<p>Distributed acoustic sensing (DAS) has emerged as a transformative technology in seismic exploration, offering significant advantages in cost efficiency, deployment scalability, and high-density spatial sampling. However, the practical interpretation of DAS data is severely hindered by complex and persistent background noise such as optical, fading, horizontal, and checkerboard noise, which arises from both environmental factors and the fundamental physical principles of DAS measurement. While many attention mechanism networks like MIRNet have shown promise in natural image restoration and possess a strong capacity for characterizing diverse noise patterns, their typical encoder–decoder architecture often leads to the loss of shallow-level features, resulting in compromised continuity and clarity of deep reflection events in DAS data. To address this limitation, we propose SFA-PromptIR for DAS denoising that integrates a step-by-step feature fusion (SFA) mechanism with channel attention. The SFA structure progressively aggregates multi-scale features throughout the encoding and decoding stages, thereby preserving critical fine-scale information and preventing the fading of weak signals. Simultaneously, the channel attention mechanism enhances learning precision by adaptively weighting feature channels, enabling the model to focus on salient structural information and improve recovery detail. Extensive experiments on both synthetic and field DAS datasets demonstrate that SFA-PromptIR effectively suppresses a wide range of background noise while faithfully recovering useful seismic signals. The method significantly enhances the signal-to-noise ratio and structural coherence of reflection events, providing a reliable foundation for subsequent seismic interpretation and geological analysis.</p>