In military operations, unmanned aerial vehicle (UAV) swarms equipped with electromagnetic sensors and computing devices can easily capture enemy spectrum data, enabling the training of deep neural network models to predict enemy frequency-hopping sequences. Traditional methods typically rely on single nodes for data sensing and model training. However, the limited onboard resources of UAVs lead to extended model training times, and the sense-then-train approach exacerbates the overall task duration, severely impacting real-time requirements on the battlefield. To address these issues, this paper proposes a Multi-UAV Distributed Incremental Learning (MDIL) method. This method establishes parallel sensing and training task flows, continuously collects data, and updates the model. By leveraging the distributed data sensing and incremental learning of UAV swarms and considering the heterogeneity of UAVs, this method optimizes the sensing sequence and workload, enhances real-time responsiveness, and reduces sensing and training times. Experimental results show that this method effectively reduces task training time by 22.13% and overall time by 36.89%, albeit at the expense of some accuracy.