The paper proposed a real-time monitoring system based on digital twin aiming to address the low level of information visualization and the challenges in equipment data collection during the operation of the apple harvesting robot. The system established a digital twin model of the robot by considering its overall structure and kinematic analysis results. It utilized multidimensional visualization monitoring modules and the twin model to create a digital twin scene, enabling dynamic mapping and virtual-real interaction between the twin model and the physical robot. By integrating the robot's ROS software architecture and TCP communication protocol, an information bridge was established between the twin model and the apple harvesting robot. We further developed a robot twin monitoring system supporting Unity and Web terminal programs, which underwent functional testing. The test results showed that the average frame time of the system was 14.5 ms with a displacement difference less than 0.05 m for the virtual-real joints of the robot during the virtual-real synchronization process of the model and the physical entity. This study provides valuable insights into applying digital twin technology in agricultural robots.

Chlorophyll content can reflect the health status of green vegetables and promote the growth and development of vegetables. As an important vegetable crop, monitoring the growth status of Chinese cabbage is of great significance for improving yield and quality. In this study, nine color features and 24 spectral image combinations were constructed using a multispectral unmanned aerial vehicle, and the SPAD values of Chinese cabbage canopy were obtained simultaneously using a handheld SPAD chlorophyll meter. Four machine learning methods, including partial least squares, support vector regression, BP neural network, and 1D-convolutional neural network, were used to construct the Chinese cabbage SPAD value estimation model. The accuracy of the model was evaluated by the determination coefficient (R²), root mean square error (RMSE), and mean absolute error (MAE). The results showed that the prediction accuracy combining color features, visible light, and multispectral image features was higher than that of a single feature. Among them, the support vector machine-based Chinese cabbage canopy SPAD prediction model showed the highest accuracy with R²=0.785, RMSE=4.320, and MAE=3.451. The conclusion drawn from the comprehensive analysis was that selecting multiple visible light and multispectral image feature combinations as input variables and using the support vector machine model can significantly improve the accuracy of SPAD value estimation, which provids new technical support for rapid and accurate monitoring of Chinese cabbage SPAD values.

The irrigation system is an innovative technology in alleviating the problems of excessive irrigation and waste of water resources in traditional greenhouse irrigation system. The purpose is to save water resources under assurance of the suitable soil moisture environment for strawberry growth. In this paper, a kind of irrigation equipment based on STM32 chip and RS485 sensor was proposed. In view of the non-linearity, large inertia and certain hysteresis of the irrigation system, fuzzy self-adaptive PID control algorithm was put forward. Combined with the regulating valve, a real-time, automatic and accurate irrigation system was designed to adjust the irrigation amount according to the change of soil moisture. Finally, a mathematical model was established according to the actual irrigation law of greenhouse strawberry in order to verify the effectiveness of the system. By comparing the effect of traditional PID and fuzzy self-adaptive PID in simulating irrigation model through MATLAB/Simulink, the fuzzy self-adaptive PID was 7.98% less than that of the traditional PID in overshoot, and maintained with ±0.5% steady-state error. The results showed that the PID control algorithm improved by fuzzy rules applied in greenhouse strawberry irrigation system was beneficial to reduce the overshoot and steady-state error. In practice, this system can effectively avoid the waste of water resources.

The movement status of the sheep can reflect its health status, and tracking the movement trajectory of the target sheep in the farm environment is a key step to obtain its movement status. Automated tracking of a small number of targets is practical in specific situations, such as before farrowing or during disease treatment. In order to improve the accuracy of multi-target tracking, an improved DeepSORT-R multi-object tracking algorithm was proposed. In the target recognition stage, the YOLOV5-CBAM network with added attention mechanism was used to realize the detection of sheep targets. In the re-identification stage, the ResNet50 network with added attention mechanism was used to realize the identification of sheep identity. The experimental results of the proposed method for the test set showed that MOTA, MOTP and IDSW reached 76.15%, 0.208 and 7.5, respectively. In addition, the proposed method was better in the long video tracking test than the commonly used YOLOV4+DeepSORT and ByteTrack in the scores of the evaluation indicators MOTA, MOTP and IDSW. The experimental results showed that the proposed method can be used for the tracking of multi-target sheep in the actual breeding environment.

