rarfileexe/Plant-Disease-Detector

🌿 Plant Disease Detector — MobileNetV2 Fine-Tuned

38-class plant disease classification | 98.75% Test Accuracy | TensorFlow / Keras

---

📖 Model Overview

This model is a fine-tuned MobileNetV2 classifier capable of identifying 38 different plant disease categories (covering both diseased and healthy states) from leaf images. It was developed as an end-to-end deep learning project to assist farmers and agricultural experts in early crop disease detection, helping prevent large-scale crop loss through fast and accessible AI-powered diagnosis.

The model was trained on the PlantVillage (Augmented) dataset (~87,000 images, 38 classes, 14 crop species) and achieves 98.75% accuracy on the held-out test set.

• Architecture: MobileNetV2 (Transfer Learning + Fine-Tuning)
• Framework: TensorFlow 2.x / Keras (Functional API)
• Input: RGB leaf images, resized to 224 × 224
• Output: Softmax probabilities over 38 plant disease classes
• GitHub Repository: Muhammad-Hassan12/Plant-Disease-Detector
• Live Demo: Streamlit App

---

🧠 Training Methodology & Experimentation

To arrive at the best possible model, 4 experimental strategies were systematically evaluated. Each experiment built on the learnings of the previous one.

Exp	Strategy	Description	Outcome
1	Simple CNN (Scratch)	Custom CNN trained from scratch on raw images	Moderate accuracy; struggled with complex visual features
2	CNN + Data Augmentation	Same architecture with random flip, rotation, and zoom augmentation	Better generalization; significantly slower convergence
3	Transfer Learning — Feature Extraction	MobileNetV2 (frozen base) + custom classification head	High accuracy; very fast convergence (~few epochs)
4 ✅	Transfer Learning — Fine-Tuning	MobileNetV2 (top 55 layers unfrozen) + fine-tuning at lr=1e-4	Best Performance — 98.75% Test Accuracy

Why Fine-Tuning Won

Freezing the MobileNetV2 base (Experiment 3) gave strong results quickly, but the frozen ImageNet features weren't perfectly adapted to plant leaf textures. Unfreezing the top 55 layers and retraining at a very low learning rate (1e-4) allowed the model to adapt its high-level feature detectors to the specific visual patterns of diseased leaves without catastrophically forgetting the low-level ImageNet features. This is the standard recipe for domain-adapted transfer learning and it paid off here — pushing accuracy from ~97% (frozen) to 98.75% (fine-tuned).

---

🏗️ Model Architecture

Input (224 × 224 × 3)
    │
    ├── Preprocessing Layer (rescale pixels to [-1, 1])
    │
    ├── Data Augmentation Layers (active during training only)
    │       RandomFlip → RandomRotation → RandomZoom
    │
    ├── MobileNetV2 Base (pre-trained on ImageNet)
    │       Top 55 layers unfrozen for fine-tuning
    │       Remaining layers frozen
    │
    ├── Global Average Pooling 2D
    │
    ├── Dense(128, activation='relu')
    │
    └── Dense(38, activation='softmax')  →  38-class output

Key design decisions:

• Preprocessing baked into the model — no external normalization needed at inference time; just pass raw PIL/NumPy images.
• Augmentation inside the model graph — augmentation layers are automatically disabled during model.predict(), so inference is clean and deterministic.
• Global Average Pooling instead of Flatten — reduces parameters and improves regularization.

---

📊 Performance

Metric	Value
Test Accuracy	98.75%
Dataset	PlantVillage (Augmented)
Test Set Size	~10% of ~87,000 images
Number of Classes	38
Number of Crop Species	14

Training history (accuracy & loss curves) are available in the Training_Graphs folder of the GitHub repository.

---

🌿 Supported Classes (38 Total)

The model classifies leaf images into the following 38 categories across 14 crop species:

#	Class
1	Apple — Apple Scab
2	Apple — Black Rot
3	Apple — Cedar Apple Rust
4	Apple — Healthy
5	Blueberry — Healthy
6	Cherry — Powdery Mildew
7	Cherry — Healthy
8	Corn — Cercospora Leaf Spot / Gray Leaf Spot
9	Corn — Common Rust
10	Corn — Northern Leaf Blight
11	Corn — Healthy
12	Grape — Black Rot
13	Grape — Esca (Black Measles)
14	Grape — Leaf Blight (Isariopsis Leaf Spot)
15	Grape — Healthy
16	Orange — Haunglongbing (Citrus Greening)
17	Peach — Bacterial Spot
18	Peach — Healthy
19	Pepper (Bell) — Bacterial Spot
20	Pepper (Bell) — Healthy
21	Potato — Early Blight
22	Potato — Late Blight
23	Potato — Healthy
24	Raspberry — Healthy
25	Soybean — Healthy
26	Squash — Powdery Mildew
27	Strawberry — Leaf Scorch
28	Strawberry — Healthy
29	Tomato — Bacterial Spot
30	Tomato — Early Blight
31	Tomato — Late Blight
32	Tomato — Leaf Mold
33	Tomato — Septoria Leaf Spot
34	Tomato — Spider Mites / Two-Spotted Spider Mite
35	Tomato — Target Spot
36	Tomato — Tomato Yellow Leaf Curl Virus
37	Tomato — Tomato Mosaic Virus
38	Tomato — Healthy

