Using lossy compression in federated learning (FL) is an effective approach to address the high communication overhead in wireless networks and resource-constrained scenarios, especially amid the growing interest in large models. Current lossy compression schemes are mostly designed based on the sparsity characteristics of the FL transmission parameters, where sparsification is a crucial technique to ensure the sparsity of transmitted parameters. However, among the existing two sparsification techniques, the threshold-based sparsification faces the challenge of threshold selection, while top- \(k\) sparsification suffers from the issue that smaller parameters, due to temporal correlations, are unable to participate in aggregation for long periods. Furthermore, most of the existing compression schemes do not consider the problem of sparse data completion. Some works that propose compensation schemes encounter issues such as long compensation cycles, staleness, or cold start problems. In this paper, we propose an innovative sparsification mechanism that dynamically generates sparsification thresholds to assist in compressing local model updates at the client side. We also design a novel compensation scheme to recover sparse data, and through a carefully designed aggregation algorithm, we prevent aggregation update from being converge toward the threshold. Our approach reduces communication overhead while mitigating the model performance degradation caused by missing information, ultimately improving the convergence speed.