Convolutional Neural Networks for Semantic Segmentation of Micro-Pores in SEM Based Images of Shales

Abstract

Geoscientists have been increasingly exploring unconventional shale plays. Shales have complex micro-porosity controlling their permeability and mechanical properties. Because of this complexity, quantitative study of shale pore-space is challenging. One way to study pore spaces is by using ultra-high resolution images that are collected using Scanning Electron Microscopy (SEM). Unfortunately, the quality of SEM images is still relatively poor and noisy compared to optical microscopy. For this reason, pore spaces in SEM images cannot simply be identified by a threshold based method (i.e., segmentation). Semi-automatic segmentation algorithms such as the top-hat filter are ineffective because of noise. So far, labeling shale pore spaces requires great effort from tracing each void space by hand. Recently, convolution neural networks (CNN) have demonstrated success in image recognition. Deep stacks of convolutional filters and nonlinear transfer functions allow the networks to identify complex features. Here, we show the results of a trained CNN for detecting pore spaces in shale SEM images. Shale cores were collected by the International Ocean Discovery Program in the Nankai Trough, Japan. The stratigraphic column shows a successive sequence of turbidites on the accretionary wedge. The samples were taken from four different lithological units. Each unit has unique pore types (e.g. interparticle pores and fractures). Twenty-one SEM images were taken by combining the secondary electrons and the backscattered electron signals in order to reduce the charging effect. Seven greyscale images of size 1024x800 square pixels at the resolution of ~9 nm are randomly selected as training data. Pore spaces are then manually segmented with the region growing algorithm. Normalization, subsampling, and image augmentation are applied to the dataset in order to avoid overfitting during the training process. After that, we train the CNN, which is designed based on the SegNet network. The architecture consists of convolutional downsampling units followed by convolutional upsampling units, similar to autoencoder neural networks. The trained network is efficient as it yields 94% accuracy of the testing dataset. In addition, the network takes ~10 seconds to segment an image. This technique leads to a more reliable method to analyze shale sample quantitatively. The algorithm performance on SEM images from other geological settings has to be further investigated.

AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019