zero effect, Didham et al., 2020). Another practice is to focus on sampling periods that are assumed to adequately cover insect diversity (Ozaki et al., 2011;Scalercio et al., 2012). However, such temporal snapshots, whether based on daily or on annual data collection, imply a loss of information (Ozaki et al., 2011), because phenological shifts and species occurring outside the general phenology peak will remain undetected (snapshot & groundhog effect, Didham et al., 2020). Monitoring networks that collect data on a daily basis, such as the Rothamsted insect survey (The Insect Survey, 2021), are a notable and rare exception. Yet even this dataset, made up of >10 million records, and thus probably more accurate than any other, cannot provide complete species inventories at the spatial resolution relevant for conservation (Sánchez-Fernández et al., 2021).
A limitation of manual species identification is the steadily decline in the number of taxonomic experts (Engel et al., 2021), such that the number of sampling points cannot be increased indefinitely. Machinebased automated species determination, especially artificial intelligence (AI)-based image recognition, reduces both the workload and the cost associated with monitoring, thus allowing the number of sampling points to be scaled to the required spatial and temporal extent and resolution (Didham et al., 2020;Montgomery et al., 2021). Furthermore, non-lethal monitoring systems can provide real-time information about the state of populations while eliminating the ethical concerns related to lethal sampling (Didham et al., 2020;Montgomery et al., 2021).
Establishing automated monitoring systems requires several development steps. Automated monitoring systems must be as costefficient as possible to achieve the needed spatial coverage and resolution (Hahn et al., 2022). For AI implementation, training data needs to be collected and labelled in sufficient amounts, which can be expensive, especially under in situ conditions (Høye et al., 2022).
Moreover, the development from prototype to fully functional monitoring tool requires an adaptive process whose applicability and functionality in the real world is constantly re-evaluated and adjusted accordingly (Hahn et al., 2022). While all of these considerations are crucial in themselves, first a basic proof of concept must demonstrate that the automated system accurately captures the response of the target group to changes in the environment. The efficiency of newly designed traps should therefore be compared with that of conventional traps. Such proof of concept then sets the stage for further investment.
Generally, every change in the design of a trap affects its capture efficiency (Preti et al., 2021) and therefore the conclusions regarding abundance, species richness and composition. For example, flight interception traps consisting of only one collecting jar at the bottom, rather than an additional collecting jar at the top, might underestimate taxa that tend to move upwards after collision (Knuff et al., 2019). Even in traps targeting single species, slight adjustments such as the width of the trap opening or the addition of a rain cover can affect the number of captured individuals (Boetzl et al., 2018;Burner et al., 2021;Guarnieri et al., 2011). In extreme cases, design-related differences in community and/or abundance data can bias ecological conclusions (Saunders & Luck, 2013), especially when the effectiveness of the trap system depends on environmental conditions and species traits (Burner et al., 2020).
In most monitoring systems, including lethal ones, the probability of detecting an individual depends on its activity (Didham et al., 2020;Holyoak et al., 1997;Wölfling et al., 2016). This dependency is likely to be reinforced in camera-based, non-lethal systems, in which individuals are recorded only when they enter the camera's field of view.
Since, unlike in lethal systems, individuals are not permanently removed, they might be counted several times if they leave and reenter the detection area. In such cases, variations in activity, and thus movement, might change not only the probability of detection but also the likelihood of duplicated counts. This could in turn lead to discrepancies in abundance trends determined by conventional vs. camera-based monitoring systems. Accordingly, we address the detection of variation in abundance, which is a fundamental ecological pattern often used to assess insect communities (van Klink, Bowler, et al., 2022;Winfree et al., 2015) as the first baseline for further development of features like species identification. Specifically, we investigated whether an automated moth trap (AMT), which attract moths to a screen via ultraviolet (UV) light and then photographs them (e.g. Bjerge et al., 2021;Hogeweg et al., 2019), leads to abundance trends consistent with those determined using conventional (lethal) light traps.
Moths are one of the most diverse insect groups and they are closely tied to their ecosystems (Fox, 2013). Accordingly, they have a long history in monitoring studies (Sánchez-Fernández et al., 2021) and are often the commonly targeted taxon in new identification algorithms (Chang et al., 2017;Poremski, 2017;Wu et al., 2019). However, moths are highly sensitive to trap design (Brehm & Axmacher, 2006;Fayle et al., 2007;Hausmann et al., 2020). For example, differences in relative abundance were observed between manual sampling, in which species are recorded directly when they land, and funnel traps, which are not necessarily entered by all moths (Brehm & Axmacher, 2006). An AMT, on the other hand, captures the moths directly, but only if they land on the screen photographed by the camera.
In this study, we deployed an AMT and a conventional bucket trap at the same sites throughout the late summer, shortly after the phenological peak of moths. We reasoned that, if the two trap types are equally efficient in capturing local moth abundance, they should show a similar temporal decline in moth individuals. In addition, by deploying the AMT under realistic in-field conditions in a relatively remote area, we were able to assess its usability while also identifying its weaknesses, which can then be addressed in subsequent iterations.

