Computers and Electronics in Agriculture最新文献

Competition for nutrients between intra-row weeds and cultivated vegetables is a major contributor to reduced crop yields. Compared with manual weeding, intelligent robots can improve the efficiency of weeding operations. Developing real-time and reliable robotic systems for weed control in vegetable fields is a significant challenge due to the complexity of real-time identification, localization, and classification of vegetables as well as various weed species. The main purpose of this study was to propose a high-performance, lightweight deep learning model and an intra-row weed severity classification algorithm for real-time lettuce identification and weed severity classification. In this study, a scaling factor (τ = 0.5) was chosen to lightweight the YOLOv7 model. A new Multimodule-YOLOv7-L lightweight model was then developed by combining ECA and CA attention mechanisms, ELAN-B3 and DownC modules. The overall performance of the Multimodule-YOLOv7- L was better than that of other deep learning models, including YOLOv7, YOLOv7-Tiny, YOLOv8m, YOLOv5n-Cabbage, SE: YOLOv5x, YOLOv5s_Ghb, MST-YOLO_CBAM, Citrus-YOLOv7, Pineapple-YOLOv7, MS-YOLOv7 and CBAM-YOLOv7. The precision, recall, [email protected], F1-score, model weight and FPS of the Multimodule-YOLOv7- L model were 97.5 %, 95.7 %, 97.1 %, 96.6 %, 18.4 MB and 37.3 FPS (Image resolution about 3000 × 3000), respectively. An intra-row weed severity classification algorithm based on the Multimodule-YOLOv7-L model was proposed for use in a new mechanical-laser collaborative intra-row weeding robot. The developed algorithm achieved a classification accuracy of 100 % in eight lettuce weed scenarios, with the processing time of a single image ranging from 4-13 ms. The results of this study provided valuable reference for the development of intelligent robots for intra-row weed control. The algorithm proposed in this article can be obtained at https://github.com/H777R/The-intra-row-weed-severity-classification-algorithm.git.

We propose a morphological characterization of the endocarp of the fruit of the almond tree, Prunus amygdalus (Batsch), using computer vision techniques to extract features in 3D almond endocarp meshes with the objective to describe the diversity of the crop in a systematic and unambiguous form. All the proposed descriptors are quantitative and easily computable, allowing fast and objective assessments of the morphological variations between almond varieties. We collect and 3D-scan a total of 9510 almond endocarps to obtain such meshes, to which we apply an affine transformation so that they are positioned in a standardized reference where meaningful physical measures can be taken. Complex descriptors derived from the geometry of the endocarp are then introduced to identify richer features. The use of 3D, compared to simply taking 2D images, allows for a more accurate and complete description of the endocarp shape. In particular, the contour and apex shapes, keel development, markings on the surface, and symmetry of the endocarp are analyzed and given quantitative measures. The validity of the presented morphological descriptors is finally tested on 2610 endocarps from the collected dataset, corresponding to 36 autochthonous almond varieties from the island of Mallorca (Spain) and 14 international reference varieties, all with well documented characteristics. Numerical results show that the proposed descriptors agree with human-made shape classifications of the studied varieties with a coincidence of $\text{75.0}\text{\%}$ for contour shape, $\text{76.0}\text{\%}$ for apex shape, and $\text{80.0}\text{\%}$ for keel development. Visual comparisons of the extracted features also show that they are coherent with commonly used guidelines for the morphological characterization of the almond endocarp. We conclude that the use of 3D imaging approaches for the description of the almond endocarp is a promising alternative to traditional methods, providing a reliable way to deal with ambiguity and helping reduce biases and inconsistencies caused by subjective visual evaluations.

Accurate fruit detection and segmentation based on deep learning is the key to successful harvesting robot operations, but the complex background of orchards, light and branch shading, and fruit overlap lead to low detection and segmentation accuracy and high complexity of existing methods. To address these problems, an improved green fruit segmentation method based on FCOS is proposed in this study. Firstly, its segmentation function for green fruits is realized by adding segmentation module. Then, the FCOS head network is improved by adding the Border-attention module (BAM) to detect the boundary of green fruits with higher accuracy. In addition, the features of mask branch and edge segmentation branch are fused in the segmentation module, and the appearance commonality is learned by modeling the pairwise affinity between all pixels of the feature map using non-local affinity-parsing, and finally the segmentation prediction results are output by combining the feature map of fruit shape and appearance commonality. The experimental results show that this model achieves 81.2% segmentation accuracy on apple dataset and 77.9% segmentation accuracy on persimmon dataset with relatively low guarantee complexity, which exceeds most current segmentation models. Meanwhile, this model has high robustness and can be used for the detection and segmentation work of other green fruits and vegetables in orchards, while effectively extending the application of agricultural equipment.

