Wheat is one of the most extensively cultivated crops in global food security. Wheat yellow dwarf disease can be caused by the barley yellow dwarf virus complex (BYDVs). The major viral threat can often occur year-round, widespread in the Huang-Huai wheat region. A substantial risk has posed to the quality and yield during production. Infected plants typically exhibit leaf yellowing, stunted growth, delayed maturity, and shriveled grains. Manual detection can be performed on the preliminary diagnoses using visual symptoms. It is often required to provide costly equipment or specialized expertise, thus limiting the large-scale field screening. In this study, DSFNet, a field-scale detection model, was proposed using drone-acquired multispectral imagery, according to the multi-stage development of wheat yellow dwarf disease. Two components were also integrated - dual-stream adaptive synergistic fusion (DASF) and spatiotemporal joint feature mapping (SFFM) - to achieve the precise identification of the diseases. Its encoder was incorporated with the DASF module. convolutional neural networks (CNNs) and Transformers were combined to dynamically balance the local and global information. The CNN branch was focused on fine-grained features to detect the contour and structural variations, such as lesion edges, textures, and colors. In contrast, the Transformer branch employed self-attention mechanisms to capture long-range dependencies. Semantic relationships among regions were obtained to reveal the spatial distribution patterns of disease on a global scale. An adaptive weight allocation in the DASF module was used to fuse CNN local extraction with Transformer global context. Rich feature representations combined spatial details with semantic depth. The decoder integrated the SFFM module into the spatial and frequency-domain information. Spatial features were preserved to incorporate the frequency-domain analysis. The regions with substantial grayscale variation were captured to enhance the multi-scale feature detection. Finally, the multi-scale features were fused with the original features for the precise segmentation via a multi-layer perceptron (MLP). A series of field experiments was conducted on the multispectral image dataset, including multiple growth stages, diverse lighting conditions, and multi-plot samples. DSFNet’s performance was compared against mainstream models, including SegFormer, UNet, DBFormer, SFFNet, DECSNet, DeepLabV3+, and PSPNet. The DSFNet outperformed all baseline models, with a mean pixel accuracy (mPA) of 92.66% and a mean intersection over union (mIoU) of 86.77%. The mPA values were improved by 3.96, 2.21, 3.65, 2.82, 3.61, 0.28, and 1.95 percentage points, with the mIoU gains of 5.77, 5.53, 5.43, 5.23, 4.49, 2.25, and 1.90 percentage points, respectively. The high accuracy of the DSFNet was maintained on an independent test dataset, particularly with the mPA and mIoU values of 89.71% and 82.56%. There were improvements of 6.35-10.16 percentage points over the rest models. The high robustness and exceptional transfer of DSFNet were also achieved under complex and real-world scenarios over the various time zones. Overall, an efficient and scalable solution was presented for the large-scale field detection of wheat yellow dwarf disease. The optimal feature fusion, frequency-spatial interaction, and adaptive semantic modeling can be combined to provide a practical framework in precision agriculture. The finding can also offer technical support for the rapid, accurate, and intelligent monitoring of crop health in the prevention and control of wheat viral diseases.

