Build a Raspberry Pi Plant Disease Detection Camera: On-Device Computer Vision for Early Blight and Pest Spotting
Raspberry Pi 4 or Raspberry Pi 5 board
Raspberry Pi Camera Module 3
Weatherproof camera enclosure
Outdoor-rated USB-C power supply
Adjustable camera mount arm/clamp
MicroSD card, A1/A2 rated, 32GB+
Silica gel desiccant packs
Waterproof cable glands
Prices and availability as of Nov 22, 8:53 AM ET and subject to change. The price shown on Amazon at the time of purchase applies.
By the time leaf blight or a pest infestation is obvious to the eye walking past a garden bed or greenhouse row, it's often already spread. A fixed camera checking plants on a schedule and flagging anything that looks off catches problems earlier than an occasional walk-through does โ and a Raspberry Pi with a basic vision model is enough hardware to do it without a cloud subscription or a commercial agricultural sensor package. This project builds a weatherproof camera station that watches a bed or greenhouse row and reports likely trouble spots.
What This Project Actually Does (and Doesn't)
This builds a monitoring and triage tool, not a lab diagnosis. The camera and on-device model flag visually distinctive symptoms โ leaf discoloration patterns, spotting, wilting, visible pest damage โ well enough to tell you "go look at bed 3" instead of discovering a problem days later. It won't reliably distinguish between similar-looking diseases, won't catch problems with no visible symptoms yet, and works best as an early-warning system that still ends with you walking over and looking at the actual plant. Treat its output as a prioritized checklist, not a diagnosis.
How It Works
A Raspberry Pi with a camera module captures images on a schedule (every 15-60 minutes during daylight is typical), runs them through a lightweight on-device classifier trained on common leaf disease and pest-damage patterns, and logs or alerts when a frame scores above a confidence threshold for "looks abnormal." Running inference on-device means no images or notifications depend on an internet connection or a subscription service, and nothing about your garden leaves your network unless you choose to forward alerts somewhere.
Software Setup
- Flash Raspberry Pi OS (Lite is sufficient since this runs headless) and enable the camera interface and SSH during initial setup.
- Install a lightweight inference runtime. TensorFlow Lite or a comparable lightweight runtime runs well enough on a Pi 4/5 CPU for infrequent, scheduled image classification โ you don't need a dedicated accelerator like a Coral USB or Hailo HAT for a once-every-15-minutes capture rate, though adding one (covered elsewhere on this site) shortens inference time if you want more frequent checks.
- Use or fine-tune a plant disease classification model. Several open, pre-trained plant leaf disease datasets and models exist publicly for common crops; a pre-trained general model gives a reasonable starting point, but accuracy improves meaningfully if you fine-tune on your own photos of healthy plants in your actual setup, since lighting, camera angle, and plant variety all affect real-world accuracy more than dataset benchmarks suggest.
- Script the capture-classify-alert loop. A simple Python script on a cron schedule (or a systemd timer) captures a frame, runs it through the model, logs the result with a timestamp and confidence score, and sends a notification โ a webhook into Home Assistant or a simple MQTT message, matching the automation patterns this site's other ESP32/Pi garden projects already use โ when confidence crosses your chosen threshold.
- Tune the alert threshold deliberately. Start conservative (fewer false alarms, possibly missing a borderline case) and loosen the threshold once you've seen a week or two of real output and know what a false positive looks like for your specific setup.
Camera Placement and Lighting
Consistent framing matters more than camera quality for this application โ mount the camera on a fixed arm or tripod pointed at the same bed section each time rather than moving it between captures, since a model trained or tuned on consistent framing degrades quickly against inconsistent angles and distances. Diffuse, consistent lighting (shoot in similar conditions across captures, or add a simple fixed light source for a greenhouse setup) reduces false positives caused by shadows and glare being misread as discoloration.
Weatherproofing
Outdoor placement needs a genuinely sealed enclosure, not just a "resistant" one โ moisture inside a camera housing fogs the lens and eventually kills the electronics. Use a proper IP-rated enclosure with a clear lens window, seal all cable entries with cable glands rather than drilled holes and silicone alone, include a desiccant pack that you can swap out, and route the power cable with a drip loop so water running down the cable doesn't follow it straight into the enclosure.
Limits and Honest Expectations
On-device models at this scale will produce false positives (a shadow, a water droplet, a naturally senescing older leaf flagged as diseased) and false negatives (early-stage problems with subtle symptoms missed entirely). Treat this as a tool that redirects your attention faster than a random walk-through would, not as a substitute for actually knowing your plants and checking anything it flags in person before acting โ pulling or treating a plant based solely on an automated alert, without visual confirmation, risks acting on a false positive.