The Learning to Rank Method Based on Multi-Source Feature Fusion for High-Similarity Space Object Association in Optical Observation

With the continuous growth in the number and density of space objects, frequent occurrences of highly similar orbital parameters and spatial positions have led to performance bottlenecks for traditional dynamic association methods in high-similarity multi-candidate scenarios. To address this, this paper proposes a multi-source fusion feature association method based on a Learning to Rank (LTR) framework and a photometric prior information database. For the space object association task in high-similarity multi-candidate scenarios, this method reframes the association problem as an intra-group candidate ranking problem. Under the Pairwise paradigm, a ranking model is constructed using the XGBoost (Extreme Gradient Boosting) algorithm to explore the relative superiority relationships among candidate targets. Simultaneously, the photometric prior information of candidate targets is introduced, fusing orbital dynamics, photometry, observation geometry, and cross-features to construct a 45-dimensional feature space. Experimental results show that under the constructed high-similarity multi-candidate optical observation scenarios, the full-feature model achieved a Mean Reciprocal Rank (MRR) of 0.9738 and a Hit Ratio at 1 (HR@1) of 95.02% on the test set. Although photometric information occupied the top 5 positions in feature importance, when photometric features were disabled, the model’s HR@1 (94.24%) still significantly surpassed the traditional Mahalanobis distance association method (53.58%). Furthermore, when relying solely on photometric features, the model achieved an HR@1 of 78.35%. These comparative analyses validate the significant advantages of the Learning to Rank model and demonstrate the independent discriminative capability and potential for further refinement of photometric prior information in complex scenarios.