Various types of crop diseases have posed a serious threat to the high-quality development of crops in agricultural production. A timely identification of disease types can greatly contribute to the prevention and control of crop diseases. However, the convolutional neural networks (CNN) recognition of crop leaf diseases cannot fully meet the large-scale production in recent years, due to the numerous parameters, high computational complexity, and poor real-time performance. There is a high demand to improve the accuracy of lightweight CNN models under the relatively complex background of crop leaf images in natural environments. In this study, a lightweight re parameterized leaf diseases identification network (RLDNet) was proposed to identify crop leaf diseases. Firstly, reparameterization was introduced into the MobileNetV2. The inverted residual blocks were reparameterized to construct the reparameterized inverse residual (RIR) block for the high inference speed. A multi-branch architecture was adopted during the training phase. The diversity of the feature space was enriched to improve the representation learning of the network. The reparameterization was used to equivalently convert the multi-branch architecture into a single path structure for the inference after training. The output of the transformed inference model was the same as the original multi-branch. The parameter identity transformation was achieved to improve the inference speed of the model with high recognition accuracy. There was a decrease in the number of output channels and the stacking times of core modules. A shallow and narrow network structure was obtained to enhance the extraction of shallow features, such as the disease colors and textures, indicating the reduced number of model parameters. Secondly, a lightweight ultra lightweight subspace attention module attention mechanism (ULSAM) was added to the last two RIR blocks of the model. Leaf disease characteristics were combined to more effectively distinguish between backgrounds and prospects for better target classification. The interference was reduced in the disease areas from the complex backgrounds. Finally, the DepthShrinker pruning was utilized to learn the importance of the activation function in the model. The redundant parameters were removed to further reduce the space occupation. The lossless lightweight and high accuracy of the model were achieved using knowledge distillation. The recognition accuracy of 99.53% was achieved in the RLDNet on the PlantVillage dataset, whereas 98.49% on the self-built leaf disease dataset with a parameter size of 0.65 M. The inference time for a single-leaf disease image was 2.51 ms. RLDNet shared a performance similar to the Transformer-based models (such as MobileVit-S) and CNN-based models (such as ResNet18) on both datasets, with a significant reduction in parameter, respectively, compared with MobileVit-S and ResNet18. The higher recognition accuracy and lighter weight were obtained in the RLDNet, compared with the lightweight models, such as MobileNetV3 and ShuffleNetV2. The improved model can be expected to effectively identify the leaf diseases in the complex backgrounds, with less parameter memory and faster inference speed. The findings can also be applied in the practical production and application of smart agriculture.

Answer selection is one of the most important tasks during natural language processing in the downstream tasks, such as question-answering systems, and search ranking. The most relevant answer can be selected to the given question from a candidate answer pool, which is usually regarded as a relevance ranking task. However, the current models of answer selection cannot discover the deep semantic relationships between questions and answers using the limited information in the text of the question-answer pairs. Fortunately, knowledge graph can be expected to serve as the background knowledge, in order to enhance the deep semantics of the answer selection model. It is still lacking on the multi-modal background knowledge support, because the answer selection models can rely solely on their own information. In this research, a multi-modal knowledge graph enhanced answer selection model was proposed, including the embedding layer, representation learning layer, knowledge graph enhancement layer, and output layer. Among them, the Glove model was used to obtain the word embeddings for the question-answer texts in the embedding layer. Furthermore, a ComplEx-based method (complex embedding) was designed to learn the entity embeddings for the multi-modal knowledge graph. The image entity information was considered to extract the image feature representations using the Vision Transformer (VIT). Bi-directional long short-term memory (Bi-LSTM) was used for the representation learning of question-answer texts in the representation layer. The context-guided multi-modal knowledge graph question and answer vector representations were obtained using context-guided attention mechanism. In the knowledge graph enhancement layer, the interaction attention mechanism was used to fuse the semantic representation of the question-answer texts with the background knowledge features that provided by the multi-modal knowledge graph, particularly for the feature representations of the multi-modal knowledge graph enhanced question and answer. The feature representations of the knowledge graph enhanced question and answer were concatenated with the additional semantic features in the output layer. The softmax function was used to predict the probability distribution of answer labels for a given question. Taking the grape planting as an example, the multi-modal entity linking was realized using the longest common subsequence algorithm. The entity recognition was also implemented to extract the knowledge using the Bert-LSTM-CRF framework and Bert pre-training model. The reference of knowledge graph was collected from the literature and experts. Finally, a multi-modal knowledge graph was constructed in the grape planting field. A grape planting question and answer dataset was also constructed using grape forums, smart agricultural platforms, agricultural managers, and agricultural benefit networks as data sources, followed by text cleaning and dataset expansion. Experimental results show that the better performance of the model was achieved to obtain more information using the multi-modal knowledge graphs. Specifically, the mean reciprocal rank and mean average precision reached 85.02% and 84.21%, respectively, in the grape question answering dataset. The mean reciprocal rank and mean average precision increased by 2.57 and 3.96 percentage points, respectively. The answer selection model with the knowledge of multi-modal knowledge graph can be expected to improve the better performance of answer selection model. The embedding representation with attention mechanism can be utilized to enhance the background knowledge from the multi-modal knowledge graph. The finding can provide a technical basis for the downstream applications of multi-modal knowledge graphs, such as the search and question answering.

