Real-Time Incremental Federated Learning for Weather Prediction with UAV based Data Collection

This article presents a comprehensive framework for real-time incremental federated learning aimed at enhancing weather prediction accuracy through UAV-based data collection. In our proposed system, UAVs are deployed across multiple regions within cities to gather diverse and comprehensive weather data. The collected data is streamed in real-time to central city servers, which aggregate the data and transmit it to a state-level server. At the state server, data from multiple cities are combined, and models are incrementally trained on newly received data in a step-wise manner. Each incremental step involves training on 80% of the new data and testing on the remaining 20%. Our key contribution is the integration of a weighted federated learning approach, where periodic federated learning steps utilize the weighted FedAvg algorithm. The weights are determined based on the accuracy of each model, allowing for more effective aggregation of model updates. Experimental results demonstrate a significant improvement in prediction accuracy, achieving a 24% increase compared to baseline independent machine learning models trained separately at the state level.