Ninety-Two Per Cent of Nine Species: Automated Insect Identification, and the Conditions That Take It Apart

1. Introduction: the number and the question under it

Automated identification is arriving in pest monitoring, attached to accuracy figures in the low nineties. Those figures are real. The question is what they are figures for.

1.1 What this paper does

Reads the reported accuracies alongside the conditions they were obtained under, identifies the specific ways performance degrades, and sets out what a purchaser should ask. It is not an argument that the technology does not work, and §14 says where it plainly does.

2. The best result located

The headline figures, with their context.

The best trained model measured an average precision of 92.7 per cent in classifying the species correctly with an average recall of 93.8 per cent, and the authors state we know of no study with such high classification performance. They report achieving a high precision above 90 per cent for detection and classification across nine different insect species.¹

2.1 What the study was for

The stated motivation is that it is crucial to understand trends in insect species, both for agriculture and wildlife conservation, and the authors note the method could contribute to pest monitoring with camera-equipped traps without killing rare insect species.¹

That non-lethal property is a real advantage this article should not skip past. A sticky card kills everything that lands on it, including beneficial and uncommon species, and a camera that counts without capturing removes that cost entirely.

2.2 Precision and recall are not the same claim

Precision is how often an identification was correct. Recall is how many of the insects present were found at all.

Reporting both is good practice and is what allows the two error types to be distinguished. This journal made the same point about detection rate and false-positive rate in canine scent work: a single accuracy figure conceals one of a pair.

3. What nine species means

The scope condition that governs everything else.

The nine species were bees, hoverflies, butterflies, and beetles in the dataset description.¹

3.1 Why the number of classes is the difficulty

A classifier choosing among nine options is solving a different problem from one choosing among several hundred. Confusable pairs multiply as classes are added, and the groups above are morphologically distinct at a glance.

A structural pest programme needs to separate species that a trained entomologist distinguishes under magnification: small flies that share a drain, stored product beetles that share a genus, ants that differ in social structure and therefore in treatment. None of those are nine easily separated groups.

The ant case is the sharpest. This journal found that supercolonial and territorial species require opposite baiting strategies, and the two are separated by characters no camera at trap distance resolves.

3.2 The confusion structure matters more than the count

Nine classes is easier than thirty-five, but the more important property is how similar the classes are to each other. Bees, hoverflies, butterflies and beetles differ in wing structure, body shape and posture.

A model separating a carpet beetle from a larder beetle, or Drosophila from Megaselia, is working within a genus or a family where the distinguishing features are setae, wing venation and antennal segments, often requiring magnification a trap camera does not have. Class count is a proxy for difficulty and a poor one.

3.3 The honest framing

We are not criticising the study for having nine classes. It was a biodiversity monitoring paper and nine was the appropriate scope.

The criticism is of the transfer. A figure obtained on nine distinct groups should not be quoted as the accuracy of automated identification generally, and §20 makes that the first question to ask.

4. The dataset behind it

What producing that result required.

The models were trained on an open access image dataset containing over two million images with 29,960 annotated bees, hoverflies, butterflies, and beetles against varied floral backdrops.¹

4.1 Two million images, thirty thousand insects

The ratio is worth noticing. Over two million images yielded 29,960 annotated individuals.¹

Most frames contain no insect. A camera monitoring a flower or a trap is overwhelmingly recording nothing happening, and the system's first job is discarding the empty majority. That is a different engineering problem from classification and it is where most of the compute goes.

4.2 The annotation cost

Nearly thirty thousand individually annotated insects is a very large amount of expert labour, and it is the input that is hardest to replicate for a new target.

A company wanting the same performance on the species that matter in a food premises would need a comparable annotated set for those species, on the substrate they will be photographed against. That is the real barrier, and it is not a computing barrier.

5. The problem the authors say they solved

Stated plainly by them, and it is the most useful sentence in the paper.

