Semi-Supervised Federated Learning for Assessing Building Damage from Satellite Imagery

Accurate and timely building damage assessments are crucial for effective disaster response. However, traditional damage assessment methods heavily rely on manual evaluations by experts, which are labor-intensive and time-consuming. Recent research leverages machine learning (ML) and satellite remote sensing techniques to streamline the process. A major challenge of this method lies in the unlabeled nature of satellite imagery, which makes traditional ML frameworks impractical. Additionally, downloading the high-resolution satellite imagery for centralized ML is hindered by limited bandwidth and sporadic connectivity between the low Earth orbit (LEO) satellites and ground server. To address these challenges, we propose a novel semi-supervised federated learning framework named Semi-FedDA. It utilizes a small amount of labeled data on the ground server and a large amount of unlabeled data on the satellites to efficiently train a building assessment model without manual labeling. Moreover, this framework leverages intra-plane inter-satellite links (ISLs) to implement intra-orbit aggregations, which can significantly reduce the communication cost. We conduct extensive experiments on the real-world dataset. Numerical results show that our proposed framework can reduce training time by up to 94% compared with baselines, without sacrificing model accuracy.