Subsurface salt bodies provide good confined storage space for oil and natural gas, and effective identification of various types of geologic bodies is the core content of oil and gas exploration. However, current data-driven salt body interpretation methods generally suffer from poor generalization. In the face of the inherent differences between synthetic seismic data and field seismic data, under the premise that field salt body markers are difficult to obtain, it becomes crucial to rationally apply field seismic data and enable the model to deeply learn the correlation features between the two types of data. We proposed an unsupervised embedding-based method for dense prediction of 3D salt bodies (3D CLP-gb), presented contrastive loss for field-oriented seismic data, and demonstrated that unsupervised learning of diverse salt body augmentations with high-frequency amplitude features of other geologic bodies as positive and negative samples of each other can help to solve the challenge of the generalizability of the salt body to field data segmentation. The experimental results show that with only 1.56% of the tagged data used, 3D CLP-gb as an upstream task in combination with the SOTA model is able to extract a more complete distribution of salt bodies from the field seismic body, achieving an IOU value of 87.5% for the SEAM synthetic validation set, and a generalization performance of 85.8% for the F3 field dataset.
|