Accurately calculating the damage volume and making clear the interconnected effects of the physical and chemical properties of apples on mechanical damage are crucial steps in reducing the possibility of apple damage. Tests have been conducted on apples at different maturity levels, including measuring the firmness, moisture content, water-soluble pectin (WSP) content, soluble solids content (SSC) of the flesh, and elastic modulus of the apple flesh and peel. Transient collisions were performed using a pendulum device to create damage zones under specific impact energies. Then, the X-ray micro-computed tomography (Micro-CT) was utilized to quantitatively analyse mechanical damage volumes, the effects of apple tissue characteristics and impact energy on the damage volume were analysed in detail. The results indicated that higher-maturity apples were more susceptible to mechanical damage, and Micro-CT measurements were more accurate when the impact energy ≥ 0.05 J, while the empirical formula showed greater deviation; the curvature radius at the impact point can be considered as a latent variable influencing the apple damage volume. Furthermore, a damage volume prediction model, based on bruise area calculated by the empirical formula, WSP content of the flesh, and elastic modulus of the apple flesh and peel, was established. With a testing dataset without anticipate in model training for verification, the developed model achieved a coefficient of determination of 0.9782, indicating that the model can predict damage volume effectively and reduce errors associated with the empirical formula, particularly at higher impact energies. The research can provide insights into potential applications in apple industry practices to reduce the mechanical damage.

Sclerotinia is a worldwide disease that often occurs at all growth stages of rapeseed, and can lead to 10 %∼70 % yield decline. It will also drastically reduce the oil content of seeds, which greatly increases the risk and difficulty of rapeseed cultivation. In order to address the problems of traditional chemical-based Sclerotinia detection methods such as complex operation, environmental pollution, plant damage and low efficiency, this study innovatively combined the two architectures of SNN (Spiking Neural Network) and LSTM, and proposed a spatial-spectral joint detection model Spiking-LSTM for the HSI segmentation. In the design process of the model, spiking neurons were used instead of gating functions in traditional LSTM units, while the back propagation of errors was solved by using gradient surrogate function. The experimental results show that when input data from the infected area, the neurons in hidden layer of the trained model exhibited distinctly regular spiking signals. Compared with the mainstream models, the Spiking-LSTM based on spatial-spectral data fusion has better performance in the evaluation parameters such as mAP, ClassAP, mIoU, FWIoU and Kappa coefficient. Its Sclerotinia detection mAP reached 94.3 % and was able to accurately extract the infected areas at the early-stage of infection. With essentially the same structure, the Spiking LSTM not only has higher detection accuracy but also, for the same HSI input, requires only one-fifth of the theoretical energy consumption compared to the traditional LSTM. This paper establishes the basis for the construction of large-scale SNN models, and also provides a reference for the application of SNNs in different fields.

Deep learning-based methods for accurately identifying plant diseases can be effective in improving crop yields. However, the effectiveness of these methods heavily relies on the availability of large-scale manually labeled datasets, which present technical and economic challenges. Few-shot learning methods can be generalized to new categories with a small number of samples, which is very promising in the field of plant disease recognition despite the limited sample size. However, due to the complexity of real-life scenarios, distribution of disease leaves, and significant intra-class variability and inter-class similarity resulting from crop species and disease species, existing methods perform poorly in the field of plant disease recognition. To address the above problems, this paper proposes an improved multi-scale attention fusion with discriminative enhancement deep nearest neighbor neural network MAFDE-DN4 based on DN4. Our approach makes three main contributions. First, we have designed a bidirectional weighted feature fusion module (BWFM) to enhance the aggregation of fine-grained features and enhance the network’s representation of complex disease images. Second, to tackle the issue of sparse feature descriptors being vulnerable to irrelevant noise in small sample conditions, a episodic attention module (EA) has been developed to produce scene category-relevant attention maps. This effectively mitigates the influence of irrelevant background information. Finally, we introduce additional spacing between category margins to enhance the original softmax loss function, amplify the inter-class differences to reduce the intra-class distances, and add L2 regularization constraint terms to stabilize the training process. To simulate different real-world scenarios, we set up different dataset settings. Under the 1-shot task and the 5-shot task, our method achieves 57.5% and 81.41% accuracy under the within-domain strategy and 36.54% and 51.23% accuracy under the cross-domain strategy. The experimental results show that our method outperforms existing related work in the field of plant disease recognition, whether it is a dataset with a single background or a field dataset with a complex background in a real scenario. MAFDE-DN4 based on Few-shot learning requires substantially less data on new categories of plant diseases.

