Plant-Disease ClassifierEfficientNet-B0 transfer learning
Early plant-disease detection is a diagnosis problem farmers can't afford a specialist for. A phone camera and a small model go a surprisingly long way.
The problem
Field photography is hostile to classifiers — inconsistent lighting, cluttered backgrounds, and visually similar diseases across different species.
The approach
EfficientNet-B0 with transfer learning on an 80/20 split, and a deliberate augmentation regime to simulate field conditions: horizontal and vertical flips, rotations to 30°, 20% zoom, and contrast and brightness jitter. Learning rate, batch size and epoch count were tuned systematically rather than by feel, with the effect of each recorded.
The outcome
95–97% test-set accuracy across six disease categories, with a per-class confusion analysis showing where the remaining errors cluster. Graded 92 — Distinction.