This project leverages deep learning to estimate the intensity of tropical cyclones using infrared satellite imagery from the INSAT-3D satellite. A web-based interface built with Django allows users to upload cyclone images and receive real-time predictions of cyclone intensity, measured in knots.
- Source: INSAT3D Infrared Cyclone Dataset (2013–2021) – Kaggle
- Input Format: Infrared
.pngimages of cyclones over the Indian Ocean - Labels: Cyclone intensities (in knots)
All stages of model training—from preprocessing to evaluation—are implemented in a Jupyter Notebook.
-
Data Loading
- CSV containing image filenames and corresponding intensity labels
-
Preprocessing
- Resized all images to 224×224
- Normalized pixel values
- Converted to PyTorch tensors
-
Model Architecture
- Backbone: EfficientNet-B0 pretrained on ImageNet
- Final layer adapted for regression
-
Training Configuration
- Loss Function: Mean Squared Error (MSE)
- Optimizer: Adam
- Learning Rate Scheduler: ReduceLROnPlateau
- Includes Early Stopping and Model Checkpointing
-
Evaluation Metrics
- Mean Absolute Error (MAE): ~9.2 knots
- Root Mean Square Error (RMSE): ~10.7 knots
- R² Score: ~0.84
| Metric | Value |
|---|---|
| MAE | 9.2 |
| RMSE | 10.7 |
| R² | ~0.84 |
An interactive Django web application allows users to upload infrared cyclone images and get predicted intensity values.
- Browser-based image upload
- Server-side preprocessing and inference using the trained model
- Clean, minimal frontend interface
- Real-time display of predicted cyclone intensity
- User uploads an image through the frontend form
- The image is stored and passed to a prediction script
- The model loads
best_model.pt, processes the image, and performs inference - Predicted intensity is returned and displayed to the user
git clone https://github.com/your-username/Cyclone-Intensity.gitpython -m venv venv
venv\Scripts\activate # For Windows
pip install -r requirements.txtcd Cyclone-Intensity
python manage.py runserverNavigate to http://127.0.0.1:8000/ to access the web interface.
- Python 3.10+
- PyTorch
- torchvision
- Pillow
- Django 5.x
- scikit-learn
- matplotlib, pandas, numpy
Install all dependencies using:
pip install -r requirements.txt- Expand dataset with high-resolution, multi-basin satellite imagery (e.g., GOES, MODIS)
- Experiment with advanced architectures (EfficientNetV2, Swin Transformer)
- Implement regularization techniques and data augmentation
- Optimize hyperparameters using cross-validation
- Connect with satellite APIs (NOAA, NASA Earthdata, ISRO MOSDAC)
- Schedule automatic image ingestion using Celery or cron
- Cyclone localization using object detection models (e.g., YOLOv8)
- Alert system (email/SMS/WhatsApp) for critical intensity thresholds
- Integration with mapping libraries (e.g., Mapbox, Leaflet.js) for visualization
- Combine with meteorological data using hybrid ML models (e.g., LSTM, GNN)
- [INSAT-3D Dataset – ISRO via IMD]
- PyTorch, EfficientNet, Django
- Kaggle Community


