Federated Broad Learning for Unmanned Aerial Vehicle Clusters in Water Monitoring

The swift advancement of Unmanned Aerial Vehicles (UAVs) has garnered considerable interest in their use for water monitoring. However, these applications remain challenging, especially for the task of UAV water surface contaminant monitoring due to the variability in target sizes and the frequent occurrence of contaminants in different morphologies and patterns than before. In addition, for UAV clusters, there is an urgent need for a method to optimize collaborative learning strategies among UAVs to fully utilize cluster data and knowledge. This paper proposes a UAV-based Federated Broad Learning (UFBL) scheme in water surface contaminant monitoring. First, we introduce a Pyramid Pooling Module (PPM) to achieve multi-size feature extraction. Subsequently, to address the problem of variable morphologies and patterns of emerging contaminants, an incremental training mode by increasing the data is proposed. It improves the model’s performance for new data by adding data to the training. Additionally, this paper proposes a BL-oriented Federated Learning (FL) architecture to leverage the collaborative potential of UAV clusters. The architecture unifies the random parameters across individual UAV BL networks, thereby assigning a shared value to the final model weight. Moreover, to address discrepancies in training progress and target characteristics among UAVs, this paper proposes an adaptive weight filter module. It can dynamically select and combine model weights based on performance metrics, therefore improving knowledge transfer among UAVs. Simulation experiments demonstrate that the proposed improvements and architecture significantly enhance training efficiency and monitoring accuracy in UAV water surface contaminant monitoring tasks, reducing training time by approximately 25.2% to 56.7%.