Crop disease model hits 97.3% and runs offline on old phones

What happened

A farmer walks along a row of cactus pads. One pad has white fluff on it. Is that an insect? Is that mould? Or is it just an old scar from last year?

The answer changes what you do next. In the Tigray region of Ethiopia, that cactus is food. It is called beles. It feeds people in the hungry months before the main harvest. The paper says farming supports more than 80 percent of people in the region.

The usual answer came from an extension worker who walked out to look. The paper says conflict damaged that support chain, along with roads and power. So a team built a phone app that gives the answer with no internet at all.

The test

The paper is “Automated Plant Disease and Pest Detection System Using Hybrid Lightweight CNN-MobileViT Models for Diagnosis of Indigenous Crops”. The authors are Tekleab G. Gebremedhin, Hailom S. Asegede, Bruh W. Tesheme and colleagues. They work at Mekelle University, in the Mekelle Institute of Technology, in Ethiopia.

First they built a dataset. They took 3,587 field photos of cactus-fig pads. The photos are in three groups. 1,500 pads are marked affected. 1,500 are marked healthy. 587 are marked no cactus, and show only soil, sky or weeds. The photos were taken at test plots at Mekelle University and Adigrat University. They were shot outdoors, with dust, shadow and glare left in.

Then they trained three models on the same photos. One is a small custom network with 1.2 million settings. One is EfficientNet-Lite1, with 4.6 million. One is MobileViT-XS, with 2.3 million, which mixes two designs. Before this, the models saw 26,394 public photos of tomato, potato, maize and apple, to learn what a plant looks like.

The test set held 1,195 photos the models never saw. The team also ran three-fold cross-validation.

The result

MobileViT-XS was the most accurate, at 97.3 percent. EfficientNet-Lite1 scored 90.7 percent. The small custom network scored 89.5 percent.

But size and speed run the other way. The small network is 4.8 MB and takes 42 milliseconds per photo. MobileViT-XS is 9.3 MB and takes 68 milliseconds. EfficientNet-Lite1 is the biggest at 19.0 MB.

Why did the mixed model win? Because of what the plain networks got wrong. An old healthy pad has rough scars. A rotting pad has rough patches too. The paper reports that this one mix-up caused 68 percent of the small network’s errors. MobileViT-XS looks at the whole photo at once, so it can see that pest wax comes in clusters while scars sit alone. It threw out 94 percent of the false alarms the plain networks raised.

All three models were right about background photos 99 percent of the time. The finished models were shrunk and put inside an Android app, written in Tigrigna and Amharic, that runs with no signal.

What it means

ACCURACY AND SPEED ARE TWO DIFFERENT PRODUCTS.

That is the useful lesson here. The team did not pick one model. They kept the fast one for quick scanning and the accurate one for the hard cases. You can copy that pattern in any tool that must run on a cheap phone.

The second lesson is about data. Global crop datasets are full of flat leaves like apple and maize. A cactus pad is thick, waxy and covered in spines. Models trained on flat leaves do poorly on it. Someone had to go out and photograph 3,587 cactus pads. That work, not the model design, is the part nobody else had done.

The third lesson is language. The app speaks Tigrigna and Amharic. A tool that a farmer cannot read is a tool that a farmer will not use.

Business ideas from this paper

1. A paid photo labelling service for one local crop that no global dataset covers, delivering a few thousand clean field photos with agreed labels. Who buys it: agritech startups and university groups that need local training data. A price to test: 500 dollars for 1,000 labelled field photos. A one-week test: shoot and label 100 photos of one crop, then show them to three agritech teams and ask for a purchase order.
2. An offline model packaging service that shrinks a client’s trained model, wraps it in a local language app, and ships it for old Android phones. Who buys it: NGOs and government extension programmes working where the signal drops. A price to test: 2,000 dollars per app, plus 100 dollars a month for updates. A one-week test: take one open model, ship a working offline demo, and hand the phone to two programme managers.
3. A short training course on the fast model plus accurate model pattern, taught with the authors’ public code and dataset. Who buys it: computer science departments and junior developer teams. A price to test: 25 dollars a seat for an online session of thirty. A one-week test: run one free session, then count how many sign up for the paid second part.

How sure can you be?

Be careful with this one. The engineering is real. The proof is thin in three ways.

First, the label set is coarse. The three groups are affected, healthy and no cactus. Affected covers both the cochineal insect and fungal rot. So the model tells you that something is wrong. It does not tell you which of the two you have, and the treatment differs.

Second, the headline number is reported in two ways. The abstract and the conclusion call 97.3 percent a mean cross-validation accuracy. The results table lists the same figure under a test set heading. Those are not the same measurement. The paper does not clear this up.

Third, no farmer trial is reported. The app is built and the models are inside it. The paper does not report a single field session with real users, and it does not say which device the 42 and 68 millisecond timings were measured on.

The authors are open about what comes next. They want to add teff, sorghum and the fall armyworm. They want the fast model and the accurate model working together in one flow. They want to try low cost drones.

We score this paper 4 out of 10. It is a preprint. A shorter version was presented at a regional conference in Mekelle in February 2025. It rests on one dataset. The strong point is openness, because the code and the photos are public on GitHub and Kaggle.

What would settle it? A trial where farmers use the app on their own phones for one season, and someone checks the calls against a plant doctor.

Do this today

If you are building any tool for a place with weak signal, measure two things and write both down. Measure accuracy, and measure the size of the file a user must download.

Source: Automated Plant Disease and Pest Detection System Using Hybrid Lightweight CNN-MobileViT Models for Diagnosis of Indigenous Crops, December 2025. arXiv:2512.11871 · arxiv.org (preprint · not yet peer reviewed).

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.