Fed-RSSC: A Semi-Decentralized Federated Framework for Remote Sensing Scene Classification

Abstract

High-resolution remote sensing (HRRS) scene classification is critical in various applications. HRRS data usually contain sensitive geographical and environmental information, such as the locations of critical assets, urban planning details, and military bases. To safeguard such private information, local governments have implemented regulations and policies that govern the sharing of HRRS data. However, existing scene classification methods typically rely on centralized training and assume that data are directly shared with a centralized server, posing significant privacy risks. To address these concerns, we propose federated remote sensing scene classification (Fed-RSSC), a novel framework enabling the collaborative training of a joint model while ensuring data remains localized. We further demonstrate that federated learning (FL) is an effective approach to tackling privacy issues in HRRS scene classification. Moreover, to reduce high communication overhead, Fed-RSSC, a semi-decentralized architecture, is designed with a local consensus aggregation (LCA) strategy based on device-to-device (D2D) communication. Consequently, Fed-RSSC significantly reduces reliance on direct communication between the server and clients, thereby enhancing both communication efficiency and scalability. Extensive experiments on the NWPU-RESISC45, AID, and UC-Merced datasets validate the effectiveness and scalability of Fed-RSSC, demonstrating its superiority in scene classification.