Deterministic Sampled Diffusion Transformer With Two-Objective Optimization For DAS-VSP Data Progressive Recovery

Abstract Distributed acoustic sensing (DAS) arrays are capable of acquiring downhole geological information with high spatial-temporal resolution, thereby facilitating further advances in deep and high-precision seismic exploration. However, noise sources with varying characteristics degrade the signal-to-noise ratio (SNR) of DAS vertical seismic profiling (DAS-VSP) data, which directly influences subsequent inversion, imaging, and interpretation tasks. Currently, several deep denoising networks demonstrate superior performance compared to traditional methods. However, certain methods are limited by model performance, leading to issues such as missing signal axes and leakage of useful signal energy in the denoising results. Furthermore, these methods typically perform single-step denoising mapping, which makes the denoising process difficult to interpret.To address these challenges, we propose a deterministic sampling diffusion transformer (DSDiffformer), which combines the detail recovery ability of the diffusion model and the long-range information modeling capability of the transformer. First, a deterministic sampling strategy is applied to accurately recover signal details and accelerate the sampling efficiency of the reverse process. Second, a two-objective optimization strategy incorporates the sampling process into the optimization range to reduce the loss of useful signals. Additionally, our DSDiff-former is based on an iterative denoising approach and offers better interpretability of the denoising process compared to conventional networks. Experiments and comparisons were conducted on synthetic and field records to verify its effectiveness and superiority in multiple noise suppression and weak signal recovery.