AI Agriculture Advisory System

Crop Health, Disease and Irrigation Advice

Design a service that watches farmers' fields from satellite images, diagnoses crop diseases from phone photos (even offline), and tells each farmer when and how much to irrigate, with advice that is careful about what it does not know.

Overview

A farm advisory service answers three questions for a farmer: is my crop healthy, what is this spot on the leaves, and should I water today. It draws on three kinds of data: free satellite images that show plant vigor across every field every few days, photos the farmer takes with a phone, and weather forecasts plus, on some farms, soil moisture sensors.

The users shape the design as much as the data. Many are smallholders on basic phones with patchy mobile data, reading or listening in a local language. So the app must work offline, keep messages short, and offer voice. Advice has consequences: a wrong irrigation call wastes water or stresses a crop, and a confident wrong diagnosis leads to the wrong spray. The system should prefer well-established agronomy (the FAO's standard method for crop water needs) over a black-box model where it can, and say 'ask an extension officer' when it is unsure.

Questions to ask first

How many farmers and fields? — 2 million farmers, 4 million fields, mostly around 1 hectare
Which satellite data? — Free Sentinel-2 imagery: 10 m pixels, a new image about every 5 days
Which crops and diseases? — A few staple crops per region, their common diseases
Do farmers have sensors? — Some; about 100,000 soil moisture sensors in total
Connectivity? — Often offline for days; the app must diagnose common diseases without a network
Languages? — Several local languages, with text and voice

What you will be able to do

Turn satellite images into a per-field health signal, and handle clouds.
Design disease diagnosis that works offline on a basic phone and knows when to defer to a person.
Compute irrigation advice from the FAO-56 water balance and a weather forecast.
Detect regional outbreaks from individual diagnoses and alert the right farmers.
Size the system and explain why most of it is batch work, not request traffic.

Key concepts

Field boundaries (GeoJSON)
NDVI from satellite bands
Cloud masking
Anomaly against a baseline
Image classification on the phone
Offline-first sync
FAO-56 water balance
Outbreak detection by area
Human expert in the loop

Functional requirements

Register a field by drawing its boundary on a map, with crop and planting date
Show each field's crop health over time from satellite images, and alert on an unusual drop
Diagnose a disease from a leaf photo, online or offline, with a confidence and the next step
Give a daily irrigation recommendation per field: irrigate today or not, and how much
Alert farmers growing a crop when a disease is spreading nearby
Deliver advice in the farmer's language by app, SMS or voice
Let an agronomist review uncertain diagnoses and correct them

Non-functional requirements

Offline diagnosis for the most common local diseases, with results synced when online
New satellite data reflected in field health within 24 hours of availability (our target)
Irrigation advice sent before the working day starts
Outbreak alerts within a few hours of the cluster being detected (our target)
Uncertain answers say so and point to a person; no silent low-confidence advice
Small downloads: the app and its model fit on a low-end phone

Interview questions

How do you turn satellite images into crop health for one field?
How does disease diagnosis work without a network?
Why not just trust a model trained on a big public dataset?
How do you decide whether a farmer should irrigate today?
How do you detect a disease outbreak and who do you tell?
Where is the load in this system?
How do you keep advice safe?