Sen Yang

Accurate localization of potato bud eyes supports automated seed-tuber cutting, but small target size, low contrast, and cultivar-dependent appearance complicate detection. This study proposes FFCR-YOLO (Frequency Fusion and Channel Recalibration YOLO), a bud-eye detector based on YOLOv8n. A total of 4460 original images from eight cultivars were acquired on a sorting machine, with one image per independent tuber. After partitioning by physical tuber, offline augmentation produced 13,380 derived images, yielding a dataset of 17,840 images. Each original image and its derivatives remained in the same subset. Tuber morphology and estimated bud-eye surface density were analyzed to describe cultivar variation. The model incorporates C2f with Frequency-Band Fusion (C2f-FBF) to combine low- and high-frequency feature components, a Feature-map Transformer Encoder (FTE) to model global context, and Channel-Recalibrated Cross-scale Fusion (CRCF) to fuse features at different resolutions and recalibrate their channel responses. MPDIoU is used for bounding-box regression with corner-distance constraints. On the mixed test set of 446 original and 1338 derived images from 446 independent tubers, FFCR-YOLO achieved 98.4% precision, 92.4% recall, 95.3% F1-score, 97.4% mAP@0.5, and 78.5% mAP@0.5:0.95. Relative to YOLOv8n, recall and mAP@0.5 increased by 12.3 and 4.4 percentage points, respectively. The model had 4.3 million parameters and a forward-pass time of 3.29 ms/image on an NVIDIA RTX 3090 GPU at a batch size of 1, excluding image preprocessing and NMS postprocessing. In the leave-one-cultivar-out (LOCO) evaluations, training used original images of the other seven cultivars; FFCR-YOLO achieved macro recall of 88.40% and macro mAP@0.5 of 95.33%. These results support FFCR-YOLO as a visual perception component for automated seed-tuber cutting, with improved bud-eye detection across the evaluated cultivars.

Adapting monocular metric depth to agricultural imagery requires balancing target-domain accuracy against the scope of model updates and embedded execution cost. We evaluate SARE-Depth on CitrusFarm using a frozen teacher-guided Depth Anything V2 Small base. Depth-Preserving Residual Head Refinement (DP-RHR) learns image-level log-scale and gated local corrections, followed by fixed depth-conditioned Range-Expert (RE) refinement. Before adaptation, fixed-checkpoint evaluations show persistent overestimation under both square and aspect-preserving input. On 750 internal-test images with common FP32 references and a 30 m cap, scalar calibration reduces absolute relative error (AbsRel) from 4.782 to 0.476. DP-RHR reaches 0.325 with 93,123 updated parameters; complete metric-head fine-tuning achieves 0.220 ± 0.006 across three runs with 2,728,513 updated parameters. Subsequent RE refinement provides smaller, domain-dependent gains with reduced threshold accuracy, while separate Dingxi examples reveal object-distance underestimation. Cumulative Jetson Orin NX latency increases from 26.87 ms for the base to 31.49 ms with DP-RHR and 35.08 ms for the complete path. These comparisons relate adaptation scope to accuracy and measured execution cost in this agricultural case study.

Leaf chlorophyll content (LCC), leaf area index (LAI), and above-ground biomass (AGB) are important growth parameters for characterizing potato growth and predicting yield. While deep learning has demonstrated remarkable advancements in estimating crop growth parameters, the limited availability of field data often compromises model accuracy and generalizability, impeding large-scale regional applications. This study proposes a novel deep learning model that integrates multi-task learning and few-shot learning to address the challenge of low data in growth parameter prediction. Two multi-task learning architectures, MTL-DCNN and MTL-MMOE, were designed based on deep convolutional neural networks (DCNNs) and multi-gate mixture-of-experts (MMOE) for the simultaneous estimation of LCC, LAI, and AGB from Sentinel-2 imagery. Building on this, a few-shot learning framework for growth prediction (FSLGP) was developed by integrating simulated spectral generation, model-agnostic meta-learning (MAML), and meta-transfer learning strategies, enabling accurate prediction of multiple growth parameters under limited data availability. The results demonstrated that the incorporation of calibrated simulated spectral data significantly improved the estimation accuracy of LCC, LAI, and AGB (R² = 0.62~0.73). Under scenarios with limited field measurement data, the multi-task deep learning model based on few-shot learning outperformed traditional mixed inversion methods in predicting potato growth parameters (R² = 0.69~0.73; rRMSE = 16.68%~28.13%). Among the two architectures, the MTL-MMOE model exhibited superior stability and robustness in multi-task learning. Independent spatiotemporal validation further confirmed the potential of MTL-MMOE in estimating LAI and AGB across different years and locations (R² = 0.37~0.52). These results collectively demonstrated that the proposed FSLGP framework could achieve reliable estimation of crop growth parameters using only a very limited number of in-field samples (approximately 80 samples). This study can provide a valuable technical reference for monitoring and predicting growth parameters in other crops.

