<p>The intrinsic absorption of seismic waves in viscoelastic media attenuates high-frequency components, thereby reducing resolution in thin-layer structures. Although deconvolution is the most common approach to resolution enhancement, conventional algorithms amplify both signal and noise, degrading the signal-to-noise ratio (SNR). To address this, we introduce a structure-constrained method for high-frequency recovery in the frequency–space (f–x) domain. First, we exploit the statistical predictability of seismic records in the f–x domain by fitting an autoregressive–moving-average (ARMA) model, from which we derive a prediction operator and its associated prediction-error filter (PEF). This operator characterizes spatial coherence in seismic signals and distinguishes seismic signals from random noise. We then embed this operator as a structural constraint within a multichannel sparse-inversion framework. By enforcing spatially adaptive regularization guided by predicted structural trends, our approach suppresses noise amplification during high-frequency recovery and enhances the fidelity of the reconstructed signal. Extensive tests on both synthetic and field datasets demonstrate that the proposed method significantly improves seismic resolution while preserving a high SNR.</p>