Detecting and locating apples are important for picking robots and orchard management. Although fully-supervised object detection (FSOD) methods have achieved impressive apple detection performance, the required annotated large-scale datasets are expensive. To balance the annotation cost and detection performance, a contrastive-learning based apple detection method with sparsely annotated dataset (CLAD-SAD) is proposed, in which a Similarity-based Group Pseudo Label Generating (SGPLG) module is exploited to take advantage of unlabeled instances by generating pseudo labels based on the similarity between existing instances and unlabeled regions. Furthermore, a contrastive learning loss function within the heatmap prediction branch is used to suppress noise from pseudo labels. In addition, Aspect Ratio Filtering Module (ARFM) is added to filter out the low-quality prediction boxes with imbalanced aspect ratios. By this, the proposed CLAD-SAD markedly improve the apple detection performance with limited annotated dataset. Experimental results show that our proposed method improves mAP and mAR by 6.7% and 3.7%, respectively, on sparsely annotated datasets compared to the traditional FSOD models. This indicates that the proposed method is feasible to apple detection with limited annotated dataset, which addressing the challenge of traditional models' reliance on large-scale annotated datasets.

Automatic non-contact body measurement of livestock is crucial for precision farming, yet existing methods rely on complex multi-view camera systems or suffer from incomplete point cloud representation and inaccurate keypoint detection from single-view acquisitions. This paper presents GoatPoinTr, a joint point cloud completion and keypoint detection network built upon AdaPoinTr for single-view goat body measurement. An online back-projecting masking strategy with viewpoint jittering is proposed to realistically simulate single-view acquisition during training by projecting complete point clouds onto virtual camera planes and retaining only visible surface points. Outlook Attention is introduced to improve fine-grained geometric detail recovery, particularly for thin structures such as limbs. A keypoint detection head is further integrated into the network through a cross-optimization mechanism that enables bidirectional refinement between point cloud completion and body measurement keypoint prediction within each forward pass. A goat point cloud benchmark comprising 5,712 samples from 82 goats with 10 manually annotated body measurement keypoints is established to support point cloud completion and body size estimation research. Extensive experiments demonstrate that GoatPoinTr achieves CD-l1 of 9.326 and CD-l2 of 0.264, reducing CD-l1 by 13.4% and CD-l2 by 33.3% compared to AdaPoinTr. The mean absolute percentage errors across seven body dimensions range from 2.66% to 5.55%, comparable to multi-view methods and meeting livestock measurement accuracy standards.

Precise identification of individual cows is a fundamental prerequisite for comprehensive digital management in smart livestock farming. While existing animal identification methods excel in controlled, single-camera settings, they face severe challenges regarding cross-camera generalization. When models trained on source cameras are deployed to new monitoring nodes characterized by divergent illumination, backgrounds, viewpoints, and heterogeneous imaging properties, recognition performance often degrades dramatically. This limits the large-scale application of non-contact technologies in dynamic, real-world farming environments. To address this challenge, this study proposes a cross-camera cow identification framework based on disentangled representation learning. This framework leverages the Subspace Identifiability Guarantee (SIG) theory in the context of bovine visual recognition. By modeling the underlying physical data generation process, we designed a principle-driven feature disentanglement module that decomposes observed images into multiple orthogonal latent subspaces. This mechanism effectively isolates stable, identity-related biometric features that remain invariant across cameras, thereby substantially improving generalization to unseen cameras. We constructed a high-quality dataset spanning five distinct camera nodes, covering heterogeneous acquisition devices and complex variations in lighting and angles. Extensive experiments across seven cross-camera tasks demonstrate that the proposed method achieves an average accuracy of 86.0