This study addressed the challenge of machine vision-based weed detection for precision herbicide application, a task complicated by the diversity of weed species, ecotypes, and variations in growth stages. We propose an indirect approach that segments crops and classifies the remaining green objects as weeds. A novel, lightweight segmentation network was developed to reduce computational demands without compromising accuracy. The model, with a size of just 2.64 MB, achieves an impressive mean Intersection over Union (mIoU) of 97.9 %, with a recall of 93.4 %, and a precision of 97.6 %, while also enhancing inference speed. Subsequently, improvements were implemented using the image processing method for extracting green plants. A crop mask was generated using a segmentation algorithm, and a mask expansion mechanism was introduced to rectify errors in the initial phase of crop segmentation. A cost-effective threshold adjustment operation was applied to eliminate the environmental influences on the detection results. The results indicate that the weed detection method completely avoided the complexity related to the variations in species, ecotypes, growth stages, and densities of weeds across different fields and realized accurate, effective, and reliable weed detection in cabbage.

A novel fully convolutional autoencoder (convAE) was introduced to analyze X-ray radiography images of ‘Braeburn’ apples and ‘Conference’ pears with and without disorders for online sorting purposes. The model was solely trained on either apple or pear samples without disorders and outperformed a traditional autoencoder (AE) across multiple test sets. We evaluated our approach using the area under the curve (AUC) as an evaluation metric. A cross-test experiment further demonstrated consistent performance between a model trained on apple data for classifying pear fruit (accuracy: 71 %) and a pear-specific model (accuracy: 70 %). We also evaluated models trained on simulated X-ray radiographs with real ones, and vice versa. For instance, under scenario of training on real data and testing on simulated X-ray radiographs, an accuracy of 80 % for detecting disordered non-consumable pear was achieved. This work provides valuable insights into anomaly detection for apples and pears with several disorders.

The use of machine vision technology for recognizing cow behavior plays a crucial role in daily management, health monitoring, and breeding and reproduction in dairy farming, making it an essential component of modern smart agriculture. This paper presents a novel dual-stream network model, the Motion Focus Global-Local Network (MFGN), for analyzing cow video data. The dual-stream network consists of a global spatiotemporal feature stream and a fine motion feature stream. The global spatiotemporal feature stream extracts key frames to remove redundant information and utilizes a Transformer network for global spatio-temporal feature extraction, reflecting the dynamic changes in cow videos and the temporal relationships between video frames. The fine motion feature stream, based on frame differencing of cow videos, uses focal convolution to capture subtle movements of cows, enhancing the focus on minor behavioral changes. To evaluate the performance of the proposed model, video data samples were collected from eight cows marked on their bodies and heads at an Australian farm site (CSIRO Armidale), including a total of 1715 video sequences across three behavior categories. The model achieved recognition accuracies of 98.1% for drinking, 95.5% for grazing, and 49.3% for other behaviors, with an overall average recognition accuracy of 79.4%, representing a 7.4% improvement over the classic TSN model. Overall, the MFGN network effectively extracts and integrates global spatiotemporal features with fine motion features from cow video data, modeling both the overall sequence characteristics and focusing on local motion details, achieving precise behavioral recognition. This research not only enhances the accuracy of cow behavior recognition but also provides new technological means for precise management in modern smart agriculture, with broad industry application potential to improve farm efficiency, profitability, and disease control and prevention.

In large-scale laying hen farms, daily inspection of dead hens is a relevant task to monitor the health of the flock and prevent disease spreading. The current manual inspection method used in caged hen farms is inefficient, costly, and particularly difficult for high-rise cages. To address this issue, a dead caged hen detection method based on improved You Only Look Once version 7 (YOLOv7) was proposed in this study, which was optimized to improve detection performance and speed in complex farming environments, such as cage wire mesh occlusion and crowded hen occlusion. First, the Convolutional Block Attention Module was used to enable the model to learn target features accurately. Second, the Distance Intersection over Union Non-maximum Suppression and repulsion loss were introduced to improve crowded hen occlusion and reduce missed detections. Additionally, to facilitate the deployment of the proposed method on mobile devices, the MobileNetv3 lightweight network was used to replace the backbone of YOLOv7. Furthermore, the lightweight model was trained using the knowledge distillation method to enhance its performance. Finally, a comparison experiment of different object detection networks and an ablation experiment were conducted to evaluate the proposed method. The experimental results reveal that the improved YOLOv7 model proposed in this study performs optimally. Its precision, recall, F1 score, and [email protected] for the dead hens in the test set are 95.7 %, 86.8 %, 0.910, and 86.2 %, respectively. Compared with the original YOLOv7 model, precision, recall, and [email protected] were increased by 6 %, 10 %, and 13.4 %, respectively. The model parameters and Giga Floating-point Operations were decreased by 31.95 % and 60.56 %, respectively, resulting in a detection speed increase of 43 Frames Per Second. Furthermore, with the assistance of an inspection robot, the proposed dead hen detection model was deployed in the actual farming environments. Compared with methods proposed by other researchers, the proposed model is more suitable for complex actual farming environments and achieves higher detection accuracy, which can offer a reference for automated caged hen detection.