🌱 SMART-SPRAY-X
AI-Driven Pesticide Optimization System
Empowering Farmers with Intelligent Disease Detection
A Solution for Smart India Hackathon 2026
🌐 Live Demo
⚠️ Note: Backend is hosted on Render free tier — first request after inactivity may take ~30-45 seconds to wake up.
📖 Overview
SMART-SPRAY-X is an AI-driven initiative for Precision Agriculture, developed for the Smart India Hackathon 2026. It addresses the critical issue of indiscriminate pesticide use by farmers.
Unlike traditional methods where pesticides are sprayed uniformly, SMART-SPRAY-X uses intelligent disease detection to determine:
- Whether a plant is infected.
- What disease it has.
- How severe the infection is — powering precise pesticide dosage recommendations via the built-in Disease Severity Index.
This "Spot-Spray" approach drastically reduces chemical usage, lowers costs for farmers, and protects the environment from toxic residue.
🏗️ Architecture
┌─────────────────────────────────────────────────────────────┐
│ User Browser │
│ (Vercel - React + Vite) │
│ https://smartspray-x.vercel.app │
└─────────────────────────┬───────────────────────────────────┘
│ REST API (HTTPS)
▼
┌─────────────────────────────────────────────────────────────┐
│ FastAPI Backend │
│ (Render - Docker Container) │
│ https://plant-disease-9v9x.onrender.com │
│ │
│ ┌─────────────────┐ ┌──────────────┐ ┌────────────────┐ │
│ │ 5-Layer Leaf │ │ MobileNetV2 │ │ Severity Index │ │
│ │ Validator │ │ (38 classes) │ │ HSV Segment. │ │
│ │ HSV + OOD + AR │ │ TF/Keras │ │ + Spot-Spray │ │
│ └─────────────────┘ └──────────────┘ └────────────────┘ │
└─────────────────────────────────────────────────────────────┘
🛡️ Key Features
| Feature | Details |
|---|
| 🌿 Multi-Class Detection | Identifies 38 different plant diseases across crops (Apple, Corn, Grape, Potato, Tomato, etc.) |
| ⚡ High Accuracy | MobileNetV2 fine-tuned on 87K+ images achieving ~96% validation accuracy |
| 🛑 5-Layer Non-Leaf Rejection | Solid background detection · Aspect ratio analysis · Foliage coverage · OOD confidence threshold · Top-2 margin check |
| 📊 Disease Severity Index | HSV symptom segmentation estimates affected leaf area on a 5-level scale (Very Low → Critical) |
| 💊 Pesticide Recommendations | Spot-spray optimization for each of the 38 disease classes |
| 📈 Interactive UI | Glassmorphism dark emerald React frontend with drag & drop upload |
🛠️ Technology Stack
| Layer | Technology |
|---|
| Frontend | React 18, Vite, Lucide Icons |
| Backend | FastAPI, Uvicorn, Python 3.10 |
| AI/ML | TensorFlow/Keras (MobileNetV2), NumPy |
| Deployment | Vercel (Frontend), Render + Docker (Backend) |
| Image Processing | PIL, Pure NumPy HSV |
📂 Dataset Details
- Total Images: ~87,000 RGB
- Classes: 38 (Apple Scab, Black Rot, Late Blight, etc.)
- Input Size: 128×128×3
- Split: 80% Training / 20% Validation
⚙️ Local Development
Backend
cd "Plant Disease"
conda activate tensorflow
python start_backend.py
# → http://localhost:8000
Frontend
cd frontend
npm install
npm run dev
# → http://localhost:5173
API Endpoints
| Method | Endpoint | Description |
|---|
GET | /api/v1/health | Health check + model status |
POST | /api/v1/predict | Upload leaf image → disease prediction + severity analysis |
GET | /api/v1/classes | List all 38 disease classes |
GET | /docs | Swagger API documentation |
📊 Disease Severity Analysis
POST /api/v1/predict returns a structured severity object alongside the existing prediction fields. Severity estimates how much of the leaf is visibly affected (not how confident the classifier is), derived from the image itself:
{
"is_valid_leaf": true,
"crop": "Apple",
"disease": "Cedar Rust",
"status": "Diseased",
"confidence_percentage": 94.0,
"recommendations": { "...": "..." },
"severity": {
"available": true,
"score": 67,
"label": "Severe",
"level": "severe",
"affected_area_percentage": 67.2,
"affected_regions": 2,
"largest_region_fraction": 0.81,
"confidence": 0.91,
"explanation": "An estimated 67.2% of the leaf area shows visible disease symptoms (necrotic/yellow tissue or pale mildew). Symptoms are spread across 2 separate regions."