Accurate body measurement is essential for dairy goat breeding and growth assessment, yet existing automated methods are limited by slow processing and difficulty in handling geometric distortions caused by non-standard postures. To address these challenges, a lightweight integrated approach for dairy goat body measurement is proposed. Based on our proposed keypoint detection model LEKDNet (Lightweight Edge Keypoint Detection Network), this model integrates Star Block, C2PSM, and Slim-Neck modules to achieve real-time inference on edge devices with only 1.88 million parameters, maintaining 98.7 % accuracy. By combining detected keypoints with depth images, six body sizes: withers height (WH), hip height (HH), chest width (CW), hip width (HW), body length (BL), and chest depth (CD) are extracted. Measurement errors were observed to increase and consistency to decrease under non-standard postures when evaluated on the GBSK2025 dataset of 77 dairy goats, indicating that pose variations exert a notable effect on measurement accuracy. To address this, 16 skeletal keypoints from the body of dairy goat are further extracted to compute pose feature sets, and a stacking ensemble regression model comprising Neural Network, XGBoost, and Random Forest as base regressors with a Gradient Boosting meta-learner is employed to predict and correct posture-induced measurement errors. After calibration, the mean absolute errors of WH, HH, CW, HW, BL, and CD were reduced to 1.78 cm, 1.53 cm, 1.15 cm, 1.12 cm, 1.73 cm, and 1.22 cm. These results indicate that the proposed calibration method achieved consistent error reduction across all body sizes, thereby enhancing measurement stability and reliability. To implement the aforementioned proposed method in practical pasture applications, a channel-based measurement system equipped with top-mounted and side-mounted depth cameras is developed. This integrated system achieves cumulative delays of only 2.5–3.8 s per cycle for image acquisition, body measurement, and error correction. These findings demonstrate that the proposed method provides a reliable and efficient solution for automated body measurement in the precision management of dairy goats.

Automated monitoring of livestock behavior is critical for enhancing animal welfare, optimizing productivity, and reducing operational costs in modern farming systems. While traditional manual observation methods are labor-intensive, subjective, and costly, recent advancements in artificial intelligence offer promising alternatives. This study presents a novel approach to livestock behavior analysis by adapting state-of-the-art multimodal large language models (LLMs) for video interpretation. We developed and evaluated a customized framework based on the MiniGPT4-Video architecture. Through adaptive multimodal fusion technology, efficient frame sampling strategy, and domain-specific fine-tuning mechanism, it is deeply integrated with LLaMA to form MiniGPT4-Video LLaMA, and fine-tuned on a comprehensive farm animal behavior dataset. Our system demonstrates superior performance in recognizing and describing key behavioral patterns (e.g., feeding, locomotion, resting) compared to conventional video captioning methods. The proposed solution offers a robust, scalable, and cost-effective platform for real-time livestock monitoring, with potential applications in precision livestock farming and automated welfare assessment.

Existing methods for detecting cotton boll diseases frequently exhibit high rates of both false negatives and false positives under complex field conditions (e.g., lighting variations, shadows, and occlusions) and struggle to achieve real-time performance on edge devices. To address these limitations, this study proposes an enhanced YOLOv11n model (YOLOv11n-ECS) for improved detection accuracy. A dataset of cotton boll diseases under different lighting conditions and shooting angles in the field was constructed. To mitigate false negatives and false positives encountered by the original YOLOv11n model during detection, the EMA (efficient multi-scale attention) mechanism is introduced to enhance the weights of important features and suppress irrelevant regions, thereby improving the detection accuracy of the model. Partial Convolution (PConv) is incorporated into the C3k2 module to reduce computational redundancy and lower the model’s computational complexity while maintaining high recognition accuracy. Furthermore, to enhance the localization accuracy of diseased bolls, the original CIoU loss is replaced with Shape-IoU. The improved model achieves floating point operations (FLOPs), parameter count, and model size at 96.8%, 96%, and 96.3% of the original YOLOv11n model, respectively. The improved model achieves an mAP@0.5 of 85.6% and an mAP@0.5:0.95 of 62.7%, representing improvements of 2.3 and 1.9 percentage points, respectively, over the baseline YOLOv11n model. Compared with CenterNet, Faster R-CNN, YOLOv8-LSW, MSA-DETR, DMN-YOLO, and YOLOv11n, the improved model shows mAP@0.5 improvements of 25.7, 21.2, 5.5, 4.0, 4.5, and 2.3 percentage points, respectively, along with corresponding mAP@0.5:0.95 increases of 25.6, 25.3, 8.3, 2.8, 1.8, and 1.9 percentage points. Deployed on a Jetson TX2 development board, the model achieves a recognition speed of 56 frames per second (FPS) and an mAP of 84.2%, confirming its suitability for real-time detection. Furthermore, the improved model effectively reduces instances of both false negatives and false positives for diseased cotton bolls while yielding higher detection confidence, thus providing robust technical support for intelligent cotton boll disease detection.

