Sell an offline crop disease check to East African agro-dealers

The finding this rests on

A team at Mekelle University in Ethiopia took 3,587 field photos of cactus-fig pads. They trained three small models on those photos. The best one, MobileViT-XS, was right 97.3 percent of the time. It is 9.3 MB and takes 68 milliseconds per photo. It runs on a phone with no signal, inside an app written in Tigrigna and Amharic. We score that paper 4 out of 10. Read it at Crop disease model hits 97.3% and runs offline on old phones.

A second team did the same kind of work on beans. They are at the Nelson Mandela African Institution of Science and Technology in Arusha, Tanzania. They trained a model on 100,000 bean leaf photos. Adding tiny nudges built to fool the model lifted accuracy from 97.4 percent to 99.4 percent. A fourth group for junk photos scored an F1 of 0.9970. We score that paper 5 out of 10. Read it at Training with attack images pushed bean disease accuracy to 99.4%.

How strong is the evidence? The engineering is solid. The field use is not proven. Neither paper reports one farmer using the app for real. Each rests on one dataset, built by the same team that tested it. Therefore this idea is a bet that the accuracy holds when a stranger holds the phone.

Who has this problem

Picture an agro-dealer with one shop in a market town in the Njombe region of Tanzania. That is one of the three regions where the bean photos were taken. He sells seed, fertiliser and spray. About sixty farmers walk past his door on a market day.

His morning goes like this. A woman arrives with three bean leaves in a plastic bag. The leaves have brown spots. She wants to know what to buy. He looks at the bag. He is not a plant doctor. He guesses rust, and he sells her a bottle.

Two weeks later the crop is worse. She tells her neighbours the spray did not work. He has lost a bottle of goodwill and a customer.

The old answer was an extension officer who walked out to look. The Ethiopian paper says conflict damaged that support chain in Tigray, along with roads and power. The same paper says farming supports more than 80 percent of people in that region. Waiting for a visit is not a plan.

What you would sell them

Sell the dealer an offline Android app for the shop counter. He points the phone at the leaf, and in under a second the app says healthy, affected, or no crop in this photo. It works with the signal off, because the model sits on the phone.

On Monday morning he does not guess. He takes the photo in front of the farmer, shows her the screen, and sells on the answer. When the app says healthy, he says so, and she remembers that he did not sell her something she did not need.

What to charge

Test it at four dollars a month for one shop. That is a hypothesis, not a price. You will only know after you ask for money.

Here is the reasoning. Do not compare the app to software, because the dealer buys none. Compare it to one bottle of spray on his own shelf. Walk into the shop and read the label price. If your monthly fee is less than one bottle, he only has to be right about one customer a month to break even.

Test a second price at the same time. Offer a cooperative or a seed company 300 dollars a year to cover every dealer in one district. Both prices are guesses to be checked. Charge the first one on day one, not after a free trial, because a free trial tells you nothing about willingness to pay.

How you would build the first version

Build nothing new for a month. The Ethiopian team put their source code on GitHub and their photos on Kaggle. The Tanzanian team put their bean dataset in a public repository. Start with one crop that already has public photos.

Buy one cheap Android phone. Train the small model first, because at 4.8 MB and 42 milliseconds it is the one that fits an old handset. Copy two habits from the papers. Keep a fast model for scanning and an accurate model for hard calls. Add a group for junk photos, so the app can say there is no leaf here rather than name a disease.

Do not build a cloud dashboard. Do not build logins, an iOS version, or advice on which chemical to buy. Do not collect your own photos yet. Your first version is one screen, one button and one answer.

The one-week test

1. Day one. Download the public code and dataset. Get the model running on one phone, offline, in your kitchen.
2. Day two. Take 100 photos of real leaves in one field. Write down what the app said and what the farmer said. Count the disagreements.
3. Day three. Sit in one agro-dealer shop for the whole day. Use the app on every customer who brings a leaf. Say nothing about buying it.
4. Day four. Ask that dealer for four dollars for the next month. Take cash or mobile money on the spot.
5. Day five. Ask four more dealers in the same town, in person, for the same four dollars.

The answer at the end of day five is yes or no. Yes means at least two of the five dealers paid you money. No means none did, or they all said come back later. Come back later is a no.

What would kill this

• The accuracy does not survive the field. This is the big one. Both papers tested on held back photos from the same farms, in the same months, taken by the same team. The Tanzanian authors say so themselves, and warn that their data may not apply to other regions or crops. A stranger with a cracked lens in the wind is a different test.
• The labels are too coarse to act on. The Ethiopian app sorts photos into affected, healthy and no cactus. Affected covers both an insect and a fungal rot, and the treatment differs. An answer of something is wrong does not tell the dealer which bottle to sell.
• The headline number is soft. In that paper, 97.3 percent is called a cross-validation mean in one place and a test set score in another. Those are different measurements. Assume the true field number is lower.
• The dealer does not want an honest answer. He earns money when he sells a bottle. A tool that says healthy costs him that sale. If that is how he thinks, your buyer is the cooperative or the seed company instead.
• Your crop is not their crop. One paper is about cactus-fig pads. The other is about beans. If you want maize, you have no model and no photos, and collecting a few thousand labelled field photos is the slow part of this business.

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.