Abstract: High-quality denoising is crucial for refining images of adverse geological bodies in tunnel seismic surveys. However, traditional methods relying on fixed thresholds or frequency-domain filtering yield suboptimal denoising results, while machine learning algorithms are constrained by limited generalization capability and real-time performance, rendering them inadequate for meeting engineering requirements. To address these issues, a joint optimization denoising model was constructed in this study to directionally suppress residual noise. Building upon the framework of compressed sensing, a joint optimization mechanism which couples signal reconstruction with noise suppression was established by incoporating the deterministic prior features inherent in the source scanning signals. Comprehensive evaluations involving multidimensional quantitative metrics and spectral analysis demonstrate that the proposed method consistently outperforms traditional filtering, wavelet transform, curvelet transform, and total variation (TV) regularization techniques in terms of signal-to-noise ratio (SNR), structural similarity index (SSIM), and root mean square error (RMSE). The method shows clear advantages in structural fidelity preservation in complex geological environments and noise suppression efficacy. It exhibits particular strength in structural edge sharpening, thin-layer signal recovery, and artifact suppression—key factors in improving the identification accuracy of adverse geological bodies. Finally, geophysical prospecting was conducted in a subway tunnel in Jiangsu Province, where karst cavities as adverse geological bodies were simulated ahead of the tunnel face. Seismic data processed by the proposed method yield high-quality imaging results of the tunnel structure, outperforming conventional imaging methods. Such results further verify that the proposed method can significantly enhance the imaging resolution of adverse geological structures ahead.