In complex orchard environments, the phenotypic heterogeneity of different apple leaf diseases, characterized by significant variation among lesions, poses a challenge to traditional multi-scale feature fusion methods. These methods only integrate multi-layer features extracted by convolutional neural networks (CNNs) and fail to adequately account for the relationships between local and global features. Therefore, this study proposes a multi-branch recognition framework named CNN-Transformer-CLIP (CT-CLIP). The framework synergistically employs a CNN to extract local lesion detail features and a Vision Transformer to capture global structural relationships. An Adaptive Feature Fusion Module (AFFM) then dynamically fuses these features, achieving optimal coupling of local and global information and effectively addressing the diversity in lesion morphology and distribution. Additionally, to mitigate interference from complex backgrounds and significantly enhance recognition accuracy under few-shot conditions, this study proposes a multimodal image-text learning approach. By leveraging pre-trained CLIP weights, it achieves deep alignment between visual features and disease semantic descriptions. Experimental results show that CT-CLIP achieves accuracies of 97.38

Body size measurement plays a crucial role in dairy cow breed selection and milk production. Employing intelligent systems for periodic assessments of body size empowers farmers to gauge the nutritional status of cows. The study introduces an end-to-end intelligent approach for the automatic measurement of cow body size via keypoint detection. Introducing Cow Keypoint-Net (CowK-Net), a one-stage dairy cow keypoint detection network. To improve the interaction of cow features at the channel level, we created the Keypoint Refine Machine (KPRM), designed to balance channel and spatial information through separate pathways effectively. Moreover, we devised an efficient hybrid encoder to interact the information across different scales. This encoder combines Convolutional Neural Network (CNN) based cross-scale fusion with Transformer-based intra-scale interaction, thereby optimizing the keypoint processing and integration. Customizing the loss function to the specific characteristics of the cow dataset ensures effective supervision of the keypoint prediction process. Additionally, we transformed the pixel coordinates of keypoints into three dimensions (3D) space, enabling automated measurement of body size. Field testing on a production farm revealed CowK-Net's accuracy, achieving an impressive 92.8%, surpassing existing keypoint detection methods. Notably, the hybrid encoder matched the accuracy of a Transformer-based encoder while reducing the number of parameters by 18%. Compared to manual measurements, our method demonstrated mean relative errors of 2.8%, 6.7%, 4.1%, and 4.4% for oblique body length, body height, hip height, and chest depth, respectively. The CowK-Net demonstrates its efficacy in measuring cow body size, laying solid foundation for the development of body measurement devices.

Detection and tracking of individual plants in the field are crucial for precision agriculture, serving as foundational technologies for tasks like precision fertilization, precision weeding, and targeted irrigation. Existing models in agricultural scenarios tend to be highly complex, which hinders their deployment on embedded or small mobile devices. This study proposes a method for individual plant detection and tracking based on improved YOLOv8 model. First, the Partial Convolution (PConv) module is introduced to reduce the computational load and significantly decrease memory consumption. Additionally, the Spatial Pyramid Pooling-Fast (SPPF) is replaced with the Simplified SPPF (SimSPPF) structure to enhance the model's ability to recognize plants at different scales. Furthermore, a Normalization Attention Module (NAM) is added further to improve the model's attention to the object plants. Building on this, we integrate the StrongSORT object tracking algorithm with YOLOv8 to achieve precise tracking of individual plants in dynamic scenarios (e.g., varying light and background). Experimental results show that our method achieves a mAP50 of 93.60% in detection tasks, with a 22.58% reduction in model size, outperforming mainstream algorithms. In tracking tasks, the model achieves a maximum MOTA of 94.55% in real farmland scenarios. Furthermore, field tests were conducted using a Jetson Orin NX-based mobile platform under varying lighting conditions and in unseen cotton fields. The model maintained robust performance across these scenarios, demonstrating its generalization ability and supporting robotic-based intelligent weeding and precision task management in real-world agricultural environments.

