Suppressing migration image artifacts using a support vector machine method

Abstract

Reverse time migration (RTM) can produce high-quality images of complex subsurface structures when using seismic data acquired by a reasonably dense data acquisition geometry. However, RTM produces significant image artifacts when using data from a sparse data acquisition geometry because of incomplete cancellation of migration “smiles.” These artifacts obscure migration images of actual geology, leading to possible misidentification of important geologic features of interest. A specularity filter based on the semblance equation is commonly used in the dip-angle angle-domain common image gather (ADCIG) to preserve signals while suppressing image artifacts. In dip-angle ADCIG, the signals are assumed to have higher semblance scores because they are horizontally more coherent than the artifacts. However, this assumption fails when the image artifacts are severe. We have developed a new approach to suppressing migration image artifacts using a support vector machine (SVM) method. We first develop multiple criteria to distinguish between the signals and artifacts in the dip-angle ADCIG, rather than using only the semblance criterion. We then calculate the weights using a supervised SVM method. The weights approach one for valid signal points, and approach zero for artifact points. Finally, we apply the weights to the dip-angle ADCIG to preserve the effective signals and suppress the image artifacts. We verify the effectiveness of our method, denoted as SVM filtering, using numerical tests on synthetic and field data to produce migration images with improved signal-to-noise ratios and reduced aliasing artifacts.

Wavefield Reconstruction Inversion via Physics-Informed Neural Networks

“…For example, the support vector machine (SVM) is a widely used technique for classification and regression [21]. SVM has shown its effectiveness in classifying seismic facies [22], identifying passive seismic source type [23], separating surface waves modes [24], suppressing artifacts in seismic images [25] and passive source images [26]. Neural network (NN) and its variation forms (e.g., deep/convolutional/recurrent NN) are becoming more powerful in pattern recognition, image processing and image segmentation for large-scale data.…”