Automatic channel detection using deep learning

Introduction

Channels are important geologic features for hydrocarbon exploration. However, manual interpretation of channels in seismic images is a time-consuming and subjective process. Numerous methods, such as using coherence attributes, sweetness attributes, and steerable pyramid, have been proposed for helping channel detection in seismic (Wu, 2017; Mathewson and Hale, 2008; Hart, 2008).

Seismic coherence and other edge-detection algorithms, such as Sobel filter, can be used to highlight channel boundaries (Phillips and Fomel, 2017; Wu, 2017; Kington, 2015). The directional structure-tensor-based coherence method computes the seismic coherence attribute using eigenvalues of the directional structure-tensors constructed from directional derivatives perpendicular and parallel to the seismic structures (Wu, 2017). These edge-sensitive methods can detect channel edges easily but do not indicate channel thickness (Liu and Marfurt, 2007). Sweetness is another seismic attribute for channel detection, and is defined as the ratio between reflection strength and the square root of instantaneous frequency (Hart, 2008). Sand channel bodies generally create stronger, broader reflections than the surrounding shale. Mathewson and Hale (2008) propose steerable pyramid filters to enhance the channel features by partitioning the seismic image with respect to scale and orientation.

All of these seismic attributes focus on detecting the channel boundaries but not the geobodies. We propose to adopt an encoder-decoder convolutional neural network to directly detect 3D channel geobodies without human interpretation on precomputed seismic attributes. The encoder-decoder neural network automatically learns useful features for channel detection. We propose to train the network using a 3D labeled synthetic dataset and then use trained parameters to predict channel bodies in 3D seismic field datasets.

While conventional methods for automatic channel picking lack uncertainty analysis, our proposed method can also provide a quantitative uncertainty analysis. Bayesian SegNet samples the posterior distribution of class probabilities at test time using dropout layers (Kendall et al., 2015). The network estimates the mean and variance of the distribution, which can be used to model the uncertainty and provide information to evaluate the risk of decision-making based on interpretation.

Automatic channel detection using deep learning