Accurate and rapid detection and counting of winter wheat seedlings are often required to obtain key seedling information, including emergence density and emergence rate, under complex field conditions. However, conventional detection is limited by variations in soil background and illumination, due to generally small, densely distributed, and easily occluded seedling targets. This study aims to propose an improved lightweight YOLOv8n model for the detection and counting of winter wheat seedlings in fields. Visible-light images were acquired from a winter wheat experimental field using near-ground Internet of Things (IoT) devices, unmanned aerial vehicles (UAVs), and smartphones. Images were collected at four representative emergence stages, namely 5, 9, 14, and 20 days after sowing, including newly emerged, one-leaf, two-leaf, and three-leaf seedlings under sunny and cloudy conditions. A multi-source dataset of winter wheat seedlings was obtained with 4 550 images after image cleaning, stratified dataset partitioning, and data augmentation. The dataset was then divided into 3 185 training images, 910 validation images, and 455 test images. A YOLOv8n-GC-S (ghost-convolutional block attention module-small object detection layer YOLOv8n) model was then developed. 1) The original YOLOv8n backbone was replaced with the lightweight GhostNet backbone. Additional feature maps were generated after low-cost linear transformations. Thereby, the number of model parameters was reduced by 28.1% compared with redundant convolution. 2) A convolutional block attention module (CBAM) was introduced between the backbone and neck. Channel and spatial attention were used to enhance seedling-related features and then suppress interference from soil and crop residues under complex backgrounds. The CBAM increased mAP@0.5 by 9.4 percentage points, compared with the GhostNet lightweight model. 3) A small-object detection branch was added with a feature-map size of 160 × 160 to retain shallow spatial features. The small-scale seedlings were represented to strengthen weak visual features. The mAP@0.5 increased by 12.6 percentage points, compared with the GhostNet lightweight model. Ablation experiments were conducted to evaluate the contribution rates of each modification. A comparison was also made with mainstream object detection models under the same dataset and experimental conditions. The experimental results showed that the precision, recall, and mAP@0.5 of the YOLOv8n-GC-S model reached 92.1%, 80.0%, and 86.2%, respectively, which were 2.3, 5.9, and 2.9 percentage points higher than those of the baseline YOLOv8n model. The final model contained 2.8 million parameters with a weight file size of 5.7 MB, which was reduced by 12.5% and 8.1%, respectively, compared with the baseline model. Furthermore, the missed-detection rate decreased from 8.0% to 4.7% under cloudy conditions, whereas it decreased from 12.7% to 7.8% under sunny conditions. The average missed detection rate decreased from 10.4% to 6.3% under both illumination conditions, with a reduction of 4.1 percentage points. Multi-temporal counting showed that there was an excellent agreement between model-predicted and manually counted seedling numbers. The coefficient of determination ( R²) ranged from 0.88 to 0.94 over the four emergence stages, while the root mean square error (RMSE) ranged from 4.73 to 7.61 seedlings. The best counting agreement was obtained at the one-leaf stage, with an R² of 0.94 and an RMSE of 4.73 seedlings. In addition, emergence density and emergence rate were calculated from the seedling numbers after detection. The emergence rate after estimation was highly consistent with the manually measured ones, with an R² of 0.93 and an RMSE of 2.52%. The YOLOv8n-GC-S achieved a favorable balance between detection performance and a lightweight model. Small seedlings of winter wheat can be detected and counted under varying illumination and complex backgrounds. The findings can also support seedling counting and emergence-rate assessment in field environments.

