Abstract: Soil pores plays a significant role in promoting crucial processes related to soil ecological functions. However, due to the lack of non-destructive and non-intrusive methods and systems for analyzing the spatial structure of multiple types of pores, studying the relationship between pore structure and functional evolution was extremely challenging. Among these, accurate segmentation of pore types and ranges was fundamental to the research. In this study, an improved UNet-VAE network method was proposed to achieve soil multi-category pore segmentation for the first time, providing technical support for studying the relationship between pore structure and ecological function evolution. Taking typical black soil as the research object, the Simplified Convolutional Network (SCN) method was used to segment soil pores and obtain three-dimensional data of soil pores. Based on the segmented pore dataset, a combination of automatic segmentation and manual correction was used to obtain four types of soil pore structure ground truth. Based on the 3D Unet network, a multi-scale fusion attention module was proposed to filter out redundant information generated by convolutional learning. Local attention was used to learn spatial features of small-scale pores (irregular pores and spherical pores), and global attention is used to extract feature information of large-scale pores (cracks and biological pores), to fuse multi-scale features of different categories of pores and improve the segmentation accuracy of different categories of pores. Meanwhile, commonly used segmentation networks in literature, such as 3D Unet network, Segresnet network, VNet network, and UNetR network, were used to achieve multi-category pores segmentation and compared with the proposed method. The experimental results showed that for large-scale cracks and biological pores, UNetR was difficult to learn features due to the high requirement of the Transformer for the number of datasets. Convolutional networks such as 3D UNet, Segresnet, and VNet lack the ability to learn global information and large-scale features, and classify cracks with obvious planar features as biological pores. For small-scale irregular pores, Segresnet, VNet, and UNetR all misclassified them as cracks. Except for the proposed improved network, all other networks exhibit under segmentation phenomenon. Comparing these five methods, the improved UNet-VAE method can accurately segment the pore range and determine the pore category. The improved UNet VAE achieved the best performance among all methods in four categories of pores, with average accuracy, precision, recall, and F1 values reaching 93.83%, 84.75%, 84.88%, and 84.60%, respectively. Compared with the suboptimal VNet method, the average accuracy, precision, recall, and F1 value have increased by 3.32%, 5.06%, 8.97%, and 8.63%, respectively. Especially for irregular pores, the accuracy, recall, and F1 value had increased by 4.88%, 15.46%, 15.70%, and 15.50%, respectively. In summary, the improved UNet-VAE had achieved high-precision three-dimensional segmentation of multiple categories of pores, with good feature learning ability for all four categories of pores, solving the problems of difficult classification of intersecting pores, low segmentation accuracy of single category pores, and low automation level of existing three-dimensional segmentation methods. This article will provide a data basis for the precise quantitative characterization of soil pore structure and a scientific basis for revealing the role of soil pore evolution in ecosystems.