Study sites
Our study was part of the Nature 4.0 framework (www.natur40.org) and was conducted in the 'Marburg Open Forest', a 1.5 km 2 university-owned forest area near Caldern, in Hesse, Germany. The forest is a typical beech-dominated managed forest embedded in an agricultural landscape. The 16 trap sites were selected from the existing Nature 4.0 sites and were evenly distributed over the research area, thus covering small-scale climatic, structural, and ecological differences. Each site included a beech tree (50-120 years-old) in its centre and is separated from the other sites by least 50 m. An overview map can be found in Appendix Figure S1.1.

Conventional traps
The conventional light traps (Appendix Figure S2.3) consisted of a super-actinic UV light tube (Sylvania blacklight F15W/T8/BL 368, 12 V, 15 W, wavelength peak: 368 nm) that was used to attract the moths, which were then directed via a funnel into the attached 10-l bucket. The bucket contained a closed chloroform jar from which a piece of cloth was drawn through a small hole in the lid to act as a wick, thus filling the bucket with the killing agent. Pieces of egg carton were added to allow the caught moths to rest, thus also reducing stress and wing damage. The UV light was powered by a 12 V lead battery. A light switch (Kemo Germany M197) attached to the battery automatically turned the light on at sunset and shut the light off at sunrise. In the morning, the traps were emptied and all moths were collected and stored in a freezer for later sorting.

Automated moth traps
A prerequisite of our AMT was that it should operate in a real-world environment and in a variety of settings. Therefore, the design requirements of the AMT were as follows: • modular and flexible, allowing both extensibility and adaptability for further applications • reproducible, such that the AMT could be easily assembled by other researchers • cost efficiency, to allow simultaneous instalments of multiple AMTs • robust, to allow continuous operation of the trap in outdoor settings • easy to use and configurable, to support long deployment periods while minimising the work load Following these requirements, we adopted the basic design of Bjerge et al. (2021) but used our own software and customised the hardware to meet our needs: The AMT consists of several power consumers (LED-UV light, LED screen, Raspberry Pi, ring light, and camera). To achieve a comparable runtime of the two trap types without access to a power grid and to reduce energy consumption as well as overall acquisition costs, the 15 W super-actinic UV light tube (176 €) in the conventional trap was replaced in the AMT by a LED-UV tube (www.entosphinx.cz; 37.12 LED/UV lamp, 9.6 W, wavelength from 395 to 405 nm, 33 €). Below the LED-UV tube, we installed a waterproof, white, cloth-lined LED screen illuminated by four rows of 30 cool-white LEDs (6000 K, 240 lumens/m). The LED screen acted as a resting site for the moths. The illuminated screen allowed for standardised photos , which were taken by an oppositely placed camera box with an attached ring light for illumination. Both the camera and the ring light were directly connected to the Raspberry Pi computer. The Raspberry Pi's software was programmed to trigger a customizable schedule, according to which photos were taken and the light (UV and LED screen) was controlled.
A 5MP Raspberry Pi-camera, v1.3 was chosen to further reduce costs.
The whole system was mounted on a tripod to avoid disturbance by animals, foliage, and puddling. To enable the longest possible operation time, only the UV light and LED screen were powered by a 12 V battery. The Raspberry Pi, which is susceptible to voltage fluctuations, as well as the ring light were powered by a 5 V power bank inside the camera box to ensure stable voltage even when the LEDs were being powered. This setup also avoided energy loss during voltage conversion, since the Raspberry Pi requires a 5 V power supply. To ensure the reproducibility of the design, the trap was constructed from offthe-shelf hardware components. All of the components, their prices, and the measured power consumption are listed in Appendixes S3 and S4 (for AMT's architecture and a detailed description of the trap design, see Appendix Figure S2.3).
The system is modular, with a sensor box at its core that switched the UV light and the LED screen on and off by means of relays. Both the UV light and the LED screen were connected to the sensor box via waterproof cables and connectors, so that all circuitry was contained in the sensor box. The specific UV lamp or LED screen could thus be replaced without having to change the sensor box itself. Only the ring light, which was also controlled by a relay, was permanently connected to the sensor box. To ensure robustness, all components, except the ring light, were waterproof. Ease of use was ensured by using standardised plugs and reverse-polarity-protected sockets for the individual components. In addition, the software was designed to be configurable and to start automatically.
Our software implementation was based on an extended version of the software of Gottwald et al. (2021), which was built using the pimod tool (Höchst et al., 2020)