Accurate extraction of phenotypic parameters from cigar tobacco plants is essential for scientifically assessing plant growth status, guiding precision breeding, and optimizing cultivation management. However, cigar plants exhibit complex structural characteristics, such as severe inter-layer occlusion and tightly overlapping leaves, which often lead to point cloud adhesion and blurred instance boundaries. To address these challenges, this study proposes a novel method for instance segmentation and phenotypic parameter extraction based on 3D point clouds. First, we develop an enhanced 3D point cloud instance segmentation model, termed PnP-3D-PSegNet, which integrates geometric structure with global contextual semantics. Specifically, a PnP-3D module is embedded into the encoder of the baseline PSegNet architecture to improve the model’s ability to capture fine-grained structural features in complex plant point clouds. This design enables more precise discrimination of individual plant organs under challenging conditions. Second, to tackle the issues of blurred instance boundaries and leaf adhesion, an improved Mean Shift clustering algorithm is introduced. We design an instance-aware clustering strategy that combines edge-aware feature integration with density–distance weighted constraints. By incorporating an edge perception mechanism based on local curvature and variations in surface normals, edge confidence is estimated and integrated with instance features, thereby guiding the clustering process in an enhanced feature space and effectively reducing the influence of unreliable boundary points. This significantly alleviates the boundary ambiguity problem in instance segmentation. Furthermore, to mitigate the incorrect assignment of points between adjacent leaves under adhesion conditions, a weighted criterion based on local point density and the distance to cluster centers is employed, further improving segmentation accuracy in densely overlapping regions. Experimental results demonstrate that the proposed PnP-3D-PSegNet model achieves notable improvements over the baseline PSegNet. Specifically, the model reaches 93.59% in mean coverage, 95.77% in mean weighted coverage, 97.21% in mean precision, and 98.55% in mean recall, representing increases of 3.39, 3.56, 3.65, and 2.78 percentage points, respectively. These results highlight the effectiveness of the proposed feature fusion and optimization strategies. Based on the segmentation outputs, six key phenotypic parameters are further extracted, including plant height, leaf number, leaf length, leaf width, leaf area, and canopy width. Comparisons with manual measurements show strong agreement, with coefficients of determination (R²) of 0.96, 0.97, 0.92, 0.95, 0.98, and 0.97, respectively. The corresponding root mean square errors (RMSE) are 1.20 cm for plant height, 0.82 cm for leaf length, 1.13 cm for leaf width, 1.26 cm for canopy width, 11.83 cm² for leaf area, and 0.32 leaves for leaf count. These results confirm the high accuracy and reliability of the proposed method. Additionally, experiments conducted on a tomato point cloud dataset further demonstrate the strong cross-crop generalization capability of PnP-3D-PSegNet. Overall, this study establishes a complete pipeline from point cloud segmentation to phenotypic parameter extraction. The proposed feature fusion and boundary optimization strategies not only effectively address the problem of leaf adhesion in cigar tobacco plants but also provide valuable technical insights for 3D phenotypic analysis of other complex crops.

Accurate perception of tomato inflorescences and flower states can often be required to support key operations in greenhouse tomato production, such as pollination and topping at the flowering and fruiting stage. However, inflorescences and flowers are characterized by small target size, dense spatial distribution, complex backgrounds, and frequent occlusion by leaves and stems in practical greenhouse environments. Single-stage detection models cannot simultaneously realize stable inflorescence localization and high-precision flower state recognition when operating on whole-plant images. In this study, a two-stage cascaded framework of visual perception was developed to detect tomato inflorescence and flower states using an improved YOLO version 11 network. A “spatial localization followed by fine-grained recognition” strategy was adopted to decompose the overall perception task into two sequential subtasks. In the first stage, an inflorescence detection model was constructed to enhance the baseline YOLO version eleven network with a Deformable Large Kernel Attention mechanism and a Dynamic Head detection structure. Deformable Large Kernel Attention Mechanism also employed large convolutional kernels with deformable convolution to capture long-range contextual information between inflorescences and adjacent peduncles, while morphological variations were also considered at different growth stages and plant structures. The dynamic head module incorporated scale- and spatial-aware feature modeling, thereby enabling the detector to robustly handle inflorescences of varying sizes for the complex regions, where inflorescences and leaves overlapped. The output precise spatial regions corresponded to inflorescences, which served as reliable regions of interest for subsequent analysis. In the second stage, a flower state recognition model was designed to operate exclusively within the inflorescence regions in the first stage. A lightweight backbone network, MobileNetV4, was adopted to reduce computational complexity for inference efficiency, while preserving feature representation. An Adaptive Task-aligned Focal Loss function was introduced to balance sample distribution among different flower developmental states. This loss function dynamically adjusted category weights, according to classification difficulty and sample frequency, thereby enhancing recognition performance for the minority and easily confused flower states under occlusion and cluttered backgrounds. Experiments were conducted on a greenhouse tomato image dataset at multiple growth stages and complex environments. In the inflorescence task, the first-stage model achieved substantial improvements in performance, compared with the baseline network, with the precision, recall, mean average precision at an intersection-over-union threshold of 0.5, and F1-score increasing by 4.02, 5.25, 8.49, and 4.66 percentage point, respectively. These results demonstrated that the attention and detection head enhancements significantly improved small-target detection stability in complex scenes. In the flower state recognition task, the second-stage model further improved precision, recall, mean average precision at the same threshold, and F1-score by 5.24, 2.97, 5.31, and 4.07 percentage point, respectively, indicating stronger identification for fine-grained flower state classification. The cascaded framework achieved an average processing speed of 38.4 frames per second under a single-input condition, fully meeting the real-time requirements of continuous greenhouse monitoring and online agricultural operations. The cascaded framework effectively balanced detection accuracy and computational efficiency to decouple spatial localization from fine-grained recognition. The reliable inflorescence and flower state recognition from whole-plant images can provide a practical visual perception for pollination, topping, and intelligent operations in greenhouse tomato production.