Most applications are developed for images captured in lab conditions, against solid backgrounds or by manual photography. Until now, deep learning models have not permitted the detection of multiple insect species against complex vegetated backgrounds.¹

Their contribution is described as solving the more difficult challenge of detecting multiple small insects in large images.¹

5.1 What that tells us about the rest of the literature

That as of that paper, most published work was on plain backgrounds under controlled photography.

Anyone reading an accuracy figure from this field should establish which of those two situations produced it, because the authors of the best result are themselves saying the distinction is the hard part.

6. The honest account of what remains hard

A later pipeline paper describes the difficulty without softening it.

Accurate image recognition analysis is challenging, particularly for images containing small insects against complex backgrounds with diverse vegetation communities. And even when insects can be detected in images, identifying their taxonomy remains difficult, particularly in footage with low image resolution, light conditions, and distances from the plants, and in cases where insects appear blurry or only partially visible.²

6.1 Every one of those applies indoors

A structural pest environment is not a flower meadow, but it presents the same list. A camera in a warehouse or a plant room works in poor and variable artificial light, at whatever distance the mounting allows, against backgrounds of stacked product, pipework and machinery.

Small insects against complex backgrounds is a fair description of a sticky card in a cluttered room, and low resolution at working distance is guaranteed by the size of the target. Nothing about moving indoors relaxes the conditions the authors name.

6.2 The value of a paper saying this

It is an AI pipeline paper listing the conditions under which AI pipelines fail. That is the opposite of the pattern this journal found in canine detection marketing, and the authors deserve credit for it.

7. Detection against identification

The distinction the previous section turns on.

Detecting that an insect is present in an image and determining which species it is are separate tasks with separate difficulty, and the quoted passage says explicitly that the second remains hard even when insects can be detected.²

7.1 Why detection is the easier half

Detection asks whether this region of the image differs from background in ways characteristic of an animal. That signal survives blur, partial occlusion and low resolution reasonably well, because it depends on coarse properties.

Identification depends on fine properties, which are exactly what degrades first as resolution falls. The two tasks fail at different rates for the same image quality, which is why the papers report them separately.

7.2 Why this matters for pest work

Counting how many things landed on a card is useful. Knowing which species they were is what determines the treatment.

This journal's drain fly article turned entirely on species identification determining where to look for the breeding site. A system that counts accurately and identifies poorly answers the question nobody needed answered.

8. The taxonomic rank trap

A specific case worth flagging because the number looks impressive.

One system using a combination of two approaches reported counting and classification precision of 92.5 per cent and 90.8 per cent respectively, but the classification was only done to the taxonomic resolution of insect orders.³

8.1 What order-level means

Beetle, fly, moth, wasp. It is a real classification and it is not one a pest technician needs a camera for.

A figure in the nineties at order level and a figure in the nineties at species level are different achievements by a wide margin, and they are reported in the same units.

8.2 Where order-level output is still useful

Fairness requires the other side. Knowing that a card caught flies rather than beetles narrows the investigation substantially, and doing it automatically across hundreds of cards has value a technician's time does not stretch to.

The objection is not that order-level classification is worthless. It is that it should be described as what it is, because a facility manager reading 92.5 per cent will assume it refers to the identification their programme depends on.

8.3 Why the authors reported it honestly

They state the limitation in the same sentence as the number, and explain that the method is difficult to scale up to individual species classification as selecting useful features would be challenging.³

The caution is in the literature. It is in the transfer to a sales claim that it gets lost.

9. The older species-level results

What happened when earlier work attempted species on a harder set.

Analysis of 774 live individuals from 35 different moth species to determine whether computer vision could be used for automatic species identification, using data mining for feature extraction and a support vector machine for classification, achieved a classification accuracy of 85 per cent among 35 classes. A nearest neighbour algorithm based on texture, colour and shape applied to the same dataset obtained an accuracy of 79.53 per cent.³

9.1 Reading this against §2

Thirty-five classes produced 85 per cent. Nine classes produced 92.7 per cent. The comparison is not clean, because the methods and the decades differ, but the direction is what theory predicts: more classes, lower accuracy.

