由于急流、裂谷、高山等自然环境的限制，采集的地震数据会出现大间距缺失的现象，影响后续的地震数据处理和地质分析工作，需要对缺失数据进行插值重建。

为解决大间距地震数据缺失问题，提出一种基于循环特征推理的重建方法。首先缺失的地震数据经过部分卷积运算，在计算过程中根据感受野内有效特征图数据的占比，自适应地调整卷积运算结果的权重，避免在连续缺失的地震道上执行无效的卷积操作。然后采用循环特征推理的方式，逐步对缺失部分进行渐进式重建。部分卷积运算和循环特征推理交替进行，直至所有缺失数据重建完成。最后特征融合每次迭代产生的重建特征，以保证推理的准确性。为增强模型对大间距缺失区域纹理细节的学习能力，结合纹理损失和均方误差函数作为复合损失函数，进一步提高重建精度。

结果显示：(1)基于循环特征推理的方法可以有效重建大间距缺失的地震数据，信噪比在原缺失数据的14.89 dB的基础上提升至28.15 dB。(2)连续缺失30道至80道的多次重建实验中，本方法的重建结果信噪比、结构相似性、均方误差等评价指标均优于U-Net方法。采用6种不同公开数据集测试了本方法的重建效果，进一步证明了本方法的有效性。(3)对比实验探究部分卷积核大小对重建结果的影响表明，当部分卷积核大小为3×3时重建结果信噪比更高并且迭代时间更短。研究成果为大间距缺失地震数据的重建方法提供了新的解决思路。

Due to the constraints of natural environments like rapids, rifts, and high mountains, the acquired seismic data are often challenged by consecutive missing, affecting subsequent seismic data processing and geologic analysis. Hence, it is necessary to reconstruct the missing data through interpolation.

This study proposed a method for reconstructing consecutively missing seismic data based on recurrent feature reasoning. First, the missing seismic data undergo partial convolution operations, in which the weight of the convolution results is adaptively adjusted based on the proportion of valid feature map data in the receptive field, avoiding invalid convolution operations on consecutively missing seismic channels. Second, the missing parts are progressively reconstructed through recurrent feature reasoning. Partial convolution operations and recurrent feature reasoning are alternated until all missing data are reconstructed. Finally, the reconstructed features generated in each iteration are integrated through feature fusion, ensuring accurate reasoning. To enhance the model's ability to learn the texture details of consecutively missing areas, the texture loss and mean square error (MSE) functions are combined as a hybrid loss function to further increase the reconstruction accuracy.

Key findings are as follows: (1) The proposed method based on recurrent feature reasoning can effectively reconstruct the consecutively missing seismic data, with the signal-to-noise ratio (SNR) increased to 28.15 dB on top of the original 14.89 dB for the missing data. (2) In multiple reconstruction experiments focusing on 30 to 80 consecutively missing seismic channels, the reconstruction results demonstrate that the proposed method outperforms the U-Net method in terms of assessment indices like SNR, structural similarity, and MSE. The effectiveness of the proposed method is further verified by the reconstruction effects of the proposed method tested on six different public datasets. (3) As revealed by the impacts of the size of the partial convolution kernel on the reconstruction results investigated through comparative experiments, the reconstruction results manifest a higher SNR and a shorter iteration time when the partial convolution kernel measures 3×3. The results of this study provide a novel approach for the reconstruction of consecutively missing seismic data.