Seismic random noise degrades data quality and obscures reflection events critical for exploration interpretation. Existing convolutional neural network-based denoisers, including ADNet, suffer from limited receptive fields and single-scale attention, causing poor long-range reflection continuity and signal leakage at low signal-to-noise ratio (SNR). To address these challenges, we propose multi-feature-enhanced (MFE)-ADNet, a novel framework that integrates a pyramid spatial attention (PSA) module into ADNet. To the best of our knowledge, this is the first application of PSA for seismic random noise attenuation. The PSA module enables multi-scale spatial feature extraction and captures long-range channel dependencies, complementing ADNet’s attention-guided denoising mechanism to preserve weak reflections while suppressing incoherent noise. Experimental results on synthetic and field datasets demonstrated that MFE-ADNet gains of 12–15 dB relative to the original noisy data. Compared with ADNet, it provides an additional 3 dB SNR improvement and reduces the mean squared error by 0.0015, indicating substantially lower residual signal energy and reduced signal leakage. The method also attained a local similarity of 0.86–0.98, outperforming wavelet denoising, time–frequency peak filtering, and ADNet by substantial margins. Residual analysis confirms minimal signal leakage, validating the method’s reliability for practical seismic processing and subsequent geological interpretation.