PyTorch image classifier used by the eKheti agricultural decision-support project. The checkpoint covers 38 crop/fruit health and disease labels listed in labels.json.

Best validation accuracy recorded during training: 99.41%. This result is from the training pipeline's controlled validation split and must not be interpreted as field accuracy. Real farm images can differ substantially in lighting, background, leaf orientation, crop variety, and symptom overlap.

Use as a preliminary screening signal inside eKheti. Low-confidence predictions require a clearer image or expert confirmation. This model must not independently determine pesticide selection or dosage.

The model recognizes only its trained classes. Nutrient deficiencies, herbicide injury, mixed infections, unfamiliar crops, and non-leaf symptoms may be misclassified. Confirm consequential decisions through local agriculture officers, KVK specialists, or laboratory diagnosis.