• MSFFNet integrates multi-scale feature extraction, adaptive feature screening, and multi-level feature fusion for robust spectrum sensing in low-SNR environments.
• At SNR = −14 dB, the method achieves a detection probability of 0.936 with a false alarm probability of only 0.1, outperforming existing approaches.
• By addressing CNN limitations in feature extraction and utilization under low SNR, MSFFNet enhances spectrum utilization in cognitive radio networks.
• A multi-level mixed-SNR dataset is constructed to emulate real communication environments, improving the generalizability and robustness of spectrum sensing.