Garlic is one of the major agricultural products and export commodities in China. It is of great significance for the accurate early warning of garlic price fluctuations in the market. However, the time series of garlic price is often characterized by the nonlinear, non-stationary, and “sharp-peaked and heavy-tailed” distribution. Conventional early warning cannot effectively capture complex temporal features, resulting in limited warning accuracy. In this study, a VMD-TCAN framework was proposed to forecast the garlic price. The early warning model was then graded for the precise prediction and automatic warning. A price alert indicator system was also constructed to divide the warning threshold. Firstly, Variational Mode Decomposition (VMD) was introduced to mitigate the high noise and non-stationarity of the original price series. The optimal solution of variational modes was searched iteratively, and then adaptively decomposed the signal into intrinsic mode functions (IMFs) with specific sparsity and finite bandwidth. Mode mixing and end effects were also reduced in the VMD, unlike Empirical Mode Decomposition (EMD). The Particle Swarm Optimization (PSO) algorithm was employed to reduce the subjectivity in VMD parameter selection. The minimum fuzzy entropy was taken as the fitness function to dynamically determine the optimal number of modes and penalty factors. The garlic price series was thereby decomposed into eight IMFs with the center frequencies. Multi-scale components were effectively separated, such as long-term trends, seasonal fluctuations, and short-term random shocks. Secondly, a TCAN prediction model was developed using multi-source information fusion. Six external influencing factors were selected to cover the production (planting area and planting cost), circulation (inventory and export volume), and market (exchange rate and CPI). A high-dimensional input tensor was then formed, together with the VMD-derived IMF components and daily garlic prices. Causal convolution dilated using the TCAN architecture. The long-range historical dependency features were efficiently extracted to exponentially expand the receptive fields with the temporal causality. A self-attention mechanism was further introduced to globally perceive multi-source inputs. The correlation weights were dynamically assigned to the critical time steps that significantly influenced the price movements. The performance was substantially enhanced to capture the abnormal price volatility intensity and trend turning points under complex market dynamics. Finally, an integrated “forecasting–early warning” framework was established after evaluation. The Shapiro–Wilk test revealed that the garlic price volatility followed the non-normal, sharp-peaked, and heavy-tailed distribution, thereby rendering the conventional rule inapplicable. Therefore, a quantile-based method was adopted in the historical volatility distribution. Five warning levels were defined: severe negative warning, mild negative warning, no warning, mild positive warning, and severe positive warning. Continuous numerical forecasts were thus mapped into discrete risk-level signals. Daily price data were collected from the Jinxiang production area from 2005 to 2024. Experimental results show that the VMD-TCAN model achieved a root mean square error (RMSE) of 0.042 yuan/(500 g) and a coefficient of determination (R²) of 0.998, significantly outperforming benchmark models and their VMD hybrid versions. In terms of early warning performance, the accuracy and F1-score were improved by 4.86 percent points and 4.97 percent points, respectively, compared with the TCN model. More precise identification of abnormal fluctuations was realized in the garlic market. The timely and reliable price forecasts and warnings can greatly contribute to the decision-making in the garlic industry.

