Computers and Electronics in Agriculture最新文献

Accurate and rapid estimation of cabbage head diameters is critical for precise decision-making in cabbage-harvesting equipment, thereby ensuring the quality of cabbage head harvesting. However, mature cabbage heads are enveloped by layers of outer leaves, resulting in varying degrees of occlusion, which poses significant challenges for direct detection and diameter measurement of cabbage heads. To address this problem, this study proposes a method based on the keypoint of cabbage head for estimating cabbage head diameters with different degrees of occlusion in the field. An improved deep learning model, YOLOv8n-Cabbage Keypoints (YOLOv8n-CK), is introduced to accurately and rapidly detect the keypoints of cabbage heads. Specifically, to enhance the attention of the network to occluded cabbage head features in complex images, the convolutional block attention module (CBAM) is introduced in the backbone, thereby improving the accuracy of the model in detecting the keypoints of occluded cabbage heads. Moreover, to balance the accuracy and speed of the keypoint detection network, all the Conv modules of the C2f-Bottleneck structure are replaced by Ghost modules, which effectively reduces the number of parameters in the model while maintaining its accuracy and reducing the computational complexity. Based on the results of keypoints detection, the physical diameter of cabbage heads is computed by integrating the depth information of the effective keypoints using a histogram filtering algorithm. The experimental results show that for varying degrees of occlusion, YOLOv8n-CK achieves an average precision (AP_50–95) of 99.2 % in detecting cabbage head keypoints, with 12.68 % and 13.04 % reductions in the params and floating point operations per second, respectively, compared to the original model. The mean absolute percentage error of the cabbage head diameter estimation model is 4.28 ± 0.13 %, and it exhibits favorable performance even under heavy occlusion (occlusion rate >65 %). Validation on an edge computing device shows that the model achieves 142.6 frames per second, which satisfies the real-time diameter estimation requirements for cabbage heads. These findings confirm the effective in-situ measurement of cabbage head diameters in the field, offering innovative insights for the development of efficient and low-damage harvesting equipment for cabbage.

In this study, in order to solve the problem of low seeding performance of the centrifugal high-speed precision seed metering device for maize during high-speed seeding, a method of airflow-assisted seed filling and cleaning to improve seeding performance is proposed. A trapezoidal opening is designed at the bottom of the hole insert and an airflow outlet is designed on the front shell to perform airflow-assisted seed filling and cleaning processes. By combining the discrete element method and computational fluid dynamics method for simulation, the optimal trapezoidal opening length is 14 mm, and the optimal number of holes for airflow outlet 2 is 11 holes. When the seeding speed was 12 km/h, 15 km/h, and 18 km/h, the optimal inlet air pressure was 800 Pa, 800 Pa, and 900 Pa respectively. The results of the bench experiment show that when the seeding speed was 12 km/h, the omission, repeated, and pass indexes were 2.75 %, 5.12 %, and 92.13 % respectively. When the seeding speed was 15 km/h, the omission, repeated, and pass indexes were 2.02 %, 4.17 %, and 93.81 % respectively. When the seeding speed was 18 km/h, the omission, repeated, and pass indexes were 3.16 %, 3.98 %, and 92.86 % respectively. The power of the air-suction seed metering device was 471.65 W when the seeding speed was 12 km/h. The centrifugal high-speed precision seed metering device for maize with airflow-assisted seed filling and cleaning has only a power of 45.84 W when seeding at a speed of 18 km/h. It not only has a faster seeding speed, but also consumes only 9.8 % of the energy compared to traditional pneumatic seed metering device, which can reduce environmental pollution.

In recent years, the use of convolutional neural networks (CNNs) and graph convolutional networks (GCNs) has significantly advanced hyperspectral image classification (HSIC). Despite these achievements, the challenge of limited labeled samples remains a critical obstacle when employing CNNs and GCNs for hyperspectral image classification. Agricultural images often face challenges due to high spectral variability and complex spatial patterns, making accurate classification difficult. Additionally, the presence of noise and limited labeled data further complicates the analysis and interpretation of these images. Although graph convolution networks and their predecessors have greatly advanced the use of spatial relationship features, deep learning algorithms, such as convolutional neural networks (CNNs), have reduced the need for and reliance on a high number of samples through spatial feature learning and consideration. Therefore, to advance the field, a novel approach termed “dimension reduction fuzzy graph network” (DRFG) was designed. This approach is a combination of deep fuzzy-based DR, enhanced with 3D-CNN and GATs, with the application of principal component analysis (PCA) for optimized DR. The DRFG model entails two major processing stages. The initial stage involves the classification of the raw data cube using the 3D-CNN. In the second stage, the results are processed by means of an algorithm enriched by lightweight GAT-based modules. The DRFG model combines morphological features selection from fuzzy C-means (FCM) clustering and optimized DR by using PCA. Thus, the model employs the best of PCA and GATs in order to allow for optimized classification. At high-performance optimal DR, the DRFG model offers optimal multispectral imaging as well as the analysis and classification of hyperspectral data, which is sufficiently promising so as to advance the field’s needs for precision agriculture.