An 85 per cent species-level result across 35 moth species is arguably the more impressive of the two for anyone whose problem involves many confusable species.

10. Why feature-based methods stalled

The technical reason the field moved to deep learning.

Because classification was made based on manually defined features, it is unlikely the method would be easy to expand for moth species classification as selecting useful features would be challenging.³

10.1 The fruit fly exception

One reported system did something harder than species identification on a single target: a trained model performed sex prediction and discrimination based on 4,753 annotated flies for a spotted wing Drosophila monitoring system.³

Distinguishing males from females within one species is a fine-grained visual task, and it worked with under five thousand annotated examples. That suggests the binding constraint is annotated examples per distinction rather than the difficulty of the distinction itself, which is an encouraging reading of §3.2.

10.2 What changed

Deep learning removes the requirement that a person specify in advance which visual properties distinguish the species. The model derives them from labelled examples.

That is why the annotation cost in §4.1 became the binding constraint. The expertise did not disappear; it moved from defining features to labelling images.

11. The trap colour finding

The result we think is most important for this trade, and it concerns the substrate rather than the model.

Using one architecture with transparent sticky traps as training data, the model predicted the pest species on transparent sticky traps with an accuracy of at least 0.95 and on other sticky trap colours with at least 0.85 of the F1 score. Statistical analysis showed that the colour and architecture of the sticky traps significantly influenced the performance of the model.⁴

11.1 What was varied

Not the insects, not the lighting, not the camera in the primary comparison. The colour of the card the insect was stuck to.

11.2 Why the study is unusual

Most work in this field varies the model and holds the imaging conditions fixed, because the model is what the researchers are contributing. This one held the model approach roughly fixed and varied the physical setup.

That is the harder question to ask and the more useful one to answer, because the physical setup is what a purchaser controls and the architecture is not.

12. Why that result matters so much

Because it identifies a failure mode nobody would think to check.

A purchaser evaluating such a system would ask about species, about lighting, perhaps about camera resolution. Asking whether the model was trained on the colour of card their programme uses is not an obvious question, and the answer changes performance materially.

12.1 The generalisation

A model learns the whole image, including the background. Change the background and you have changed part of what it learned from, even though the insect is identical.

That is our framing rather than the authors', and it predicts that any change in the imaging context, being trap type, adhesive, card texture or mounting, is a candidate for the same effect.

12.2 The retraining question

If the model is substrate-sensitive, then every change to the physical programme is potentially a retraining event. Switching card supplier, moving to a different adhesive, changing camera: each is a candidate for the effect §11 measured.

Nobody in the purchasing conversation would think to ask about that, and we have not seen a vendor commitment to revalidate after such a change. It belongs on the list in §20 and we have put it there.

12.3 The honest counterweight

Performance on the untrained colours was still at least 0.85, which is not collapse. The model degraded rather than failed.⁴

And the authors' conclusion is constructive: that development of automatic classification on sticky traps should focus on colour and deep learning architecture to achieve good results, with future work incorporating the trap system into pest monitoring to provide more accurate and cost-effective results.⁴

13. The interaction with pest biology

The part that makes §11 a genuine problem rather than a solvable annoyance.

Pests are not equally attracted to all colours of sticky traps and different coloured sticky traps are typically used for different pests.⁴

13.1 The conflict this creates

Trap colour is chosen for the insect's behaviour. Model accuracy depends on trap colour. So optimising the trap for catching and optimising it for automated reading can point at different cards.

This journal has documented the same class of conflict before: in mating disruption, where the treatment and the monitoring instrument shared a signal, and in pheromone trapping, where trap design altered the number that was supposed to measure the population. The instrument and the thing being measured are not independent.

13.2 The same problem in a different form

There is a second version of this conflict. A model trained on clean cards will meet cards that have been in service for weeks, covered in dust, non-target insects and partial specimens.

