Multimodal Federated Learning for Air Quality Estimation With Model Personalization Over Aerial–Ground Networks

Accurate and timely air quality estimations are crucial for effective environmental management and public health protection. Most literature works primarily focus on unimodal data-based analysis, i.e., relying exclusively on spatial or temporal data. However, unimodal learning methods cannot capture the complex interdependencies between different perspectives, leading to suboptimal performance in predicting air quality. Furthermore, recent multimodal data approaches rely on centralized data collection, raising privacy concerns and delays in air quality estimations. This paper addresses these practical challenges by proposing a novel multimodal federated learning (FL) framework with model personalization tailored for aerial-ground networks, where the term ‘aerial-ground networks’ refers to the collaboration between airborne drones and mobile ground sensors. The proposed personalized model training approach enables real-time air quality estimation by integrating data from drones and ground sensors, accounting for data heterogeneity with high precision. To achieve this, we develop a collaborative model training method in a multimodal setting, allowing drones and ground sensor devices to share encoder models and build a high-quality decoder at a ground station in real-time. The experimental results from real-world datasets show that our method outperforms existing state-of-the-art methods by providing around 2.46% better prediction in air quality estimations.