Is attention all geosciences need? Advancing quantitative petrography with attention-based deep learning

Recent advances in deep learning have transformed data-driven geoscientific analysis. In particular, the adoption of attention mechanism in deep learning has received considerable interest and shown impressive results that can surpass traditional convolutional neural networks (CNN) in image analysis. The application of attention-based algorithms has been fairly limited in geosciences, hence the actual value and potential of the attention mechanism to perform image analysis remain untapped. In geosciences, petrographical image analysis provides fundamental qualitative and quantitative information to characterize different rock types and their composition. The quality and accuracy of such an analysis depend on several factors, including: (i) thin section quality; (ii) image resolution, and (iii) subject matter expertise. However, with the current technology, there is a common trade-off between image resolution and field of view of microscope cameras which could potentially hinder reliable quantitative analysis. This is compounded by the time-consuming and costly analysis to obtain highresolution images. In this study, we evaluated the performance of attention-based deep learning algorithm, Super Petrography that adopted the architecture of shifted window vision transformer, to (i) perform multiresolution image upscaling of petrographic images and (ii) provide improved quantitative petrographic analysis. Overall, the proposed model is proven superior to other conventional methods (e.g., Bicubic and Lanczos) and even CNN-based model, showing up to 30% improvement in both peak signal-to-noise ratio and structural similarity index measure. Furthermore, we observed a more accurate quantitative petrographical analysis, including grain edge detection and segmentation from these reconstructed super resolution images. This work highlights the potential application of attention-based deep learning in advancing quantitative petrography that otherwise is not possible to achieve with traditional methods. The proposed method could help mitigate the limiting effect of low-resolution images and improve the accuracy of geological or geophysical description and interpretation such as mineral and porosity segmentation, and lithology identification.