Establishing accurate velocity models is crucial for seismic imaging and interpretation.Deep learning offers a new way to build precise velocity models,but the samples available for training the velocity-modeling network are severely limited.To address this issue,we propose a method that utilizes random curves to simulate the subsurface velocity model and automatically generates a large number of samples for deep learning training.A set of random numbers is generated and interpolated to form random se-quences,which are then transformed into random curves using trigonometric functions to simulate subsurface interfaces and pro-duce a layered velocity model.The velocity model is complicated by incorporating such features as faults,velocity anomalies,and bedding angles.We adopt Deeplabv3+as the velocity prediction network.The Deeplabv3+network is optimized by adding convo-lutional layers to address the issue of blurry boundaries caused by direct output after upsampling.We applied the proposed method to both synthetic and real data,and evaluated its stability under conditions of noises,wavelet variation,and data missing.The re-sults demonstrate that our approach effectively mitigate the impact of wavelet variation,noises,and partial data absence,and show reliable generalization and robustness.
维普
万方数据
