WeatherNet ⛅ – Multi-Class Weather Image Classification

🔧 Tech Stack

---

🌤 Project Overview

This project presents a multi-class image classification pipeline built on the Multi-class Weather Dataset (MWD) using advanced deep learning methods. It compares simple feedforward classifiers with tuned dense networks, convolutional architectures, and transfer learning using MobileNet.

Key challenges addressed include:

• Handling image resizing and channel correction
• Preventing overfitting through dropout and augmentation
• Improving model generalization on confusing weather conditions like "cloudy"

---

📁 Dataset

• 📦 Source: Multi-class Weather Dataset (Mendeley)
• 📸 Classes: cloudy, rain, shine, sunrise
• 🖼️ Images: 1,125 JPG files (~280 per class)
• 📊 Data Format: CSV-based partitioning into training, validation, and test splits

---

📌 Workflow Breakdown

1. 🔍 Data Exploration

• Random partition into my_training.csv, my_validation.csv, my_test.csv
• Label count visualization and distribution checks

2. Preprocessing

• Resizing to 230x230x3
• Normalization (pixel values to [0,1])
• Data pipeline using tf.data.TextLineDataset

3. Baseline Models

• Simple Flatten → Softmax model
• Custom Dense NN (512 → 128) + Dropout
• Hyperparameter tuning using keras-tuner

4. Advanced Models

• ConvNet: 2×Conv2D + MaxPooling + Dropout
• MobileNet Transfer Learning:
  • Pretrained on ImageNet
  • Frozen base layers
  • Fine-tuned classification head

---

📈 Performance Comparison

Model	Architecture	Accuracy	Notable Insight
Task 2.1	Flatten + Dense	~77%	Strong generalization with minimal overfitting
Task 2.2	Dense + Dropout (tuned)	~73%	Lower accuracy but better confidence (low loss)
Task 3.1	Custom ConvNet	83.43%	Best performer; struggles with "cloudy" class
Task 3.2	MobileNet (frozen)	66.27%	Struggles with data adaptation

---

📉 Error Analysis & Insights

🌩️ Classification Difficulty – “Cloudy”

Despite the strong ConvNet performance, cloudy images remained most difficult due to visual similarity with rainy and low-light sunrise scenes.

✅ What Could Improve It?

• Targeted Augmentation: Add fog, blur, and gray-shade filters
• Class Weighting: Penalize cloudy misclassifications during training
• Attention Layers: Apply CBAM or SE blocks for fine-grained feature extraction

---

📚 Key Takeaways

• Task 2.1 generalizes well with simple architectures and clean data
• Dropout and tuning help regularize dense models
• ConvNets outperform transfer learning when dataset is small and highly specific
• Cloudy images require specialized treatment due to ambiguity

---

🚀 Getting Started

# Clone the repository
git clone https://github.com/your-username/WeatherNet-Multi-Class-Weather-Image-Classification.git
cd WeatherNet-Multi-Class-Weather-Image-Classification

# Install requirements
pip install tensorflow keras pandas matplotlib numpy

# Launch the notebook
jupyter notebook WeatherNet-Multi-Class-Weather-Image-Classification.ipynb

---

📂 Project Structure

WeatherNet-Multi-Class-Weather-Image-Classification/
├── WeatherNet-Multi-Class-Weather-Image-Classification.ipynb
├── training.csv
├── validation.csv
├── test.csv
├── dataset2/
│   ├── cloudyXXX.jpg
│   ├── shineXXX.jpg
│   └── ...
├── README.md
└── requirements.txt

---

✨ Author

Gishor Thavakumar 🔗 LinkedIn 💻 Portfolio 📬 GitHub

---

📜 License

This project is licensed under the MIT License. See the LICENSE file for more details.

---

"This case study serves as a strong baseline for environmental AI classification tasks. The techniques here are extensible to other small-scale image recognition challenges involving weather, emotion, and scene classification."