🌍 Classifying Potentially Hazardous Asteroids

This project applies machine learning to classify potentially hazardous near-Earth objects (NEOs) using publicly available NASA data.
The goal is to explore whether physical and orbital characteristics of asteroids can help predict NASA’s hazardous label — and to understand the strengths and limitations of such models.

This project was completed as part of an individual take-home assignment in Machine Learning.

---

🚀 Motivation

I have always been fascinated by space, partly because it is beautiful, partly because it is chaotic, and partly because there are gigantic rocks zooming around the Solar System.

NASA tracks thousands of near-Earth objects, but telescope time and attention are limited.
This project explores whether machine learning can help prioritize which asteroids deserve closer inspection, using a simplified tabular dataset.

Spoiler: machine learning is cool, but planetary defense is very hard.

---

📊 Dataset

• Source: NASA Near-Earth Object (NEO) dataset (via Kaggle)
• Observations: ~90,000 asteroids
• Target variable: hazardous (binary)
• Features used include:
  • Estimated diameter
  • Relative velocity
  • Miss distance
  • Absolute magnitude (brightness)
  • Sentry flag

Note: Absolute magnitude uses a reverse scale --> lower values indicate brighter (and often larger) objects.

---

🧠 Models Used

• Logistic Regression

  • Baseline linear classifier
  • Feature scaling with StandardScaler
  • Hyperparameter tuning via grid search

• Random Forest Classifier

  • Nonlinear, tree-based model
  • Hyperparameter tuning for depth, splits, and number of trees
  • Best overall performance on the hazardous class

Evaluation focused on ROC-AUC, precision, recall, and confusion matrices, rather than accuracy alone, due to strong class imbalance.

---

📈 Key Findings

• The dataset is highly imbalanced (~10% hazardous).
• Logistic Regression achieves high accuracy but misses most hazardous asteroids.
• Random Forest significantly improves recall and AUC for hazardous objects.
• Absolute magnitude (brightness) is the strongest single predictor.
• The model’s limitations highlight how much more information NASA uses in reality (orbital simulations, uncertainty propagation, repeated observations).

---

⚠️ Important Disclaimer

This project does not attempt to predict real asteroid impacts.

The models only try to reproduce NASA’s existing labels using a small set of features.
NASA’s real hazard assessment involves complex physics, long-term trajectory simulations, and continuous updates — far beyond the scope of this dataset.

---

🛠️ How to Run

1. Clone the repository
2. (Optional) Install dependencies:

   pip install -r requirements.txt

3. Open the notebook: ML_FinalProject_Anna_NASA.ipynb