To address the challenges of low recognition accuracy and substantial model complexity in crop disease identification models operating in complex field environments, this study proposed a novel hybrid model named ConvTransNet-S, which integrates Convolutional Neural Networks (CNNs) and transformers for crop disease identification tasks. Unlike existing hybrid approaches, ConvTransNet-S uniquely introduces three key innovations: First, a Local Perception Unit (LPU) and Lightweight Multi-Head Self-Attention (LMHSA) modules were introduced to synergistically enhance the extraction of fine-grained plant disease details and model global dependency relationships, respectively. Second, an Inverted Residual Feed-Forward Network (IRFFN) was employed to optimize the feature propagation path, thereby enhancing the model’s robustness against interferences such as lighting variations and leaf occlusions. This novel combination of a LPU, LMHSA, and an IRFFN achieves a dynamic equilibrium between local texture perception and global context modeling—effectively resolving the trade-offs inherent in standalone CNNs or transformers. Finally, through a phased architecture design, efficient fusion of multi-scale disease features is achieved, which enhances feature discriminability while reducing model complexity. The experimental results indicated that ConvTransNet-S achieved a recognition accuracy of 98.85% on the PlantVillage public dataset. This model operates with only 25.14 million parameters, a computational load of 3.762 GFLOPs, and an inference time of 7.56 ms. Testing on a self-built in-field complex scene dataset comprising 10,441 images revealed that ConvTransNet-S achieved an accuracy of 88.53%, which represents improvements of 14.22%, 2.75%, and 0.34% over EfficientNetV2, Vision Transformer, and Swin Transformer, respectively. Furthermore, the ConvTransNet-S model achieved up to 14.22% higher disease recognition accuracy under complex background conditions while reducing the parameter count by 46.8%. This confirms that its unique multi-scale feature mechanism can effectively distinguish disease from background features, providing a novel technical approach for disease diagnosis in complex agricultural scenarios and demonstrating significant application value for intelligent agricultural management.

In order to improve the performance of potato planter, reduce miss-seeding rates, enhance the overall quality of the seeding operation, and ultimately increase the yield of the potato, it is necessary to implement effective technical means to monitor and identify the miss-seeding issues during the seeding process. The existing miss-seeding detection technologies commonly use sensors to monitor, but such technologies are easily affected by factors like heavy dust and strong vibrations, resulting in poor interference resistance and adaptability. Therefore, this study aims to explore and apply deep learning algorithms to achieve real-time monitoring of the miss-seeding phenomenon in potato planter during the planting process. Considering both the lightweight of the miss-seeding detection model and its practical deployment, this study selects and adapts the YOLOv5s algorithm to achieve this goal. Firstly, the attention mechanism is integrated into the backbone network to suppress background interference and improve detection accuracy. Secondly, the non-maximum suppression algorithm is improved by replacing the original IoU-NMS with the Soft-NMS algorithm to enhance the bounding box regression rate and reduce missed detections of potato seeds due to background overlap or occlusion. Experimental results show that the accuracy of the improved algorithm in detecting miss-seeding increased from 96.02% to 98.30%, the recall rate increased from 96.31% to 99.40%, and the mean average precision (mAP) improved from 99.12% to 99.40%. The improved model reduces missed and false detections, provides more precise target localization, and is suitable for miss-seeding detection in natural environments for potato planter, providing technical and theoretical support for subsequent intelligent reseeding in potato planter.