The trap colour study varied a clean variable deliberately. Service conditions vary many at once, and none of the studies we located tested a model on cards aged in a working facility.

13.3 Which should win

Catching, in our view. A trap that reads well and catches poorly is measuring the wrong population, and the automation is downstream of the sample.

That is a judgement rather than a finding, and the practical resolution is to train the model on whichever card the entomology requires.

14. Where the technology clearly wins

The counterweight this article owes, and it is substantial.

A smartphone application for counting a mite pest had higher counting accuracy, 7.1 per cent higher for motiles and 12.7 per cent higher for eggs, and speed, 29 seconds faster per leaf, than manual counting using a magnifying lens.⁵

The stated problem is that the traditional method of manually counting pests is time-consuming and causes bias between different observers.⁵

14.1 Why this comparison is the fair one

It measured the machine against the thing it would replace, performed by a person doing the job as it is actually done, rather than against an idealised standard.

That is the comparison missing from almost everything else in this article, and §22 flags its absence as our main evidential gap. A model at 92 per cent is only interesting relative to what a person achieves on the same material.

14.2 Why counting is the right application

Counting is the task §7 identified as easier than identification, performed against a known target, on a controlled substrate, by a person who would otherwise be doing it worse and slower.

There is only one species in question, so the confusion structure of §3.2 does not arise. The system is not being asked what this is; it is being asked how many there are, having already been told what to look for.

Observer bias between people counting the same sample is a real and well-documented problem, and a consistent machine removes it even if the machine is imperfect, because it is imperfect in the same direction every time.

14.3 The consequence claimed

That quick assessment of field conditions can allow growers to implement quick management tactics and potentially reduce the amount of pesticides.⁵

Better counting leading to less pesticide is the integrated pest management argument this journal has made repeatedly, arriving here through a camera.

15. The stored product proposal

The application closest to this trade's actual work.

A proposed system would provide image-based stored product insect species identification and detection for insect monitoring in a warehouse, food facilities, and retailer environment, integrating a simple RGB camera and an artificial intelligence-based deep learning model.⁶

The stated benefit is that a real-time automated detection model would serve as a major decision support tool, which eliminates the need for rigorous procedures, calendar scheduling, periodic trap placement and retrieval, and time-consuming species identification.⁶

15.1 Why this is the right target

Stored product monitoring is high-frequency, repetitive, and currently depends on somebody physically retrieving traps and identifying the contents, which this journal's pheromone article described as the labour that makes monitoring expensive.

15.2 And why it is also the hardest

Stored product beetles are small, brown, and separated by characters that require magnification. The confusion structure in §3.2 is at its worst in exactly the application where the economic case is strongest.

That tension is worth stating rather than resolving. The place the technology would be most valuable is the place the published species-level evidence is thinnest.

16. The claim to interrogate

One sentence in that proposal deserves examination.

The system would be simple, convenient, low-cost, and require no expertise or training.⁶

16.1 Where the expertise went

It did not disappear. It moved into the training dataset, as §10.1 described, and into deciding what to do with the answer.

A system that reports the presence of a species still requires somebody who knows what that species implies about the facility, and this journal's entire stored product and drain fly coverage is about that inference rather than about the identification.

16.2 The comparison with the other detection technologies

This journal found that canine detection teams drift because nobody verifies their alerts in service. A classifier does not drift in that sense: it produces the same output for the same input indefinitely.

What it does instead is stay fixed while the world moves. New species arrive, cards change supplier, cameras get replaced. The dog degrades and the model becomes obsolete, and only one of those is visible to the person using it.

16.3 The fair reading

It means no expertise at the point of use, which is a genuine and valuable claim. A warehouse employee can photograph a trap.

We would resist only the stronger reading, which is that the expertise has been eliminated from the system rather than relocated within it.

17. The consumer applications

The version of this technology most people have actually used.

