地震反演是地震数据解释的关键技术之一。传统的地震反演方法虽然理论完备、可解释性强，但它们大都根据简化模型推导而出，需要子波估计等人工交互控制，制约了在实际应用中的准确性和灵活性。近年来被广泛研究的基于深度学习的地震反演方法具有自动化程度高、预测精度高等特点。但受限于可以作为标签数据的测井数据为一维序列，现有的基于深度学习的地震反演方法一般采用一维网络，所以不能充分利用地震数据的空间连续性特点，从而导致反演结果横向连续性不佳。针对上述问题，本文研究了如何使用高维卷积网络进行地震反演，并借助迁移学习技术，提升深度学习地震反演结果的准确性。本文的主要内容包括：（1）提出了一种基于多任务学习的二维闭环地震反演方法。本方法使用端到端二维卷积网络建立二维地震数据和二维波阻抗之间的相互映射，利用二维网络的空间表达能力，灵活地引入地震数据的空间模式约束，提高了地震反演结果的横向连续性。同时，针对地震数据低频和高频信息缺失的问题，利用多任务学习方法，同时反演宽频和低频波阻抗、波阻抗边缘三项任务，提高了断层、溶洞等关键地质结构以及波阻抗低频成分的反演效果。（2）提出了一种基于域迁移的地震反演方法。本方法首先利用简化一维闭环网络反演方法处理实际地震数据，得到一个初步的波阻抗估计，然后利用地震褶积模型合成了一个和实际地震数据统计特性近似的三维地震数据-波阻抗样本对，实现有效的深度网络的预训练。进一步利用域迁移技术，通过域适应层，对齐合成数据和真实地震数据在特征空间中的分布，提高了网络在真实地震数据上的泛化能力，并将其应用在二维地震反演中。（3）提出了一种基于对抗迁移的三维地震反演方法。在前述训练样本合成方法基础上，通过引入生成对抗网络，利用在对抗迁移过程中使用风格判别器对网络训练进行约束，使训练结果的特征风格尽可能与实际测井波阻抗相似。同时在训练过程中采用迭代式样本更新的方式，不断更新作为训练标签的合成波阻抗，并通过网络和样本的同步更新进行迭代，提升反演精度。

Seismic inversion is one of the key techniques for seismic data interpretation. Although traditional seismic inversion methods are theoretically complete and interpretable, they are mostly derived based on simplified models and require manual interactive control such as wavelet estimation, which restricts the accuracy and flexibility in practical applications. In recent years, seismic inversion methods based on deep learning have been widely studied. Deep learning methods have the characteristics of high degree of automation and high prediction accuracy. However, the existing deep learning-based seismic inversion methods generally use one-dimensional convolutional networks because the log data that can be used as labeled data are one-dimensional sequences, so the spatial continuity of the seismic data cannot be fully utilized, which leads to poor lateral continuity of the inversion results. To address the above problems, this dissertation investigates how to use high-dimensional convolutional networks for seismic inversion and improve the accuracy of deep learning seismic inversion results with the help of transfer learning techniques. The main contents of this dissertation include:(1) A two-dimensional closed-loop seismic inversion method based on multi-task learning is proposed. This method uses an end-to-end 2D convolutional network to establish the mutual mapping between the 2D seismic data and the 2D wave impedance, and improves the lateral continuity of the seismic inversion results by flexibly introducing the spatial constraints on the seismic data using the spatial expression capability of the 2D network. Meanwhile, to address the problem of missing low-frequency and high-frequency information of seismic data, a multi-task learning method is used to invert three tasks simultaneously, namely, the inversion of the broad and low-frequency wave impedance and wave impedance edge, to improve the inversion performance of the key geological structures such as faults and caves as well as the low-frequency components of wave impedance.(2) A domain adaption-based seismic inversion method is proposed. This method first processes the actual seismic data using a simplified one-dimensional closed-loop network inversion method to obtain a rough wave impedance estimation, and then synthesizes a three-dimensional seismic data-wave impedance sample pair that approximates the statistical properties of the actual seismic data using the Robinson convolution model to achieve an effective pre-training of the depth network. The domain transfer technique is further utilized to align the distribution of the synthetic and real seismic data in the feature space by domain adaptation layer, which improves the generalization ability of the network on real seismic data, and it is applied to 2D seismic inversion.(3) A 3D seismic inversion method based on adversarial transfer is proposed. Based on the aforementioned training sample synthesis method, by introducing a generative adversarial network and using a style discriminator in the adversarial transfer process to constrain the network training, the feature style of the training results is made as similar as possible to the actual logging wave impedance. Meanwhile, iterative sample update is used in the training process to generate pseudo-labels for unlabeled regions based on the current network prediction results. And the iteration of the network training and the sample update helps to improve the inversion accuracy.