Sampling/experimental design
The two trap types were tested at separate times, to avoid their mutual interference during sampling. Thus, the eight AMTs and eight conventional trap systems were randomly assigned to the 16 sites for one night. The following night, the trap types were exchanged such that sites assigned an AMT the first night received a conventional trap the following night and vice versa (two nights = one round). This routine was performed five times from July 2021 to August 2021. The AMTs were scheduled to take photos every 30 s. Activation of the lights (UV and LED screen) was scheduled to start at 10:00 PM and end at 6:00 AM, with an on time of 50 consecutive min/h followed by 10 min of lights out. This was done to test the functionality of the hourly schedule, while roughly mimicing the light phase of the conventional trap. Examples of the configuration of our specific unit and its operation schedule are shown in Appendix S5. With this schedule, the theoretical power consumption was 19.175 Wh (see Appendix S3).

Moth counting
Moth abundance in the photos taken with the AMT was determined by manually counting the number of moths in each frame.
For the total number of moths per night, only new arrivals on the LED screen were counted; in other words, when the total number of resting insects in a photo superseded the total number of resting insects in the prior photo, the moth count was raised. This method was used because movements of the insects on the LED screen made it difficult to identify individuals. In addition, the possibility that the same individual flew out of the photo, swirled around the light, and then again rested on the LED screen could not be ruled out. Accordingly, in the frame by frame counts of the insects, only the total number of currently sitting insects was considered. Nontarget insects were not counted. Due to varying image quality, very small insects could not be identified with confidence. Thus, moths smaller than approximately 0.8 cm and without an easily identifiable characteristic moth profile could not be included in the count, since the distinction from, for example, small Diptera species was no longer possible due to blurriness. Samples from conventional bucket traps were determined to the insect order and moths counted using a binocular.

