点击次数：
摘要：Identifying and characterizing fractured cavities is essential for exploring carbonate reservoirs. However, characterizing the development and distribution of fractured cavities through post-stack seismic attribute analysis remains challenging. Recently, convolutional neural networks (CNNs), such as UNet and its enhanced versions, have enabled the quantitative identification of fractured cavities. Despite these advancements, the local receptive field and weight-sharing mechanisms of these CNNs limit their capability to capture long-range features within strike–slip fault systems. In addition, neural networks are inherently affected by data uncertainty. To address these challenges, a two-step methodology is proposed. The first step utilizes a Swin-UNet transformer (UNETR) model, enhanced with an attention gate, to interpret fractured cavities. The transformer in Swin-UNETR improves the detection of fractured cavities in strike–slip fault zones, whereas the attention gate enhances the recognition of small fractured cavities by increasing their response in the feature maps. This enhanced Swin-UNETR model overcomes the limitations in modeling long-range features. In the second step, the fractured-cavity identification results are combined with seismic attributes from conventional analysis. Principal component analysis is employed both to increase the relative weight of the neural network recognition results in the attribute fusion and to reduce the uncertainty associated with any single identification method. The methodology was validated in the Shunbei area, yielding horizontal segmentation and vertical zonation of fractured cavities, as well as their characterization through fixed-grid modeling. By combining deep learning based feature extraction with seismic attributes, this approach improves the accuracy of fractured cavity identification and characterization in carbonate reservoirs.
是否译文：否