Enhanced Pseudo-Labeling and Dual Mixup Augmentation for Semi-Supervised RF-Based UAV Recognition

Abstract

Unmanned aerial vehicle (UAV) recognition has become increasingly important for securing the low-altitude economy. Radio frequency (RF) signal analysis using deep learning (DL) has emerged as a promising approach to this task. However, training robust DL models relies on large-scale labeled datasets, which are costly to acquire in practical scenarios. While semi-supervised learning (SSL) can alleviate this bottleneck by leveraging abundant unlabeled data, existing methods typically discard low-confidence samples, limiting further performance improvements. To address these issues, we propose the enhanced Pseudo-Labeling and Dual Mixup (PLDM) augmentation framework for RF-based UAV recognition. Specifically, PLDM utilizes high-confidence samples to smooth decision boundaries and lowconfidence samples to safely exploit the latent information of uncertain RF signals. Experiments on a real-world UAV dataset demonstrate that PLDM achieves 78.38% accuracy with only 5% labeled data, substantially outperforming existing methods. The results confirm the practical potential of PLDM for robust UAV recognition under label-scarce conditions.