Abstract: Soil pores can play a significant role in the crucial process related to soil ecological functions. However, it is still challenging on the relationship between pore structure and functional evolution, due to the lack of non-destructive and non-intrusive systems for the spatial structure of multiple types of pores. Among them, accurate segmentation of pore types and ranges was fundamental to the research. In this study, an improved UNet-VAE network was proposed to segment the multiple pores in soil for the first time. Taking the typical black soil as the research object, a Simplified Convolutional Network (SCN) was used to segment the three-dimensional data of soil pores. According to the segmented pore dataset, a combination of automatic segmentation and manual correction was used to obtain four types of soil pore structures as ground truth. A multi-scale fusion attention module was proposed to filter out the redundant information generated by convolutional learning using a 3D UNet network. Local attention was used to learn the spatial features of small-scale pores (irregular and spherical pores). Global attention was used to extract the feature information of large-scale pores (cracks and biological pores), in order to fuse the multi-scale features for high segmentation accuracy in the different categories of pores. Meanwhile, the commonly used segmentation networks (such as 3D UNet, Segresnet, VNet, and UNetR network) were used to compare the segmentation of the multiple pores. The experimental results showed that the improved UNet-VAE network accurately determined the range and the category of pores. Specifically, the UNetR network was difficult to learn the features, due to the high requirement of a Transformer for the number of datasets in the large-scale cracks and biological pores. Convolutional networks (such as 3D UNet, Segresnet, and VNet) failed to learn the global and large-scale features, where the cracks with obvious planar features were classified as biological pores. Furthermore, the Segresnet, VNet, and UNetR network misclassified the small-scale irregular pores as cracks. By contrast, the improved UNet VAE network achieved the best performance in the four categories of pores, with the average accuracy, precision, recall, and F1 values reaching 93.83%, 84.75%, 84.88%, and 84.60%, respectively. Compared with the suboptimal VNet, the average accuracy, precision, recall, and F1 value increased by 3.32, 5.06, 8.97, and 8.63 percentage points, respectively. Especially for irregular pores, the accuracy, recall, and F1 values increased by 4.88, 15.46, 15.70, and 15.50 percentage points, respectively. In summary, the improved UNet-VAE network was achieved in the high-precision and three-dimensional segmentation of multiple categories of pores, indicating better feature learning for all four categories of pores. The three-dimensional segmentation was achieved in the better classification of intersecting pores, high segmentation accuracy of single pores, and high automation level. Technical support was offered to explore the relationship between pore structure and ecological function evolution. This finding can also provide a data basis for the precise quantitative characterization of soil pore structure, in order to reveal the role of soil pore evolution in ecosystems.