In order to solve the problem of difficult detection caused by the heavy occlusion and overlap of kiwifruit flowers at different flowering periods, this study proposes YOLOv8-KFP, a densely distributed kiwifruit flowering period detection model based on improved YOLOv8n. Firstly, taking the lightweight YOLOv8n as the benchmark model, StarBlock is used to improve the C2f of YOLOv8n, and the "Star operation" is used to enhance the ability of feature expression, so as to reduce the amount of model parameters and calculation and improve the ability of feature extraction. Secondly, the scale sequence feature fusion (SSFF) module and the triple feature encoder (TFE) module are introduced into the neck network to fuse the feature maps of different scales, so as to enhance the ability of the model to capture the multi-scale details and alleviate the problem of false detection and missing detection of blocked and overlapping kiwifruit flowers; The DySample upsampler is used to replace the traditional interpolation method to dynamically adapt to flower targets with different scales, shapes and boundaries, alleviate the fuzzy problem of dense flower edges, and improve the target positioning accuracy. Finally, the Soft-NMS post-processing algorithm is introduced to further optimize the detection effect of dense targets by adjusting the confidence of overlapping candidate boxes to reduce false deletion. 3 853 images of kiwifruit flowers densely distributed under natural conditions were collected and labeled according to the bud stage, half open stage, full open stage and withering stage. Combined with a variety of data enhancement strategies, 300 epochs of iterative training were completed on NVIDIA A40 GPU platform. Ablation experiments show that the use of C2f_Star module to improve the backbone network, the introduction of SSFF module and TFE module to improve the neck network, the use of DySample upsampler, and the use of Soft-NMS post-processing algorithm can respectively improve the performance of the model to a certain extent, and the improved module has a significant synergy. The final model achieved the precision (P) of 89.1%, the recall (R) of 88.7%, and the mean average precision (mAP) of 92.4% at an intersection over union (IoU) ratio of 0.5, while reducing the number of parameters and floating point of operations (FLOPs) to 2.66 M and 7.6 G, respectively. The improvement effect of the feature fusion network using SSFF module and TFE module and various feature fusion networks such as BiFPN, Slim-Neck, RepGFPN, HS-FPN on YOLOv8n were compared. The YOLOv8n using SSFF module and TFE module to improve the feature fusion network achieved the best effect in precision, recall and mAP_0.5, showing its excellent performance in multi-scale feature fusion of kiwifruit flowers. Compared with the mainstream lightweight target detection models SSD、YOLOv9t、YOLOv10n、YOLOv11n and YOLOv12n, the mAP_0.5 of YOLOv8-KFP is increased by 7.0, 4.0, 5.5, 4.2 and 4.1 percentage points respectively. Compared with the more complex YOLOv8s, YOLOv8-KFP is increased by 2.7 percentage points on mAP_0.5, the amount of FLOPs is reduced by 73.2%, and the amount of parameters is reduced by 76.1%, giving good consideration to the detection accuracy and the lightweight of the model. Moreover, the visualization results show that the improved model significantly reduces the missed detection rate in complex scenes such as branches and leaves occlusion and flowers occlusion, and can more accurately distinguish flower targets in different flowering periods, and has stronger robustness in the densely distributed kiwifruit flowering period detection task. The research results provide a high-precision and lightweight detection method reference for kiwifruit intelligent pollination robot, and have practical application value for improving the level of orchard automation management.

