- Project
- AI/ML
- Web
- Solo
- 2024
Plant disease detection from leaf photos
Classical models versus CNNs on PlantVillage leaves, served through a web app with a low-confidence guard.
How it works
- 01Preprocessing and hand-made features for SVM and Random Forest
- 02MobileNet and ResNet fine-tuned on leaf images
- 03Model served from Google Cloud Functions behind an API
- 04Gradio app rejects predictions below 50% confidence
MobileNet macro-F1On 3,943 test leaves; Random Forest 0.90 and SVM 0.88 on hand-made features.
MobileNet macro-F1
On 3,943 test leaves; Random Forest 0.90 and SVM 0.88 on hand-made features.
The problem
Farmers need a quick read on leaf disease from a phone photo, and a wrong confident answer does harm.
My part
Built the full project: training, API, cloud deployment, and the Gradio web app.
What did not work
- Hand-made features topped out at 0.88-0.90 macro-F1; CNNs reached 0.993.
Limitations
- Trained on PlantVillage studio-style photos; field photos with clutter are harder.