AgriMind Ensemble Boosts Plant Disease Detection to 99.23% Accuracy
AgriMind, an innovative ensemble deep learning framework, integrates ResNet50, EfficientNet-B0, and DenseNet121 to classify multiple plant diseases. Utilizing 20,638 images from PlantVillage, covering 15 disease categories in pepper, potato, and tomato, this ensemble achieved an impressive accuracy of 99.23% on a separate test set, significantly lowering error rates by two-thirds compared to individual models (96–97%). The approach employed transfer learning with fixed ImageNet backbones and 10 epochs of head-only training to maintain a streamlined pipeline. Efforts to enhance the average towards the top validation model or to exclude any single model led to diminished results. The framework aims to facilitate automated disease detection for extension workers in Bangladesh, where manual inspections remain common, with practical mobile use relying on TensorFlow Lite optimization.
Key facts
- AgriMind ensemble combines ResNet50, EfficientNet-B0, and DenseNet121.
- Trained on 20,638 PlantVillage images across 15 disease classes.
- Ensemble accuracy: 99.23% on held-out test set.
- Individual models achieved 96–97% accuracy.
- Error rate reduced by two-thirds compared to best single model.
- Transfer learning with frozen ImageNet backbones and 10 epochs head-only training.
- Pepper and potato classes classified perfectly; tomato reached 99.01%.
- Runs at 53 FPS on NVIDIA T4 GPU.
Entities
Institutions
- arXiv
- PlantVillage
- NVIDIA
Locations
- Bangladesh