EarthAdaptNet: Deep learning architecture for Seismic Facies Analysis

I. INTRODUCTION

Interpretation of geologic features and inference of reservoir properties are key to the success of hydrocarbon exploration and production. For example, seismic images are acquired in different stratigraphic settings and are related by reflection patterns, stratigraphic settings, and depositional environments. In this article, we present an approach that exploits accurate and robust semantic segmentation (classification) of seismic images with cropped local image patches on the F3 block of the Netherlands. seismic images are acquired in different stratigraphic settings and are related by reflection patterns, stratigraphic settings, and depositional environments.

Dataset

This study uses processed seismic data collected from the F3 block in the Netherlands and Penobscot in Canada. Generating seismic images is a sophisticated process that involves data acquisition, where intense sound sources are placed between 6 and 76 m below the ground to generate sound waves. This signal is then processed using an iterative procedure to generate seismic images. The same idea applies to the crossline slices, which are images along the depth axis and perpendicular to the crossline axis. The F3 dataset included 401 crossline and 701 inline slices, with a dimension of 401 701. In a previous study [2] , the slices were interpreted and annotated, and a label mask was generated for each slice.

PROPOSED NETWORK ARCHITECTURES

In the present study, we propose state-of-the-art architectures, i.e., EarthAdaptNet and its variants, for semantic segmentation of seismic facies. The model is trained using small patches created out of large seismic sections. In this paper, we will be using the patch-based model for both studies. The proposed model for semantic segmentation is inspired by U-Net [24] and Danet-FCN3 . Building blocks of EarthAdaptNet can be broadly divided into RBs and TRBs similar to those of Danet-FCN3 but with some modifications. In the proposed architecture, RB comprises two convolutional layers, each followed by batch normalization and a downsampling residual connection of a convolutional layer. Finally, output of all 5 parallel layers is concatenated followed by another convolution with 256 filters. An important point to note in the EAN architecture is that there is no batch normalization layer in the shortcut connection. 4 RBs followed by a Global Average Pooling (GAP) layer followed by 2 fully connected layers. 3 RBs followed by a Global Average Pooling layer followed by 2 fully connected layers.

Building blocks of the proposed EarthAdaptNet, which consists of RBs and TRBs. RB comprises two convolutional (Conv) layers, each followed by batch normalization and a downsampling residual connection of the 1x1 Conv layer. TRB is similar in architecture as RB except with the use of a transposed convolutional (ConvT) layer instead of a convolutional layer. The encoder starts with a Conv Layer and is followed by the RB. Decoder starts with a TRB and the number of TRBs used is kept the same as the RB used in the Encoder and is followed by a Transposed Convolutional Layer which outputs the segmented seismic image. A 1x1 convolutional layer also exists in the middle which acts as a bridge (Bottleneck) between the Encoder and the Decoder. Skip connection is present between each RBs and TRBs.

 

Results obtained from EAN study. Top section: raw seismic section along inlines 295, crossline 620 from #Test1 and crossline 411 from #Test2 respectively. Middle section: Original labels and bottom section: Interpreted labels from 4 RB-TRB pairs with ASPP model. We can see few misclassifications for inline 295 and crossline 620.

In conclusion, we have introduced a deep learning model EarthAdaptNet (EAN) that can efficiently classify facies with patch sections and is able to achieve a classification accuracy>50% for smaller classes like Zechstein and Scruff. The architecture performs better than the patch-based baseline model. The proposed DDA approach is among one of the first applications of DDA to the study of unlabeled seismic facies. We also present and examine 3 variants of the proposed DDA architecture to understand how the components such as residual blocks, global average pooling, and fully connected layers behave in domain adaptation.