论文详情
现有基于深度学习的层位识别方法通常在地震振幅信号特征方面进行处理，而地层之间上、下位置的空间关系、不同尺度特征未得到充分关注，导致普通深度学习网络在识别多个地震层位时容易产生层位识别结果连续性不强和错层等问题。为了充分利用层位之间的空间位置关系及多尺度特征，使用MultiResBlock多尺度残差模块、CBAM注意力与UNet++，提出了基于多尺度注意力UNet++的层位识别方法(MR_CBAM_UNet++)。该方法利用MultiResBlock提取更多层位尺度特征，采用CBAM注意力模块以减少非目标层的振幅信号干扰，利用Focal Loss与Dice Loss组成的联合损失函数对网络进行训练。最后，加入唯一性约束对模型识别结果优化得到层位识别结果。在实际地震数据上的评价结果显示，MR_CBAM_UNet++模型相比于传统模型，对非层位信息的抑制能力和复杂地势下层位的识别能力均有很大提升。在测试集上，层位识别结果的准确率达到了86.19%，有效缓解了层位解释连续性不强和错层等问题，唯一性约束也使层位识别结果更贴近实际。
Common horizon identification methods based on deep learning primarily focus on seismic amplitude without sufficient attention to the spatial relationship among horizons of different scales, resulting in discontinuous and even inaccurate interpretation. To address this problem,we propose a method based on the multi-scale attention UNet++, termed MR_CBAM_UNet++, which involves MultiResBlock to extract a broader spectrum of horizon scale features, CBAM to reduce the amplitude interference of non-target signals, and a UNet++. A joint loss function composed of Focal Loss and Dice Loss is utilized for network training, and the uniqueness constraint is incorporated to refine the results of horizon identification.According to its application to actual seismic data, the MR_CBAM_UNet++ model shows significantly improved capabilities compared to traditional models in suppressing non-horizon information and identifying horizons in complex subsurface conditions.A mean pixel accuracy rate of 86.19% is achieved for the test dataset,indicating more accurate horizon interpretation with better continuity. Additionally, the results of horizon identification are more geologically significant by using the uniqueness constraint.