To address the challenges of low accuracy and difficulty in handling complex environments in tomato leaf disease recognition using traditional methods, this paper proposed an Attention-based Multi-task Tomato Leaf Disease Recognition Method (AMTDR). Firstly, ResNet18 was adopted as the backbone network, with Convolutional Block Attention (CBA) modules introduced after each residual block. Secondly, a multi-task structure was designed with disease recognition and disease severity branches. The disease recognition branch accurately identified the type of tomato leaf disease, while the disease severity branch assessed the severity of the disease. In each branch, Convolutional Triplet Attention (CTA) modules were introduced to enhance the representation capability of disease features. Experimental results demonstrated that the proposed AMTDR method achieved an accuracy and F1 score of 98.54% on a dataset containing 11 types of tomato diseases in complex environments. Compared with the ResNet50 network, the accuracy and F1-score were improved by 1.27% and 1.25%, respectively, while the parameter count and FLOPs were only 48.72% and 44.30% of ResNet50. The AMTDR method effectively identified tomato leaf diseases in complex environments, providing significant value for agricultural disease recognition.

Molecular property prediction has a wide range of applications in drug research and development. Although graph neural networks and other methods have been applied to predict molecular properties, there are still limitations in processing large-scale molecular maps and information dissemination. To solve this problem, a network model was constructed fusing graph neural network (GNN), gated circulation unit (GRU) and attention mechanism (GAGCN) in this paper to predict molecular properties. The model used GNN to represent and learn molecular graphs, and the connections between nodes and information propagation was used to capture molecular structural features. The GRU was used to model the molecular sequence, so that the timing information in the molecular sequence was captured, and the information in the sequence was adapted to retain or discard through the gating mechanism. Finally, the attention mechanism was used to learn the weights between different features, and the GNN and GRU were integrated, so that the model can make full use of the molecular structure and sequence information to improve the accuracy of molecular property prediction. The experimental results showed that the prediction accuracy of MSE, MAE and R² for LogP attributes were 0.001 0, 0.011 6 and 0.999 3, respectively. The model proposed in this paper provides technical support and reference for the research and development of new pesticides and new veterinary drugs.

Aiming at the problems of low registration accuracy of animal point clouds and the need to manually set parameters, this paper proposed a precise registration method for cow point clouds from top to side view based on improved dolphin echolocation algorithm. Two strategies were used to improve the original algorithm: Strategy 1 improved the symmetric linear coefficient function to a nonlinear function; Strategy 2 dynamically adjusted the influence radius, introduced memory matrix and reduced the range of variable values for secondary discretization. According to the characteristics of less overlapping areas of point clouds from different perspectives of cows after coarse registration, the fitness function was established based on the centroid of overlapping areas of point clouds slices after coarse registration. An improved dolphin echolocation algorithm was used to test the precise registration of 106 cow point clouds. The results showed that compared with the standard dolphin echolocation algorithm, the registration error and registration time of strategy 1 were reduced by 27% and 28% respectively. Using strategy 2, the mean registration error were reduced by 56% and the registration time was increased by 30%. The mean and median registration errors of the proposed algorithm are 0.51 cm and 0.49 cm, which were 64% and 58% lower than that of the original algorithm, respectively. The average registration time was 1.31 s, which was slightly lower than that of the original algorithm. The proposed method can accurately realize the automatic registration of the overhead and side-view point clouds of cows, which provides technical support for animal 3D point cloud reconstruction.

