Blended acquisition significantly improves acquisition effi-ciency, and deblending algorithms, particularly intelligent de -blending methods, continue to be developed to provide separated results for subsequent seismic inversion and imaging. Iterative deblending algorithms improve deblending perfor-mance and determining how to evaluate the blending noise level becomes critical. Instead of using a multilevel blending noise strategy to evaluate the blending noise level qualitatively, a su-pervised multistep deblending algorithm is developed that can evaluate the blending noise level quantitatively in multiple steps. The developed multistep method combines the iterative estima-tion-subtraction strategy based on sparse inversion and the deep learning strategy. Each deblending step handles a different level of blending noise, ranging from a strong level to a weak level as the deblending steps increase. The first step is to train a U-net to attenuate strong blending noise, and then we can obtain a rough signal estimation for predicting the blending noise to be sub-tracted. The obtained data, via the blending noise estimation and subtraction following the previous deblending step, are used as the input for the current deblending step, which attenuates weak blending noise and extracts signal leakage in a step-by-step manner. The optimized parameters of the previous deblend-ing step can initialize the current step for efficient fine-tuning based on transfer learning. After sequential blending noise es-timation and subtraction, the supervised multistep deblending algorithm with varying input can improve deblending accuracy. A thorough examination of 2D and 3D synthetic blended data demonstrates the validity of our multistep deblending method, particularly when compared with the recently proposed multi-level blending noise strategy. The 3D field blended data process-ing validates our method in terms of removing blending noise while preserving the signal.
展开▼
