🌾 AgriScan – AI-Powered Crop Disease Detection

Inspiration

Agriculture is the backbone of our food supply, yet millions of farmers lose crops each year due to late detection of plant diseases. Many small-scale farmers lack access to agricultural experts, leaving them vulnerable to crop failures. We wanted to build a fast, accessible, and intelligent solution that could help farmers detect plant diseases early — all from a simple smartphone camera. Our inspiration came from seeing how AI in healthcare has revolutionized diagnostics, and we thought:

“Why not bring this power to the fields?”

What I Learned

Model Training: We learned to train a custom CNN model using TensorFlow/Keras and transfer learning with MobileNetV2 for lightweight performance on web apps.
Dataset Handling: We explored the PlantVillage dataset and preprocessed thousands of leaf images with data augmentation (rotation, brightness, flipping).
Full-Stack AI Integration: We mastered connecting ML models with a backend API (Flask) and deploying them for real-time inference on the web.
Explainability in AI: We implemented Grad-CAM heatmaps to visualize which part of the leaf image influenced the model’s decision.

How I Built It

Data Collection & Preprocessing

Used PlantVillage Dataset with 50,000+ leaf images across various crops.
Preprocessed data using OpenCV and imgaug for image augmentation.

Model Development

Built a Convolutional Neural Network (CNN) using TensorFlow/Keras.
Trained on Google Colab using GPU acceleration.
Final model achieved 92% accuracy on the test set.

$\text{Accuracy} = \frac{\text{Correct Predictions}}{\text{Total Predictions}}$

Backend (Flask API)

Exposed a /predict endpoint that takes an uploaded image and returns the detected disease with confidence score.

Frontend (Web App)

Developed a responsive interface using HTML, TailwindCSS, and JavaScript.
Integrated real-time image upload & results visualization.

Deployment

Hosted backend on Render and static frontend on Vercel.
Optionally deployed the model on Hugging Face Spaces (Gradio) for live demos.

Challenges I Faced

Model Overfitting: Initial models overfit due to limited data variety, solved by heavy augmentation and dropout layers.
Lightweight Deployment: Serving large ML models on the web was tricky — we optimized using quantization and MobileNetV2.
Real-World Testing: Some images from farmers were noisy (blurred or bad lighting), so we added preprocessing filters before prediction.
SPA Routing Issues: We fixed routing on deployment using a JSON rewrites rule:

  { "rewrites": [ { "source": "/(.*)", "destination": "/" } ] }

What’s Next for AgriScan?

Adding offline mode (PWA) with TensorFlow.js for on-device inference.
Expanding the model to detect nutrient deficiencies and pest infestations.
Integrating multi-language support for rural farmers.
Building a community dashboard to track crop disease trends globally.