This article proposed a pear blossom recognition method based on improved YOLOv5s in natural environments to address the low recall rate caused by dense pear blossoms, severe occlusion, and small targets. This method first added a small target detection layer. The small target detection layer was a shallow output feature layer added to the CSPDarknet backbone feature extraction network, and further feature fusion was performed on this shallow feature layer in the PANet enhanced feature extraction network. This enhanced the ability to extract shallow features and detailed information. Secondly, the CBAM attention module was introduced into the PANet network to improve the expression ability of important features. The experimental results showed that the improved YOLOv5s network model in this paper could reduce the missed recognition rate. The accuracy, recall, F₁ value, and mAP of the improved model were 91.62%, 83.05%, 87.12%, and 94.06%, respectively, which were 0.16%, 1.55%, 0.93%, and 0.61% higher than that of the original model. In addition, better recognition results could be achieved on pear flower images of three varieties: Xueqing, Yali, and Qiuyue. This model has strong generalization ability and provides technical support for machine intelligent thinning of pear orchards.

In order to reduce the invalid genes in high-dimensional genes and mine the key genes for abiotic stress resistance of crops, a multi-objective genetic programming based method was proposed to mine the key genes for abiotic stress resistance of crops. A method for feature selection and classification of high-dimensional data based on multi-objective genetic programming (MONSGP) was proposed by combining genetic programming (GP) and multi-objective optimization algorithms NSGA-II. The method aimed to optimize recall, accuracy, and number of features to obtain the optimal Pareto solution set. The optimal classifier selection strategy was used to select the classifier with the best performance and a small number of features from the optimal solution set as the optimal classifier. Experiments on 9 NCBI high-dimensional gene datasets have showed that MONSGP can achieve better classification performance with fewer features compared to standard GP classification algorithms and three latest GP classification algorithms. GO functional enrichment analysis confirmed that the genes screened by MONSGP were related to abiotic stress and had biological significance.

In order to solve the problems of strong subjectivity, slow speed and low efficiency in manual identification of leaf color of Chinese cabbage, this study proposed a method of rapid and accurate classification and quantification of leaf color of Chinese cabbage by combining multi-spectral image processing with machine vision technology. The results showed that the original spectral data extracted by the 19-channel multispectral system contained more comprehensive and accurate information, and the SVM classification model established by the system showed the best classification effect. The accuracy of training set was 98.24%, and the accuracy of verification set was 87.18%. The continuous projection algorithm (SPA) was used to extract characteristic wavelength for analysis, and the Chinese cabbage samples collected by a 5-channel camera were selected to continue to continuously study the quantification of leaf color of Chinese cabbage. By extracting the RGB, HSV, LAB nine color feature values for data processing, the color of Chinese cabbage leaves can be accurately quantized by 0-100 values.

The poverty-returning risk of the poverty alleviation population is a major factor on the results of poverty eradication and rural revitalization. Accurate prediction of the potential poverty-returning risk of the poverty alleviation population plays a crucial role in guiding the implementation of policies, allocation of resources, and risk assessment. This paper proposed a prediction method based on Stacking ensemble learning for the poverty-returning risk of the poverty alleviation population. taking The monitoring data after desensitization of the poverty alleviation households in Province H was analyzed to identify and filter the key features that significantly affect the poverty-returning risk after correlation analysis and importance ranking of the data features, whose key features were adopted in inter-model correlation analysis of the independent models such as Random Forest(RF), Naive Bayes(NB), Support Vector Machine(SVM), etc. The Stacking ensemble learning prediction model was conducted with RF meta-learner using eXtreme Gradient Boosting(XGBoost), Adaptive Boosting(AdaBoost) and SVM that have lower correlation and higher prediction accuracy. The model was trained and validated by dividing 412 919 data into training and validation sets in 7:3, and the model effect was evaluated using accuracy, precision, recall and F1-Score. The experimental results showed that all evaluation indexes of the poverty-returning risk prediction model based on Stacking ensemble learning were better than that of a single model, and its prediction accuracy was improved by 3.64%, 10.96%, 3.15%, 2.29%, and 5.41% compared with RF, NB, SVM, XGBoost, and AdaBoost, respectively, and finally reached 95.65%, which verified the effectiveness of the method proposed in this paper. The study provided new solution ideas for consolidating and expanding the results of poverty eradication and improving the timeliness of returning to poverty dynamic monitoring and warning.

