Computers and Electronics in Agriculture最新文献

Non-contact body size measurement has become a hot research topic in intelligent livestock farming. In regard to body size measurement of large livestock, such as cattle, collecting a substantial number of point clouds is frequently involved. The direct calculation of all point clouds for body size measurement can be impacted as point clouds of different body parts may interfere with each other, which poses huge challenges for the positioning of key points and induces inaccurate positioning, resulting in measurement errors. The accuracy of body size measurement can be improved by segmenting point clouds of different body parts from each other, key measurement points can be precisely located, thus enhancing the accuracy of body size measurement. In this paper, we propose a network model initially trained for pig point cloud segmentation based on the Kernel Convolution integrated with Graph Attention Mechanism (KCGATNet for short), which, through transfer learning techniques, can also be used to achieve successful segmentation of various cattle point clouds using only 7 training samples. The model utilizes two core modules, Kernel Convolution (KC) and Point-based Graph Attention Mechanism (P-GAT), to extract local neighborhood features of point clouds. When using pig body point clouds as training data, it achieved precise segmentation of the head, ears, limbs, torso, and tail of pigs through a downsampling-upsampling architecture. On the test set of pig point clouds, Overall Accuracy (OA) reached 98.1% and mean Intersection over Union (mIoU) was up to 90.5%. Furthermore, when this model served as a pre-trained model and underwent transfer learning using 7 sets of annotated data of Simmental cattle, it achieved a mIoU of 90.1% on a test set of 93 Simmental cattle, 89.6% on a test set of 439 dairy buffalo, 90.2% on a test set of 103 Hereford cattle, and 90.0% on a test set of 119 Black Angus cattle. The experimental outcomes fully demonstrate the robustness of the proposed livestock point cloud segmentation model, KCGATNet. With transfer learning of a small sample size, it can reliably perform point cloud segmentation on a wide range of different breeds of quadrupedal livestock, saving a significant amount of time spent on manual annotation and improving the efficiency of livestock point cloud segmentation models.

Tree fruit harvesting is a labor-intensive operation. Multi-level fruit catching and retrieval (MFCR) systems have been proposed for the mass harvesting of soft fruits using trunk shaking. However, overcoming excessive fruit damage as fruits fall through the canopy is very challenging and requires optimization of several aspects of an MFCR’s design. In this work, we present a novel computer-aided design approach for optimizing design parameters of MFCR systems. A simplified index that utilizes geometry and simple mechanics is developed to quantify the interference between an MFCR’s catching booms and tree branches. Also, an index is introduced that utilizes geometry and simplified collision kinematics to represent accumulated damage on fruits falling through the canopy. These two indices are used in a case study to determine the optimal solution – and its sensitivity – for the number of layers of an MFCR system that maximizes marketable fruit collection over twenty digitized pear trees. In conjunction with more elaborate machine-tree-fruit interaction models, the proposed methodology can be used to optimize the design of fresh-fruit mass harvesters that utilize multi-level catching and retrieval systems.

Greenhouse vegetables have become increasingly important in global crop production due to their ability to be cultivated out of season and ensure a year-round supply of vegetables. With the rapid advancement of “phenomics”, accurately measuring the phenotypic information of greenhouse vegetables is crucial for enhancing both their yield and quality. Over the past two decades, various technologies have been developed for phenotypic detection of fruits, vegetables, and other crops, based on the interaction between electromagnetic waves and matter. While some articles have investigated these applications, there is a lack of a systematic review specifically focused on the phenotypic detection of greenhouse vegetables. In this review, RGB imaging, Multispectral/Hyperspectral imaging, Chlorophyll fluorescence imaging, Thermal imaging, Raman imaging, X-ray imaging, Magnetic resonance imaging, and Terahertz imaging are collectively referred to as spectrum imaging technologies. We provide a comprehensive review of the origins, research progress over the past twenty years, and current challenges of spectrum imaging in the field of greenhouse vegetable research. It focuses on identifying the most suitable spectrum imaging technologies for detecting four categories of phenotypic traits: biochemical, physiological, morphological, and yield-related traits. Additionally, we highlight the issues that need optimization in the practical application of these technologies and the bottlenecks faced in different trait studies. Finally, based on existing research, we propose several potential solutions and future research directions to maximize the utility of spectrum imaging technologies in the phenotypic detection of greenhouse vegetables.

