C-ELTSNet: a novel network for the classification of UAV complex signal under low signal-to-noise ratio conditions

The widespread adoption of drones in complex scenarios has raised significant security and privacy concerns, highlighting the need for reliable drone monitoring and classification. However, research under low signal-to-noise ratio (SNR) and complex environmental conditions remains limited. Consequently, we propose a deep neural network specifically designed to address the drone classification problem in low SNR environments, named the Channel-Element Learnable Threshold Shrinking Network (C-ELTSNet). The network takes the real and imaginary components as two separate input channels and comprises four residual shrinking blocks (RSBs), each containing a multi-channel adaptive threshold module and a convolutional layer. The adaptive threshold module generates data-driven threshold maps and applies element-wise soft-thresholding to progressively filter out noise at each feature element, thereby enhancing feature quality. This design improves classification accuracy and robustness in low-SNR environments, while remaining more lightweight than other widely used methods. Furthermore, the micro-Doppler signature dataset of multi-rotor drones is simulated with the Martin–Mulgrew physical model, covering SNRs from −15 dB to 15 dB at 5 dB intervals. On the simulated dataset, C-ELTSNet achieves a peak classification accuracy of 97.9%. Across different SNR levels, its average accuracy reaches 82.5%, which is 2% higher than the other baseline models. On the real-world dataset, C-ELTSNet attains a mean accuracy of 92.5%. Experimental results demonstrate that the proposed algorithm achieves higher accuracy, stronger robustness, and lower computational complexity compared with several state-of-the-art baselines.