A Simplified Convolutional Network for Soil Pore Identification Based on Computed Tomography Imagery

Advances of X-ray computed tomography (CT) in soil micromorphology have motivated researchers to examine the internal structure of soil, particularly the geometry and spatial distribution of the pores. The effectiveness of CT data to characterize pore structures depends on how accurately the soil grayscale image is converted to the binary image. Therefore, the objective of this study was to propose a simplified convolutional network (SCN) to automatically identify the solids and pore structures. To establish the pore identification model and assess the performance of the SCN, we obtain the correction image of pore structures by manual labeling. With automatic learning of the shallow features and the deep features of pore structures, the SCN can accurately identify the irregular boundary and complex structure from the complex hierarchical organization of soil. Compared with four commonly used identification methods in the literature, promising results were obtained with soil samples under different physical conditions. The SCN achieves an identification accuracy of 99.82%, an identification precision of 99.61%, and an identification recall of 99.93%, which are 1.20, 22.47, and 0.82% higher than that of the suboptimal method (fuzzy C-means method), respectively. Overall, the experimental data illustrate that the SCN method can accurately identify the pore structures for soil CT imagery under different physical conditions. Moreover, this paper introduced state-of-the-art artificial intelligent technology into the soil field, which will provide an intelligent technique to the soil micromorphology tool set.