Abstract
Estimating the shape of space objects helps infer key characteristics such as object type, mass, and potential operational status, thereby providing critical decision-making support for space threat assessment. This paper proposes a Physics-Informed Transformer deep learning model for space object shape classification based on photometric time series data. The model achieves a classification accuracy of 89.22% on a hybrid dataset combining simulated and measured data, outperforming eight classical models with an average accuracy improvement of 3.61%. Ablation experiments demonstrate that the introduction of physical gating yields an average accuracy improvement of 1.43%. Using evaluation metrics including the confusion matrix, PR curve, ROC curve, and confidence distribution histogram, we demonstrate that the model possesses high accuracy, strong robustness, and good interpretability.