Abstract: To elucidate the mechanisms of reactive solute transport in heterogeneous rocks, it is essential to characterize their spatially heterogeneous structures. This study focuses on granite samples from Beishan in Gansu Province, China. Mineral composition segmentation is performed using both of the grayscale threshold segmentation and intelligent recognition method based on deep residual networks. Combined with geostatistical approaches, the semivariogram function is applied to establish the spatial correlation structures of sorption coefficients within the heterogeneous granite. The results indicate that the errors between the mineral segmentation results and the actual mineral components in the granite samples are less than 5%. Moreover, the intelligent recognition method can preserve the mineral continuity better. Among the major mineral components of the granite, the correlation lengths of the sorption coefficient for biotite are 3.75, 4.92, and 4.14 mm in the x, y, and z directions, while for feldspar they are 7.88, 6.15, and 5.94 mm, and for quartz, 5.25, 5.45, and 5.28 mm, respectively. These findings demonstrate that there is significant anisotropy in the granite's sorption properties.

Key words: mineral facies identification, granite, deep residual network, semivariogram function, multi-scale；spatial correlation structure