🌦️ Weather Image Classification — TensorFlow

A weather-phenomenon classifier based on a single photograph, built on Transfer Learning, Fine Tuning, and Grad-CAM on the EfficientNetB0 architecture (TensorFlow / Keras). A project presenting a full, production-style pipeline: from data, through training, to evaluation, model explainability, and a ready-to-use prediction function for new images.

---

📖 Project Description

The model recognizes 11 weather phenomena from a single photograph: dew, fogsmog, frost, glaze, hail, lightning, rain, rainbow, rime, sandstorm, snow.

The goal of this project is to demonstrate how TensorFlow/Keras can be used for image classification by covering the entire workflow—from data exploration and augmentation, through two-stage training (transfer learning followed by fine-tuning), comprehensive evaluation using multiple performance metrics, to model explainability with Grad-CAM and a production-ready prediction pipeline for new data.

🧠 Model Architecture

Input (224, 224, 3)
        ↓
Data Augmentation (RandomFlip, RandomRotation, RandomZoom, RandomContrast, RandomBrightness)
        ↓
EfficientNetB0 (base, ImageNet weights)
   Phase 1: fully frozen (transfer learning)
   Phase 2: last 30 layers unfrozen, low learning rate (fine-tuning)
        ↓
GlobalAveragePooling2D
        ↓
Dropout (0.3)
        ↓
Dense (11, softmax)

Training strategy — two phases:

Phase	Base model	Learning rate	Goal
1. Transfer Learning	frozen	1e-3	training the classification head
2. Fine Tuning	last 30 layers unfrozen	1e-5	precise adjustment of high-level features

🛠️ Technologies Used

• TensorFlow 2.x / Keras — model building and training
• EfficientNetB0 (tf.keras.applications) — transfer learning base
• Data Augmentation — Keras layers (RandomFlip, RandomRotation, RandomZoom, RandomContrast, RandomBrightness)
• Callbacks: EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, TensorBoard
• Grad-CAM — model decision explainability (Explainable AI)
• scikit-learn — precision, recall, F1-score, confusion matrix, classification report
• Pandas / Matplotlib / Seaborn — data analysis and visualization
• Kaggle API — dataset download

📊 Results

Metric values are generated automatically in the notebook (Sections 14–20) after training and depend on the specific training run and hardware.

Metric
Test Accuracy
Precision (macro)
Recall (macro)
F1-score (macro)

Full results (confusion matrix, per-class classification report, training curves) are automatically saved to the results/ folder:

• results/metrics_summary.csv
• results/classification_report.csv
• results/confusion_matrix.png
• results/training_curves.png

Saved predictions on new images (Section 27) go to the predictions/ folder.

🚀 Setup Instructions

1. Clone the repository

git clone https://github.com/AIResearchForge/Weather-Image-Classification.git
cd Weather-Image-Classification

2. Virtual environment and dependencies

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Kaggle API key (to download the dataset)

1. Create a free account at kaggle.com.
2. Account → API → Create New API Token — this downloads a kaggle.json file.
3. The notebook will interactively prompt for KAGGLE_USERNAME and KAGGLE_KEY (getpass) — you'll find the values in the downloaded kaggle.json file. They are never hard-coded anywhere.

4. Run the notebook

jupyter notebook notebooks/Weather_Image_Classification_TensorFlow.ipynb

The notebook can also be run directly in Google Colab (recommended — free GPU access): just upload the .ipynb file and set Runtime → Change runtime type → GPU.

5. Predict on your own image(Section 27)

A full "file-to-result" inference pipeline — with automatic saving of the prediction result (image + label + confidence) to the predictions/ folder, simulating how results might be archived in a real application.

How to actually use this function on your own image — three ways:

1. Google Colab (easiest): run the cell below — a "Choose Files" button will appear, letting you pick an image from your computer. It will be automatically uploaded to the notebook and classified.

2. Local Jupyter / any environment: upload your image to the images/ folder in the repository, then in a new cell call:
predict_new_image("images/my_photo.jpg", loaded_model, loaded_class_names)
(replacing the filename with your own).

3. Image already present on the server/Colab disk — simply provide the full path to it as the first argument of the function.

📁 Repository Structure

Weather-Image-Classification/
├── docs/
├── images/
├── models/
├── notebooks/
│   └── Weather_Image_Classification_TensorFlow.ipynb
|   └── Weather_Image_Classification_TensorFlow_example.ipynb                 
├── predictions/        
├── results/           
├── .gitignore         
├── LICENSE
├── README.md
└── requirements.txt

🗺️ Development Roadmap

• Larger EfficientNet variants (B3–B7) or newer architectures (ConvNeXt, ViT)
• Combine with additional weather datasets (WEAPD and others) — a larger, better-balanced data pool
• Multi-label classification (co-occurrence of several phenomena in one image)
• Temporal sequence analysis from monitoring cameras (CNN+LSTM, video transformers)
• Export to TensorFlow Lite / ONNX for edge devices (weather stations)
• Deployment as a REST API (FastAPI + Docker) with Grad-CAM in the API response
• Active learning for the least confident predictions
• Production model monitoring (e.g. MLflow) — data drift detection

A full description of the model's limitations and possible development directions can be found in Sections 29–30 of the notebook.

📄 License

This project is released under the MIT license. The Weather Dataset is subject to the license set by its author — see Kaggle — jehanbhathena/weather-dataset.

👤 Author

AIResearchForge — a TensorFlow/Keras-based project for image classification, featuring Transfer Learning, Fine-Tuning, and model explainability using Grad-CAM.