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.

Fine-grained detection of bird species has been confined to the high cost of data annotation during imaging. The low accuracy of detection can also be attributed to the diversity of bird species under the complex environments in wetlands. In this study, a semi-supervised CST-based algorithm was proposed to detect the fine-grained images of the birds in the wetland scenes, termed SSBY-CST (semi-supervised bird detection with CNN and Swin Transformer). The unlabeled samples were also utilized to enhance the fine-grained performance during bird detection in complex environments with minimal labeled data. This framework also facilitated the intelligent digital management of agroforestry ecosystems. The core contributions of this research were as follows. 1) Data Collection and Dataset Construction. The images were first collected from 14 monitoring stations in Beijing, China. A dataset was then created under different wetland environments. The dataset included 17 species of birds, thus offering reliable data support for the robustness of the model. The variability and diversity were obtained to train a robust detection model under real-world conditions, particularly in wetland habitats where environmental conditions were challenging and variable. A single-stage Semi-Supervised Learning Framework was also proposed using Pseudo-Labeling. A teacher-student model was constructed using the Yolov5 backbone network. The unlabeled data was efficiently utilized to improve the performance of detection. A dual-threshold pseudo-label assignment was introduced to replace the traditional single-threshold during training. The unsupervised loss function was optimized to effectively reduce the impact of low-quality pseudo-labels. The overall accuracy of the model was improved to minimize the reliance on the large amounts of annotated data. The labeled and unlabeled data were combined for the generalization and robustness of the improved model, particularly for the condition with scarce labeled data. 2) Dual-Channel Convolution Module (CST). A dual-channel convolution module (CST) was developed to distinguish between different bird species and wetland backgrounds. The convolutional neural networks (CNN) were integrated with the Swin Transformer for the local and global feature extraction. Fine-grained details of bird species were also distinguished from the complex background of wetland environments. The CST module was enhanced by both the local and global feature fusion. Diverse patterns were effectively captured from the input data. 3) Performance Improvement. The experimental results demonstrated that the SSBY-CST algorithm was achieved with 77.5% accuracy and 58.2% mAP@0.5 for the bird detection of over 17 species in complex environmental settings, particularly with only 100 labeled images. There were significant improvements of 17.4 percentage points in precision and 15.5 percentage points in mAP@0.5, compared with the state-of-the-art YOLO models. A remarkable performance was enhanced using limited annotated data. Notably, the mAP@0.5 reached 95.7% and 94.5%, respectively, for the species (such as the Black Stork and Siberian Gull), which were improved by 24.9 percentage points and 14.3 percentage points over the baseline. The semi-supervised learning significantly improved the detection performance in the fine-grained bird species, even under the conditions of limited labeled data. 4) Ablation Study and Effectiveness Validation: An ablation study was conducted to analyze the effect of the dual-threshold pseudo-label assignment and the CST module on the performance of the improved model. The effectiveness of each improvement was validated. Specifically, the dual-threshold pseudo-labeling technique was performed better to refine the assignment of pseudo-labels, and then correctly detect the low-confidence targets. Additionally, the CST module was enhanced to extract the features from both local regions and global contexts. A great contribution was gained to better understand the relationship between bird species and their wetland environments. In conclusion, the semi-supervised learning and self-attention mechanisms were performed the best to detect the fine-grained bird species. The SSBY-CST framework was also suitable for the detection of birds in agroforestry and wetland environments, where the high-quality annotated data was limited. This work can greatly contribute to the bird monitoring systems in the intelligent management of agroforestry ecosystems. Geographical location information can be integrated to improve the overall monitoring of bird species in dynamic habitats. Moving forward, the diversity of the dataset can be further extended to refine in complex environments. Additionally, the multi-modal data (including images and geographical data) can be explored for more robust and scalable bird detection for ecological conservation.