MaRNet: A UAV Radio Frequency Recognition Method Combining Mamba With ResNet

In the paper, we investigate the radio frequency (RF) signal detection of an unmanned aerial vehicle (UAV) under low signal-to-noise ratio (SNR) that combines a Mamba module with the ResNet framework, namely MaRNet. First, short-time fourier transform (STFT) is employed to obtain time-frequency spectrograms, and the elbow method is utilized to determine the optimal sliding window length—balancing short-term local and overall spectral features to provide more discriminative inputs. Next, the network fully exploits ResNet's stable training in deep architectures and efficient feature extraction capabilities, while the Mamba module's multi-level state space model (SSM) and feature fusion mechanism significantly enhance the capture of long-range dependencies and latent correlations in complex time-frequency features. Consequently, the proposed method can shorten per-sample processing time and greatly improve detection accuracy. Finally, an efficiency score is introduced to jointly consider both accuracy and inference time, enabling a quantitative evaluation of various windowing strategies. Experimental results show that under low-SNR conditions, the proposed method can achieve an F1-score of 98.2%, an accuracy of 99.1%, both recall and mean average precision of 96.7%, and it also demonstrates strong robustness in recognizing 9,000 unlabeled STFT spectrograms—outperforming other comparison models overall.