Smartphone apps for taxonomic identification often offer artificial intelligence image identification functionalities for species classification, and many are constantly improving as algorithms and computing power advance. Two of the most popular have millions of downloads, allowing users to record, identify, and share observations of wild organisms, with the community actively curating observations by adding identifications to those made by others, enhancing data accuracy.⁷

17.1 Why this is the version that matters for households

A homeowner photographing an insect on their kitchen counter is using this technology, whether or not anyone in this trade calls it a pest monitoring system. The identification they get shapes whether they call anybody.

That makes consumer app accuracy a pest control question. A confident wrong answer sends somebody to the wrong remedy, and this journal has documented what misidentification costs in spiders, in small flies and in wood damage.

17.2 The mechanism worth noticing

Accuracy on those platforms is not purely machine accuracy. Human community curation is part of the pipeline.

Which means a person's experience of the app being reliable may reflect a combination of the model and subsequent correction by expert users, and the unaided model performance is a different quantity.

18. What those platforms are for

A limitation stated by researchers using them.

Biodiversity assessments using mobile apps are usually considered noninvasive and expand the possibilities of conservation monitoring in near-real time. However, the main purpose of such portals is the opportunistic recording of species occurrences rather than structured ecological monitoring.⁷

18.1 What opportunistic data cannot support

An absence in opportunistic records is uninformative. Nobody photographed a German cockroach in a building this month, which could mean there were none or that nobody looked.

Structured monitoring makes absence meaningful, because the sampling effort is known. That is the property a pest programme needs most and the one an identification tool does not supply.

18.2 The distinction that applies to pest work too

Opportunistic recording answers what is here. Structured monitoring answers how much, where, and whether it is changing.

A pest programme needs the second, and this journal's articles on trap catch interpretation and on rat counting both concluded that unstructured observation does not substitute for a designed sampling scheme. An identification tool improves one step of that scheme and does not replace it.

19. The pattern this journal keeps finding

Worth naming because it has now appeared in four technologies.

Remote rodent monitoring, canine scent detection, bed bug monitors and now automated identification have each been validated under controlled conditions, deployed into uncontrolled ones, and marketed on the controlled figure.

19.1 Why the pattern recurs

Because controlled validation is how a technology gets developed, and it should be. Nobody can build a classifier without a clean dataset, and nobody can train a dog without knowing where the target is.

The failure is not in validating under control. It is in the step where a figure obtained that way becomes a performance claim for the uncontrolled setting, and nobody funds the study that would measure the difference.

19.2 What is different here

The researchers are saying so themselves. The best-performing paper states that most prior work used solid backgrounds,¹ the pipeline paper lists the degrading conditions,² the order-level study reports its own limitation,³ and the trap colour study was designed specifically to find out whether the substrate mattered.⁴

This is a literature being honest about its own boundary conditions. The gap, when it appears, is between that literature and how a product built on it gets sold.

20. What to ask a vendor

How many classes was that measured over. Nine and thirty-five gave very different numbers.¹³

At what taxonomic rank. One 92.5 per cent figure was insect orders.³

On what background. Solid backgrounds were the norm before recently.¹

On which trap. Colour significantly influenced performance.⁴

Precision or recall, or both. They are different errors.¹

Who verifies the answers in service. The feedback problem this journal identified for detection dogs applies identically.

What happens if we change cards or cameras. Substrate changed performance significantly.⁴

21. Where we would and would not use it

Counting a known target on a known card. The clearest win, and it beat a person with a lens.⁵

Reducing observer variation. Consistency is valuable even where accuracy is imperfect.⁵

Triage rather than decision. Flagging cards that need a human to look, not replacing the human.

Not for species-level calls that determine treatment. Not on the current published evidence for confusable sets.

Not as a substitute for a sampling design. Identification is one step in a scheme, not the scheme.⁷

22. Limitations and open questions

This field moves quickly. The results here span from older feature-based work to recent pipeline papers, and performance reported next year may be materially better.²

Most of this literature is agricultural and ecological. Pollinator monitoring, crop pests and biodiversity survey. We located little validation on structural pest species specifically, and the stored product work is a proposal rather than a field result.⁶

We have read abstracts and summary sections. The performance figures come from abstracts and author summaries rather than from full methods sections, so we cannot assess how the test sets were separated from training data, which is the question that most affects whether a reported accuracy is meaningful.