Statistical analyses
The analyses were aimed at testing whether the number of moths captured by the AMTs were comparable to the number of moths captured by the conventional traps, thus yielding similar seasonal patterns in the number of individuals.
First, a variance-covariance matrix was used to evaluate whether the number of individuals captured by the AMT covaried with the number of individuals captured by conventional traps across sites and sampling events. We then tested whether the two trap systems differed in depicting abundance trends. Therefore, we took advantage of the phenology of moths. In Europe, moths usually peak in abundance and diversity around July (Roth et al., 2021;Sánchez-Fernández et al., 2021). The later the sampling time thereafter, the fewer moth individuals in the area. We thus used the Julian day, which well-reflected the time lapse between the rounds and thereby possible seasonal trends, as an independent variable to model the number of individuals using a generalised linear mixed effect model (GLMM). We centred the Julian day such that its intercept was in the middle of the sampling period, which thus reflects the differences in abundance. The trap type was included as an interaction term to test for differences in the systems in depicting the temporal trend in abundance. To account for the nonindependence of data points due to the repeated sampling per plot, we also added the Plot ID as a random intercept: Model : Number of Individuals $ Trap type Ã centred Julian day ð Þ þ1 j PlotID ð Þ : Because our count data displayed overdispersion, we fitted the model using a negative-binomial error structure using the lme4-package (Bates et al., 2015) and goodness-of-fit was assessed using the DHARMa-package (Hartig & Lohse, 2022). Note that ideally one should also account for potential random differences within the slope, for example, because of slightly different community compositions.
However, including the centred Julian day as a random slope led to over-parameterization for the given number of observations, which made inferential procedures inappropriate (Bates et al., 2015). An autoregressive error structure to account for temporal nonindependence of the data points has also been omitted for the same reasons. Lastly, prediction intervals were calculated using the ciTools packages (Haman & Avery, 2020) for all plot, day and trap combinations and averaged across plots for each Julian day and trap.
To estimate the reliability of the traps and to check whether the two traps would perform similarly well when the decline becomes less steep, we also performed a power analysis with the simr-package (Green et al., 2022). The power analysis first simulated new Y values for different slopes, based on the parameters of the original model.
The model was then recalculated with these Y values and the p-value was extracted. The whole process was repeated 1000 times.
We used this method to compare the statistical power of AMTs To mirror the main model, we kept the Plot ID as random factor. Note, however, that this resulted in variance components close to zero.
Though this should not affect the results, we also tested the models without the inclusion of Plot ID, leading to comparable results.
The estimate of the slope was then exchanged with the desired slope (from 3% change, which is close to the observed slope, to 0.1% change). The other parameters of the models, that is, intercept, SEs, and covariance matrices, were retained. The simulations were initially based on the observed data points. Additionally, we simulated the statistical power for the planned sampling events (that is, without trap failures) using the simr::getData-command. All analyses were conducted using R (R Core Team, 2021) with RStudio (RStudio Team, 2021) and all figures were created using the package 'ggplot2' (Wickham et al., 2021).

RESULTS
The AMTs worked completely fault-free and delivered data in $40% out of a total of 80 sampling events (5 rounds Â 16 plots). In the remaining events, failures resulted in partial to total data loss. The Although they covaried, the overall variance of the latter was 2.3 times higher than that of AMTs (Appendix Figure S6.1).
Sampling conducted later in the year resulted in significantly fewer recorded individuals overall, reducing the average number from 592 (CON) and 303 (AMT) individuals recorded in the first round of sampling to 82 (CON) and 49 (AMT) recorded individuals in the last round of sampling. Yet the two trap types did not differ in depicting the seasonal decline of moths in the study area (Table 1), which was $85% for the studied time interval (Figure 2).
In our setting, both trap types had the same statistical power to reliably detect a decline of more than 2% per day ( Figure 3).
The threshold of 0.8 statistical power was exceeded at a decline

DISCUSSION
Automation can facilitate the monitoring of insects at high temporal and spatial resolution, but automated monitoring must be able to track processes and changes in nature (Collett & Fisher, 2017). In addition, these systems must be user-friendly, easy to maintain, cost efficient, allow for long operation times, and sufficiently robust to withstand environmental conditions, such as rain, dew, and wide temperature fluctuations (Preti et al., 2021). In this study we tested the ability of an AMT to depict ecological patterns, as well as its in-field applicability. While our field test demonstrated the comparability of AMTs and conventional traps in terms of capturing trends in moth abundance, the need for improvements in the AMT's design was also identified.

