A capacitive soil probe, a $5 ESP32, and a tiny ternary model — that's a complete, offline, no-cloud crop sensor that decides when to irrigate, when frost is coming, and when conditions favour pests, right on the chip in the field.
Why edge AI fits agriculture
Farms are exactly where the cloud fails: patchy connectivity, no power for a gateway, thousands of cheap nodes, and data you'd rather keep local. An on-device classifier needs no network, draws almost nothing, and costs a few dollars per node.
What the agriculture head does
- Reads a short window of soil-moisture, air temperature, and humidity from real sensors (capacitive soil probe + DHT22).
- Classifies the field state on-chip: healthy, needs-irrigation, frost-risk, pest-favourable, or sensor-fault.
- Fires a relay (pump/valve) or an alert when it matters — no server in the loop.
- Knows when it's unsure and abstains, instead of raising a false alarm.
Train it on YOUR field
The shipped model is a physics-grounded demo. To make it real for your crop and soil, flash the logger firmware, record each condition for a few minutes, and run one training command — you get a custom, signed model trained on your own data, flashed back in minutes.
Honest scope
Two of the five sensor channels use documented derivations unless you add a soil-temperature and leaf-wetness sensor; on-silicon throughput is measured on an ESP32-WROOM-32. It's a real device path, not a lab mock-up — but it's a tool you calibrate to your field, not an oracle.