Fourier transform mid-infrared spectroscopy (FT-MIRS) technique has been extensively employed for performance measurement of dairy cows and dairy herd improvement (DHI), but different milk analyzers have shown significant differences in the sensitivity, laser intensity, and stability of FT-MIRS determination, which cannot be directly integrated and applied in phenotype prediction and relevant studies. Existing literature has reported several FT-MIRS calibration methods such as piecewise direct standardization (PDS) and retroactive percentile standardization (RPS), achieving good standardization results. However, these methods require to be optimized because they take no account of the collinearity and redundancy of the spectrum.

Therefore, this study established an improved agglomerative clustering piecewise direct standardization (ACPDS) method. This study used 432 standard milk samples prepared by the standard laboratory within 4 months (based on the standard sample preparation procedures in the International Dairy Federation Guidelines for the Application of Mid-infrared Spectroscopy) and carried out FT-MIRS measurements and data collection on 9 instruments in 5 DHI laboratories. Meanwhile, the new method established in this study together with the existing methods of single wavelength standardization (SWS) and PDS were adopted to standardize the spectra collected on 9 instruments. The reproducibility, computation time, memory usage, and repeatability of the milk component prediction models were verified and compared.

The results revealed that ACPDS exhibited significant advantages over SWS and PDS, with a higher level of spectral reproducibility, and there was a significant advantage in the repeatability of the milk component prediction models but no significant increase in memory usage. The impact of its application across regions, months, and years was insignificant. In addition, based on the respective characteristics of ACPDS and the existing two methods, application strategies have been proposed for these three methods, providing new technologies and laying the foundation for the FT-MIRS-based milk component prediction models, widespread performance measurement of dairy cows in different instruments and at different times, and comparative analysis on the traits and phenotypes of dairy cows as well as their joint breeding in China and even the world.

This paper introduces a Multi-Object Tracking (MOT) framework for agricultural applications that estimates global positions in pixel coordinates using the local feature matching transformer — LoFTR. We design an efficient tracker that augments the capabilities of a state-of-the-art tracking algorithm by incorporating a novel association strategy based on spatial information of targets leaving and returning the camera field of view. We evaluate our framework using the publicly available LettuceMOT benchmark dataset and an adapted version of the AppleMOTS benchmark dataset that we denominate AppleMOT. Our experimental results demonstrate that our method outperforms cutting-edge algorithms for robotic plant tracking in the LettuceMOT dataset. The evaluation metrics show average improvements of up to 25% compared to the best publicly available results, demonstrating the benefits of our spatial association approach. For the AppleMOT dataset, we obtained bounding-box-based MOT evaluation metrics comparable to the segmentation-based (MOTS) counterparts presented in the original AppleMOTS paper. These findings highlight the effectiveness and potential of our approach in addressing the unique challenges posed by agricultural environments.

The precise analysis of individual fruit tree canopy information for accurate navigation and spraying operations of plant protection machinery is important for intelligent orchard management. However, in the complex environment of an orchard, it is quite challenging to simultaneously accomplish the detection, localization, and segmentation of tree canopies to enable precise spraying. Fortunately, advancements in high-performance unmanned aerial vehicle (UAV), sensors, and deep learning algorithms have made it possible to quickly extract and analyze tree information from complex backgrounds. In this study, we proposed a comprehensive operational framework based on UAV data and deep learning algorithms to accurately obtain apple tree information, thereby enabling variable targeted spraying. First, the Max-Relevance and Min-Redundancy (mRMR) algorithm was used to select three features (RVI, NDVI, SAVI) to create fused images to enhance tree canopies from the background environment, and the enhanced images were then utilized to generate a labeled sample dataset. Secondly, leveraging the labeled dataset, the YOLO-Fi model was developed. Using this optimal model, precise detection, localization, and segmentation of fruit trees in the experimental area were conducted. Our results showed that the YOLO-Fi model achieved optimal results (FPS = 370, mAP_50-95 (B) = 0.862, mAP_50-95 (M) = 0.723, MIoU = 0.749). Subsequently, based on the segmented areas of the fruit tree canopies, a variable spraying prescription map was generated, contributing to a 47.92% reduction in spraying volume compared to direct spraying. Finally, the ant colony algorithm was employed to design the shortest path for the plant protection UAV to traverse over each fruit tree within the experimental area, leading to a 2.04% reduction in distance compared to the conventional UAV flight path. This research can provide a comprehensive scheme for UAV-based precision management in orchards, encompassing tree canopy monitoring, analysis, localization, navigation, and precise spraying.