Perturbations always affect livestock during the breeding process, including harmful diseases. Researching the impact of disease outbreaks on pig herds is extremely important so that disease control measures can be applied early. However, conducting practical experiments on disease outbreaks is extremely expensive. Precision feeding systems (PFS) for pigs use data on the animal’s own feed intake to calculate the appropriate amount of feed for each individual. This helps increase productivity and product quality while contributing to reducing waste generation in the environment. Daily feed intake (DFI) and cumulative feed intake (CFI) data can be automatically collected and estimated from the PFS, which can help detect or predict disease outbreaks. In this article, we introduce an advanced simulation model of the PFS for pigs and the integration of disease outbreak models into this system. A disease outbreak simulation application within the pig herd raised by the precision feeding method is also developed for running high-performance experimental simulations. The results of the simulation scenarios are analyzed and compared with data from a real-world experiment to assess the accuracy of the application. The correlation coefficient values of DFI in all scenarios fall within the range of 0.25 to 0.5, suggesting almost no correlation between simulated DFI and actual DFI. The overall average correlation coefficient of CFI for all scenarios is 0.764, falling within the strong correlation range. It can be concluded that the simulation accurately represents CFI values compared to reality.

Large language models (LLMs) have revolutionized various scientific fields in the past few years, thanks to their generative and extractive abilities. However, their applications in the Agricultural Extension (AE) domain remain sparse and limited due to the unique challenges of unstructured agricultural data. Furthermore, mainstream LLMs excel at general and open-ended tasks but struggle with domain-specific tasks. We proposed a novel QA benchmark dataset, AgXQA, for the AE domain to address these issues. We trained and evaluated our domain-specific LM, AgRoBERTa, which outperformed other mainstream encoder- and decoder- LMs, on the extractive QA downstream task by achieving an EM score of 55.15% and an F1 score of 78.89%. Besides automated metrics, we also introduced a custom human evaluation metric, AgEES, which confirmed AgRoBERTa’s performance, as demonstrated by a 94.37% agreement rate with expert assessments, compared to 92.62% for GPT 3.5. Notably, we conducted a comprehensive qualitative analysis, whose results provide further insights into the weaknesses and strengths of both domain-specific and general LMs when evaluated on in-domain NLP tasks. Thanks to this novel dataset and specialized LM, our research enhanced further development of specialized LMs for the agriculture domain as a whole and AE in particular, thus fostering sustainable agricultural practices through improved extractive question answering.

Pesticides are commonly applied by using agricultural nozzles to generate droplets during delivery process. Initial spray atomization characteristics including droplet size and velocity are important factors that affect the pesticide utilization rate. Exploring efficient methods for atomization measurement is helpful to deeply understanding nozzle sprays. In this study, droplet size and velocity of a flat-fan nozzle were measured with phase doppler interferometry (PDI), and sub-area statistics method was adopted to establish a fitting model for atomization characteristics analyse. The results demonstrated that the distribution patterns and value contrasts of droplet size and velocity in different sub-areas visually reflect the nozzle atomization characteristics under varying spray pressures. The quantized model of droplet size and velocity within spatial sub-areas of spray atomization revealed significant differences in droplet size and velocity at various positions within the atomization area. Near the edge of the initial atomization zone, droplet size increases while velocity exhibits a decreasing trend. Additionally, the coefficient of determination for the x-axis position within the atomization zone, in relation to droplet size and velocity, was above 90%. The PDI with the sub-area statistical method employed in this study offers a fine-grained approach for investigating nozzle atomization characteristics.

Accurate quantification of hyperspectral features altered by plant disease infection is pivotal for effective disease management. However, the sensitivity of hyperspectral features to plant disease progression remains elusive, primarily because these features are often influenced by plant growth and environmental factors in addition to the specific disease. This study explores the sensitivity of biophysical and spectral features as indicators for maize adaptation to leaf spot disease. Using high-resolution UAV hyperspectral imaging, we captured maize adaptation dynamics over 30 days post-infection. We evaluated the sensitivity and importance of hyperspectral features for disease monitoring, including biophysical parameters retrieved by the PROSAIL model, and spectral features, including spectral reflectance, vegetation indices (VIs), and wavelet features (WFs). Our findings reveal that WFs first indicate disease response as early as 6 days after infection (DAI), followed by VIs at DAI 8, and variations in chlorophyll content (C_ab) become apparent by DAI 10. The C_ab, plant senescence reflectance index (PSRI), and normalized photosynthetic reflectance index (NPRI) are determined to be important features at the early stage of the disease. Our experimental results show that the different feature sets are complementary at the early and severe stages of the disease. Our classification models integrating C_ab, VIs, and WFs showed higher overall accuracy than models using only spectral features or VIs, with a maximum improvement of 9.36 %. However, these feature sets are redundant in the mild and initial severe disease stages, where models using only spectral features achieve the highest overall accuracy of 86.21 %. This study underscores the novel insights by offering an understanding of plant responses to disease infection and enhancing early detection strategies.

