On the role of machine learning in satellite internet of things: A survey of techniques, challenges, and future directions

Abstract

The drive towards an interconnected world via satellites is reshaping the landscape of communication technolo-gies. This survey comprehensively reviews studies in the Satellite Internet of Things (SIoT) domain, focusingon the role of Machine Learning (ML) techniques. Indeed, the global data collection scale in SIoT is ideallysuited for data-intensive and sophisticated ML approaches. We highlight the innovative use of ML to addressspecific SIoT challenges, aiming to identify current trends, methodologies, and outcomes. We consideredtheoretical, practical, and experimental research, organizing existing publications into a new taxonomy thatintersects ML and SIoT categories. Our taxonomy reveals that Deep Learning (DL), Reinforcement Learning(RL), and Federated Learning (FL) are widely applied to address radio access schemes, resource and networkmanagement, and application-specific issues. This survey identifies critical gaps in current research on MLapplications in SIoT, such as the lack of differentiation between space-based and ground-based processing,insufficient integration of SIoT-specific metrics, and the oversight of limited computational resources onorbiting satellites. These issues raise concerns about the feasibility and efficiency of proposed solutions. Wepropose promising research directions based on the derived insights to effectively bridge the gap between MLresearchers and industrial SIoT entities.