Reliable individual identification is essential for precision livestock management, but sheep facial appearance changes substantially during growth, weakening non-contact recognition. This study evaluated whether early-life facial images can support longitudinal identification at later growth stages. A monthly face dataset of 30 Small-tailed Han sheep was collected from 1 to 12 months of age. Images from months 1–8, 9–10 and 11–12 were used for training, validation and independent testing, respectively. An identity-preserving generative alignment framework was developed to reduce growth-related appearance mismatch while maintaining discriminative facial information. On the independent test set, the proposed method achieved 95.6% Top-1 accuracy and 93.9% macro-F1, outperforming convolutional, augmentation-based, feature-alignment, CycleGAN-based and Transformer baselines. Verification analysis achieved an area under the receiver operating characteristic curve (AUC) of 0.978, an equal error rate (EER) of 4.2%, and a true accept rate (TAR) of 94.5% at a false accept rate (FAR) of 1 × 10⁻⁴. Genuine similarity decreased with growth interval but remained separated from impostor similarity. Robustness tests covering time gaps, device changes, viewpoint shifts and image perturbations showed remaining deployment challenges. A mobile prototype achieved 32-bit floating point (FP32) and 16-bit floating point (FP16) latencies of 210 and 150 ms.

The automatic feeding device is crucial in grassland livestock farming, enhancing feeding efficiency, ensuring regular and accurate feed delivery, minimizing waste, and reducing costs. The shape and size of pellet feed render it particularly suitable for the delivery mechanism of automated feeding troughs. The uniformity of pellet flow is a critical factor in the study of automatic feeding troughs, and optimizing the movement characteristics of the pellets contributes to enhanced operational efficiency of the equipment. However, existing research often lacks a systematic analysis of the pellet size characteristics (such as diameter and length) and flow behavior differences in pellet feed, which limits the practical application of feed troughs. This study optimized the angle of repose and structural parameters of the feeding trough using Matlab simulations and discrete element modeling. It explored how the stock bin slope and baffle opening height influence pellet feed flow characteristics. A programmable logic controller (PLC) and human–machine interface (HMI) were used for precise timing and quantitative feeding, validating the design’s practicality. The results indicated that the Matlab method could calibrate the Johnson–Kendall–Roberts (JKR) model’s surface energy. The optimal slope was found to be 63°, with optimal baffle heights of 28 mm for fine and medium pellets and 30 mm for coarse pellets. The experimental metrics showed relative errors of 3.5%, 2.8%, and 4.2% (for average feed rate) and 8.2%, 7.3%, and 1.2% (for flow time). The automatic feeding trough showed a feeding error of 0.3% with PLC-HMI. This study’s optimization of the automatic feeding trough offers a strong foundation and guidance for efficient, accurate pellet feed distribution.

Accurate assessments are vital for the effective conservation of desert steppe ecosystems, which are essential for maintaining biodiversity and ecological balance. Although geostatistical methods are commonly used for spatial modeling, they have limitations in terms of feature extraction and capturing non-linear relationships. This study therefore proposes a novel remote sensing framework that integrates geostatistical methods and machine learning to predict the Shannon–Wiener index in desert steppe. Five models, Kriging interpolation, Random Forest, Support Vector Machine, 3D Convolutional Neural Network and Graph Attention Network, were employed for parameter inversion. The Helmert variance component estimation method was introduced to integrate the model outputs by iteratively evaluating residuals and assigning relative weights, enabling both optimal prediction and model contribution quantification. The ensemble model yielded a high prediction accuracy with an R² of 0.7609. This integration strategy improves the accuracy of index prediction, and enhances the interpretability of the model regarding weight contributions in space. The proposed framework provides a reliable, scalable solution for biodiversity monitoring and supports scientific decision-making for grassland conservation and ecological restoration.

Precise recognition of individual ovine specimens plays a pivotal role in implementing smart agricultural platforms and optimizing herd management systems. With the development of deep learning technology, sheep face recognition provides an efficient and contactless solution for individual sheep identification. However, with the growth of sheep, their facial features keep changing, which poses challenges for existing sheep face recognition models to maintain accuracy across the dynamic changes in facial features over time, making it difficult to meet practical needs. To address this limitation, we propose the lifelong biometric learning of the sheep face network (LBL-SheepNet), a feature decoupling network designed for continuous adaptation to ovine facial changes, and constructed a dataset of 31,200 images from 55 sheep tracked monthly from 1 to 12 months of age. The LBL-SheepNet model addresses dynamic variations in facial features during sheep growth through a multi-module architectural framework. Firstly, a Squeeze-and-Excitation (SE) module enhances discriminative feature representation through adaptive channel-wise recalibration. Then, a nonlinear feature decoupling module employs a hybrid channel-batch attention mechanism to separate age-related features from identity-specific characteristics. Finally, a correlation analysis module utilizes adversarial learning to suppress age-biased feature interference, ensuring focus on age-invariant identifiers. Experimental results demonstrate that LBL-SheepNet achieves 95.5% identification accuracy and 95.3% average precision on the sheep face dataset. This study introduces a lifelong biometric learning (LBL) mechanism to mitigate recognition accuracy degradation caused by dynamic facial feature variations in growing sheep. By designing a feature decoupling network integrated with adversarial age-invariant learning, the proposed method addresses the performance limitations of existing models in long-term individual identification.

Intelligent sheep behavior monitoring is vital for modern husbandry, but faces severe challenges in natural pastures due to high-density flock occlusion. Traditional 2D skeleton-based networks often suffer from depth ambiguity and feature collapse, misclassifying static tremors as dynamic displacement. To overcome this, we propose a robust multimodal dual-stream framework using skeletal and local visual fusion. The architecture features an upstream spatial perception stage utilizing YOLOv11m-Pose. To reduce annotation costs and improve robustness, we introduce an Active Hard-Example Mining mechanism, explicitly retaining difficult samples with severe overlapping or edge truncation. For downstream behavioral decisions, a multimodal dual-stream architecture processes the targets. The Spatio–Temporal Kinematic Stream employs a Kinematic Denoising Engine, incorporating a 1D Gaussian filter and displacement dead-zone gate to purify 2D coordinates before feeding them into a BiLSTM network. Concurrently, the Spatial Visual Stream uses a ResNet-50 backbone on cropped RGB patches to capture essential spatial context, addressing the limitations of pure coordinates. Finally, a weighted Softmax layer integrates both streams. Experiments on a complex real-world dataset validate this approach. A baseline kinematic-only model achieved just 69.05% overall accuracy and 68.18% walking precision. In contrast, our dual-stream fusion network achieved 93.26% overall accuracy, elevating walking precision to 97.14% and the eating F1-score to 94.29%. By effectively decoupling similar static and dynamic behaviors, this study demonstrates the indispensability of local visual features, establishing a high-precision baseline for smart livestock monitoring.