The knowledge within wheat production chain data has multiple levels and complex semantic relationships, making it difficult to extract knowledge from them. Therefore, this paper proposes a fine-grained knowledge extraction method for the wheat production chain based on ontology. For the first time, the conceptual layers of ploughing, planting, managing, and harvesting were defined around the main agricultural activities of the wheat production chain. Based on this, the entities, relationships, and attributes in the conceptual layers were defined at a fine-grained level, and a spatial–temporal association pattern layer with four conceptual layers, twenty-eight entities, and forty-two relationships was constructed. Then, based on the characteristics of the self-constructed dataset, the Word2vec-BiLSTM-CRF model was designed for extracting the knowledge within it, i.e., the entity–relationship–attribute model and the Word2vec-BiLSTM-CRF model in this paper were compared with the four SOTA models. The results show that the accuracy and F1 value improved by 8.44% and 8.89%, respectively, compared with the BiLSTM-CRF model. Furthermore, the entities of the pest and disease dataset were divided into two different granularities for the comparison experiment; the results show that for entities with “disease names” and “pest names”, the recognition accuracy at the fine-grained level is improved by 32.71% and 31.58%, respectively, compared to the coarse-grained level, and the recognition performance of various fine-grained entities has been improved.

The high performance of deep learning networks relies on large datasets and powerful computational resources. However, collecting enough diseased training samples is a daunting challenge. In addition, existing few-shot learning models tend to suffer from large size, which makes their deployment on edge devices difficult. To address these issues, this study proposes a pruning-based lightweight few-shot learning (Prune-FSL) approach, which aims to utilize a very small number of labeled samples to identify unknown classes of crop diseases and achieve lightweighting of the model. First, the disease few-shot learning model was built through a metric-based meta-learning framework to address the problem of sample scarcity. Second, a slimming pruning method was used to trim the network channels by the γ coefficients of the BN layer to achieve efficient network compression. Finally, a meta-learning pruning strategy was designed to enhance the generalization ability of the model. The experimental results show that with 80% parameter reduction, the Prune-FSL method reduces the Macs computation from 3.52 G to 0.14 G, and the model achieved an accuracy of 77.97% and 90.70% in 5-way 1-shot and 5-way 5-shot, respectively. The performance of the pruned model was also compared with other representative lightweight models, yielding a result that outperforms those of five mainstream lightweight networks, such as Shufflenet. It also achieves 18-year model performance with one-fifth the number of parameters. In addition, this study demonstrated that pruning after sparse pre-training was superior to the strategy of pruning after meta-learning, and this advantage becomes more significant as the network parameters are reduced. In addition, the experiments also showed that the performance of the model decreases as the number of ways increases and increases as the number of shots increases. Overall, this study presents a few-shot learning method for crop disease recognition for edge devices. The method not only has a lower number of parameters and higher performance but also outperforms existing related studies. It provides a feasible technical route for future small-sample disease recognition under edge device conditions.

Potato diseases and pests have a serious impact on the quality and yield of potatoes, and timely prevention and control of potato diseases and pests is essential. A rich knowledge reserve of potato diseases and pests is one of the most important prevention and control measures; however, valuable knowledge is buried in the massive data of potato diseases and pests, making it difficult for potato growers and managers to obtain and use it in a timely manner and to develop the potential of knowledge. Therefore, this paper explores the construction method of a knowledge graph for automatic knowledge extraction, which extracts the knowledge of potato diseases and pests scattered in heterogeneous data from multiple sources, organises it into a semantically related knowledge base, and provides potato growers with professional knowledge and timely guidance to effectively prevent and control potato diseases and pests. In this paper, a data corpus on potato diseases and pests, called PotatoRE, is first constructed. Then, a model of ALBert-BiLSTM-Self_Att-CRF is designed to extract knowledge from the corpus to form a triplet structure, which is imported into the Neo4j graph database for storage and visualisation. Furthermore, the performance of the model constructed in this paper is compared and verified using the datasets PotatoRE and People’s Daily. The results show that compared to the SOTA models of ALBert BiLSTM-CRF and ALBert BiGRU-CRF, the accuracy of our model has been improved by 2.92% and 3.12%, respectively, using PotatoRE. Compared to the Bert BiLSTM-CRF model on two datasets, our model not only improves the accuracy, recall, and F1 values, but also has a higher efficiency. The model in this paper solves the problem of the difficult recognition of nested entities. On this basis, through comparative experiments, the TransH model is used to effectively represent the constructed knowledge graph, which lays the foundation for achieving inference, extension, and automatic updating of the knowledge base. The achievements of the thesis have made certain contributions to the automatic construction of large-scale knowledge bases.

