Aurora-FM

<div align="center"> <h1 style="text-align: center;">Aurora-FM: Foundation Model for All-Sky Auroral Imagery</h1>

A self-supervised foundation model for understanding auroral dynamics through all-sky imagery

</div>

---

🌌 Overview

Aurora-FM is a foundation model specifically designed for analyzing all-sky auroral imagery from the THEMIS all-sky imager network. Built on the SimCLR self-supervised learning framework, Aurora-FM learns rich representations of auroral phenomena without requiring extensive manual labeling, making it a powerful tool for space physics research and auroral morphology studies. Two versions of the model are provided: one that maps all-sky images to 512-dimensional vectors in a latent space where proximity corresponds to morphological or structural similarity, and another that has been finetuned on the manually labeled Oslo Auroral THEMIS (OATH) dataset (Clausen et al. 2018) to classify all-sky images into the six categories Arc, Diffuse, Discrete, Cloudy, Moon, and Clear/No Aurora.

Key Features

• 🎯 Optimized for THEMIS Data: Trained on THEMIS all-sky imager data
• 🔄 Transfer Learning Ready: Easily finetune for other all-sky imaging systems
• 🏆 State-of-the-Art Performance: When finetuned on labeled data, achieves SOTA results on auroral classification benchmarks
• 🔬 Morphology Detection: Identifies complex auroral structures including beads, omega bands, and more
• 📊 Self-Supervised: Trained using SimCLR, reducing dependence on labeled datasets

---

🚀 Quick Start

quickstart.py contains utilities for loading and classifying images using the finetuned model. The included Jupyter notebook demonstrates their use. The Python library THEMISim uses the model that has not been finetuned and demonstrates how the latent representations learned by Aurora-FM facilitate downstream tasks such as similarity search.

---

📊 Model Details

Architecture

Aurora-FM employs a ResNet-based encoder architecture trained with a variant of the SimCLR contrastive learning framework (Chen et al. 2021). This approach enables the model to learn meaningful representations of auroral phenomena through self-supervision.

Training Details:

• Primary Dataset: About 1m images taken during periods of high auroral activity from the THEMIS All-Sky Imager Network
• Training Framework: SimCLR (Simple Framework for Contrastive Learning of Visual Representations)
• Encoder Architecture: ResNet-18
• Input Resolution: 256×256 pixels
• Embedding Dimension: 512

The full details of the training process are documented in the JGR paper cited below.

---

🔬 Training Methodology

Aurora-FM was trained using the SimCLR framework, which learns representations by maximizing agreement between differently augmented views of the same image. This approach is particularly well-suited for auroral imagery due to:

1. Natural Variations: Captures the inherent variability in auroral displays
2. Invariance to Reflections: Handles reflections of images
3. Scale Robustness: Handles different spatial scales of auroral structures
4. Limited Labels: Achieves strong performance without extensive manual annotation

Data Augmentation Pipeline

• Random flips over vertical and/or horizontal axes
• Random cropping and resizing
• Color jittering (preserving aurora characteristics)

📚 Citation

If you use Aurora-FM in your research, please cite:

@article{https://doi.org/10.1029/2024JH000292,
    author = {Johnson, Jeremiah W. and Öztürk, Doğacan Su and Hampton, Donald and Connor, Hyunju K. and Blandin, Matthew and Keesee, Amy},
    title = {Automatic Detection and Classification of Aurora in THEMIS All-Sky Images},
    journal = {Journal of Geophysical Research: Machine Learning and Computation},
    volume = {1},
    number = {4},
    pages = {e2024JH000292},
    keywords = {aurora, annotation, machine learning, THEMIS, all-sky images},
    doi = {https://doi.org/10.1029/2024JH000292},
    url = {https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2024JH000292},
    eprint = {https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2024JH000292},
    note = {e2024JH000292 2024JH000292},
    year = {2024}
}

---

📝 License

This project is licensed under the Creative Commons 4.0 License - see the LICENSE file for details.

The THEMIS data has its own usage and attribution policy, please refer to https://themis.ssl.berkeley.edu/roadrules.shtml.

---

🙏 Acknowledgments

• NASA: This project is supported by NASA awards #80NSSC22K0642 and #80NSSC24K0237.
• THEMIS Mission: We acknowledge NASA contract NAS5-02099 and V. Angelopoulos for use of data from the THEMIS Mission. Specifically: S. Mende and E. Donovan for use of the ASI data, the CSA for logistical support in fielding and data retrieval from the GBO stations, and NSF for support of GIMNAST through grant AGS-1004736.
• SimCLR: Chen et al. 2021 for the foundational self-supervised learning approach.
• Space Physics Community: For valuable feedback.

---

<div align="center">

Built with ❤️ for the space physics community

</div>