The integration of advanced technologies, such as soil proximal sensing, remote sensing, and machine learning, has revolutionized agricultural practices, particularly for corn yield prediction. This interdisciplinary approach harnesses the power of cutting-edge sensors to gather high-resolution data on soil conditions coupled with remote sensing technologies that provide a comprehensive view of crop health and environmental factors. This study aimed to evaluate the feasibility of accurately predicting corn (Zea mays) yield at the management zones (MZs) level using the fusion of visible and near-infrared spectroscopy (Vis-NIRS)-derived soil properties, remote sensing-derived crop spectral indices, and machine learning algorithms. Clustering analysis was used to develop MZs to implement variable-rate nitrogen fertilization (VRNF) in a drip-irrigated corn field. Site-specific models to forecast corn yield at the MZs level were developed using Sentinel 2A-derived spectral indices and machine learning regression algorithms. Partial least squares Vis-NIR spectral regression modelling for MZs development achieved high accuracy in terms of the coefficient of determination (R²) which was ranged from 0.60 to 0.99 in cross-validation and from 0.52 to 0.78 in online validation. The developed corn yield prediction models demonstrated moderate efficacy, as evidenced by the R² values ranging from 0.50 to 0.71. Further research should include supplementary spectral crop canopy indices and the application of alternative deep and machine learning approaches to improve the accuracy of the prediction models.

Conventional cotton harvesters are efficient but heavy causing soil compaction. They normally perform one harvesting pass, but since cotton bolls mature over two months, the early opened bolls must wait for later ones to be harvested, exposing their fiber to weather and degrading fiber quality. A swarm of small, lightweight robotic cotton harvesters can address these issues. This study presents field tests and evaluations of an innovative robotic cotton harvester prototype. A stereovision camera in conjunction with the YOLOv4-tiny algorithm was used for cotton boll detection and localization. The picking system included a 3-DOF (degree of freedom) linear robotic arm, a three-finger end-effector, and an agile control algorithm. The performance rates of detection, localization, and picking systems were 78.1 %, 70.0 %, and 83.1 %, respectively, with an average cycle time of 8.8 s. Collecting cotton bolls orientation data proved that they tend to stay their faces upward causing difficulty in picking the rear part of the bolls in 40.5 % of cases. Controlling the illumination, developing more robust detection and localization systems, increasing the arm’s DOF, enhancing the end-effector’s operating speed, and its adaptability to different boll orientations can improve the robot’s performance in terms of the picking ratio of the seed cotton and speed. The dataset, including field images, annotations of cotton bolls, and the best training weights, is publicly available at: https://github.com/hussein-pasha/Robotic-Cotton-Harvester. A video demonstration of the harvester being tested in the field is available at: https://youtu.be/IztKk3E7zSc.

The basic process of corn sowing includes seed selection, land preparation, fertilization, sowing, and soil compaction. Soil compaction is an important step in the sowing process, playing a crucial role in protecting the seeds, promoting germination and root development, and providing a stable growth environment for corn. Currently, mainstream soil compaction devices used in corn sowing employ non-active adjustment structures, which cannot regulate the amount of soil covering and the compaction force for individual seeds during the sowing process, making it difficult to ensure consistent sowing depth. To address these issues, this study investigates the soil compaction device on a corn planter and proposes a soil compaction device that utilizes a binocular structured light camera to detect the opening depth of the planter and flexibly adjust the soil covering and compaction force for each seed. Experimental evaluations of the device’s performance were also conducted. The design of the sowing depth consistency control system includes the selection and application of the design, motor, gearbox, binocular structured light camera, dust removal device, user interface, electric-driven soil compaction device, and control system. The experimental results showed that when the system detects a variation in trench depth of around 2 cm, the average response time of the system is 2.23 s with a standard deviation of 0.042 s. When the system detects a variation in trench depth of around 4 cm, the average response time of the system is 4.68 s with a standard deviation of 0.078 s. This suggests that the system’s response time fluctuates within 0.1 s, indicating good stability of the system. The average error of the planter’s opening depth, as measured by the binocular structured light camera, is approximately 6 mm, the success rate of detection can be maintained above 70 % under different trench depths. The dust removal device’s performance meets the requirements of the detection system. The research demonstrates that the sowing depth consistency control system developed in this study can accurately detect the planter’s opening depth during operation and adjust the soil covering, compaction force appropriately based on the depth information provided by the soil compaction device.