(Lesson 07.01 · 20 minutes · No sensor · Final project)

A real case, the olive leaf.

OliveSense recognizes olive tree diseases out in the countryside, with no network, in less than half a second. Here you'll follow the project from the question to the decision, with the choices it called for and what's still missing.

What you learn

How the place of use shapes the choices in an Edge AI project
Why you split a problem into several small models
How a diagnosis becomes a useful decision
What it means, in a real project, to say what's missing

You need to know

If you don't remember one of these, the link takes you to the chapter that explains it.

How do you want to follow it

Preparing the scene…

01 The problem
Olive trees, a weak signal and a field logbook to fill in every evening.

In Puglia, olive growing is made up of lots of small farms. Growers often have a mid-range phone, and out in the fields they have no coverage. Diseases are spotted late, Xylella comes with strict regional rules, and the field logbook is filled in by hand, in the evening. That's where OliveSense comes from, and it's a project by the author of this course.

In practice. Always start from who will use the model, and where. Here, the average phone and the countryside without a network decided almost everything else.
02 The question
Thirteen conditions, each with its own symptoms.

The question isn't healthy or sick. Before taking any photos, the work was to document thirteen conditions of the olive tree, with their symptoms and thresholds, from peacock spot to olive knot, from the olive psyllid to Xylella. Without this list, you don't know what to label. And a model can't be more precise than its classes^01.03.

In practice. Write the classes as a guide for a person. If a person can't tell them apart by reading the guide, the model won't be able to either.
03 The data
The photos are taken by people who know olive trees.

The photos were taken in the olive groves, one by one. Healthy leaves and leaves with spots, a trunk with olive knot, psyllids on the new shoots. To collect and label them from the phone there's Edgy Studio, an app built just for this. The dataset^01.05 is the longest part of the project, and it isn't finished yet.

In practice. Collect the data with the same kind of phone and in the same conditions where the model will be used. Photos downloaded from the internet look useful, and they often teach the wrong thing.
04 Four models
Four small models in a row, instead of one big one.

Four models run in a row on the phone, four or five megabytes each. The first one checks that the photo is good. The second works out which part of the tree it's looking at. The third looks for an anomaly^01.04. The last one says which disease it is. All together, the answer arrives in less than half a second, without a connection.

In practice. Splitting a problem into small steps makes each model simpler to train and to check. If the diagnosis is wrong, you know which step to look at.
05 The decision
The diagnosis alone isn't enough.

Knowing there's peacock spot doesn't tell you what to do. The system combines the diagnosis with the weather, with the Sentinel-2 satellite indices for the plot, and with the treatments already done. Then it compares how widespread the disease is against an economic threshold, and suggests an intervention, with the right dose and the right timing.

In practice. A model answers a narrow question. It becomes valuable to the person using it when the answer turns into a decision and a concrete action.
06 Phone and cloud
The phone decides on its own when it's sure.

If the confidence^02.01 is high, the phone confirms on its own. If it's low, and there's a network, it passes the case to a bigger model in the cloud^01.01. That way almost all the photos stay in your pocket, and data traffic and costs stay low. It's the division of labor of edge AI^01.01, applied to a real case.

In practice. Choose the threshold for asking the cloud for help by looking at the wrong cases, not the average. The cloud is there for the hard cases.
07 What's missing
A year of photos, and a trial in the field.

Today the project is a working prototype, ready for a controlled trial. It still needs a whole year of photos, because every condition has to be seen in every season and in every kind of weather. The next step is a trial with a few farms and a research institute. That way we'll see how the model does on real trees.

In practice. Saying honestly what's missing is worth more than any promise. A model tested over a single season doesn't know yet what autumn looks like.

01 / 07

(Go deeper)

A few more numbers.

(01)

Why a row of models

Each model in the row does one narrow job. The photo check throws out blurry or dark shots right away, before spending any computation and before giving a diagnosis to a photo you can't use. Recognizing the part of the tree tells the next steps what they're looking at, because a leaf, a branch and a trunk get sick in different ways.

The anomaly step is there to catch anything that doesn't look healthy, even when it isn't one of the thirteen conditions. Only at the end comes the disease classifier, which works on photos that are already good and already framed. Each of the four can be improved on its own, without touching the others.

13olive tree conditions documented
4models in a row on the phone
4-5 MBthe size of each one
< 0.5 sfor the answer, with no network

(02)

The field logbook

Anyone who uses plant protection products has to record every treatment. A 2023 European regulation requires these records to be kept in electronic form from 2026. In Italy the data goes through SIAN, the national agricultural information system.

OliveSense fills in the logbook on its own, starting from the diagnosis and the chosen intervention, in a format that SIAN accepts. For growers this is often the most concrete reason to use the app, even before the diagnosis.

(03)

How it's built

The app is written in Flutter and works without a network. The models are in TensorFlow Lite format and run on the phone. The data sits in a local database, Isar, and syncs with Supabase when the connection comes back. There's also a free web version, built with Next.js, that doesn't ask for an account and doesn't collect any data.

The plot maps use Leaflet, and the vegetation indices come from Sentinel-2 images from the European Copernicus program. For the future there's also a node with sensors to leave out in the field.

(Try it on your phone)

The real case, applied to yours

Go back over the choices behind OliveSense with your own project in mind. Answer each point in a line or two. The card stays saved in your browser, and you'll need it in the next lesson.

(Quiz)

4 questions to close the lesson.

When you have answered them all, the lesson counts as done and its stone on the map turns yellow. Answering is what counts, not the score.

(01)

Why does OliveSense have to work without a network?
(02)

Why four models in a row instead of just one?
(03)

When does OliveSense ask the cloud model for help?
(04)

Why do you need a whole year of photos?

New words, and where they are explained

Models in a rowchapter 4

Sources and reading

To close the lesson, answer the quiz

← Responsible AI Back to the map

(Next lesson · 07.02) Your project