The paper introduces SATSense, a multi-satellite collaborative framework for spectrum sensing in Low Earth Orbit (LEO) satellite networks. The authors address the challenges of dynamic spectrum sharing and accurate spectrum sensing in the presence of variable channel conditions and limited spectrum resources. Key contributions include: 1. **Graph Learning-based Collaborative Algorithm**: SATSense models the sensing data from multiple satellites as a graph to characterize the correlations among RF channels, using graph neural networks (GNNs) to improve the effectiveness of model training and mitigate the impact of channel heterogeneity. 2. **Hybrid Data Compression**: A hybrid data compression method based on sub-Nyquist sampling and autoencoders (AEs) is proposed to reduce the amount of transmitted sensing data. Sub-Nyquist sampling reduces the burden on the ADC, while AEs further compress the data, ensuring timely transmission to the ground station (GS). 3. **Contrastive Learning for Packet Loss**: A contrastive autoencoder (CAE) is integrated into the AE structure to compensate for packet loss during data transmission. CAE minimizes the distance between embeddings with packet loss and ideal embeddings, enhancing the reliability of data analysis at the GS. The paper evaluates the performance of SATSense through extensive experiments, demonstrating its superior accuracy and efficiency compared to conventional deep learning algorithms in spectrum sensing. The results show that SATSense outperforms the baseline deep cooperative sensing (DCS) algorithm, especially under low SNR and high packet loss rates. The framework also demonstrates robustness against packet loss and the ability to handle a large number of collaborative satellites, making it suitable for complex satellite communication scenarios.The paper introduces SATSense, a multi-satellite collaborative framework for spectrum sensing in Low Earth Orbit (LEO) satellite networks. The authors address the challenges of dynamic spectrum sharing and accurate spectrum sensing in the presence of variable channel conditions and limited spectrum resources. Key contributions include: 1. **Graph Learning-based Collaborative Algorithm**: SATSense models the sensing data from multiple satellites as a graph to characterize the correlations among RF channels, using graph neural networks (GNNs) to improve the effectiveness of model training and mitigate the impact of channel heterogeneity. 2. **Hybrid Data Compression**: A hybrid data compression method based on sub-Nyquist sampling and autoencoders (AEs) is proposed to reduce the amount of transmitted sensing data. Sub-Nyquist sampling reduces the burden on the ADC, while AEs further compress the data, ensuring timely transmission to the ground station (GS). 3. **Contrastive Learning for Packet Loss**: A contrastive autoencoder (CAE) is integrated into the AE structure to compensate for packet loss during data transmission. CAE minimizes the distance between embeddings with packet loss and ideal embeddings, enhancing the reliability of data analysis at the GS. The paper evaluates the performance of SATSense through extensive experiments, demonstrating its superior accuracy and efficiency compared to conventional deep learning algorithms in spectrum sensing. The results show that SATSense outperforms the baseline deep cooperative sensing (DCS) algorithm, especially under low SNR and high packet loss rates. The framework also demonstrates robustness against packet loss and the ability to handle a large number of collaborative satellites, making it suitable for complex satellite communication scenarios.