<p>The utilization of Low Earth Orbit (LEO) satellites for Earth observation missions has gained significant traction with the rapid evolution of satellite network technologies. However, this approach is hindered by the inherent technical constraints and limited link bandwidth of LEO satellite systems, which restrict the efficient transmission of large-scale image data to ground stations for further processing. Additionally, the deployment of satellites across different epochs has led to considerable heterogeneity in their energy states and computational capacities, rendering them unsuitable for concurrently executing tasks that demand high energy consumption and extensive computational resources. These limitations present a formidable challenge to the effective implementation of machine learning algorithms in resource-constrained satellite environments. To address these issues, we propose an approach of cross-satellite data distillation and model collaboration for LEO satellite constellations via semi-federated learning, namely DDMC-SFL. Our approach incorporates a novel DP-KIP data distillation, which not only ensures differential privacy protection for the distilled data but also optimizes the utilization of distributed data and computational resources. By integrating the strengths of both centralized and federated learning paradigms, the proposed framework introduces a dual-class satellite categorization strategy: The first category of satellites is responsible for locally extracting distillation datasets and transmitting them to ground stations for model training, while the second category conducts local model training and uploads the resulting model parameters to ground stations. Extensive experiments have been conducted that include comparative analyzes and model training to validate the framework. The experimental results demonstrate that the proposed framework significantly outperforms existing methods in terms of improving model accuracy and accelerating convergence speed.</p>