Label-Efficient and Lightweight Spectrum Prediction for UAV-Based Spectrum Sensing: A Critical Review

Radio-spectrum prediction can support proactive channel verification, sensing scheduling, and access decisions in unmanned aerial vehicle (UAV) systems. However, existing evidence remains fragmented across UAV-oriented prediction, label-efficient learning, and deployment-oriented efficiency. This article presents a structured critical review of studies identified in IEEE Xplore, Scopus, and the Web of Science Core Collection from database inception to 24 August 2026. Studies were included when they evaluated the prediction of a future spectrum-related condition and were excluded when they addressed only current-state sensing, static spectrum mapping, UAV detection or classification, localization, or superseded study versions. The final corpus comprised 76 retained publications, including 63 original studies and 13 background references. The original studies included 14 direct UAV-related prediction studies, 38 transferable radio-spectrum studies, and 11 UAV-scenario studies. The review shows that transfer learning currently provides the most consistent support for reducing target-domain data requirements when related source bands, sensing stations, or radio environments are available. Self-supervised and other unlabeled-data methods are particularly relevant to UAV missions that can continuously collect spectrum traces but cannot obtain extensive labels, whereas meta-learning remains promising but lacks a standardized support-query evaluation protocol for UAV spectrum prediction. Generative augmentation can expand limited training data, but its effectiveness depends on whether the generated samples preserve the temporal, spectral, spatial, and propagation characteristics of the target environment. Lightweight architectures, online learning, model compression, knowledge distillation, FPGA implementation, and embedded execution provide complementary efficiency mechanisms, but their benefits should be distinguished from one another. Embedded prediction-related processing has been demonstrated on Raspberry Pi and software-defined-radio platforms; however, no end-to-end validation of a UAV-mounted predictor during flight was identified. Overall, the evidence suggests a conditional trade-off among target-domain data requirements, prediction generalization, adaptation cost, and deployment efficiency rather than a universal conflict between few-shot learning and lightweight models. The principal research gap is the limited joint validation of UAV-acquired data, target-domain adaptation, efficient inference, uncertainty-aware decision making, and onboard hardware under representative flight conditions.