Crop Disease Photo Triage That Knows When to Ask an Agronomist, on Budget Android Phones
Most crop disease apps give a confident answer for every photo, even blurry ones, the wrong crop or a disease they never saw in training. Build an on-device classifier for a few crops that abstains when unsure and sends those photos to an agronomist review queue; the agronomist's answers become new training labels. The project measures how much accuracy improves when the app is allowed to say "I don't know".
Intermediate
A small team, or one strong student willing to learn something new.1 to 2 semesters · 2 students
7 to ship
4 optional extrasSuggested stack
What you should ship
- A small classifier for two or three crops common in your region (for example cotton, wheat, tomato or cassava), trained on public data plus your own field photos
- A quantised model that runs on a phone with 2 to 3 GB of RAM, with measured speed and memory
- Photo quality checks before classification: blur, darkness, leaf not in frame, wrong crop
- Calibrated confidence and an abstain rule tuned on a validation set
- An agronomist review queue on the web for abstained photos, with a simple labelling screen
- A retraining loop that adds reviewed photos and reports the change in accuracy
- An evaluation on field photos, separate from lab-style photos
If you have time left
- Offline queueing of photos for review when there is no signal
- Answers in Urdu, Hindi, Bangla, Hausa or Swahili, written with an agronomist
- Active learning that asks agronomists to label the photos that teach the model most
- A map of reviewed cases for extension officers
The problem: confident and wrong in the field
Farmers and extension workers across South Asia and Africa now photograph sick plants and ask apps what is wrong. Models trained on clean photos of single leaves on plain backgrounds do well in papers and badly in a real field, where photos are blurry, leaves overlap, and pests, nutrient problems and sunburn look like disease. Worse, these models always answer. A wrong diagnosis can lead to the wrong spray, wasted money and resistance.
The honest design is a model that answers when it is sure and asks a human expert when it is not.
What already exists and the gap
- Plantix, developed by the Berlin company PEAT, is widely used in India and elsewhere to diagnose crop problems from photos.
- PlantVillage Nuru, from Penn State's PlantVillage team, runs offline on phones for crops such as cassava in East Africa.
- Government and NGO extension helplines answer farmers' questions by phone, with limited staff.
The gap for a student team is not another classifier. It is a careful study of abstention and human review: how much of the traffic the model can handle alone, how accurate it is on what it keeps, and how fast it improves from expert labels.
Inspiration
The PlantVillage dataset, released by David Hughes and Marcel Salathé (arXiv, 2015), made crop disease classification a standard benchmark. Later work on PlantVillage Nuru (Mrisho and colleagues, Frontiers in Plant Science, 2020) compared the app with extension officers in Tanzania and showed that field conditions matter. On the method side, "Selective Classification for Deep Neural Networks" by Yonatan Geifman and Ran El-Yaniv (NeurIPS 2017) gives the theory for trading coverage against accuracy.
What you build
- Farmers and extension workers take a photo, get either a clear answer with next steps or "sent to an expert, answer within a day".
- Agronomists work through a short queue of hard cases on the web.
- The team sees coverage, accuracy and how both change after each retraining round.
The technically interesting part
- Domain shift. Hold out field photos you collected yourself and never train on lab photos alone. Report both numbers.
- Calibration. Raw softmax scores are overconfident. Use temperature scaling and check calibration curves before trusting any threshold.
- Out-of-scope inputs. Photos of the wrong crop, a hand or a road must be caught before the classifier.
- Small phones. Quantise the model and measure latency and memory on a real budget phone, not only an emulator.
How to evaluate it
- Risk-coverage curve: accuracy on the photos the model keeps, against the share it keeps.
- Field versus lab accuracy for the same classes.
- Out-of-scope rejection rate on a set of irrelevant photos.
- Improvement per review round: accuracy after adding 100, 200 and 400 reviewed photos.
- On-device latency and memory on at least one low-end phone.
Suggested plan
- Weeks 1 to 3: choose crops with an agronomist, collect field photos, set up public data.
- Weeks 4 to 7: baseline model, calibration and abstain rule.
- Weeks 8 to 10: phone app with quality checks and the on-device model.
- Weeks 11 to 13: review queue and retraining loop.
- Weeks 14 to 15: evaluation, report and demo.
Path to a product
Agri input companies, contract farming firms and extension services need diagnosis tools they can trust and that improve with their own experts' knowledge. A triage model plus review queue is a sellable service, and the abstention work is a strong story for machine learning roles.
Scope warning
Do not give pesticide doses or product recommendations in the app; leave treatment advice to the agronomist. Keep to two or three crops and a handful of diseases. Ask permission before photographing anyone's fields.
Learn what you need on Sefism
- Machine learning roadmap: training, calibration and honest evaluation.
- Mobile roadmap: running models and handling images on phones.
- MLOps roadmap: the retraining loop and model versions.
- PyTorch: Learn the Basics: training the classifier.
- Android Developers documentation: camera and on-device inference.
Get help with this idea
The common failure is reporting lab accuracy and calling it done. Tauseef can help you set up field test data and a fair abstention study. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.