Abstract:

Objective: A multi-scale fusion method based on a cycle generative adversarial network for soil CT/SEM images (multi-scale fusion based on CycleGAN, MSF-CycleGAN) is for the first time proposed in order to address the obligatory dependence on paired datasets in super-resolution reconstruction of soil CT images. By integrating micro-structural priors from SEM imagery, the approach substantially enhances soil CT image fidelity while circumventing the prohibitive expense of high-resolution CT scanning, thereby offering a novel technical pathway and broad application prospects for intelligent forestry and agricultural management. Method: Firstly, this study constructed a dataset of soil CT and SEM images, encompassing soil column samples across three moisture levels and multiple freeze-thaw cycles, providing a data foundation for subsequent image resolution enhancement. Secondly, a CycleGAN model suitable for cross-domain transformation was introduced for the multi-scale fusion task of soil CT/SEM images. This model leveraged the different imaging characteristics of CT and SEM images to supplement low-resolution soil CT images. To improve the resolution of the generated images, upsampling modules were added to generator A, and downsampling modules were added to generator B. To enhance the authenticity and consistency of the images, the bilinear interpolation method was introduced to upsample the original images and then incorporate them into the calculation of the identity loss function. Finally, an improved ConvNeXt model was employed to segment the pores in the fused high-resolution images, construct a three-dimensional pore model and extract parameters. Result: Under the same experimental conditions, MSF-CycleGAN was qualitatively and quantitatively compared and analyzed with the super-resolution reconstruction algorithms. It was proved that the multi-scale fusion of MSF-CycleGAN could improve the resolution of soil CT images and the effect is more remarkable than that of the super-resolution reconstruction. In addition, ablation experiments were conducted to analyze the contributions of the introduction of SEM images and the improvement of the identity loss function to the overall performance, and the effectiveness of both designs was verified. The multi-scale image fusion experiments on soil columns with moisture contents of 50%, 75%, and 100% indicated that the fused three-dimensional models exhibited more fine pores, with pore quantity increasing twofold. The pore volume fraction increased by 3.93%, and the roundness of pores increased by 3.64%, while the fractal dimension increased by 0.55%. Quantitative experiments on pore parameters based on the high-resolution fused image segmentation quantification under freeze-thaw cycles conformed more closely to existing research patterns, confirming that the proposed method accurately and effectively enhanced the resolution of soil CT images. Conclusion: This study demonstrates the feasibility of improving the resolution of CT images through multi-scale fusion of soil CT and SEM images, providing a new technical approach to reduce the cost of acquiring high-resolution images and promoting the refinement and intelligent development of soil structure research.

Key words: soil, multi-scale fusion, cycle-consistent GAN, soil CT images, SEM images