🌾 Crop AI — Intelligent Crop Classification & Quality Grading

Computer Vision for 5-Crop Recognition + Multi-Crop Quality Grading 🤖🌱

Crop AI is a deep-learning agricultural computer-vision project using a two-stage pipeline:

1. Crop classification: Banana, Guava, Maize, Rice, or Wheat.
2. Quality grading: a second EfficientNetB2 model grades supported crop images into A / B / C / D quality classes.

Current quality-model release: EfficientNetB2 at 260×260, trained on 4,295 labelled images across 18 available crop-quality classes. The best validation accuracy recorded during training was 91.46%.

🚀 Pipeline

📷 Crop Image
      ↓
🖼️ 224 × 224
      ↓
🧠 EfficientNetB0 Crop Classifier
      ↓
🌾 Banana / Guava / Maize / Rice / Wheat
      ↓
🖼️ 260 × 260
      ↓
🧠 EfficientNetB2 Quality Model
      ↓
🏷️ A / B / C / D

🧠 Released Models

Model	Architecture	Classes	Recorded Accuracy
Crop Classification Champion	EfficientNetB0	5 crops	91.76% test
Multi-Crop Quality Champion	EfficientNetB2	18 available crop-quality classes	91.46% validation

Crop Classification Champion

model/CROP_MODEL_CHAMPION_91_76_TEST.keras

• Input: 224 × 224 × 3
• Classes: Banana, Guava, Maize, Rice, Wheat
• Recorded test accuracy: 91.76%
• EfficientNetB0 transfer learning with augmentation and a 5-class softmax head.

Multi-Crop Quality Champion

model/QUALITY_MODEL_B2_260_BEST.keras

• Architecture: EfficientNetB2
• Input: 260 × 260 × 3
• Best recorded validation accuracy: 91.46%
• Dataset: 4,295 images
• Available quality classes:
  • Banana: A / B / D
  • Guava: A / B / D
  • Maize: A / B / C / D
  • Rice: A / B / C / D
  • Wheat: A / B / C / D
• Banana-C and Guava-C are not present in the current training dataset, so this release is an 18-class model rather than a complete 20-class (5 crops × 4 grades) model.

Quality labels are project-defined visual grading categories, not official agricultural certification grades.

📊 Dataset

Crop Classification Dataset

Crop	Images
Wheat	4,000
Rice	4,000
Maize	4,000
Banana	1,194
Guava	1,000
Total	14,194

Split: 80% train / 10% validation / 10% test, stratified with random_state=42.

• Train: 11,355
• Validation: 1,419
• Test: 1,420

The raw dataset is not included in this repository.

Quality Dataset

The current quality training set contains 4,295 images across 18 available crop-quality classes. The current distribution is intentionally capped at up to 250 images per class where available; Banana and Guava do not currently have Class C images in the prepared dataset.

🧪 Quick Inference

git clone https://github.com/prabhtheone/Crop-AI.git
cd Crop-AI
pip install -r requirements.txt
python src/predict.py path/to/your/image.jpg

A notebook demo is also included at:

notebook/Crop_AI_Inference_Demo.ipynb

📁 Repository Structure

Crop-AI/
├── model/
│   ├── CROP_MODEL_CHAMPION_91_76_TEST.keras
│   └── QUALITY_MODEL_B2_260_BEST.keras
├── notebook/
│   └── Crop_AI_Inference_Demo.ipynb
├── src/
│   └── predict.py
├── .github/workflows/ci.yml
├── README.md
├── LICENSE
├── requirements.txt
├── CITATION.cff
├── CONTRIBUTING.md
├── SECURITY.md
├── .gitattributes
└── .gitignore

The .keras files are managed with Git LFS. Raw datasets and private credentials are excluded from Git.

🛠️ Tech Stack

• Python
• TensorFlow / Keras
• EfficientNetB0
• EfficientNetB2
• NumPy
• Pillow
• Pandas
• Scikit-learn
• Matplotlib
• Google Colab
• Git LFS

📈 Evaluation Notes

The crop-classification 91.76% figure is a recorded test-set accuracy.

The quality-model 91.46% figure is the best validation accuracy recorded during the 18-class training run. It should not be presented as a held-out test accuracy.

Real-world performance can change with lighting, backgrounds, camera quality, crop varieties, image source, and other domain-shift conditions.

⚠️ Limitations

• The current quality model has 18 available crop-quality classes because Banana-C and Guava-C are missing from the prepared dataset.
• A complete 20-class A/B/C/D model requires real Banana-C and Guava-C labelled images.
• Quality grades are project-defined visual categories, not official agricultural certification.
• This is a research/learning prototype, not a professional agricultural diagnosis system.
• Larger and more diverse real-world testing is still needed.

🔮 Roadmap

• Multi-crop classification
• EfficientNetB0 transfer learning
• Crop evaluation
• Multi-crop quality model
• A/B/C/D grading scheme
• EfficientNetB2 260×260 quality model
• Add real Banana-C and Guava-C data
• Train complete 20-class A/B/C/D model
• Held-out test evaluation for quality model
• Larger real-world robustness evaluation
• Web/mobile deployment
• Explainable AI / visual attention

🤝 Contributing

Bug reports, experiments, documentation improvements, and pull requests are welcome. See CONTRIBUTING.md.

📄 License

MIT License — see LICENSE.

⭐ Support

If you find Crop AI useful for learning or experimentation, consider starring the repository and sharing constructive feedback.

Built as a student project exploring AI, computer vision, and agriculture. 🌾🤖