Using Machine Learning to Optimize Near-Earth Object Sightings Data at the Golden Ears Observatory

This research project focuses on improving Near-Earth Object (NEO) detection using advanced machine learning techniques, particularly Vision Transformers (ViTs). The study addresses challenges such as noise, limited data, and class imbalance. The ViT model, initially designed for natural language tasks, was adapted for image processing to effectively capture complex patterns and relationships in astronomical data.   The methodology involved preparing a curated dataset of NEO images, resizing them to 128x128 pixels, and organizing them into triplet sequences. ViTs processed these sequences, leveraging self-attention and feed-forward neural networks to distinguish NEOs from other objects. Multiple learning rates and batch sizes were tested, revealing the optimal combinations for stability and accuracy.   The results revealed distinct behaviors associated with varying learning rates. Notably, the learning rate of 0.001 consistently demonstrated stable convergence and high accuracy across different batch sizes. In contrast, a learning rate of 0.01 exhibited significant fluctuations in the loss function, indicating challenges in training stability. Conversely, a learning rate of 0.0001 showcased relatively low and consistent loss values during training. These insights highlight the potential of the ViT model in enhancing NEO detection by capturing temporal and spatial patterns effectively. Furthermore, the study emphasizes the significance of larger and diverse datasets, addressing class imbalances, and enhancing model transparency for guiding future research.   In conclusion, ViTs have shown promise in enhancing NEO detection and classification, offering insights into celestial object dynamics and contributing to planetary defense efforts. Collaboration, data sharing, and fine-tuning model parameters have led to improved accuracy and understanding in this field.