With the rapid deployment of ground-based optical surveillance networks, the volume of short-arc optical data has surged exponentially. Cataloging space objects utilizing these short arcs constitutes a fundamental pillar of Space Situational Awareness (SSA), serving as a necessary prerequisite for critical downstream tasks such as conjunction assessment and anomaly detection. The cataloging process generally encompasses the initialization of new objects and the maintenance of existing orbits across Low Earth Orbit (LEO), Medium Earth Orbit (MEO), Geosynchronous Earth Orbit (GEO), and High Elliptical Orbit (HEO). Faced with a massive influx of new arcs, the ability to rapidly and accurately classify the orbital regime of these objects without prior information represents a critical bottleneck for conserving computational resources. Moreover, conventional empirical workflows may generate spurious track associations that subsequently pass through least-squares Precision Orbit Determination (POD) undetected, leading to catalog corruption. In parallel, given that observation arcs inherently constitute time-series data, the rapid advancement of deep learning sequence models, particularly Transformers and the emerging Mamba architecture, offers revolutionary paradigms for extracting complex dynamic patterns. Leveraging these innovations, this work introduces intelligent framework for orbit regime classification and track association, representing a synthesis of data-driven approaches and orbital dynamics.

First, this study implements an end-to-end deep learning model, specifically a Transformer Encoder, that takes optical arcs directly as input to output orbital regimes. This approach circumvents traditional dependencies on angular rate thresholding or the necessity of performing Initial Orbit Determination (IOD) for type identification, achieving an overall classification precision of 96.60%. Specifically, leveraging high-fidelity simulations of a global network of over ten stations tracking hundreds of real-world objects, we generate a dataset containing over 100,000 noisy optical arcs ranging from 30 to 600 seconds uniformly distributed across all orbit regimes. We apply functional smoothing and resampling to transform raw arcs into fixed-length multi-channel sequences, integrating time, angles and derivatives up to the second order, and a goodness-of-fit metric indicative of noise levels to enrich kinematic representation. We construct a Transformer-based model utilizing multi-head self-attention to capture temporal dependencies to execute the classification task. Rigorous evaluation on nearly 20,000 samples yields precision rates for LEO, MEO, GEO, and HEO of 93.63%, 98.21%, 97.66%, and 97.01%, respectively. The corresponding recall rates are 97.68%, 99.60%, 99.96%, and 89.14%. Furthermore, to circumvent the quadratic complexity of Transformers, we adopt the Mamba architecture built upon State Space Models (SSM). Its linear-time complexity and continuous-time formulation naturally align with the differential equations governing orbital dynamics, ensuring robust feature extraction even under non-uniform sampling.

Second, to mitigate spurious track associations, we construct a hybrid physics-aware classifier via an ensemble learning framework, strategically fusing the robustness of XGBoost with frontier deep learning paradigms. Unlike traditional methods such as the Admissible Region or Covariance-Based Track Association (CBTA), we construct a high-dimensional feature space by extracting metrics from the POD results of combined arcs, including solved orbital states, post-fit residual statistics, and iteration convergence metrics. Simulation in dense LEO scenarios (over 70,000 samples) demonstrates that the model achieves an accuracy exceeding 90% in assessing association validity, with an Area Under the Curve (AUC) of 0.97, effectively minimizing the risk of catalog corruption. Moreover, this classifier is demonstrated to be inherently scalable to associations between Uncorrelated Tracks (UCT). Crucially, this framework provides the verified metric backbone for multi-track association, enabling global association of pre-classified candidates via graph clustering algorithms.