Edge-AI Farm Monitoring Station

Problem

Farm sites are 80 km from the city, with 4G dropping 3-4× a day. Cloud-based solutions go blind during outages. Owners need pest detection within minutes, not hours.

Architecture

ESP32 (soil moisture/EC/pH) + Luckfox Pico RV1106 (1080p cam, 0.5-TOPS NPU) → YOLOv8-nano INT8 RKNN (2.1 MB) → local MQTT broker → 80 Wh solar battery. Offline: local SQLite log, delta-sync when 4G returns.

Stack & rationale

• Rockchip RV1106: NPU runs YOLOv8n at 12 fps, board is ~$8.
• INT8 YOLOv8-nano: 6.1 MB → 2.1 MB, −1.8% accuracy, 3× fps.
• Solar + LiFePO4: 14-day backup with zero sun.

Results

• 7 pest classes (leaf-roller, brown planthopper, red mite…) — mAP@0.5: 0.86
• 23 outbreaks caught in 6 months, 2 false positives
• Pesticide use −34% by spraying right place, right time
• 99.2% uptime (a 3-day storm with no sun still ran fine)

Lessons

Cheap NPU + tightly-quantized model beats a powerful cloud GPU when the network is the bottleneck. Quantization-aware training matters — post-training quant loses more accuracy.

Security & Compliance

• Secure boot + signed firmware: only Ed25519-signed binaries can run on the device.
• mTLS between edge and gateway; cert rotation every 90 days.
• in production firmware — OTA updates only, via signed channel.