Spectrum Surveying for Machine Learning-Assisted UAV Communications

Abstract

The use of unmanned aerial vehicles (UAVs) has become widespread, spanning applications from surveillance to wireless communications infrastructure. Remarkably, it has been proposed that UAVs can be utilized as aerial base stations (ABSs) for providing wireless connectivity in areas with poor coverage. Radio maps, which provide channel metrics like received signal strength (RSS) across a geographic area, are instrumental to this and other applications. In many setups, it is convenient to construct a radio map by collecting measurements at various locations using a mobile robot, such as a UAV. The development of such a technology, here referred to as spectrum surveying, requires addressing several challenges. Specifically, current methods rely on predetermined measurement locations, which results in long surveying times. Besides, the current understanding of the radio map estimation (RME) problem from a theoretical and practical standpoint is still highly limited. For this reason, the contributions of this thesis span three main aspects: efficient algorithms for active sensing, theoretical analysis of the RME problem, and empirical evaluation of RME algorithms using real data collected in spectrum surveying missions. In the realm of active sensing, a scheme is devised to determine the trajectory of a UAV based on past measurements to efficiently estimate radio maps. This involves the development of novel approaches for joint RME and uncertainty mapping. To this end, an online Bayesian learning algorithm and a data-driven uncertainty mapping technique based on deep neural networks (DNNs) are proposed. Additionally, trajectory planning algorithms are designed to guide UAVs towards the most informative measurement locations, optimizing the efficiency of spectrum surveying missions. Theoretical analysis sheds light on the complexity of RME by studying the spatial variability of power maps and evaluating the performance of various interpolation algorithms. Insights gleaned from theoretical analysis inform the development and refinement of practical RME (and, therefore, spectrum surveying) methodologies. Empirical investigation involves the collection and analysis of large-scale measurement datasets using a spectrum surveying platform developed in this work. This real-world data is utilized to evaluate the performance of existing RME algorithms, providing valuable insights into their strengths and limitations. A comprehensive comparison of traditional estimators and DNN-based approaches reveals the potential for hybrid estimators that combine the advantages of both paradigms.