SATSense is a multi-satellite collaborative framework for spectrum sensing, designed to address the challenges of dynamic spectrum sharing in Low Earth Orbit (LEO) satellite networks. With the deployment of large-scale LEO satellite constellations, such as Starlink and Kuiper, the competition for limited spectrum resources between terrestrial and non-terrestrial networks has intensified. Traditional single-satellite spectrum sensing is insufficient due to variable channel conditions, leading to unstable performance. To overcome these challenges, SATSense employs a collaborative approach that leverages data from multiple satellites to enhance spectrum sensing accuracy. The framework first constructs a graph based on sensing data from each satellite to characterize correlations among RF channels. Graph learning is then used to explore these correlations, improving the effectiveness of model training. To reduce data volume, SATSense introduces a hybrid data compression method combining sub-Nyquist sampling and autoencoder (AE). This approach reduces the amount of data transmitted to the ground station (GS) for fusion. Additionally, contrastive learning is integrated into the AE to compensate for missing packets, mitigating the impact of packet loss on spectrum sensing. Extensive experiments demonstrate that SATSense achieves efficient spectrum sensing performance and outperforms conventional deep learning algorithms in accuracy. The framework addresses three main challenges: (1) heterogeneity in RF channels, (2) the large volume of data generated by wideband sampling, and (3) packet loss during satellite-to-ground transmission. SATSense's graph-based collaborative algorithm, hybrid data compression, and contrastive learning-based packet loss compensation enable accurate and reliable spectrum sensing in dynamic environments. The framework's contributions include the first spectrum sensing framework that jointly considers satellite data downlink and data fusion in GS, the use of graph learning to handle RF channel heterogeneity, and a hybrid data compression method to ensure timely transmission of sensing data. The integration of contrastive learning into AE further enhances the framework's robustness against packet loss. Overall, SATSense provides a comprehensive solution for efficient spectrum sensing in LEO satellite networks.SATSense is a multi-satellite collaborative framework for spectrum sensing, designed to address the challenges of dynamic spectrum sharing in Low Earth Orbit (LEO) satellite networks. With the deployment of large-scale LEO satellite constellations, such as Starlink and Kuiper, the competition for limited spectrum resources between terrestrial and non-terrestrial networks has intensified. Traditional single-satellite spectrum sensing is insufficient due to variable channel conditions, leading to unstable performance. To overcome these challenges, SATSense employs a collaborative approach that leverages data from multiple satellites to enhance spectrum sensing accuracy. The framework first constructs a graph based on sensing data from each satellite to characterize correlations among RF channels. Graph learning is then used to explore these correlations, improving the effectiveness of model training. To reduce data volume, SATSense introduces a hybrid data compression method combining sub-Nyquist sampling and autoencoder (AE). This approach reduces the amount of data transmitted to the ground station (GS) for fusion. Additionally, contrastive learning is integrated into the AE to compensate for missing packets, mitigating the impact of packet loss on spectrum sensing. Extensive experiments demonstrate that SATSense achieves efficient spectrum sensing performance and outperforms conventional deep learning algorithms in accuracy. The framework addresses three main challenges: (1) heterogeneity in RF channels, (2) the large volume of data generated by wideband sampling, and (3) packet loss during satellite-to-ground transmission. SATSense's graph-based collaborative algorithm, hybrid data compression, and contrastive learning-based packet loss compensation enable accurate and reliable spectrum sensing in dynamic environments. The framework's contributions include the first spectrum sensing framework that jointly considers satellite data downlink and data fusion in GS, the use of graph learning to handle RF channel heterogeneity, and a hybrid data compression method to ensure timely transmission of sensing data. The integration of contrastive learning into AE further enhances the framework's robustness against packet loss. Overall, SATSense provides a comprehensive solution for efficient spectrum sensing in LEO satellite networks.