Unsupervised dealiased seismic data interpolation based on differentiable dynamic time warping distance constraint

Existing unsupervised seismic data interpolation methods are mostly based on the deep image prior (DIP) framework. These methods typically use L2 loss between the sampled data and the interpolated results at corresponding sampled positions. However, due to the lack of direct regularization on the interpolated data. They are less effective at interpolating seismic data with large-slope events, which exhibit strong aliasing in the F-K domain. To address this issue, we propose a novel unsupervised dealiased seismic data interpolation method. This method leverages the local similarity of seismic data and employs a differentiable dynamic time warping (DTW) distance to measure the similarity between interpolated trace and neighboring sampled traces. The differentiable DTW distance is incorporated as a regularization term in the loss function of a traditional DIP framework. Experimental results show that, compared to the Spitz method and the deep seismic prior (DSP) approach, our method exhibits the strongest dealiased capability and achieves the best interpolation performance.