Physics-aware latent neural ODEs with multi-task gradient surgery for cross-domain light curve classification of resident space objects

This is the authors' abstract. We don't add key points for this paper.

Accurate classification of resident space objects (RSOs) from ground-based photometric light curves is fundamental to space situational awareness(SSA). However, this process remains hampered by sparse, non-uniform optical sampling and the pronounced domain shift between simulated and empirical data. To address these challenges, we present a unified framework coupling a physics-aware latent Neural Ordinary Differential Equation (Phy-ODE) with cross-domain multi-task adversarial adaptation to identify five RSO categories. Unlike traditional domain adaptation methods that distort under severe sparsity, our Phy-ODE models light curve evolution as a continuous-time latent process driven by rigid-body rotational dynamics. This integration enables reliable continuous-curve reconstruction and robust feature extraction, even across extended observational gaps. Building upon this, we construct a cross-domain multi-task adversarial network comprising a domain discriminator, a label classifier, and a physics-assisted regression branch. Through domain adversarial learning, our approach minimizes the feature distribution discrepancy between the simulation and observation domains, while utilizing conditional entropy constraints to enhance the discriminative power of target domain predictions. Furthermore, to eliminate gradient conflicts between the primary classification task and the auxiliary physical task, we employ the Projecting Conflicting Gradients (PCGrad) strategy, effectively mitigating negative transfer. Extensive experiments are conducted on a dual-modal dataset containing 50,000 simulated sequences and measured observations from the Mini-Mega TORTORA(MMT-9) system. Results demonstrate that our method reliably restores specular glares and non-stationary variations under extreme sparsity. In a fully unsupervised domain adaptation setting, the proposed framework achieves a classification accuracy of 92.6% and an F1 score of 92.2% across the five RSO categories, significantly outperforming baseline methods such as LSP+RF, Phased-LSTM, and Vanilla Neural ODE. Ultimately, the integration of continuous-time modeling, physics-based priors, and cross-domain multi-task optimization emerges as a highly effective and robust approach for the recognition of non-cooperative space objects.