MFHSformer: Hierarchical sparse transformer based on multi-feature fusion for soil pore segmentation

Soil is a crucial component of the Earth’s surface, playing a significant role in maintaining ecological balance and promoting sustainable development. The automatic segmentation of pores in soil CT images is essential for exploring the soil’s internal structure. However, due to the complex structure and blurred boundaries of soil pores, existing methods cannot accurately and automatically identify pore structure, thereby adversely affecting the characterization of the soil’s internal structure. Therefore, this study aimed to develop a Hierarchical Sparse Transformer model based on Multi-feature Fusion (MFHSformer) to improve the segmentation accuracy of soil pore structure. The proposed method extracted multi-scale features while reducing the computational complexity of the model through the multi-scale sparse self-attention module, which was used to improve the recognition ability of complex and variable pore structures. Meanwhile, the feature-complementary hierarchical resampling block was employed to fuse local features extracted by convolution and global features extracted by the self-attention mechanism, further enhancing the identification of blurred pore boundaries. Compared with deep learning methods, the MFHSformer method showed higher pore segmentation accuracy (99.40 %), recall (86.30 %), and harmonic means (85.51 %). Specifically, the recall and harmonic means were 9.81 % and 3.82 % higher, respectively, than those of the second-best method (ConvNext). This study demonstrated that the proposed MFHSformer method could automatically and accurately segment complex pores in soil CT images, providing an intelligent technique for comprehending soil internal structure.