Fine-grained space object classification with Convolution-Boosted LSTM using light curves: A new method and a large scale dataset

Ground-based light curves provide a cost-effective means of monitoring and identifying space objects, playing a critical role in space situational awareness. However, fine-grained classification of space objects from light curves remains challenging due to significant intra-class variability, high inter-class similarity, and severe class imbalance. To address these challenges, we introduce SOLID-50, a large-scale real-world light curve dataset comprising 104,034 samples from 50 subcategories, covering satellites, rocket bodies, and space debris. Further, we propose Convolution-Boosted LSTM (CoBo-LSTM), a deep learning framework that enhances temporal modeling with diversified convolutional feature extraction. CoBo-LSTM integrates three key innovations, including a set of diversified convolutional kernels to capture multi-scale photometric patterns, four heterogeneous pooling operators that extract both global structures and local variations, and a hybrid loss function to enhance robustness against class imbalance and improve feature discriminability. Experimental results on SOLID-50 demonstrate that our proposed CoBo-LSTM achieves a macro-averaged precision of 84.49% on coarse-grained classification and 44.63% on fine-grained classification, significantly outperforming state-of-the-art methods, such as ResNet, RLNet and LC-VAE.