Defective seeds significantly affect seed quality and pricing, and their sorting and removal is an important part of seed quality detection. At present, the seed quality detection is mainly completed by manual operation, which is inefficient and subjective. Aiming at the need for rapid and accurate identification of corn seeds in appearance quality detection, this paper proposed an improved YOLOv5 target detection model with input of multi-spectral RGB+NIR+NIR1 imaging information of corn seeds to identify and classify appearance quality of corn seeds. By changing the Spatial pyramid pooling (SPP) structure in the YOLOv5 backbone network CSPDarknet, the efficiency of network model detection was improved, and the attention mechanism was used to strengthen the feature information fusion in the feature extraction network to improve the accuracy of network model detection. The test results showed that the comprehensive evaluation index F1 value of the improved model YOLOv5+SE+SPPF reached 96.71%, the mAP value reached 96.96%, the average time for each image detection was about 0.28 s, and the average time for each seed detection was about 20 ms, which provided a reference for achieving efficient and accurate seed quality detection and optimal grading, and can be applied to the intelligent seed sorting equipment.

In order to meet the needs of for model lightweight and high-precision detection for embedded devices, this paper proposed a group-raising pig target detection algorithm based on the improved YOLOv8n model. First, the C2fFB structure was introduced in the backbone network to reduce the amount of memory access and redundant calculations. Then the new Neck network was constructed with the BiFPN structure and the C2fSC module was introduced to further achieve deeper feature fusion and reduce of the spatial redundancy and channel redundancy of the fusion. Finally, SIoU was used to replace the original CIoU to improve the accuracy of the model. The experimental results showed that the F1 score, precision, recall rate, and average precision of the proposed algorithm were improved compared with the original algorithm by 3%, 1.8%, 3.5%, and 1.5%, respectively. And the number of parameters, calculation amount, and model size were reduced by 46.84%, 27.16%, and 44.71% respectively. Therefore, the algorithm model in this paper provides an efficient target detection solution for the intelligent breeding of group-raising pigs.

At present, the determination of the social hierarchy of the natural mating caged breeder roosters mainly relies on manual observation, which is time-consuming and labor-intensive. Therefore, this study took natural mating caged roosters as the research object and proposed a method based on the Improved Gray Wolf Optimization Algorithm (IGWO) combined with LGBM to identify the social hierarchy of chickens based on the activity quantity analysis. Firstly, the nine-axis inertial sensors were used to obtain the chicken behavior data from the sliding window extracted the chicken’s activity information, that was characterized by the combined acceleration and combined angular velocity of totaling 44-dimensional time-domain and frequency-domain features, which characterize. The nonlinear convergence factor and the competition strategy of the first wolf were introduced to improve the optimization ability of the Gray Wolf Optimization Algorithm, and dimensionality reduction of activity features was performed to reduce redundancy and improve the model recognition effect. The results showed that the IGWO-LGBM model accurately identified the social hierarchy of chickens. Among them, the precision, recall, and F1-score of social hierarchy recognition were 84.71%, 84.59%, and 84.57%, respectively, and the model accuracy was 84.57%, which were improved by 3.80%, 3.65%, 3.67%, and 3.64%, respectively, compared with the original dataset. The features selected after dimensionality reduction were used as the activity features. After clustering statistics according to activity intensity and fitting by the least squares method, it was found that the chicken social hierarchy was positively correlated with the proportion of high activity behavior and negatively correlated with the proportion of low activity behavior, which enriched the research on social hierarchy. This study is helpful for rapid identification of the social hierarchy of breeder roosters and provides a method for automatic identification of the social hierarchy of breeder roosters.