The moth dataset figures are reported second-hand. They reach us through the related work sections of later papers rather than from the original studies.³⁸

We have not compared against human expert accuracy. The relevant question is whether the machine beats the available alternative, and except for the counting study in §14 we do not have that comparison.

Sections 3.1, 7.1, 12.1, 13.1, 16.1 and 18.1 are our reasoning. The class-count argument, the pest-work consequence, the background-learning generalisation, the biology conflict, the relocated-expertise reading and the sampling-design distinction are ours rather than sourced findings.

Our commercial position. APC Exterminators does not sell an automated identification product and does not currently use one, which gives us an interest in the sceptical reading. Section 14 and §15 are included because the evidence for the technology in its proper applications is genuinely good.

23. Conclusion

The strongest camera-trap result located reports 92.7 per cent precision and 93.8 per cent recall, across nine species, from a dataset of over two million images containing nearly thirty thousand annotated individuals.¹ Earlier species-level work on thirty-five moth species reached 85 per cent.³ A 92.5 per cent figure that circulates was classification to insect order.³ And a study that changed nothing but the colour of the sticky card found the trap colour significantly influenced how well the model performed.⁴

The researchers are candid about all of this. They say most prior applications used solid backgrounds, that taxonomy remains difficult at low resolution and partial views, and that their order-level method would not scale to species.¹²³ Where the technology is pointed at counting a known target, it beat a person with a magnifying lens on both accuracy and speed.⁵

So the useful summary is narrow and favourable. This works now for counting, for consistency, and for triage. It does not yet work as an unsupervised species call on the confusable sets that determine what gets treated, and the figure most likely to be quoted at you was measured on nine species against a background chosen by the people who measured it. Ask which nine.

