An Intelligent Method for Sandstone Identification Combining Deep Learning with Laser Raman Spectroscopy

Sandstone identification serves as an essential foundation for sedimentary geological research and oil-gas exploration and development. At present, sandstone identification mostly relies on manual mineral identification under the microscope, or adopts deep-learning-based image recognition technology to achieve automatic particle segmentation and mineral classification. Traditional manual identification suffers from prominent subjectivity and low efficiency, making it difficult to meet the requirements of high-throughput standardized analysis for massive samples. In contrast, pure image-recognition methods are constrained by optical information and cannot fundamentally address the challenge of distinguishing minerals with similar optical properties. To address these issues, this paper proposes an intelligent sandstone identification method integrating deep learning and laser Raman spectroscopy: image recognition is first adopted to segment framework grains in sandstone thin-section images and implement preliminary grain classification, molecular spectral verification is then conducted on screened single-crystal grains and low-confidence grains via Raman spectra, and collaborative judgment of the two techniques ultimately realizes accurate identification of sandstone components. For Raman spectrum acquisition, a “funnel model” is proposed to hierarchically screen sandstone component grains, and point scanning is used to replace traditional matrix area scanning, drastically cutting the quantity of grains awaiting detection and spectrum acquisition time. At the data fusion stage, a multi-feature point voting scheme and matching rules for image and Raman results are designed to maximize valid data utilization while securing identification accuracy. Application verifications reveal that compared with pure image recognition, the average confidence of sandstone component identification rises by 0.07–0.11 with an overall average above 0.9; taking TIMA test results as the benchmark, the identification accuracy of quartz and feldspar reaches 87.2% and 99.3% respectively, with the coefficient of determination (R²) of 0.989. This method complements the high-throughput strength of deep learning and the fingerprint-level precision superiority of laser Raman spectroscopy, effectively resolving the confusion of optically similar minerals, achieving automatic, standardized and high-precision intelligent identification of sandstone components, and delivering efficient technical support for sedimentary geological research and oil-gas exploration and development.