Six months in the field with the smallholder vision kit
What we learned deploying crop-disease detection on sub-$50 hardware across forty-one farms in Punjab — including everything that broke.
Dr. Zara Siddiqui
Assistant Professor · Perception & Robotics
In March we placed forty-one vision kits with cotton farmers across three districts. Each kit is a camera, a microcontroller board and a solar cell in a 3D-printed housing. The model detects leaf curl and bacterial blight and sends a daily summary by SMS.
What worked
- Detection recall for leaf curl held at 91% in the field, within three points of validation.
- Farmers used the SMS summaries; the cooperative’s agronomist changed her visit schedule based on them.
What broke
- Dust. Lenses fogged within weeks. A cheap hood and a monthly wipe schedule fixed it.
- Heat. Two boards failed above 48°C. We moved the electronics under the panel.
- Novel objects. The model flagged a plastic bag as blight for a week. The review loop caught it; a retraining cycle with 200 new images fixed it.
The dataset from this deployment is now part of the public release accompanying our CVPR workshop paper, and the hardware design is open.
- vision
- agriculture
- edge
Get the next note by email
One considered letter a month on research, admissions and what we are building.
More in Research
UrduQA: why we built a benchmark that models cannot have memorised
Static benchmarks decay the moment they enter training data. Here is how we built one that stays honest, and what it tells us about the state of Urdu language technology.