🌿 FASALAI — Crop Disease Detection for Indian Farmers

Detect crop diseases from a leaf photo. Hindi output. Offline-first. No internet needed.

---

Problem Statement

India loses ₹50,000+ crore annually to crop diseases. Over 86% of Indian farmers are small or marginal — they lack access to agronomists, can't afford lab testing, and often rely on guesswork until it's too late.

The core failure: Disease identification requires expert knowledge that most farmers simply don't have. By the time symptoms are visible and an expert is consulted, 30–70% of yield is already lost.

Pain Points

• Rural areas have no internet connectivity for cloud-based diagnosis tools
• Existing tools output in English only — inaccessible to most farmers
• Expert consultation costs ₹500–2000+ per visit — unaffordable at scale
• Lab results take 3–7 days — crops don't wait

---

Our Solution — FasalAI

FasalAI is an offline-first, Hindi-language crop disease detector that runs entirely on a smartphone with no internet connection.

A farmer takes a photo of a diseased leaf → FasalAI identifies the disease → gives Hindi treatment advice in under 100ms.

Key Differentiators

Feature	FasalAI	Other Tools
Works offline	✅ Yes	❌ Requires internet
Hindi output	✅ Full treatment advice	❌ English only
Inference speed	✅ <100ms on CPU	⚠️ 2–5 seconds (cloud)
Model size	✅ 4MB (INT8)	❌ 100MB+
Environmental context	✅ Temp + Humidity aware	❌ Image only
Cost to farmer	✅ Free	❌ Subscription / data cost

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        FasalAI Model                            │
│                                                                 │
│  Input Image (224×224)                                          │
│        ↓                                                        │
│  EfficientNet-B0 (Pretrained ImageNet)                          │
│  [Blocks 0-4 frozen | Blocks 5-8 fine-tuned]                    │
│        ↓                                                        │
│  CBAM Attention Module                                          │
│  ┌──────────────────────┐                                       │
│  │ Channel Attention    │ ← Weights 1280 feature channels       │
│  │ Spatial Attention    │ ← Weights 7×7 spatial regions         │
│  └──────────────────────┘                                       │
│        ↓           ↓                                            │
│  Avg Pool      Env Branch (Temperature + Humidity)              │
│  (1280-dim)    → Linear(2→32) → ReLU → Dropout                  │
│        ↓           ↓                                            │
│  Concat: [1280 + 32] = 1312-dim                                 │
│        ↓                                                        │
│  Classifier: Dropout(0.4) → FC(1312→512) → ReLU →               │
│              Dropout(0.3) → FC(512→38)                          │
│        ↓                                                        │
│  Softmax → Top-1 Class + Confidence                             │
│        ↓                                                        │
│  hindi_output.py → Hindi Disease Advice                         │
└─────────────────────────────────────────────────────────────────┘

Why This Architecture?

• EfficientNet-B0: Best accuracy/size tradeoff. Proven on medical imaging. 5.3M parameters vs 25M for ResNet-50.
• CBAM Attention: Forces the model to focus on disease lesions, not background soil/sky. Improves accuracy by ~2-3% on leaf datasets.
• Environmental Branch: Temperature and humidity are strong predictors of certain diseases (e.g., Yellow Rust thrives below 20°C with >75% humidity). Even without real sensors, these inputs act as soft contextual priors.
• INT8 Quantization: Reduces model from 15MB → 4MB with <0.5% accuracy drop. Runs on mid-range Android phones.

Environmental Data — No Sensor? No Problem

Since real IoT sensors aren't always available:

• Default: Model uses neutral values (25°C, 65% humidity) — equivalent to a standard Indian kharif season day
• Demo mode: User adjusts sliders manually based on current weather
• Production path: Designed to integrate with cheap DHT11 sensors (₹50) or weather APIs
• Dataset augmentation: Training env features are randomly sampled per realistic Indian season ranges

---

Dataset

PlantVillage Dataset — the gold standard for plant disease classification

Property	Value
Total images	~54,000
Disease classes	38 (across 14 crop types)
Image format	RGB JPG, 256×256
Crops covered	Tomato, Potato, Wheat, Corn, Apple, Grape, Pepper, Cherry, Peach, Orange, Soybean, Strawberry, Blueberry, Raspberry
Source	Kaggle: emmarex/plantdisease

---

Quick Start

Prerequisites

git clone https://github.com/Chandanrohit/crops_disease_detector.git
cd fasalai
pip install -r requirements.txt

Run Demo (Mock Mode — No Model Needed)

# Works immediately, no training required
streamlit run demo.py
# Opens http://localhost:8501

Train From Scratch

# 1. Download dataset
# Place PlantVillage in ./data/PlantVillage/

# 2. Train (GPU recommended — or use Google Colab)
python train.py

# 3. Export to ONNX
python export_onnx.py

# 4. Quantize (optional, for production)
python quantize.py

# 5. Run real inference
# Edit demo.py: set USE_MOCK = False
streamlit run demo.py

Train on Google Colab (Recommended)

Open notebooks/FasalAI_Colab.ipynb in Google Colab for a step-by-step guided notebook with GPU support, live training graphs, and automatic Drive checkpointing.

CLI Inference

python inference.py --image leaf.jpg --temperature 22 --humidity 78

---

Repository Structure

fasalai/
├──  notebooks/
│   └── FasalAI_Colab.ipynb        ← Full Colab training notebook
├──  checkpoints/
│   ├── crop_disease_model.onnx     ← FP32 inference model