References

1. Accurate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Transformation. doi:10.1371/journal.pstr.0000051. Principal source. Used for the statement that computer vision and deep learning are revolutionising entomology but that most applications are developed for images captured in lab conditions, against solid backgrounds or by manual photography, and that until this work deep learning models had not permitted the detection of multiple insect species against complex vegetated backgrounds; for the result that the best trained model measured an average precision of 92.7 per cent in classifying the species correctly with an average recall of 93.8 per cent, with the authors stating they know of no study with such high classification performance; for the reported high precision above 90 per cent for detection and classification across nine different insect species; for the description of the open access image dataset containing over two million images with 29,960 annotated bees, hoverflies, butterflies and beetles against varied floral backdrops; and for the authors' statement that their study solves the more difficult challenge of detecting multiple small insects in large images and could contribute to insect pest monitoring with camera-equipped traps without killing rare species. https://journals.plos.org/sustainabilitytransformation/article?id=10.1371%2Fjournal.pstr.0000051
2. InsectDCT: A generalized pipeline for detection, taxonomic classification, and tracking of insects in camera-trap recordings. Preprint. Used for the statements that automated monitoring of insect pollinators with camera traps and trained deep learning algorithms provides novel data for ecological studies, that efficient and accurate image recognition analysis of recorded images or videos is challenging particularly for images containing small insects against complex backgrounds with diverse vegetation communities, and that even when insects can be detected in images identifying their taxonomy remains difficult, particularly in footage with low image resolution, light conditions and distances from the plants, and in cases where insects appear blurry or only partially visible. https://www.biorxiv.org/content/10.64898/2026.07.07.736939.full.pdf
3. An Automated Light Trap to Monitor Moths (Lepidoptera) Using Computer Vision-Based Tracking and Deep Learning. PubMed Central PMC7825571. Used for its related work account of an analysis of 774 live individuals from 35 different moth species using data mining for feature extraction and a support vector machine for classification, achieving a classification accuracy of 85 per cent among 35 classes; for the report that a nearest neighbour algorithm based on texture, colour and shape applied to the same dataset obtained an accuracy of 79.53 per cent; for the statement that because classification was made on manually defined features the method is difficult to scale up to individual species classification as selecting useful features would be challenging; for the report that a combined approach achieved counting and classification precision of 92.5 and 90.8 per cent but only to the taxonomic resolution of insect orders; and for the account of an automated monitoring system for spotted wing Drosophila in which a trained model performed sex discrimination based on 4,753 annotated flies. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7825571/
4. Trap colour strongly affects the ability of deep learning models to recognize insect species in images of sticky traps. Pest Management Science. PubMed Central PMC11716339. Used for the statement that pest monitoring is an important element of integrated pest management and that sticky traps with adhesive materials are among the most common monitoring tools; for the observation that pests are not equally attracted to all colours of sticky traps and that different coloured sticky traps are typically used for different pests; for the result that using one architecture with transparent sticky traps as training data the model predicted pest species on transparent sticky traps with an accuracy of at least 0.95 and on other sticky trap colours with at least 0.85 of the F1 score; for the finding from statistical analysis that the colour and architecture of the sticky traps significantly influenced the performance of the model; and for the authors' conclusion that development of automatic classification of pests on sticky traps should focus on colour and deep learning architecture, with future work incorporating the trap system into pest monitoring to provide more accurate and cost-effective results. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11716339/
5. A smartphone application for site-specific pest management based on deep learning and spatial interpolation. Computers and Electronics in Agriculture. Used for the statement that the traditional method of manually counting pests is time-consuming and causes bias between different observers; for the result that the smartphone application had higher counting accuracy, 7.1 per cent higher for two-spotted spider mite motiles and 12.7 per cent higher for eggs, and was 29 seconds faster per leaf than manual counting using a magnifying lens; and for the statement that quick assessment of field conditions with smartphone technology can allow growers to implement quick management tactics and potentially reduce the amount of pesticide used. https://www.sciencedirect.com/science/article/abs/pii/S0168169924001170
6. Real-time stored product insect detection and identification using deep learning: system integration and extensibility to mobile platforms. Journal of Stored Products Research. Used for the proposal of an image-based stored product insect species identification and detection system for insect monitoring in a warehouse, food facilities and retailer environment, integrating a simple RGB camera and a deep learning model; for the statement that a real-time and automated insect detection model would serve as a major decision support tool eliminating the need for rigorous procedures, calendar scheduling, periodic trap placement and retrieval, and time-consuming species identification; and for the claim that the system would be fast, simple, convenient, low-cost and require no expertise or training. https://www.sciencedirect.com/science/article/pii/S0022474X23001224
7. The Field Automatic Insect Recognition Device: a non-lethal semi-automatic Malaise trap for insect biodiversity monitoring, proof of concept. PubMed Central PMC11602669. Used for the account of light traps using computer vision and of automated moth traps that identify, track, count and classify video-captured moths; for the statement that the use of smart mobile devices as biodiversity monitoring tools has gained popularity thanks to smartphone apps for taxonomic identification which often offer artificial intelligence image identification functionalities and are constantly improving; for the description of two popular platforms with millions of downloads allowing users to record, identify and share observations, with the community actively curating observations by adding identifications and enhancing data accuracy; and for the caution that while biodiversity assessments using mobile apps are considered noninvasive and expand conservation monitoring possibilities, the main purpose of such portals is the opportunistic recording of species occurrences rather than structured ecological monitoring. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11602669/
8. Deep Learning Pipeline for Automated Visual Moth Monitoring: Insect Localization and Species Classification. Preprint. Used for its related work account of light traps as a commonly used method for monitoring insects, of surveys on the influence of weather, time of year and light source type on species richness and abundance, and of early automated species identification systems using the same 35-species dataset with 20 individuals per species, reporting accuracy of up to 85 per cent using support vector machines and nearest neighbour classifiers with leave-one-out cross validation. https://arxiv.org/pdf/2307.15427