---

🚀 How to Use

Option 1 — Direct Inference with TensorFlow

import tensorflow as tf
import numpy as np
from PIL import Image

# Load model
model = tf.keras.models.load_model("model_4_mobilenet_finetuned.keras")

# Class labels (must match training order)
CLASS_NAMES = [
    "Apple___Apple_scab", "Apple___Black_rot", "Apple___Cedar_apple_rust", "Apple___Healthy",
    "Blueberry___Healthy", "Cherry___Powdery_mildew", "Cherry___Healthy",
    "Corn___Cercospora_leaf_spot_Gray_leaf_spot", "Corn___Common_rust",
    "Corn___Northern_Leaf_Blight", "Corn___Healthy",
    "Grape___Black_rot", "Grape___Esca_(Black_Measles)",
    "Grape___Leaf_blight_(Isariopsis_Leaf_Spot)", "Grape___Healthy",
    "Orange___Haunglongbing_(Citrus_greening)",
    "Peach___Bacterial_spot", "Peach___Healthy",
    "Pepper,_bell___Bacterial_spot", "Pepper,_bell___Healthy",
    "Potato___Early_blight", "Potato___Late_blight", "Potato___Healthy",
    "Raspberry___Healthy", "Soybean___Healthy", "Squash___Powdery_mildew",
    "Strawberry___Leaf_scorch", "Strawberry___Healthy",
    "Tomato___Bacterial_spot", "Tomato___Early_blight", "Tomato___Late_blight",
    "Tomato___Leaf_Mold", "Tomato___Septoria_leaf_spot",
    "Tomato___Spider_mites_Two-spotted_spider_mite", "Tomato___Target_Spot",
    "Tomato___Tomato_Yellow_Leaf_Curl_Virus", "Tomato___Tomato_mosaic_virus",
    "Tomato___Healthy"
]

def predict(image_path: str):
    img = Image.open(image_path).convert("RGB").resize((224, 224))
    img_array = np.expand_dims(np.array(img), axis=0)  # shape: (1, 224, 224, 3)
    # No manual normalization needed — preprocessing is baked into the model
    predictions = model.predict(img_array)
    predicted_class = CLASS_NAMES[np.argmax(predictions)]
    confidence = float(np.max(predictions))
    return {"class": predicted_class, "confidence": f"{confidence:.2%}"}

result = predict("your_leaf_image.jpg")
print(result)
# Example output: {'class': 'Tomato___Early_blight', 'confidence': '99.12%'}

Option 2 — Run the Streamlit App Locally

# Clone the repository
git clone https://github.com/Muhammad-Hassan12/Plant-Disease-Detector.git
cd Plant-Disease-Detector

# Install dependencies
pip install -r requirements.txt

# Launch the app
streamlit run Deploy/app.py

---

📂 Repository Structure

📦 Plant-Disease-Detector
├── app.py                              # Streamlit web app
├── model_4_mobilenet_finetuned.keras   # Final deployed model
└── requirements.txt                    # Requirements and Dependencies

---

📦 Dependencies

tensorflow>=2.10
streamlit
pillow
numpy
matplotlib

Install via:

pip install -r requirements.txt

---

📋 Dataset

Property	Detail
Name	PlantVillage (Augmented)
Source	Kaggle — PlantVillage Dataset
Total Images	~87,000 RGB images
Classes	38 (disease + healthy states)
Crop Species	14
Image Type	Controlled lab-condition leaf photographs
Split	~80% Train / ~20% Validation+Test

Note on Dataset Scope: The PlantVillage dataset consists of lab-controlled images with uniform backgrounds. Real-world field performance may vary due to lighting variation, occlusion, soil on leaves, and other environmental factors. Always validate with domain experts before production agricultural deployment.

---

⚠️ Limitations & Disclaimer

• Dataset bias: Trained exclusively on PlantVillage lab-condition images. Performance on field photographs with cluttered backgrounds may degrade.
• Species coverage: Only covers the 14 crop species present in PlantVillage. Predictions on out-of-distribution crops will be unreliable.
• Intended use: This model is designed for educational and assistive purposes. It is not a replacement for professional agricultural diagnosis. Findings should be confirmed by qualified agricultural experts before taking large-scale action (chemical treatment, crop disposal, etc.).
• Not for clinical or regulatory use.

---

👤 Author

Syed Muhammad Hassan

• GitHub: @Muhammad-Hassan12
• HuggingFace: rarfileexe
• BS Computer Science, Ziauddin University, Karachi, Pakistan

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

🙏 Acknowledgements

• Kaggle for providing free GPU compute (12-hour sessions) used to train all four experimental models.
• The PlantVillage project and Hughes et al. for the open dataset enabling this research.
• Google for MobileNetV2 and the ImageNet pre-trained weights that made transfer learning possible.