Noise-to-Context: A Self-Supervised Deep-Learning Approach for Seismic Data Denoising

Summary Seismic data are inherently noisy because of environmental conditions and the acquisition system’s limitations, which can ruin subsurface imaging and hinder correct geological interpretation. Traditional denoising techniques, such as filtering and model-based methods, rely on noise behavior assumptions and are less robust in real cases with complex and nonstationary noise. While deep-learning-based techniques have obtained considerable performance in seismic denoising, most current methods are supervised and rely on clean-noisy pairs that are hard to get in practical scenarios. We propose Noise2Context (N2C), a novel self-supervised deep-learning technique for seismic data denoising that eliminates the need for paired training data. Inspired by masked signal modeling, N2C trains a deep neural network by randomly masking small seismic data regions and learning to predict missing information from its local context. This masking allows the model to learn noise-robust representations without access to labeled clean data. We employ an adapted U-Net architecture augmented with attention operations and an Atrous Spatial Pyramid Pooling (ASPP) module that can effectively capture local and global seismic structures. The model is optimized using a masked mean squared error (MSE) loss function to train the model such that training is only applied to unmasked regions to avoid learning noise. We compare our method on synthetic and field seismic data, demonstrating that N2C effectively denoises while maintaining strong generalization to new noise patterns. Compared with existing methods, our results highlight the benefits of N2C’s context-aware learning process. The experimental results confirm that N2C can effectively remove noise while preserving significant geological features, making it an ideal candidate for real-world seismic denoising applications.