Federated learning plays a significant role in utilizing multi-party data for public safety emergency detection while preserving privacy. However, due to the dispersion of public safety emergency data across different departments and the need for privacy protection, data islands have formed where data cannot be shared. Therefore, it is necessary to utilize federated learning to leverage multi-party data while protecting privacy. However, the heterogeneity of client data introduces noise during federated learning, which affects the performance of the global model. Thus, this paper proposes a federated public safety emergency detection method based on adaptive aggregation strategy (FedPSED). This method uses federated learning to train a public safety emergency detection model based on GAT and contrastive learning (PSED). When aggregating the global model, an adaptive aggregation strategy is employed to consider factors such as the topology of client data, enabling the FedPSED to utilize more high-quality data information and consider the performance of client models. This approach helps the global model parameters converge towards optimal parameters, enhancing the performance of the global model. Our proposed FedPSED has been shown to be effective through extensive experiments on multiple datasets.