论文详情
常规的生成式对抗神经网络在地震数据去噪过程中受模型限制, 地震数据有效信息还原能力差。因此, 对生成式对抗神经网络进行改进, 以U-net神经网络为基础建立更深层级的生成器神经网络, 优化模型的批标准化层和池化层, 提升特征还原能力, 搭建多尺度判别器神经网络, 提升判别器性能, 提出一种包含对抗损失、配准损失和结构信息损失的多层次综合损失函数。改进后的模型结构无需预先估计噪声, 能够实现端到端的盲去噪功能, 神经网络泛化能力强, 对数据细节的保护还原水平高。南海北部涠A地区地震数据测试结果表明, 改进后的神经网络去噪能力以及对地震有效信息的保护要优于目前常见的去噪算法的结果, 去噪过程对地震有效反射信息保护好, 地震边界信息成像质量高。与常见的去噪方法相比, 改进的生成或对抗神经网络方法在地震数据去噪中具有良好的应用效果, 去噪能力强, 在实际地震数据处理中具有良好的推广价值。
Random noises in seismic data will deteriorate data quality and have a negative impact on interpretation. In random noise reduction, it is difficult to restore effective information in seismic data using a conventional generative adversarial network. Based on the U-net network, we develop a modified generative adversarial network with optimized batch normalization and pooling layers to improve effective information restoration. A multi-scale discriminator network is established to improve the performance of the network model. A set of multi-module loss functions are formulated with feature matching loss and structural information loss. Owing to the new network structure, it is unnecessary to estimate noises in advance, and thus end-to-end blind denoising could be achieved. The model also features improved ability of generalization and data restoration. Field data tests in the northern South China Sea show improved performance of noise reduction and signal preservation compared with other denoising algorithms, leading to better imaging of boundaries. The improved generative adversarial network is a good method for seismic data denoising and could be applied to seismic data processing in additional prospects.