Specific emitter identification (SEI) is crucial in satellite communication. However, due to limited spectrum resources and hardware constraints, deploying complex models remains challenging. This paper proposes a network pruning strategy based on Twin Delayed Deep Deterministic Policy Gradient (TD3) for satellite communication emitter identification. This pruning strategy enhances the stability and reliability of the deterministic gradient strategy by adding attention mechanism layers and feedback rewards, and restores the recognition performance of the pruned channel enhanced model. In the measured ADS-B dataset of 30 categories at 20dB, the model parameters were reduced by 85% while the accuracy reached 90%.