Super-resolution reconstruction of soil CT images via a physics-guided latent-space diffusion model (<italic>invited</italic>)

ObjectiveSoil microscale pore structures play a crucial role in regulating water transport, gas diffusion, nutrient cycling, and microbial habitats. Current Computed Tomography (CT) imaging techniques face a trade-off between spatial resolution and scanning coverage, and existing super-resolution reconstruction methods often fail to simultaneously recover fine structural details and preserve physical consistency. Common problems include blurred pore boundaries, missing micropores, and reconstructed images that violate physical property constraints. To address these challenges, this study proposes a Physics-consistent Latent Conditional Diffusion for Super-resolution (PLCD-SR) of soil pore CT images.MethodsFirst, a VAE latent space is fine-tuned on high-resolution soil CT images to capture rich structural priors, providing a stable foundation for subsequent diffusion-based reconstruction. Second, a multi-condition latent space constraint is constructed by mapping low-resolution images and thresholded pore masks into the latent space, explicitly guiding the model to focus on key structural regions and suppress redundant information. Third, a physics-guided UNet denoiser integrates physical attributes such as porosity, specific surface area, and permeability into the time-step embedding, enabling collaborative modeling of pore morphology and physical properties. Finally, a latent-space reconstruction loss is introduced to ensure the generated images closely match the true distributions of pore structures and their physical attributes, promoting accurate and physically consistent reconstructions.Results and DiscussionsExperiments on soil CT datasets demonstrate that PLCD-SR achieves superior reconstruction performance compared to traditional interpolation methods (Nearest, Bilinear, Bicubic) and state-of-the-art deep learning baselines (SRCNN, RCAN, SwinIR, SRGAN, SRLGAN, SRDiff, Stable Diffusion). Specifically, PLCD-SR reaches a PSNR of 34.35 dB, SSIM of 0.87, porosity error of 0.026, two-point correlation function error of 0.023, and LPIPS of 0.18. Compared with the second-best method, SRCNN, PLCD-SR improves PSNR by 4.8 dB and reduces errors in porosity, the two-point correlation function, and LPIPS by 0.118, 0.057, and 0.22, respectively. Qualitative results further show that PLCD-SR effectively preserves micropore boundaries, reduces artifacts, and maintains realistic structural distributions. These findings indicate that the proposed method can generate high-resolution soil pore images with high visual fidelity and accurate structural representation.ConclusionsThe PLCD-SR framework achieves multimodal conditional fusion of structural priors and physical attributes based on a unified latent-space representation, enabling super-resolution reconstruction with both high-fidelity structural recovery and strong physical consistency. Through latent-space multi-conditional modeling, physical-attribute embedding, and reconstruction-loss guidance, the model enhances pore-boundary sharpness and microstructural detail while simultaneously preserving the rationality and stability of key physical indicators such as porosity, specific surface area, and permeability. The results demonstrate that the proposed method not only improves the visual quality of soil CT images but also surpasses existing approaches in terms of physical structural fidelity. This provides a highly reliable and interpretable technical route for digital soil microstructure modeling, precision agriculture monitoring, ecological process quantification, and soil carbon-cycle research.