On Smart IoT Remote Sensing over Integrated Terrestrial-Aerial-Space Networks: An Asynchronous Federated Learning Approach

Abstract

While the Internet of Things (IoT), coupled with integrated terrestrial-aerial-space networks, has revolutionized the domain of ubiquitous remote sensing for natural resource management, several research challenges emerge due to the explosion of the collected IoT data. For instance, the edge nodes in these hybrid, next-generation networks are anticipated to carry out edge computing on the collected data to provide localized computing to enable early warning systems, including forest fire occurrence and spread, earth- quake wave detection, tsunami forecasting, and so forth. While edge computing can significantly reduce the high communication time required in the traditional cloud-based remote sensing analytics, it is important to develop a lightweight training framework to obtain smart remote sensing analytics at the edge devices in the integrated network while preserving the privacy of the collected data. In this article, we propose an asynchronously updating federated learning model for the edge nodes to build local artificial intelligence models for smart remote sensing with a forest fire detection use case without the need for explicit data exchange with the cloud. This jointly preserves data privacy and also alleviates the network overhead. Extensive experimental results demonstrate the viability of our proposal in terms of significantly high remote sensing accuracy, low convergence time, and low bandwidth overhead compared to existing methods.