Robust damped multichannel singular spectrum analysis with adaptive correction

Abstract

Seismic data denoising and signal separation are critical for downstream processing tasks such as amplitude variation with offset (AVO) analysis and inversion. Multichannel singular spectrum analysis (MSSA) is a widely adopted rank-reduction technique for this purpose; however, its performance is notoriously sensitive to parameter selection. Standard MSSA relies on hard rank truncation, where sub-optimal rank selection leads to either severe signal leakage or residual noise artifacts. While damped MSSA improves stability, it introduces amplitude bias that compromises signal fidelity. To address these limitations, we propose a robust damped MSSA (RDMSSA) with adaptive correction, a two-stage framework designed to be inherently parameter-tolerant. First, we employ RDMSSA to estimate the signal subspace. By intentionally using conservative damping parameters, we prioritize the suppression of random noise and artifacts, accepting a degree of signal leakage to ensure stability. Second, we introduce an adaptive correction step that treats the residual as a leakage reservoir. Using a non-stationary least-squares adaptive filter, coherent signal energy is extracted from the residual and restored to the result. This “under-fit and repair” strategy significantly relaxes the requirement for precise parameter fine-tuning. Numerical experiments on synthetic and field data demonstrate that the proposed method achieves superior separation results compared to traditional methods. Crucially, we show that our approach maintains high signal-to-noise ratios even when initialized with sub-optimal rank, damping or windowing parameters, offering a robust and efficient workflow for industrial seismic processing.