Vectorized seismic random noise attenuation based on an unsupervised Monte Carlo deep learning framework

Seismic data acquisition is inevitably affected by various factors such as environmental conditions and equipment limitations, resulting in the presence of random noise in the data. Effective suppression of random noise to improve the signal-to-noise ratio is crucial to ensure high-quality subsequent inversion and interpretation. However, maintaining signal fidelity while denoising remains a significant challenge. To address this, we propose a seismic random noise attenuation method based on an unsupervised Monte Carlo deep learning framework with zero-lag cross-correlation regularization. The proposed method first employs a sliding window technique with edge-reflection padding to divide the original noisy data into overlapping patches, which are then unfolded into one-dimensional vectors. Then, a one-dimensional sparse representation neural network is constructed and optimized using a self-supervised learning strategy. The network takes vectors from original noisy data as input and is trained by minimizing a hybrid objective function composed of a mean squared error term and a zero-lag cross-correlation regularization term. Furthermore, a vector selection strategy based on Monte Carlo theory is introduced to accelerate the network optimization. After training, the optimized network is applied to all noisy vectors to obtain the denoised vectors, which are then reshaped and merged to produce the final denoised result. The experiments on both synthetic and field datasets indicate that the proposed method effectively suppresses random noise while preserving seismic signals. Compared with several methods, our method exhibits superior performance in both noise attenuation and signal preservation. Moreover, the Monte Carlo-based vector selection strategy significantly improves the efficiency of the proposed method.