AI crop monitoring that farmers can see, trust and control

Product AgroPulse
AgroPulse is a plant-monitoring system for small-scale farmers. Field sensors and a low-cost camera feed an AI model that looks for crop disease, and the app explains what it found with pictures, plain language and audio, leaving the final decision to the farmer.
Details
- Mobile app · field sensors · disease-detection model
- surface
- UX and system designer, UX researcher
- my role
- Figma ∙ UX research methods ∙ TouchDesigner ∙ Google Colab
- tools
- Agriculture · IoT + AI + HCI
- domain
Problem Predictions farmers can’t check
AI can flag crop disease early, but farmers with little digital experience distrust a prediction they can’t see or question, and acting on a wrong one can cost a crop.
Solution Explain, then recommend
Every alert shows where on the plant the problem is, why the system thinks so, and what other farmers did, then offers options. Irrigation can always be switched to manual.
Result A working system and prototype
A high-fidelity app, a sensor and camera setup, and a disease model that combines images with sensor data. Scenario walkthroughs and informal usability testing guided the final design.
Who it’s for
The people deciding on plant health are small-scale farmers, often with little digital experience.
Our primary users are small-scale farmers and the farm operators responsible for plant-health decisions. Agricultural extension officers, community farming groups and cooperative managers could use it later.
Four constraints shaped everything: limited experience with digital interfaces, a preference for visual cues over numbers, unreliable internet in the field, and mistakes that can mean losing a crop.
Research
Farmers want advice they can judge, not automation that decides for them.
Farmers were hard to reach, so most of our information was second-hand. We compared existing smart-agriculture dashboards, ran interpretation sessions on second-hand interviews, reasoned through real farming decisions as scenarios, and studied irrigation and crop-disease workflows.
Farmers want advice, not automation. Full automation lowers trust; they prefer recommendations they can evaluate. Pictures beat numbers. Colour, shape and imagery communicate faster than charts. Unexplained AI feels unreliable. Farmers distrust predictions that don’t show why. Control builds confidence. Manual override makes a system easier to accept.
The question
How might we make AI crop predictions something farmers can understand, trust and control?
Design principles
Farmers expressed seven fears of new technology. We turned them into five rules.
The fears were black-box AI, losing control, high-stakes risk, low digital and technical literacy, information overload, no proof that others use it, and generic advice.
Explain before you predict. Show, don’t calculate. Recommend, don’t enforce. Design for low literacy by default. Always show the system’s status clearly.
The system
Sensors and a cheap camera in the field, processing on the device.
Soil moisture, temperature and humidity sensors and a low-cost camera collect data in the field. An ESP32 microcontroller gathers the readings, captures live footage, does basic processing and manages connectivity, so the system keeps working when the internet drops.
The circuit was simulated in TinkerCad before being built with real sensors.

The TinkerCad simulation, then the sensors and camera setup.
The model
Images alone weren’t enough. Adding sensor data made predictions reliable.
The model first cleans up footage from cheap cameras (less noise, even lighting, better contrast), then isolates the leaves from sky, soil and shadow. That isolation is what lets the app highlight exactly where a problem is.
Disease detection then combines what the camera sees with the field conditions. On our test set, the combined model reached 95.56% accuracy, against 83.67% for images alone and 25.22% for sensor data alone, which cuts down false alerts.

Classification results and confusion matrices for image-only, sensor-only and combined models.
Design system
Earthy colours and one friendly typeface, so the app feels familiar outdoors.
Everything sits on a warm cream ground with deep greens for actions, so the app looks like the field rather than a dashboard. Orange and red are kept for warnings, so they stand out when something needs attention.
Colour
Sampled from the final screens. Greens carry navigation and actions; orange means “possible problem” and red means “act now”.
Type and icons
Poppins in three weights for everything, so text stays round and easy to read. Icons come from the Iconify plugin in Figma, and every section has a speaker icon for audio.
Monitor your plants
Poppins · regular, medium, bold
From research to design
Each screen answers one of the farmers’ fears.
These are the main screens of the final design and the research each one responds to.
Low literacy, many languages
The first thing you choose is your language.
Language selection comes before anything else, in six Indian languages. Login uses a phone number instead of email, since many farmers don’t have an email account. Right after login, a Go To page with six large icons takes you straight to what you need.



Show, don’t calculate
Plants that need care come first, in pictures.
Home lists every plant with a short status, and plants that need attention appear first. Weather from the sensors is shown as simple words like Mild and Humid, with the readings underneath.

Explain before you predict
The app shows where and why, not just what.
The plant’s health page shows live footage and a 3D view of the plant, the sensor readings for its bed, and the prediction with how sure the model is (Likely, Possible or Unsure). Tapping the highlighted area explains what the camera saw, and you can upload your own photo for a second look.

Recommend, don’t enforce
Alerts give reasons, other farmers’ experience and a choice.
An alert describes what was noticed and why it matters, quotes nearby farmers who saw the same thing, and offers next steps: Treat now, Wait and observe, or Ignore. Farmers can reply or leave a comment.

Control equals confidence
The farmer can always take over.
Irrigation can be switched on or off for the whole field, each sprinkler can be controlled on its own, and schedules can be edited. The app recommends a moisture level but never forces it.

Build trust over time
Farmers can see what happened after each decision.
Plant History records each action taken and whether the plant improved, stayed the same or got worse. Farmers can also log their own observations. The Help tab adds plain-language guides, an expert to ask, and a community of local farmers.

Accessibility
Accessibility was the main design requirement, not a final check.
Language selection, a clear visual hierarchy with icon-assisted navigation, audio assistance for farmers who can’t read, and as little numeric or technical language as possible.
Evaluation
Seeing the reasoning made people trust it.
We walked through scenarios (a disease detected, a false positive, a manual irrigation override), ran informal usability tests, and watched for confidence and hesitation at each decision.
Visual explanations increased trust, the 3D plant view helped people understand the disease, and manual override reduced fear of automation.
Reflections
What workedExplainable AI improved trust, visual-first design lowered the learning barrier, and manual control made people more willing to adopt it.
ChallengesBalancing simplicity with useful feedback, designing for unpredictable outdoor conditions, and avoiding automation bias so farmers keep the choice.
NextStop judging a whole field from one plant, let farmers engage further so AgroPulse grows into a community, and improve the model’s accuracy and precision.








































































