论文笔记: 临时分享

摘要: 深度学习做地震数据反演的一些文献综述. 组内分享, 会不断充实.

1. 初至波拾取

2. 分辨率提升

2.1 Seismic resolution enhancement by spectral shaping using shaping-regularized inversion

@ARTICLE{XueJ2022Spectral,
    author    = {Jiao Xue and Chengguo Cai and Hanming Gu and Hongmei Luo},
    title     = {Seismic resolution enhancement by spectral shaping using shaping-regularized inversion},
    journal   = {IEEE Geoscience and Remote Sensing Letters},
    year      = {2022},
    volume    = {19},
    number    = {},
    pages     = {7506905},
    doi       = {}
}

Xue et al \cite{XueJ2022Spectral}
频域中的反褶积通常分为三个步骤：小波估计、反卷积算子构造和数据滤波。Deconvolution in the frequency domain is normally implemented in three steps: wavelet estimation, deconvolution operator construction, and data filtering.
优点：使用新的反褶积方法
注意事项：是物理方法

2.2 Learning deep CNN denoiser prior for image restoration

@INPROCEEDINGS{ZhangK2017Prior,
    author    = {Kai Zhang and Wangmeng Zuo and Shuhang Gu and Lei Zhang},
    title     = {Learning deep {CNN} denoiser prior for image restoration},
    booktitle = {CVPR},
    year      = {2017},
    pages     = {2808--2817},
    doi       = {}
}

Zhang et al \cite{ZhangK2017Prior}针对图像数据去噪，设计了模型与辨识性的结合的方法。
注意事项：这篇论文只讨论了图像数据，而没有讨论地震数据。

3. 去噪

4. 分辨率提升与去噪一起进行

4.1 Can learning from natural image denoising be used for seismic data interpolation?

@ARTICLE{ZhangH2020Natural,
    author    = {Hao Zhang and Xiuyan Yang and Jianwei Ma},
    title     = {Can learning from natural image denoising be used for seismic data interpolation?},
    journal   = {Geophysics},
    year      = {2020},
    volume    = {85},
    number    = {4},
    pages     = {WA115--WA136,},
    doi       = {}
}

Zhang et al \cite{ZhangH2020Natural} 指出：用自然图片训练的去噪网络，可以用于地震数据去噪。
贡献：这篇论文的题目很刚很直接，论点很清晰。以前将地震数据显示为图像，再进行人为的观察与处理，其实也是用到了图像与地震数据的共同点：纹理。从这个角度可以提炼出关键科学问题：自然图像与地震数据的相似性机理探索。

4.2 Deep learning for simultaneous seismic image super-resolution and denoising

@ARTICLE{LiJ2022Simultaneous,
    author    = {Jintao Li and Xinming Wu and Zhanxuan Hu},
    title     = {Deep learning for simultaneous seismic image super-resolution and denoising},
    journal   = {IEEE Transactions on Geoscience and Remote Sensing},
    year      = {2022},
    volume    = {60},
    number    = {},
    pages     = {5901611},
    doi       = {}
}

Li et al \cite{LiJ2022Simultaneous}使用深度学习（具体为CNN）同时进行地震图像的超分辨率重建与去噪。人工合成了大量不同分辨就给、不同噪声的地震图像数据用于网络训练。损失函数结合了$l_1$损失与多尺度结构化相似度损失。另外还进行了故障检测。
优点：超分辨率重建与去噪同时进行。
缺点：理论方面创新很有限，方法也一般。

5. 反射系统反演

6. 波阻抗反演

6.1 Semi-supervised deep learning seismic impedance inversion using generative adversarial networks

@INPROCEEDINGS{MengD2020Semi,
    author    = {Delin Meng and Bangyu Wu and Naihao Liu and Wenchao Chen},
    title     = {Semi-supervised deep learning seismic impedance inversion using generative adversarial networks},
    booktitle = {IGARSS},
    year      = {2020},
    pages     = {1393--1396},
    doi       = {}
}

Meng et al \cite{MengD2020Semi}使用GAN进行地震波阻抗反演，克服仅有少量测井数据的困难。
优点：进行波阻抗的反演，任务有物理意义。使用GAN比CNN的MSE更小（0.0184 vs. 0.0332）