Field-work applicability and lessons learned
In evaluating the in-field applicability and usability of our ATM, two issues came to our attention that need to be addressed. First, the DIY  prototypes constructed from consumer-class hardware suffered problems in the field, despite prior laboratory testing. These problems resulted in data loss (Appendix Figure S7.1). As technical failures such as these can only be identified during field work and/or during longterm deployment, they demonstrate the importance of thoroughly testing trap prototypes in order to achieve a truly operational design.
For example, although the LED screen was initially thought to be beneficial for the standardisation of photos, during the course of our study it became apparent that, while the LED screen guaranteed an even background, backlighting led to underexposure of the moths themselves. The use of a flash strong enough to evenly illuminate the background without underexposing the moths would make the LED screen obsolete, thus also reducing acquisition and energy costs.

An honest evaluation of the costs and benefits
As the aim of this study was to determine whether our AMT could provide information on moth abundances and the changes therein, the generation of a fully autarkic, automated monitoring system was beyond its scope. The power supply of the AMT tested in this study consisted of portable energy storage, since the number of AMTs was limited and the traps had to be shared between plots. However, in a long-term deployment the additional investment in photovoltaic systems or, in shadier locations, shade-tolerant solar panels, would be worthwhile. A similarly constructed bat monitoring system, described in a previous study (Gottwald et al., 2021), was likewise operated autonomously. In addition, a shorter illumination time would save energy; for example, a reduction to 30 min would save $7 Wh, or even 12 Wh without the LED screen (see Appendix S3). For data transfer, rather than frequent visits to the trap sites to exchange the SD card, the research site could be remotely accessed. Data could be directly extracted using LTE sticks (Gottwald et al., 2021), or, when restricted by mobile phone reception, LoRa (Ferreira et al., 2020), which would again require additional infrastructure. A further necessity is data-processing software (a trained neural network) that would allow the Raspberry Pi to analyse and process the images in-field, thus reducing the size of the output files.
These investments in solar power and automatic data transfer will add to the current total material costs of $500 € per trap, in addition to incurring further running costs. Nevertheless, the cost of an updated version of the DIY-system presented herein would still be at the lower end of the price range of systems utilising more elaborate components, such as high-end cameras (e.g. AMMOD; Wägele et al., 2022), or custom-built cases (e.g. www.diopsis.eu; Diopsis, 2022). The cost still obviously exceeds that of conventional traps, including the model used in this study (240 €) but also those in other studies (e.g. White et al., 2016). The acquisition costs of automated monitoring systems will amortise over the long run, through a reduction of the workload both in the field and during AIbased post-processing. Our study did not address the latter as our focus at this stage was to acquire data on the number of individuals, without the uncertainties of AI, which despite showing promising results, has to be fully trained and validated (see Appendix S8). However, a full validation can be expected in the near future  and will allow automated identification to become nearly instantaneous. Admittedly, automated counting will not necessarily lead to a crucial cost savings relative to the acquisition cost of AMTs, as the post-processing costs of conventional traps are relatively low if only abundance data are considered.
The fastest amortisation and greatest entomological value would be gained by automated species identification. The manual identification of species is both time-consuming and costly (yet also rewarding), and remains the greatest challenge in large-scale monitoring using conventional traps. For example, for this study the cost of sending the moths caught during each sampling event (night/plot) using the conventional trap to a specialist for identification would have been $110 € (on average, 218 individuals, current common price 0.5 € per individual). A preliminary test showed that a slight upgrade in the camera (from RPI CAM 5 MP, v1.3, resolution 2592 Â 1944 px (6.55 €) to RASP CAM HQ 12 MP with wide-angle lens, resolution 4056 Â 3040 px (88 €), Appendix Figure S2.2) and an adjustment of the illumination would produce images in which conspicuous species and genera could be identified Korsch et al., 2021). This upgrade would provide data not only on total moth abundance but also on species abundance and composition. Further studies need then to test for differences in the species composition obtained with the AMTs vs. conventional traps.

