Extracting Fresnel zones from migrated dip-angle gathers using a convolutional neural network

Abstract

Fresnel zones are helpful for obtaining a high signal-to-noise ratio (S/N)-migrated result. A migrated dip-angle gather provides a simple domain for estimating 2D Fresnel zones for 3D migration. We develop a deep-learning-based technology to automatically estimate Fresnel zones from migrated dip-angle gathers, thus avoiding the cumbersome task of manually checking and modifying the boundaries of the Fresnel zones. A pair of 1D Fresnel zones are incorporated to represent a 2D Fresnel zone in terms of the inline and crossline dip angles because it is difficult to directly extract 2D Fresnel zones from a 2D dip-angle gather. The proposed convolutional neural network (CNN) is established by modifying VGGNet. As picking boundaries of the Fresnel zones is a regression problem, we remove the last soft-max layer from the VGGNet. The last three convolution layers and a pooling layer are also removed because the feature maps are small enough. To improve the contrast and definition, we enhance the features of the reflected events in the dip-angle gather. Data normalisation is carried out to accelerate the training process using a simple-rescaling method before training the modified VGGNet. Field data examples demonstrate the effectiveness and efficiency of the proposed method.