Microscopic Recognition of Tight Sandstone Based on Deep Learning

Traditional geological experiments, such as rock thin section identification, are often labor-intensive, time-consuming, and involve complex analytical procedures. The intelligent transformation of these methods is crucial for advancing the digitalization of geological research. Tight sandstone is primarily composed of minerals such as quartz, plagioclase, potassium feldspar, and contains metamorphic and igneous rock fragments. The analysis of casting thin sections of tight sandstone enables qualitative and quantitative identification of compositional structure, pore-throat characteristics, and filler types, offering essential experimental insights for studying and evaluating micro-reservoir properties. However, intelligent identification faces three main challenges: fine-grained mineral segmentation, discrimination of visually similar minerals, and quantitative analysis of micro-pores, all arising from the fine grain size, strong diagenetic alteration, and complex pore systems of tight sandstone. Here, we present an intelligent identification method based on deep learning for analyzing the microfabric of tight sandstone thin sections. First, a multi-angle microscopic image database of tight sandstone casting thin sections under both plane-polarized and cross-polarized light was constructed. Semantic segmentation and segment anything model (SAM) were employed to delineate mineral grain boundaries. Subsequently, a deep convolutional neural network automatically extracted optical and structural features from the thin section images, enabling mineral identification, quantification, and analysis of grain sorting and roundness. The method also supports pore-type classification and quantitative pore analysis. Validation on a test set of casting thin section images from the Linxing-Shenfu block in the Ordos Basin demonstrates a mineral grain segmentation accuracy of up to 95%, and a recognition accuracy of over 91% for major minerals such as quartz, plagioclase, and potassium feldspar. This research demonstrates that the cumbersome and repetitive manual work of mineral identification and quantitative statistics can be avoided, providing technical support for the intelligent microscopic images of tight sandstone thin sections.