• Deep learning models (PSPNet, U-Net, FPN, LinkNet) automate quartz identification in thin sections, reducing manual effort and expertise requirements.
• PSPNet achieved the highest IoU scores, demonstrating superior segmentation accuracy even in complex mineral coexistence scenarios.
• The study used a comprehensive dataset of 2470 hyperspectral images from 120 thin sections, with expert-reviewed masks ensuring robust training.
• This automated approach enhances reliability and efficiency, offering a valuable tool for geologists and advancing mineralogical analysis.