An accurate prediction of the garlic price can be one of the most important tasks in modern agriculture. However, the conventional single forecasting models are often limited to the high volatility of the garlic price data due to the complex and intertwined influencing factors. In this study, a high-precision ensemble prediction framework was proposed using deep learning fusion. Three modules were the two-stage feature selection, adaptive sub-model construction, and nonlinear optimization. Firstly, the two-stage feature selection was implemented to integrate the maximum relevance and minimum redundancy (mRMR) algorithm with the least absolute shrinkage and selection operator (LASSO) regression. The mRMR algorithm selected the features to maximize their correlation with the target variable. While the redundancy of the features was minimized to filter out the irrelevant and overlapping variables, in order to reduce the computational complexity and noise interference. The LASSO was used to further refine the subset. The L1 regularization was applied to promote the sparsity in the coefficients and the robustness to suppress the multicollinearity. The resulting feature subset contained the most informative variables that highly correlated with the garlic prices, providing a reliable input for the subsequent modeling. Secondly, two complementary sub-models were designed to solve the multi-scale variability in the garlic prices. The influencing factors and historical garlic prices were input into a bidirectional long short-term memory (BiLSTM) network. Temporal dependencies were captured in both forward and backward directions using BiLSTM. Nonlinear relationships were simulated for a stable prediction sequence. The original series of the garlic price was decomposed by variational mode decomposition (VMD) into several intrinsic mode functions (IMFs) with frequency components. Each IMF was modeled using a deep hybrid kernel extreme learning machine (DHKELM). The global and local kernel advantages were integrated to effectively learn the nonlinear patterns within each component. The prediction of the IMFs was reconstructed to form the final forecast sequence, thereby enhancing the interpretability of the low-frequency trends and high-frequency fluctuations. Thirdly, the advanced optimization was introduced to further enhance the model performance and convergence efficiency. The black-winged kite algorithm (BKA) was employed to optimize the kernel parameters and hidden-layer nodes of the DHKELM. The global search was improved to avoid the local optima. Meanwhile, Bayesian optimization (BYS) was used to fine-tune the BiLSTM hyperparameters, such as the learning rate, number of neurons, and batch size, indicating the high convergence and better generalization. Finally, the outputs of the sub-models were fused using a nonlinear ensemble structure with a convolutional neural network, bidirectional gated recurrent unit, and attention mechanism (CNN–BiGRU–Attention), instead of conventional linear averaging. The CNN was used to extract the local nonlinear interactions among the sub-model outputs. While the BiGRU was adopted to capture the long-term temporal dependencies and cross-period correlations. The attention layer dynamically assigned the adaptive weights to the important time steps and high-performing sub-models. Meanwhile, the end-to-end nonlinear prediction was fused to effectively reduce the accumulated errors for the complementary information. Experimental results demonstrate that the ensemble model achieved a mean absolute percentage error (MAPE) of 1.61%, a root mean square error (RMSE) of 0.078, and a coefficient of determination (R²) of 0.976. The R² value was improved by approximately 2%–7%, compared with the individual sub-models. The ensemble model also outperformed the conventional linear combination in terms of stability and accuracy. The forecasting precision of the garlic price was significantly enhanced after optimization. The framework can also provide a quantitative decision-making reference for the garlic market regulation. The finding can also serve as a reusable foundation for the price prediction of the agricultural commodities that are characterized by high volatility, complex interactions, and nonlinear dynamics.

Phenotypic extraction is often required for the facility-cultivated cucumber plants. The major challenges can arise from their significant morphological transformations across growth stages, frequent occlusions from the overlapping leaves and stems, greenhouse lighting, and variable suboptimal solutions. Particularly, the traditional manual phenotyping is labor-intensive, time-consuming, and subjective. It is seriously limited to the large-scale, high-throughput plant breeding and precision agriculture. In this study, a fully automated, robust, and highly accurate framework was developed and validated to quantify the key parameters of the cucumber growth. A multi-modal data acquisition was also constructed to capture the synchronized RGB and depth images. Both top-down and lateral perspectives were also selected to record the data. Complete spatial coverage of the cucumber canopy was obtained over its entire life cycle. A lightweight instance segmentation model (named YOLO-LCOS) was also developed, according to an enhanced YOLOv11n-seg architecture. The HGNetV2 network was integrated as the backbone for more computationally efficient feature extraction. A hierarchical scale-wise feature pyramid network (HSFPN) module was incorporated into the network neck for superior multi-scale feature fusion. The plant organs of vastly different sizes were effectively segmented after enhancement. An efficient multiscale attention (EMA) mechanism was embedded to adaptively prioritize salient features while suppressing the irrelevant background clutter and noise. An extraction pipeline for the phenotypic parameter was engineered after segmentation. The high-precision mask outputs from YOLO-LCOS were fused with 3D spatial information from the depth maps. Furthermore, geometric modeling techniques were then applied to quantitatively measure the critical agronomic traits, including the total leaf count, total and individual leaf surface area, plant height, and flower count. Experimental results show that the highly positive superior performance was achieved in the refined YOLO-LCOS model after improvement, compared with the baseline YOLOv11n-seg. There were the 7.8% and 8.8% reductions in the floating-point operations and the total parameters, respectively, indicating a more streamlined architecture. Remarkably, this complexity reduction coincided with a 1.5% increase in the inference speed. More critically, the segmentation accuracy was improved substantially: The mean average precision (mAP), precision, and recall increased by 8.8 percentage points, 7.1 percentage points, and 6.9 percentage points, respectively. The end-to-end phenotypic pipeline was evaluated after optimization. The results revealed that there was an exceptionally strong agreement with manual ground-truth data. The coefficients of determination (R²) were 0.95 for the leaf count, 0.96 for leaf area, 0.95 for plant height, and 0.90 for flower count, respectively. The low error metrics were achieved in: Root mean square errors (RMSE) were 0.86 for leaf count, 8.20 cm² for leaf area, 9.80 cm for plant height, and 0.81 for flower count, respectively; Mean absolute errors (MAE) were 0.50, 6.50 cm², 7.21 cm, and 0.54, respectively. Particularly, the high accuracy was observed for the complex traits, like the leaf area and plant height, indicating the pipeline's resilience to occlusions and precise spatial reasoning. In conclusion, an integrated framework was successfully developed and validated to accurately extract the key phenotypic parameters from the cucumber plants under complex facility environments. Notably, the lightweight YOLO-LCOS segmentation model was integrated with its HGNetV2 backbone, HSFPN module, and EMA mechanism. The highly effective performance was achieved in the morphological variation, occlusion, and lighting interference. Consequently, a powerful and reliable pipeline for the trait quantification was formed with the advanced image segmentation, spatial depth information, and geometric modeling. The accurate, robust, efficient, and practical approach can be expected to deploy for the real-world greenhouse. Thus, this finding can provide a solid technical foundation for the continuous, non-invasive, high-throughput digital crop monitoring. Significant promise can also be held to advance precision horticulture. Critical data can be supplied for the optimal cultivation, intervention, and breeding programs for facility cucumbers. A reliable technological solution can also be supported for data-driven decision-making for the digital monitoring and fine-grained production in the greenhouse.

