Non-destructive methods are essential for linking the microstructural geometry of porous geo-materials to their mechanical behavior, as destructive testing is often infeasible due to limited material availability or irreproducible conditions. X-ray micro-computed tomography (micro-CT) provides high resolution three-dimensional reconstructions of porous microstructures, enabling direct quantification of geometric descriptors. Recent advances in morphometric theory have demonstrated that four independent morphometric measures (porosity, surface area, mean curvature, and Euler characteristic) are required to capture the relationship between microstructure and strength, thereby forming the basis of generalized strength laws. To facilitate practical application of this framework, a Fiji plugin was developed to extract the four morphometric measures from X-ray micro-CT datasets automatically. The plugin integrates within the Fiji platform to provide reproducible, accessible, and user friendly analysis. The extracted descriptors can be readily incorporated into constitutive models and machine learning workflows for forward prediction of stress–strain behavior and inverse microstructure design. This approach supports non-destructive evaluation, accelerates materials selection, and advances the integration of imaging with predictive modeling in porous media research.
Soil classification is a foundational task for combating land degradation, safeguarding food security, and realizing sustainable land use. However, traditional field surveys and laboratory analyses are costly, inefficient, and constrained by complex terrain, making large-scale soil monitoring impractical. While computer vision-based AI technologies have emerged as a promising alternative, existing solutions face critical trade-offs: conventional lightweight Convolutional Neural Networks (CNNs) struggle to capture long-range spatial dependencies in soil images (e.g., texture continuity and structural variability), limiting classification accuracy; Vision Transformers (ViTs) excel at global context modeling but suffer from high computational complexity, hindering deployment on edge-oriented and resource-constrained devices for on-site soil detection. To address these practical challenges in soil science applications, this paper proposes a dual-branch hybrid network named SoilNet-Hybrid. The model integrates the hardware-friendly gated convolution mechanism from MambaOut with the long-range dependency capture capability of ViTs, enabling synergistic extraction of local soil structural features (e.g., pores, particle textures) and global contextual information. We compared SoilNet-Hybrid with five mainstream lightweight baselines (Efficientnet-Lite, EfficientViT_b0, Mambaout_femto, MobileNet_V3-s, and MobileViT-xxs) on two soil datasets. Experimental results show that with only 1.62M parameters, the proposed model achieves balanced accuracy of 92.67% ± 1.56% and 94.64% ± 2.45%, Macro-F1 scores of 93.02% ± 1.44% and 94.56% ± 2.57%, and overall accuracy of 93.90% ± 1.39% and 95.31% ± 2.06% across 5-fold cross-validation, demonstrating stable performance across the evaluated group-stratified 5-fold cross-validation folds under class imbalance and promising potential for edge-oriented inference.
