Pheromone Trap Moth Counting from Phone Photos for Cotton Pest Scouts

The problem: counting moths by eye

Bollworms are among the most damaging pests of cotton in South Asia. Farmers and agriculture departments use pheromone traps: a lure attracts male moths, which get stuck on a card or fall into a container. Counting the catch every few days tells scouts when a pest population is building and when spraying is justified.

Counting is done by eye. Scouts lean over dozens of traps, count small brown moths among other insects and dust, and write numbers on paper. Counts differ between scouts, the paper reaches the office days later, and nobody can see a pattern across a district in time to act. Spraying too late loses the crop; spraying too early or too often wastes money and harms the beneficial insects that keep pests in check.

What already exists and the gap

• Trapview sells automated camera traps that photograph catches and identify species, including cotton bollworm, with a forecasting service.
• Research datasets of bollworm trap images exist, including the BOLLWM dataset from cotton fields in India.
• Extension departments collect trap counts on paper or simple spreadsheets.

Automated camera traps are excellent where farms can buy and maintain hardware in every field. Most smallholder areas cannot. The gap is a tool that uses the phone the scout already carries and the trap already in the field, and turns each visit into a reliable count on a shared map.

Inspiration

"Automatic moth detection from trap images for pest management" by Weiguang Ding and Graham Taylor, published in Computers and Electronics in Agriculture in 2016, used a trap with a pheromone lure and sticky surface, captured daily images and trained a convolutional network to detect moths among other insects. Your project brings the same idea to a phone photo taken by a scout.

What you build

• Pest scouts take one photo per trap, check the count and move on.
• Extension officers see a map of all traps, counts over time and alerts when a trap crosses the threshold they set.
• Farmers can be told when catches near them are rising.

Each trap gets a short code printed on a tag, so the scout scans or types it and the photo is linked to the right trap without relying on GPS alone. The app shows the count with boxes drawn around each moth, and the scout can add or remove a box before saving. Those corrections become new training labels, so the model gets better with each season of use instead of staying frozen at the version you shipped.

The technically interesting part

• Small, overlapping objects. Moths are small, often overlap and get damaged on the card. Counting by detection works when objects are separate; you may need to compare it with density estimation for crowded cards.
• Look-alikes. Other moths, flies and debris land on the same card. The model must count only the target species, so labelling must be careful and checked by someone who knows the pest.
• Photo conditions. Bright sun, shadows, cards at an angle and dirty cards. A guided camera screen with a frame and a blur check reduces bad photos.
• Offline first. Fields often have no signal. The app must store photos and counts and sync later without duplicates.

How to evaluate it

• Counting error: average difference between model counts and careful hand counts, per card and for crowded cards separately.
• Detection quality: precision and recall for the target moth.
• Scout time: minutes per trap with the app versus by eye.
• Field test: a few scouts use the app for two weeks; record problems and missed syncs.

Suggested plan

• Weeks 1 to 3: read the paper, get the public dataset, contact an extension office or agricultural university for field access.
• Weeks 4 to 7: label photos, train and test the detector.
• Weeks 8 to 10: the app and offline sync.
• Weeks 11 to 13: dashboard, alerts and a field test.
• Weeks 14 to 15: report and demo.

Path to a product

Agriculture departments, seed and pesticide companies and large farms all run trap networks. A low-cost phone-based service with a dashboard is easier to roll out than hardware traps, and can later connect to them. The project also gives you a clear computer vision portfolio piece with real users.

Scope warning

Do not design a new trap or camera. Do not give spraying advice; show counts and let the officer's threshold drive alerts. Work with an agricultural expert on labels, because a model trained on wrong labels will count the wrong insect with confidence.

Learn what you need on Sefism

• Machine learning roadmap: data splits and evaluation for the detector.
• Mobile development roadmap: offline storage and sync for the field app.
• Ultralytics YOLO Documentation: training the moth detector.
• Label Studio: marking every moth on the training photos.
• OpenCV-Python Tutorials: blur checks and cropping the card.
• Which final year project will actually help me get a job?: presenting a field project well.

Get help with this idea

The risk is weak labels and a model that counts the wrong insects. Tauseef can help you find an expert partner, plan labelling and keep the app simple enough for field use. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session