Real-time detection of fresh corn ear height can provide a basis for dynamic adjustment of harvester header parameters, reducing mechanical damage and improving harvest quality. This study proposes a corn ear height detection model (CEHD). A YOLO-HAMDF network is developed for ear recognition, in which the core modules—TBDA, GLSA, and AQE—respectively suppress background interference, enhance contextual perception, and optimize bounding-box scoring. Depth information is incorporated to filter non-target regions and improve system robustness. In addition, a DI-DeepSORT module is designed for ear tracking, where DBC-Net and IDA-Kalman, respectively, enhance the discriminability of ReID features and enable independent-dimension adaptive noise modeling with smoothed positional updates. Experimental results demonstrate that the proposed CEHD model achieves a mean absolute error (MAE) of only 3.21 ± 0.05 cm under field conditions, indicating strong stability and practical applicability. In summary, this study presents a stable and reliable corn ear height detection system, achieves real-time monitoring of ear height, and provides data support for the dynamic adjustment of header parameters in fresh corn harvesters.

Apple yield estimation is a critical task in precision agriculture, challenged by complex tree canopy structures, growth stage variability, and orchard heterogeneity. In this study, we apply multi-source feature fusion by combining vegetation indices from UAV remote sensing imagery, structural feature ratios from ground-based fruit tree images, and leaf chlorophyll content (SPAD) to improve apple yield estimation accuracy. The DeepLabv3+ network, optimized with Convolutional Block Attention Module (CBAM) and Efficient Channel Attention (ECA), improved fruit tree image segmentation accuracy. Four structural feature ratios were extracted, visible-light and multispectral vegetation indices were calculated, and feature selection was performed using Pearson’s correlation coefficient analysis. Yield estimation models were constructed using k-nearest neighbors (KNN), partial least squares (PLS), random forest (RF), and support vector machine (SVM) algorithms under both single feature sets and combined feature sets (including vegetation indices, structural feature ratios, SPAD, vegetation indices + SPAD, vegetation indices + structural feature ratios, structural feature ratios + SPAD, and the combination of all three). The optimized CBAM-ECA-DeepLabv3+ model achieved a mean Intersection over Union (mIoU) of 0.89, an 8% improvement over the baseline DeepLabv3+, and outperformed U2Net and PSPNet. The SVM model based on multi-source feature fusion achieved the highest apple yield estimation accuracy in small-scale orchard sample plots (R2 = 0.942, RMSE = 12.980 kg). This study establishes a reliable framework for precise fruit tree image segmentation and early yield estimation, advancing precision agriculture applications.

With the progress of computer vision and machine learning, the research of object detection and pedestrian recognition has demonstrated significant performance. However, the identification studies in domestic animals, especially in the same species of domestic animals, remains a significant challenge. His study focuses on distinguishing cashmere and dairy goats, which share similar traits. Our contributions are: (1) Proposing a dynamic quadruple optimization algorithm to optimize goat images from local and global dimensions, enhancing network representation with a multi-branch structure; (2) Introducing a novel transfer learning algorithm based on goat granularity to preview dataset knowledge; (3) Validating our approach on our goat dataset and a public bird dataset. We achieved recognition accuracies of 95% for cashmere goats, 94.04% for dairy goats, and 82.48% on the public dataset, demonstrating the effectiveness of our methods for animal biometric identification.

Precise monitoring of pig behavior has become pivotal for enhancing animal welfare and breeding efficiency. However, existing studies predominantly focus on behavior recognition while neglecting environmental influences, and lack specialized image captioning models and datasets tailored for farm scenarios, hindering textual analysis of behavior-environment interactions. In this study, a multimodal image captioning model was proposed to generate semantic textual descriptions of pig behavior, thereby supporting smart decision-making in farm management. The model employs a ResNet-18 encoder to extract pig visual biometric features from RGB and depth images, coupled with an innovative decoder integrating an enhanced Long Short-Term Memory (LSTM) network and Graph Convolutional Network (GCN) for pig behavior textual description, effectively resolving the input inconsistency between training and inference phases in traditional Encoder-Decoder architectures. Additionally, a dedicated pig behavior dataset comprising 9,052 annotated images was constructed, covering four behavioral categories: standing, sitting, lying, and eating. The experimental results show that the proposed approach achieves a METEOR score of 88.25%, which is outperformed baseline models by up to 21.58%. By recognizing pig behavior and interpreting environmental context, the proposed approach introduces a practical methodology for analyzing behavior-environment interactions and facilitates the integration of LLM-embedded robotic systems into smart livestock farming.