Foodborne pathogenic bacteria is one of the main pathogenic bacteria that threaten human health in refrigerated food, and it is a must-check item in food hygiene microbiology inspection. The traditional method of bacterial identification needs to be observe and counte by naked eyes after bacterial culture for national standard time, which is time-consuming and laborious because the observer is easy to get wrong counting due to eye fatigue. Moreover, traditional methods require special reagents for bacterial detection, which is costly and requires specialized knowledge to operate.To quickly and accurately detect small bacterial targets, this article proposed a new method for detecting foodborne pathogens - CBAM^1,2,3-YOLOv7. Firstly, an industrial camera was used to replace a microscope to capture images, and the captured images of Listeria monocytogenes were inputted into the optimized model for recognition. This model added CBAM attention mechanism to the YOLOv7 model, making the model more sensitive in the channel domain and enhancing feature extraction ability. To enhance contrast, training was conducted on deep learning models, including Faster RCNN, YOLOv4, YOLOv5, and YOLOv7. Compared with YOLOv7 model, the optimized model improved the accurate mean value by 0.52% and the recall rate by 0.27%. The results suggested that CBAM^1,2,3-YOLOv7 algorithm realized the high-precision identification of Listeria monocytogenes, which has guiding significance and reference value for the rapid detection of other foodborne pathogens.

Manually picking cucumbers is time-consuming and labor-intensive. Automatic cucumber picking can greatly improve the production efficiency. However, few effective methods have been found so far to accurately detect the freshness of cucumber fruits during the automatic picking process. This paper proposed a cucumber freshness detection method based on machine vision. The method used RGB image analysis to extract color features, connected area labeling to count the number of spines, applied gray-level co-occurrence matrix to extract texture features and used support vector machine algorithm to recognize and classify freshness. The average recognition accuracy on 320 images of Ningyang cucumbers reached 98.43%, suggesting that the cucumber freshness detection method proposed in this paper is effective.

The development of deep learning technology and convolutional neural network has provided a new solution for the rapid and accurate detection of crop diseases. In this paper, potato images were collected in the field, and the UNet semantic segmentation model was used to detect potato diseases. Two backbone network models VGG16 and ResNet50 were used, whose precision was 93.00%, F1 was 92.48% and 92.77%, and MPA was 94.47% and 94.42%, MIoU was 84.79% and 84.75%. An improved UNet semantic segmentation model was proposed. The feature map was obtained by adding an attention mechanism module at the first upsampling of the network, and the feature map obtained by the attention mechanism was multiplied by the original input feature map for the next step. During the sampling network process, the final Precision, F1, MPA and MIoU were 94.83%, 92.89%, 95.96% and 85.32%, respectively. Compared with the initial network, the index was improved, which provided a more comprehensive deep learning algorithm and model research basis for the identification and detection of potato leaf diseases in natural environment.

Aiming at the complex and dynamic characteristics of knowledge requirements in the stages of vegetable production, such as variety selection, pest and disease control, and field management, an active recommendation method for vegetable production knowledge services is proposed to provide effective support for personalized agricultural management. Taking the construction of knowledge graph as the goal, the dynamic user interest modeling based on self-attention mechanism, hierarchical embedding technology representation, and multi-hop reasoning mechanism of graph neural network (GNN) are combined to screen and aggregate complex relationship information layer by layer to achieve accurate recommendation of vegetable production knowledge services. The experimental results show that the proposed method is significantly superior to the existing advanced models in terms of recommendation accuracy and performance indicators, and can adapt to the dynamic knowledge requirements of different stages of vegetable production. This study provided an efficient solution for personalized agricultural management and intelligent decision-making, and new ideas and methods for the research of agricultural knowledge service recommendation methods.