The high performance of deep learning networks relies mainly on massive data. However, collecting enough samples of crop disease is impractical, which significantly limits the intelligent diagnosis of diseases. In this study, we propose Heterogeneous Metric Fusion Network-based Few-Shot Learning (HMFN-FSL), which aims to recognize crop diseases with unseen categories using only a small number of labeled samples. Specifically, CBAM (Convolutional Block Attention Module) was embedded in the feature encoders to improve the feature representation capability. Second, an improved few-shot learning network, namely HMFN-FSL, was built by fusing three metric networks (Prototypical Network, Matching Network, and DeepEMD (Differentiable Earth Mover’s Distance)) under the framework of meta-learning, which solves the problem of the insufficient accuracy of a single metric model. Finally, pre-training and meta-training strategies were optimized to improve the ability to generalize to new tasks in meta-testing. In this study, two datasets named Plantvillage and Field-PV (covering 38 categories of 14 crops and containing 50,403 and 665 images, respectively) are used for extensive comparison and ablation experiments. The results show that the HMFN-FSL proposed in this study outperforms the original metric networks and other state-of-the-art FSL methods. HMFN-FSL achieves 91.21% and 98.29% accuracy for crop disease recognition on 5way-1shot, 5way-5shot tasks on the Plantvillage dataset. The accuracy is improved by 14.86% and 3.96%, respectively, compared to the state-of-the-art method (DeepEMD) in past work. Furthermore, HMFN-FSL was still robust on the field scenes dataset (Field-PV), with average recognition accuracies of 73.80% and 85.86% on 5way-1shot, 5way-5shot tasks, respectively. In addition, domain variation and fine granularity directly affect the performance of the model. In conclusion, the few-shot method proposed in this study for crop disease recognition not only has superior performance in laboratory scenes but is also still effective in field scenes. Our results outperform the existing related works. This study provided technical references for subsequent few-shot disease recognition in complex environments in field environments.

Potato malformation seriously affects commercial value, and its removal has become one of the core steps in the post-harvest and pre-sales process of potatoes. At present, this work mainly relies on manual visual inspection, which requires a lot of labor and incurs high investment costs. Therefore, precise and efficient automatic detection technology urgently needs to be developed. Due to the efficiency of deep learning based on image information in the field of complex object feature extraction and pattern recognition, this study proposes the use of the YOLOv3 algorithm to undertake potato malformation classification. However, the target box regression loss function MSE of this algorithm is prone to small errors being ignored, and the model code is relatively large, which limits its performance due to the high demand for computing hardware performance and storage space. Accordingly, in this study, CIOU loss is introduced to replace MSE, and thus the shortcoming of the inconsistent optimization direction of the original algorithm’s loss function is overcome, which also significantly reduces the storage space and computational complexity of the network model. Furthermore, deep separable convolution is used instead of traditional convolution. Deep separable convolution first convolves each channel, and then combines different channels point by point. With the introduction of an inverted residual structure and the use of the h-swish activation function, deep separable convolution based on the MobileNetv3 structure can learn more comprehensive feature representations, which can significantly reduce the computational load of the model while improving its accuracy. The test results showed that the model capacity was reduced by 66%, mAP was increased by 4.68%, and training time was shortened by 6.1 h. Specifically, the correctness rates of malformation recognition induced by local protrusion, local depression, proportional imbalance, and mechanical injury within the test set range were 94.13%, 91.00%, 95.52%, and 91.79%, respectively. Misjudgment mainly stemmed from the limitation of training samples and the original accuracy of the human judgment in type labeling. This study lays a solid foundation for the final establishment of an intelligent recognition and classification picking system for malformed potatoes in the next step.