Mung bean (Vigna radiata L.) is one of the most favorite crop plants worldwide. It is often required for the precise quantification of the flowering traits during crop phenotyping and breeding. Particularly, the flowering synchrony can directly influence the population trait uniformity and varietal maturity consistency. Thereby the breeding materials can be refined to evaluate the varietal adaptability. However, the conventional semantic segmentation of flowers cannot fully meet the large-scale quantitative analysis in recent years. Some challenges remain in the low contrast with the background, such as the complex inflorescence structure of the mung bean plants, the minute size of the floral organs, their similar coloration to leaves and stems, coupled with the intense lighting variations, extensive occlusions, and overlapping in field environments. Alternatively, deep learning can be expected for the image segmentation of small targets in complex agricultural scenarios. Therefore, it is very necessary for sufficient feature extraction and model generalization to avoid the spatial information loss. In this study, an IDCA-UNet model was proposed to integrate a dual channel attention (DCA) mechanism and instance normalization (IN). The high-resolution images of the mung bean plants were also captured at the flowering and pod-setting stage using an unmanned aerial vehicle (UAV) remote sensing platform. According to the classic U-Net architecture, the DCA module was employed to dynamically aggregate some features from the global average pooling and max pooling, in order to realize the adaptive feature enhancement in the channel dimension. The high-resolution features effectively improved the sensitivity and recognition for the tiny flowers. Simultaneously, the Instance Normalization layers were introduced to replace the batch normalization layers. The unstable statistical estimation was avoided to reduce the internal covariate shift during small-batch training. The robustness and generalization were enhanced under complex and variable field conditions. Experimental results demonstrate that the IDCA-UNet significantly outperformed the various models, including the ResNet-UNet, YOLOv11, and DeepLabV3+ in the mung bean flower segmentation task, with the mIoU, mAP, and F1 scores of 88.95%, 93.47%, and 93.79%, respectively. There were the improvements of 1.62, 1.48, and 0.72 percentage points, compared with the VGG16-UNet benchmark model with the Focal loss. The high-precision segmentation was further quantitatively analyzed to study the dynamic growth patterns of the mung bean plants at the flowering and pod-setting stage. The flowering dynamics exhibited the "Λ"-shaped growth curve with a concentrated peak flowering period after daily flower count statistics and fitting, indicating the strong flowering synchrony within the population. Therefore, important evidence can be provided to assess the varietal maturity, consistency, and suitability for mechanical harvesting. The mung bean flower segmentation can also offer a practical cross-platform application for the organ-level crop phenotyping. Multi-source remote sensing data and temporal analysis can be integrated for environmental forecasting in the future. Thereby, the finding can serve as the robust theoretical and technical support to accurately assess the crop growth status, particularly for the high yield and efficiency breeding in precision agriculture.

An accurate and rapid identification of sheep periparturient behaviors can often be required to prevent the potential health risks and production abnormalities in modern animal husbandry. It is crucial to safeguard the ovine welfare for reproductive efficiency, thus reducing neonatal lamb mortality. However, some challenges still remain in the accurate recognition of the behavior during the periparturient period. Particularly, there is a high similarity between the behavioral traits and environmental interferences, such as the complex illumination and cluttered backgrounds in sheep farming. In this study, an advanced recognition was proposed to integrate an enhanced YOLOv8n-Pose key-point model with a backpropagation (BP) neural network. Specifically, an additional P2 detection layer was incorporated into the network architecture, in order to improve the precision of the key-point detection. The fine-grained and small-scale features were captured to accurately localize the anatomical key points in the complex behavioral scenarios. The detection layer was added for a higher degree of spatial resolution. Particularly, there were subtle movement variations in the periparturient sheep. Furthermore, a multi-scale attention block (MAB) module was introduced into the framework, in order to mitigate the feature representation in dynamic environments. A dynamic weighting module was employed to interactively learn the global and local spatial dependencies. Consequently, the robustness and generalization performance of the improved model was achieved under heterogeneous illumination. The MAB module effectively prioritized the most discriminative feature regions, thereby reducing the impact of the background noise and occlusions commonly observed in practical farming environments. The L1-norm channel pruning was systematically implemented to reduce the excessive parameters in the practical deployment constraints. The parameter compression was effectively optimized to eliminate the redundancy in the refined model. An optimal combination was achieved to balance computational efficiency and performance retention. The pruning was utilized to maintain the model integrity using structured sparsity, in order to significantly reduce the computational overhead. The real-time livestock monitoring was realized as suitable for edge computing. A multidimensional dataset of the behavioral feature was constructed to accurately extract the 12 key-point coordinates. Five joint angle parameters were integrated with two pairs of the key-point relative distance metrics, and the key-point detection confidence scores. The dataset was obtained with a 32-dimension feature vector. These feature representations were extracted to serve as the input into a BP neural network for the precise classification of the periparturient behaviors. The BP neural network was trained using adaptive learning. The complex spatiotemporal dependencies were effectively captured among the extracted features. The high classification accuracy was achieved after extraction. A series of experiments were conducted to evaluate the performance of the improved model on a self-developed dataset of periparturient sheep. The results demonstrated that the improved YOLOv8n-Pose model achieved a notable 4.6 percentage point increase in the mean average precision (mAP50) and a 6.7 percentage point improvement in mAP50:95 for the key-point detection, compared with the baseline architecture. Moreover, the BP neural network exhibited outstanding performance in the classification. An F1-score of 95.7% was obtained to distinguish the critical periparturient behaviors. The superior efficacy of the key-point recognition was obtained to identify the periparturient behavior. Ultimately, the robust technical framework greatly contributed to the intelligent livestock systems. Full automation and precision monitoring were enhanced in sheep farming.