Abundance and abundance trends
Before AMTs can be used in monitoring schemes, their ability to accurately capture ecological patterns must be ensured. A quantitative assessment is the first step in analysing variations in the ecological patterns of insects, such as those arising from changes in the environment and in the viability of populations (Gaston, 2000;Hausmann et al., 2020). However, trap counts only provide an estimate of the total number of individuals (Mistro et al., 2012).
Abundances detected with AMTs could be over-or underestimated because the counting mode of AMTs is fundamentally different from that of conventional traps and because it may attract different species due to differences in insect settlement behaviour (Brehm & Axmacher, 2006;Wölfling et al., 2016). Indeed, in this study, AMTs Conversely, a duplicated count bias and thus an over-estimation is also possible, as some moths may have flown away during the 10 min without light and then returned to the screen after the light had been switched on again. These duplicated counts might have contributed to the few instances in which the AMT captured a higher fraction of the total population than the conventional traps. Yet, these instances also coincided with the operation of the AMTs during nights that were slightly warmer than those during paired conventional trap operation (Appendix Figure S7.2), indicating that the effect of weather conditions exceeds the effect of the trap type. Without marking and tracking individual moths, it is not possible to completely disentangle the effect of changes in ambient weather conditions between trapping events and potential inaccuracies in the counting method.
Most methods are subject to an inherent error that affects the estimation of absolute population size-especially for insects whose population sizes vary considerably (Didham et al., 2020). Yet, as long as this error remains constant over the collection period, it should not affect the ability to depict and compare trends. Given that there was no difference in slope, AMTs were just as good at detecting large changes in population size as conventional traps. Conversely, it could just as well be said that both trap types are similarly poor at detecting temporal declines of abundance, considering that only decreases above 40% could be detected. However, one must bear in mind that this was a small experimental set up, designed to test the functionality of an AMT. For this purpose, a definite trend (in this case seasonal decrease) was needed to test for slope differences that would not be biased by snapshot effects (Didham et al., 2020).
Neither slopes nor statistical power of the observed data points differed between the two traps, indicating that AMTs could be used for more intensive monitoring. In order to perceive and predict smaller trends, it is necessary to sample repeatedly over a long period of time and at several sites, regardless of trap type. While spatial scalability is dependent on the amortisation rate, if the above robustness and data transfer issues are resolved, the AMT would allow for higher temporal scalability than the conventional trap. Not only would it have a lower depletion effect, but it would also reduce the interval in which data is collected without increasing the effort. Further, this should avoid a 'groundhog effect', that is, the phenological mismatch between sampling and activity period (Didham et al., 2020).

Potential and outlook
The  (Stuijfzand et al., 1999) could be monitored in a similar, non-lethal, fashion using AMTs. Finally, the modularity of the AMT allows for its easy adjustment to specific needs, for example, the use of stronger UV LEDs, depending on whether the aim is to compare demarcated locations or obtain qualitative assessments over a wider area.

Conclusion
This study provided a first demonstration of the comparability of a low-cost DIY AMT with a conventional trap in capturing significant trends in moth abundance. Clearly, AMTs and automated species identification cannot replace the expertise of entomologists, as some species cannot be identified from photographs alone. However, an automated system can perform 'taxonomically trivial tasks ' (van Klink, August, et al., 2022) and thus serve as a complimentary tool in species detection and identification. Perhaps most importantly, AMTs can be deployed in the field for long periods of time without the ethical issues associated with conventional moth sampling.
Overcoming the technical barriers that hinder the implementation of ATMs and fully automated monitoring will require a high level of expertise. Cooperation and knowledge-sharing between different working groups and interdisciplinary fields of expertise are essential (Hahn et al., 2022). Both successes and failures should be communicated, to minimise development time and resources. Accessibility and functionality should be reviewed in parallel with each design change, to ensure that the AMTs can be operated and the data analysed by non-specialists. In this regard, initiatives aimed at coordinating knowledge-transfer (e.g. wildlabs.net; Wildlabs, 2022) are highly valuable. With the full automation of the monitoring system, the AMTs will provide the fine-grained monitoring needed to detect temporal changes in insect populations. Although there is still a long way until full automation is achieved, its potential applications as well as the first proof of concept presented in this study indicate that it is worth taking these steps.