Planetary Defense: AI for Asteroid Hazard Classification

This Master's thesis, completed for the MSc in Artificial Intelligence at Universidad Internacional de La Rioja (UNIR) and developed within the framework of the European-funded ADELA Project (B0036-2223-624-ADELA: Applications of Deep Learning for Astrophysics), investigates the application of advanced Machine Learning and Deep Learning techniques to classify asteroids based on their potential threat to Earth. The research involved a comprehensive evaluation of 14 different models, ranging from traditional algorithms to cutting-edge ensemble, boosting, and neural network architectures, using a NASA dataset of Near-Earth Objects (NEOs).

The primary goal was to optimize the identification and classification of potentially hazardous asteroids (PHAs), with a critical focus on minimizing false negatives – failing to identify a genuine threat. The study rigorously compared models based on metrics such as AUC-ROC, F1-Score, precision, recall, and computational efficiency, considering their suitability for early warning systems.

Key contributions include an exhaustive comparative analysis, insights into feature importance for asteroid hazard prediction, and a strong foundation for future research in planetary defense systems. Gradient Boosting, XGBoost, and MLP with CUDA emerged as top performers.