A knowledge graph is often required in the field of sheep diseases, particularly for disease prevention, control, and intelligent diagnosis. The common challenges still remained in the sheep disease texts, such as the ambiguous semantic boundaries, overlapping entity roles, and complex relational semantics. In this study, a knowledge graph was constructed using the CaRoMHPE model (CasRel-based Model Combined with RoBERTa, Multi-scale Cross-Attention Mechanism, and Hybrid Position Encoding in Multi-Head Attention). Firstly, a dataset was constructed with 9 types of entities and 8 types of relationships, according to the features of sheep disease corpora. The key entities and their relationships were also covered throughout the entire process, from disease diagnosis to treatment. Substantial data support was then provided for the entity-relationship extraction tasks. Secondly, the Robustly Optimized BERT Approach with Whole Word Masking (RoBERTa-wwm-ext) was used to replace the conventional Bidirectional Encoder Representations from Transformers (BERT) as the text encoder. The Cascade Relational Triple Extraction (CasRel) framework was established to enhance the contextual semantics in order to handle the complex linguistic structures. A multi-scale cross-attention mechanism was introduced after the subject annotation module. Taking the subject embeddings and full-sentence embeddings as the inputs, the interaction scores were computed using multi-head attention. Multi-source information was fused through residual connections and layer normalization. Furthermore, the high-level semantic representations were extracted via a Feed-Forward Network (FFN). The FFN consisted of two fully connected layers with a GeLU activation function in between. The multi-scale feature fusion module was obtained in the low-dimensional features after linear projection. The original and compressed features were concatenated along the last dimension. And then the original dimensionality was restored after the linear transformation, thereby integrating the global and local semantic information. There was a more refined characterization of semantic relationships between entities in order to enhance the perception of the long-range dependencies and local semantic features. Additionally, a Hybrid Position Encoding (HPE) was incorporated into the multi-head attention mechanism in order to combine the absolute and relative position encoding. There were the relative position bias and Rotary Position Encoding (RoPE). Among them, the relative position bias dynamically adjusts the bias matrix shape using sequence length and maps positional relationships into the attention scores between queries and keys. The RoPE applied the sine and cosine functions to the even and odd dimensions of query and key vectors for the rotational transformation, and then explicitly encoded the positional information into the attention calculation. More accurate modeling of the structural relationships between words enhanced the perception of the positional information. The better performance significantly improved the entity boundary recognition and role differentiation. Experimental results show that the precision, recall, and F1 scores were achieved in 94.70%, 94.04%, and 94.37%, respectively, in the knowledge extraction task, which was improved by 9.14, 9.21, and 9.18 percentage points over the original CasRel model. The significant optimization was realized in the entity-relationship extraction for the sheep disease domain. Finally, the semantic embedding vectors were generated by the RoBERTa model. Cosine similarity was calculated for the synonym discrimination and entity fusion using the structured triples extracted from the text. The knowledge fusion corpora were obtained for different entity types. Synonymous expressions were effectively identified and unified to eliminate duplicate entities and potential ambiguities. No duplicates were verified after retrieval. The fused triples were stored in a Neo4j graph database, resulting in a standardized, consistent, and well-structured sheep disease knowledge graph. This finding can provide reliable data support for the subsequent intelligent diagnosis and treatment of sheep diseases.