Accurately assessing the force exerted on fresh waxy corn kernels provides a foundation for adjusting harvester parameters, minimizing mechanical damage to the corn, and improving harvest quality. The research proposes a detection method based on an improved gramian angular field-based Siamese network (Siamese MGAD-Net), which, combined with hyperspectral imaging technology, enables high-precision prediction of force applied to corn kernels. The method integrates an improved gramian angular field (MGAF) with a Siamese network architecture. The image encoding process is optimized over traditional GAF by introducing multi-scale sliding window normalization and Gaussian weighting mechanisms. The prediction module incorporates a pyramid structure, attention mechanism, and a multi-scale dilated convolution downsampling module (MDCDS). At the end of the Siamese network, the processed data from the improved gramian angular summation field (MGASF) and gramian angular difference field (MGADF) are fused at the end of the Siamese network, enhancing the model’s generalization capability and prediction accuracy. The experimental results demonstrate that the Siamese MGAD-Net achieved Rp2 values of 0.9543, 0.9580, 0.9570, and 0.9647 on three single-variety datasets and a combined dataset of three varieties, respectively. Compared to traditional machine learning and CNN models, it demonstrated remarkable robustness and reliability, particularly when handling low-discrimination spectral features and small-scale datasets, effectively improving the accuracy of corn kernel force prediction. This study provides significant theoretical and technical support for the further development of force prediction technology for fresh waxy corn kernels and the optimization of harvester parameters for fresh waxy corn.

Continuous development of convolutional neural networks has shown good performance for fine-grained image classification by identifying fine features in high-resolution images.However, in the real world, many images are due to camera or environmental restrictions. Low resolution images with fewer fine features result in a dramatic reduction in classification accuracy.In this study, a twophase Data Augmentation guided Decoupled Knowledge Distillation (DADKD) framework is proposed to improve classification accuracy for low-resolution images.In the proposed DADKD, one phase is data augmentation that generates a composite image and corresponding labels. Another stage is knowledge distillation, which minimizes differences between high-resolution and low-resolution image features. The proposed DADKD validated on three fine-grained datasets (i.e Stanford-Cars, FGVC-Aircraft, and CUB-200-2011 datasets). Experimental results show that our proposed DADKD achieves 88.19%, 78.98% and 80.33% classification accuracy on these three datasets, state-of-the-art methods such as SnapMix and Decoupled Knowledge Distillation (DKD).The method proposes a viable solution for fine-grained classification at low resolution.

