Comparative Analysis of Machine Learning Methods for Asteroid Detection and Risk Assessment

We evaluated the performance of machine learning algorithms for near-Earth asteroid detection and assessed their risk potential. Our approach overcomes shortcomings in conventional detection methods by leveraging massive astronomical datasets to identify small or dark near-Earth asteroids. A full machine learning stack was designed, using six algorithms - Support Vector Machines (SVM), Decision Trees, Random Forests, Convolutional Neural Networks (CNNs), LightGBM, and Extreme Gradient boosting (XGBoost). This investigation analyzed hundreds of thousands of astronomical data points from numerous space-borne surveys, incorporating parameters such as orbital elements, size, albedo, and spectral classifications. This process included normalization and feature selection using Principal Component Analysis (PCA) Data. The models were trained on the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 0 \% - 2 0 \%}$</tex> data split (K-fold crossvalidation) and were validated with usual metrics (accuracy, precision-recall curves). XGBoost showed the best results (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5 \%}$</tex> accuracy & 94% precision for asteroid detection), followed by LightGBM (94% accuracy, 93% precision). CNNs got 91.5 % F1score and were best for image-based detection. Statistical analysis showed a significant improvement (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{p}&lt;0.01$</tex>) over classical detection strategies. A total number of previously undetected small and dark asteroids were found, and increased detection efficiency for non-elliptical objects. The results provide a strong foundation for the use of ML in planetary defense mechanisms. The improved ability to detect challenging targets, especially small and dark asteroids, allows for better early warning systems. This work lays the groundwork for developing real-time automated systems capable of predicting and responding to asteroid collisions, decreasing response times for hazard assessment and mitigation planning.