Tea planting has developed rapidly in China in recent years. It is of ever-increasing importance to integrate massive and complex knowledge in a structured manner, particularly with the development of the internet. The diverse and scattered sources of tea planting have resulted in fragmented knowledge. Knowledge graphs can be expected to greatly enhance the manageability and accessibility of data. However, it is also still lacking in the complete knowledge graphs of tea planting. It is of great significance to promote the informatization and modernization of the entire tea industry. In this study, a systematic and structured knowledge graph of tea planting was constructed on the orderly knowledge resources for the tea planting field. A tea-planting knowledge graph completion (TPKGC) model was integrated into the relationship context and path for the missing relationships between entities in the tea-planting knowledge graph. The improved model consisted mainly of three parts: a relationship context layer, a relationship path layer, and a fusion output layer. In the relationship context layer, the TeaConAggr (tea context aggregate) module was first used to aggregate the hop relationship contexts of entities. After that, the hop relationship contexts were summarized to obtain the relational context of entity pairs using a multi-layer relational messaging mechanism. In the relationship path layer, the relationship path aggregation module was used to aggregate the relationship paths among entities. The relationship path learning module was used to learn the paths. In the fusion output layer, multiplication attention and Softmax function were used to fuse the relationship context representation from the relationship context layer with the relationship path feature from the relationship path layer, in order to obtain the weight of each path. Subsequently, the weight was calculated with the relationship path feature to obtain the aggregated representation of the relationship path. Finally, additive attention was used to process the relationship context representation. The relationship path that aggregated representation was obtained for the predicted ranking of relationships between entity pairs, in order to achieve the goal of knowledge graph relationship completion. Data collection was constructed a knowledge graph in the field of tea planting. There were two major types of resources: books and the internet. The dataset was defined as 16 entity and 16 relationship types to represent various entities and their complex relationships in the tea planting knowledge graph. The TPKGData dataset was also developed for the tea-planting knowledge graph. The experimental results showed that the performance of the improved model was achieved in 85.40%, 80.95%, and 90.08% on the Mean Reciprocal Rank, Hits@1, and Hits@3 metrics, respectively, which were 2.56, 2.63, and 4.25 percentage points higher than those of Shallom model. The effectiveness of the model was fully met to predict the missing relationships between entities in the tea-planting knowledge graph. In addition, the comparative experiments were also conducted on the publicly available datasets FB15K-237 and WN18RR, in order to further verify the generalization of the model. At the same time, the TPKGC model also achieved the best performance on these datasets. Therefore, the model was performed outstandingly on the specific domain datasets, indicating better generalization. The stable performance was achieved in the different datasets of the knowledge graph. The TPKGC model demonstrated excellent performance and generalization potential in the task of relationship completion within the tea-planting knowledge graph. The finding can also provide important guidance for the construction of knowledge graphs.

Annotating entities and relationships have been found with the high cost in the joint extraction tasks of entity relationship. There is the strong correlation with the professional fields. It is crucial to improve the extraction performance of the model in the scenarios of low resource. In this study, the reinforcement learning-based joint extraction model was proposed for the entity relationship. The model included the following two modules: entity recognition and relationship extraction. The extraction efficiency and generalization were improved to jointly extract the entities and relationships in the low resource scenarios. The feature extractor was used to convert the text into feature representations with the richer semantics, which were shared by entity recognition and relationship extraction modules. Entity recognition was realized to utilize the CRF for sequence labeling. The output entity labeling sequence was traversed to locate the entity boundaries, and then connect the entity features for the entity embedding vectors. The limited labeled and abundant unlabeled data was considered in the low resource scenarios. Reinforcement learning was used to train the relationship extraction module. The input of the relationship extraction module included [sentence features, entity embeddings, and relationships] for the labeled data, and [sentence features, and entity embeddings] for the unlabeled data. The improved model was trained to simulate the pseudo labels that generated by unlabeled data in the gradient direction of labeled data, in order to maximize the similarity of the average gradients between them. There was the