It is imperative to meet the growing food demands of our expanding global population while safeguarding the Earth’s finite natural resources. This challenge becomes even more pressing for resource-constrained farmers residing in low- and middle-income countries (LMICs), who disproportionately bear the brunt of food insecurity. In response to this critical issue, the Predictive Optimal Water and Energy Irrigation (POWEIr) controller is a promising solution. The POWEIr controller was designed as an affordable precision irrigation controller for solar-powered drip irrigation (SPDI) systems and offers an avenue to widen access to SPDI and precision agriculture for low-income farmers. The POWEIr controller creates energy- and water-efficient irrigation schedules that aim to reduce overall system costs. Employing simple yet effective physics-based models alongside minimal sensors to maintain cost-effectiveness, the controller’s accuracy has, until now, remained unexplored. This paper investigates the sensitivity of the POWEIr controller’s optimized irrigation schedules to user and weather sensor accuracy errors in inputs, while also assessing their impact on simulated crop yields. The results reveal that, under the tested scenarios, opting for a low-cost weather station over a high-quality counterpart could potentially save farmers over $900 with negligible consequences to crop yields. This conclusion held steadfast across diverse crop and soil types. The most significant factor affecting the optimal irrigation schedule was found to be changes in the crop coefficient, pointing to the need for calibration of the controller. This research underscores the POWEIr controller’s capability to optimize irrigation schedules through the use of cost-effective sensors and minimal calibration efforts. In doing so, it opens the door to greater adoption of precision irrigation technology and sustainable irrigation practices among farmers in LMICs. Ultimately, this progress has the potential to catalyze sustainable agriculture intensification on a global scale, moving us closer to a more food-secure and environmentally responsible future.

Edge-AI based AIoT technology offers significant benefits advantages in modern poultry management by optimizing farming operations and reducing resource requirements. To address the challenge of developing a highly accurate and lightweight edge-AI enabled detector that can be deployed within memory-constrained edge environments, this study propose an innovative real-time, compact and highly accurate edge-AI enabled detector, based on improved FCOS-Lite and designed to detect chickens and their health status using a highly resource-constrained edge-AI enabled CMOS sensor. The proposed FCOS-Lite detector leverages MobileNet as the backbone to achieve a compact model size. To mitigate the issue of reduced accuracy in compact edge-AI detectors without incurring additional inference costs, we propose a gradient weighting loss function for classification and introduce a CIOU loss function for localization. Furthermore, a knowledge distillation scheme is employed to transfer critical information from a larger teacher detector to the FCOS-Lite detector, enhancing performance while preserving the compactness. Experimental results demonstrate the proposed detector achieves a mean average precision (mAP) of 95.1% and an F1-score of 94.2%, outperforming other state-of-the-art detectors. The detector operates efficiently at over 20 FPS on the edge-AI enabled CMOS sensor, enabled by int8 quantization. These results confirm that the proposed innovative approach leveraging edge-AI technology achieves high performance and efficiency in a memory-constrained environment, meeting the practical demands of automated poultry health monitoring with low power consumption and minimal bandwidth costs.

Rising global population of the world is resulting higher demand for buckwheat being a high-quality food crop. In this study, the multi-stage shearing drum working mode was proposed to solve the problems of easy entanglement, blockage in threshing process due to inconsistency of ripening period, decrease grain breakage rate and loss rate and increase threshing efficiency. The designed threshing unit incorporates key components: frame, feeding wheel, main and secondary threshing drums, discharge unit, and concave plate. FEA and modal analysis were integrated to assure robust structural performance and stability of the threshing components within set limits, which were validated by indoor testing that confirmed the threshing drum’s working frequency did not cause resonance. Single Factor Method identifies optimal conditions: 600 rpm, 7 mm, 1.2 kg/s for minimal grain breakage; 700 rpm, 9 mm, 1.2 kg/s for lowest grain loss. A three-factor, three-level orthogonal experiment validates these findings. In conclusion, optimal results are achieved with a drum speed of 600 rpm, feeding rate of 1.2 kg/s, and a threshing gap of 9 mm thus, minimizing both grain loss and breakage rates.