Multichannel Sandstone Thin Sections Identification Based on Improved DeepLab V3 Plus Neural Network.

Area of Science:

• Geoscience
• Petroleum Geology
• Artificial Intelligence in Earth Sciences

Background:

• Tight sandstone gas reservoirs store natural gas in small pores between sand grains.
• Accurate characterization of pore structure and mineral composition is vital for efficient gas extraction.
• Existing automated methods for sandstone thin section analysis suffer from low accuracy and speed.

Purpose of the Study:

• To develop a more accurate and efficient method for identifying mineral components and pore structures in sandstone thin sections.
• To enhance the automated analysis of geological thin sections using deep learning.
• To improve reservoir evaluation and gas production prediction through precise mineral identification.

Main Methods:

• Amalgamation of cross-polarized light (CPL) and orthogonal polarized light (XPL) images into multichannel data.
• Utilizing an enhanced DeepLab V3 Plus model with an attention mechanism for semantic segmentation.
• Training multiple networks and selecting optimal architectures and datasets for mineral identification.
• Calculating mineral sizes for precise classification and naming of sandstone thin sections.

Main Results:

• The multichannel identification method achieved high recognition accuracy, with Mean PA of 89.8% and Mean IOU of 81.2%.
• The enhanced model demonstrated superior performance compared to previous automated identification approaches.
• The attention mechanism significantly improved the identification accuracy of the semantic segmentation network.

Conclusions:

• The novel multichannel approach offers a more precise method for identifying mineral composition and pore structure in sandstone thin sections.
• This technique is crucial for accurate reservoir evaluation and predicting oil and gas production.
• The method shows potential for application in identifying and categorizing other types of geological thin sections.