Company case
John Deere: Computer vision
What the published sources confirm
See & Spray uses boom-mounted cameras and machine learning to detect weeds and activate individual nozzles, so herbicide is applied only where weeds are seen.
Business problem and AI concept
- Problem
- Broadcast spraying applies herbicide to the whole field even where there are no weeds.
- Concept
- Real-time object detection on edge hardware with a physical action.
Reported results
Deere reported an average 59% herbicide mix saving across more than 1 million acres in the 2024 season, compared with broadcast spraying.
Deere news release, 18 Sep 2024. Caveat: Manufacturer figure; depends on crop, weed pressure and settings.
A three-year University of Arkansas field trial found 43–59% herbicide reductions and noted that low sensitivity settings increased the weed seed bank.
UADA news, 10 Mar 2025. Caveat: Independent trial; shows trade-offs between savings and weed control.
Suggested workflow
Written for learning, not the company's system. This flow shows how you could build something similar. It is not a description of John Deere's private architecture.
Step 1: Collect and label field images across seasons and lighting
Step 2: Train a weed-versus-crop detector
Step 3: Test on fields not seen in training
Step 4: Optimise the model for edge inference
Step 5: Convert detections into nozzle decisions with a sensitivity setting
Step 6: Monitor missed weeds and chemical use
Suggested technologies and alternatives
| Layer | Suggested | Alternatives |
|---|---|---|
| Model | PyTorch detector (for example a YOLO-family model) | TensorFlow |
| Edge runtime | ONNX Runtime or TensorRT | OpenVINO |
| Data | Labelled image store with versioning | DVC |
| Monitoring | Per-field metrics dashboard | Grafana |
Small example
- Input
- Camera frame of a soybean row with two broadleaf weeds.
- Output
- Two bounding boxes labelled weed (0.91, 0.84) → nozzles 14 and 15 on for 30 ms (simulated).
Rebuild it yourself
Train a weed detector on a public agricultural image dataset and simulate nozzle decisions at different sensitivity thresholds.
Definition of done
- Separate test fields, not random images from the same field
- Report recall of weeds at each threshold
- Show the savings-versus-missed-weeds trade-off curve
Security and privacy risks
- Missed weeds reducing yield
- Model failing in new lighting or crops
- Safety of physical actuation (simulation only in this project)
Sources
Listed by title and publisher. This guide contains no external links; search for the title if you want to read the original.
- See & Spray customers see 59% average herbicide savings in 2024. Deere & Company, 18 Sep 2024. Company press release.
- Precision agriculture research measures effectiveness of See & Spray technology. University of Arkansas Division of Agriculture, 10 Mar 2025. Independent research.