increase in the diversity and richness of the data, particularly for the better generalization with the less risk of overfitting. Meanwhile, the generated pseudo labels were reduced the dependence on a large amount of labeled data and lower annotation costs. More importantly, the gradient simulation was also balanced the sample distribution of different relationship categories in the dataset, especially in the cases of imbalanced relationship categories. The effectiveness of the model was verified to compare the mainstream models of low resource relationship extraction in the apple cultivation corpus (ATC). The results showed that the F1 score of the model was 88.71%, when the proportion of labeled data reached 30%, indicating the significantly improved model than the rest baselines. In addition, the entity relationships model was effectively extracted from the public dataset TACRED in the low resource scenarios. The proportion of unlabeled data was changed in the ATC and TACRED datasets. The experiments showed that the F1 performance varied on the fixed 10% labeled data and 10%, 30%, 50%, 70%, and 90% unlabeled data. The improved performance was achieved to add the unlabeled data for training. The optimal F1 performance was consistently achieved in the different proportions of unlabeled data. The effectiveness of the gradient simulation module was verified through ablation experiments. The relationship extraction model without gradient simulation module was basically the same as the Self-Trained_BERT model. There was an average F1 decrease of 6.12% in the Self-Trained _BERT model using different proportions of labeled data. The improved performance of the relationship extraction module was attributed to the gradient simulation module, which was improved the quality of pseudo labels. Finally, principal component analysis was used to demonstrate the gradient descent direction of the relationship extraction module for the labeled and pseudo labeled data, representing the quality of pseudo labeled data. The gradient simulation module was also added to gradually approach the ideal local minimum, although the optimization direction of pseudo label data initially fluctuated greatly. The effectiveness of the gradient simulation module was further proved to generate the high-quality pseudo labels. Therefore, the proposed model can effectively extract the entity relationships in the low resource scenarios, indicating the high generalization and performance

Uneven brightness and noise images have posed a great challenge on the recognition rate and accuracy of visual tasks, due mainly to the mutual occlusion of kiwifruit flowers in natural environments. In this article, the image enhancement was proposed to detect the kiwifruit flower under low light using improved generative adversarial network (GAN). The research object was taken as the flowers at the Meixian Kiwi Experimental Station of Northwest Agriculture and Forestry University, Shaanxi Province, China. Digital cameras and mobile phones were used to capture the images of flowers in different open states at some periods. All uniform pixel images were selected to create a dataset. Three steps were divided in the improved model of image enhancement. Firstly, GAN was used to optimize and improve the model. The channel and spatial attention were combined to perform the residual connections in the generator section. Among them, ECANet was used to more accurately extract the image features under the brightness distribution of specific areas in the channel attention. SAM was used to focus the specific regions in the image using lightweight computational operations in the spatial attention, in order to achieve the local attention. As such, the generator was obtained to extract the brightness distribution of kiwifruit low-light flower images. Secondly, a Swin Transformer block was added at the connection between the upsampling and downsampling of the generator. The Swin Transformer block was used as the standard multi-head self-attention evolution of the original Transformer. Therefore, the global modelling of the Transformer was also utilized to split the input image. The dependency relationship among each module of the image was then calculated to improve the global modelling and feature extraction of image details in the generator. Finally, the pattern collapse that encountered by single discriminators was reduced to adaptively enhance the local regions and global lighting. The generator was then provided with the required adaptive adjustment. A dual discriminator mechanism was utilized to enhance the perception of image details. The accuracy of image evaluation was improved to generate the clearer images in the generator network. On the self-built dataset, the PSNR and NIQE of the improved model were 7.09 and 10.36, respectively. Meanwhile, a comparison was then made on the five models, such as RetinexNet, Enlighten GAN, Zero DEC, Cycle GAN, and Diffusion Low Light. The peak signal-to-noise ratio of the improved model was 0.11db higher than that of EnlightenGan. The quality assessment of natural image was 12.41% lower than that of Zero DCE. Therefore, the low-light image enhancement model was performed the better quality, lower distortion, better naturalness, and smaller color deviation after enhancement. This improved model can also be applied in the practical production to develop the computer vision in smart agriculture.