AI · 2022 · Built alongside the KJSCE study by Vora and Padalia
Apple Leaf Disease Classification
Which of these leaves is sick, and with what?
An ensemble of Xception, InceptionResNet and MobileNet that reads scab, rust, frogeye leaf spot, powdery mildew and complex infections off apple leaf photos from the Plant Pathology 2021 dataset, including leaves carrying more than one disease.
Problem
Apple orchards are scouted by eye for scab, rust, frogeye leaf spot and powdery mildew, and one leaf can carry more than one of them. Diagnosis comes late and is easy to get wrong.
Approach
- Plant Pathology 2021 (FGVC8): 18,632 leaf photos from an unsprayed Cornell AgriTech orchard, 1,355 of them showing several diseases; resized to 256×256 and split 80/20.
- Augmentation to mimic the field: rotation, flips, translation, zoom and shear.
- Five backbones with a dense, batch-norm and dropout head and sigmoid outputs for multi-label prediction; the best ensemble votes across Xception, InceptionResNet and MobileNet, with Xception breaking ties.
Stack
- Python
- TensorFlow
- Keras
- OpenCV
Outcome
- Multi-label classification across six leaf conditions, including leaves with more than one disease
- The reference ensemble reports 87.31% accuracy and an F1 score of 0.90 on 3,726 test images