Have you ever wondered how farmers could spot plant diseases early, even before the damage spreads too far? Early detection can save crops, boost food supply, and even make farming more sustainable. That’s exactly what this project is about: using AI to detect plant diseases from leaf images.

In this post, I’ll walk you through how I built a leaf disease classifier using the PlantVillage dataset and a modern deep learning model called EfficientNet. Don’t worry if some terms sound new—I’ll explain as we go!

🛠️ GitHub Repository: 👉 View the Code on GitHub
( https://github.com/debabratapruseth/Plant_disease_AI/tree/main)

---

Why This Matters

Plant diseases aren’t just a farmer’s problem—they affect food security worldwide. If crops are infected and go unnoticed, yields drop, prices rise, and food shortages can follow.

By using machine learning to automatically detect diseases from leaf photos, farmers can:

• Diagnose problems earlier
• Treat crops faster
• Reduce waste and loss
• Grow food more sustainably

---

The Dataset: PlantVillage

For this project, I used the PlantVillage dataset, which is freely available through TensorFlow Datasets.

• Images: ~54,000 leaf photos
• Categories: 38 types (examples: Tomato – Late blight, Apple – Black rot, Potato – Late blight)

This variety makes the model capable of recognizing many different plant species and diseases.

---

The Model: EfficientNet

Instead of starting from scratch, I used transfer learning—which means taking a model trained on a huge dataset (like ImageNet) and fine-tuning it for plant disease detection.

I started with EfficientNet-B0 (a smaller version of the model) and then upgraded to EfficientNet-B4 for better accuracy.

Here’s what I did to improve results:

• ✅ Two-stage training – First, train only the top layers. Then, fine-tune the deeper layers.
• ✅ Focal loss – A special loss function that helps when some classes have fewer examples.
• ✅ Mixed precision training – Speeds up training on GPUs while using less memory.

The result? B4 with these tweaks outperformed B0 and gave more accurate predictions.

---

Features at a Glance

• Transfer learning with EfficientNet (B0 → B4)
• Handles class imbalance with focal loss
• Efficient tf.data pipeline with lightweight augmentations
• Training in two stages for stability and better performance
• Evaluation with classification reports, confusion matrix, and accuracy metrics
• Explainability using Grad-CAM (to see which parts of the leaf the model looks at)

---

Results & Next Steps

The upgraded EfficientNet-B4 model performed significantly better than the baseline. But there’s always room to experiment!

Some future ideas include:

• Training specialized models per plant species and combining them
• Trying stronger image augmentations (random crop, cutout, color jitter)
• Using different model architectures like Vision Transformers (ViT)
• Analyzing model failures with Grad-CAM heatmaps

---

How You Can Try This Yourself

🛠️ GitHub Repository: 👉 View the Code on GitHub
( https://github.com/debabratapruseth/Plant_disease_AI/tree/main)

The project is beginner-friendly—you can run it right inside Google Colab with GPU.

Steps to run:

1. Open the notebook in Colab.
2. Enable GPU runtime.
3. Run the cells in order (setup → dataset → model → training → evaluation).

That’s it—you’ll have your very own plant disease detector running in the cloud! 🌿

---

Final Thoughts

AI in agriculture is more than just a tech experiment—it has real-world impact. By catching plant diseases early, we can help farmers, improve crop yields, and contribute to global food security.

This project is just the beginning, and I encourage you to try it out, play with the code, and maybe even improve it. Who knows—your version could one day help farmers in the field!