}
}
How the score is computed (backend/app/services/severity_service.py):
- Symptom segmentation — pure NumPy HSV analysis separates healthy-green foliage from necrotic/yellow tissue and pale mildew patches.
- Affected area = weighted symptom pixels ÷ total foliage pixels → a percentage of the leaf, not the whole image.
- Region analysis — connected components on the symptom mask report how many separate lesions exist and the share held by the largest one.
- Model verdict gate — a
Healthy classification discounts the measured area (healthy leaves naturally show veins/edges/senescence that color analysis can't fully separate from disease); a Diseased classification reports it directly.
- Severity confidence (separate from the score) reflects trust in the estimate: diagnosis confidence, top-2 margin, foliage coverage, and whether symptoms were actually measurable.
Severity scale (configurable via SEVERITY_THRESHOLDS):
| Score | Label | Indicator |
|---|
| 0 – 20 | Very Low | 🟢 |
| 21 – 40 | Mild | 🟡 |
| 41 – 60 | Moderate | 🟠 |
| 61 – 80 | Severe | 🔴 |
| 81 – 100 | Critical | 🔴 |
Limitations — The CNN is a classifier, not a segmenter; affected area is estimated from color, not measured by the network. Diseases with shape-based or very subtle symptoms (early leaf curl, mosaic mottling) are under-estimated. Color segmentation cannot separate true symptoms from sunburn, dirt, or lighting artifacts. When symptoms cannot be measured the module returns "available": false rather than fabricating a score.
🧠 Model Architecture
| Component | Detail |
|---|
| Base Model | MobileNetV2 (ImageNet pre-trained, fine-tuned) |
| Input | 128×128 RGB → internal Rescaling(1/127.5, offset=-1) layer |
| Feature Head | GlobalAveragePooling2D → BatchNormalization → Dense layers with Dropout |
| Output | Softmax over 38 classes |
| Optimizer | Adam |
| Model Size | ~13 MB (.keras format) |
🛑 5-Layer Validation System
To prevent false diagnoses (e.g., rejecting a dog photo on a green background), every uploaded image passes through 5 independent checks before the model is even consulted:
| Check | What it does | Catches |
|---|
| 1. Solid Background Detection | Color variance in green regions; rejects >70% uniform green with low variance | Green screen photos, artificial scenes |
| 2. Aspect Ratio Analysis | Bounding box of foliage; rejects objects with aspect ratio >6.0 | Elongated non-leaf objects |
| 3. Foliage Coverage | Requires ≥8% green/brown leaf-colored pixels | Documents, sky, water |
| 4. OOD Confidence | Rejects top-1 softmax <40% | Out-of-distribution inputs, non-plant images |
| 5. Top-2 Margin | Rejects when gap between top-1 and top-2 <8% | Ambiguous / confused predictions |
📸 How It Works
Upload Leaf Image
│
▼
┌─────────────────────────┐
│ 5-Layer Leaf Validator │ ← Rejects non-leaf, adversarial, ambiguous inputs
└─────────────┬───────────┘
│ ✅ Valid
▼
┌─────────────────────────┐
│ MobileNetV2 Classifier │ ← 38-class disease prediction + confidence margin
└─────────────┬───────────┘
│
▼
┌─────────────────────────┐
│ HSV Severity Engine │ ← Estimates affected leaf area → Very Low to Critical
└─────────────┬───────────┘
│
▼
┌─────────────────────────┐
│ Spot-Spray Recommender │ ← Pesticide type + dosage guidance
└─────────────────────────┘
🔮 Future Roadmap
👥 Team
Smart India Hackathon 2026 Participant
Developing innovative solutions for sustainable agriculture.
Made with ❤️ for SIH 2026 · SMART-SPRAY-X © 2026