Nitrogen (N), phosphorus (P) and potassium (K) are important macronutrients to crops, and hence, in situ, timely and non-destructive estimation of their contents and distinguishing N, P, and K status is of critical prominence in precision farming for rational use of fertilizers. The main goal of this study was to proposes an accurate model to monitor leaf N, P, and K contents (i.e., LNC, LPC and LKC) utilizing canopy hyperspectral data of summer maize. Twelve field experiments were conducted over three consecutive growing seasons (2020-2022) at different sites (Yuanyang, Wen and Fangcheng county) in Henan, China, using different N, P, and K application rates, growing stages, cultivars and ecological sites. The in situ canopy raw hyperspectral (R) were acquired over a wavelength range from 325 to 1075 nm (the visible and near-infrared region). Continuous wavelet transform (CWT) was used to process the collected spectral reflectance; partial least square (PLS) and lambda-lambda r(2) (LL r(2)) models were applied to analyze the relationships between LNC, LPC, and LKC and the spectral reflectance. Results showed that CWT transformation technique can significantly improve the prediction accuracy of summer maize LNC, LPC, and LKC, and the best decomposition scales are CWT-1, CWT-3, and CWT-1. The CWT-PLS model for LNC, LPC, and LKC prediction in the three decomposition scales yielded a relatively higher accuracy compared to the canopy R based on the full range hyperspectra, however, the prediction accuracy varied greatly among the three nutritional status, the effect of the LNC was the best, LKC was the second. The coefficient of determination of the validation datasets (R-val(2)) were 0.821, 0.732 and 0.773 for LNC (CWT-1-PLS), LPC (CWT-3-PLS), and LKC (CWT-1-PLS) prediction, and the relative percentage deviations (RPDval) were 2.176, 1.900, and 2.041, respectively. Eventually, ten bands centred at 405, 517, 560, 660, 685, 735, 750, 770, 838 and 875 nm; ten at 442, 479, 575, 630, 700, 730, 795, 838, 858 and 870 nm; and ten at 479, 540, 597, 653, 695, 755, 808, 858, 870 and 890 nm were selected as effective wavelengths for predicting the LNC, LPC and LKC values. The newly-developed CWT-PLS models for LNC (R-val(2) = 0.780, RPDval = 1.730), LPC (R-val(2) = 0.704, RPDval = 1.434), and LKC (R-val(2) = 0.722, RPDval = 1.725) also provided relatively accurate estimations (RPD > 1.40) based on field experiment validations using the effective wavelengths. The findings will provide theoretical basis and effective methodologies for using the in situ canopy hyperspectral technique to accurate and nondestructive estimation of LNC, LPC, LKC and analyzing N, P, and K nutrient stresses of summer maize.

With the increasing deployment of agricultural robots, the traditional manual spray of liquid fertilizer and pesticide is gradually being replaced by agricultural robots. For robotic precision spray application in vegetable farms, accurate plant phenotyping through instance segmentation and robust plant tracking are of great importance and a prerequisite for the following spray action. Regarding the robust tracking of vegetable plants, to solve the challenging problem of associating vegetables with similar color and texture in consecutive images, in this paper, a novel method of Multiple Object Tracking and Segmentation (MOTS) is proposed for instance segmentation and tracking of multiple vegetable plants. In our approach, contour and blob features are extracted to describe unique feature of each individual vegetable, and associate the same vegetables in different images. By assigning a unique ID for each vegetable, it ensures the robot to spray each vegetable exactly once, while traversing along the farm rows. Comprehensive experiments including ablation studies are conducted, which prove its superior performance over two State-Of-The-Art (SOTA) MOTS methods. Compared to the conventional MOTS methods, the proposed method is able to re-identify objects which have gone out of the camera field of view and re-appear again using the proposed data association strategy, which is important to ensure each vegetable be sprayed only once when the robot travels back and forth. Although the method is tested on lettuce farm, it can be applied to other similar vegetables such as broccoli and canola. Both code and the dataset of this paper is publicly released for the benefit of the community: https://github.com/NanH5837/LettuceMOTS.

The body size of livestock is an essential phenotypic trait in genetic breeding, gene improvement, health screening, and animal welfare. To develop a non-contact automatic system for measuring goat body traits, we propose a point-cloud segmentation model based on an improved PointStack, which segments the automatically acquired three-dimensional (3D) point-cloud data of goats into different parts, including the head, front legs, hind legs, chest, abdomen, hip, and tail. The segmented point cloud, along with the physiological features of the goat, is then used to locate the corresponding key points for body size measurement. A novel method for key point localisation is proposed that includes coordinate normalisation, retrieval of key clusters, key point adjustment, optimisation of the traveling salesman problem, and edge detection. These methods were designed to reduce discrepancies at crucial points of body features, thereby facilitating the precise computation of the body size parameter in goats. In this work, 326 point clouds representing the upright posture of 55 goats were used for segmentation and body size measurement testing. The proposed segmentation model achieved a mean intersection over union of 89.21% and accuracy of 94.54%, outperforming comparative models. In the body traits measurement experiment, mean absolute percentage errors for body length, body height, chest width, chest girth, hip height, and hip width were recorded as 3.24%, 2.54%, 5.43%, 3.08%, 2.16%, and 4.59%, respectively. In summary, the proposed automated measurement method demonstrates high accuracy, strong robustness, and holds significant potential for widespread application.