Solar greenhouses have been widely used to provide abundant fruits and vegetables in the cold regions of China in winter. The microclimate in the greenhouse is crucial to the crop quality and yield. But the greenhouse environment can often be regulated for dehumidification and carbon dioxide supplementation using low-cost natural ventilation. The low indoor temperatures have resulted from the significant heat loss. In this research, active ventilation was proposed to reduce the heat loss for the better suitability of winter greenhouse environments using the spatial distribution of temperature and humidity. Firstly, the differences in temperature and humidity were quantitatively analyzed in the different regions of the greenhouse. 28 sensors of temperature and humidity were deployed inside and outside the greenhouse. The data was collected for 115 days during winter and spring, respectively. A field test was performed on the winter greenhouses in Shandong Province, China. The results show that there were significant differences in temperature and humidity in the different areas. The upper-middle areas shared the higher temperatures and lower humidity, while the main growth area of the crops, i.e., the lower-middle area, had the lower temperatures and higher humidity. The maximum accumulated temperature difference between the upper and lower areas during winter can reach up to 300℃, which is 30% of the average accumulated temperature, and the relative humidity difference during daylight hours can be up to 20 percentage points. This heterogeneous distribution was attributed to the substantial heat dissipation and the low dehumidification efficiency that was caused by the replacement of warm, dry air at the top with outside air during natural ventilation. Secondly, an active ventilation dehumidification was proposed to utilize the temperature and humidity heterogeneous distribution. The axial fans were installed at the bottom of the greenhouse. The direction of airflow was adjusted to remove the cold and humid air from the bottom, while the top air was replenished with outside air, in order to realize the efficient convective exchange of natural ventilation. Furthermore, 4 axial fans were installed at the bottom of the greenhouse. The maximum ventilation capacity was achieved at 5 300 m³/h when each fan with a power of 550 W. The uniformity of the airflow field was also achieved inside the greenhouse during the ventilation process. A 70-meter-long corrugated pipe was laid along the east-west direction of the greenhouse, where evenly spaced openings faced the greenhouse. The wet and cold air was preferentially removed from the lower part of the greenhouse while retaining the dry and warm air in the upper part, in order to improve the dehumidification and insulation. Thirdly, an evaluation system was constructed for the insulation and dehumidification performance of the ventilation. The indicators included the ventilation rate, daily average temperature-humidity ratio, and return on investment. The active ventilation rates were finely adjusted from 0 to 30 m³/(m²∙h), compared with the less controllable natural ventilation subjected to the indoor and outdoor climates. The daily average temperature-humidity ratio with the active ventilation was consistently higher than that with the natural ventilation. The better meteorological adaptability and stability were obtained under different weather conditions, such as sunny, cloudy, and rainy. In particular, the daily average temperature-humidity ratio increased by 33.1%-32.9% in sunny, whereas, the absolute humidity in the air decreased by about 1 g/kg. The indoor temperature increased by about 2.0-2.7 ℃, indicating the better performance of the insulation and dehumidification. The active ventilation still provided significant dehumidification and thermal insulation performance in cloudy weather. The relative humidity was reduced by about 15 percentage points, whereas, the average temperature increased by about 2-3 ℃. However, both ventilation were affected by the outside climate. The average temperature was about 2 ℃ lower in the greenhouse, compared with sunny weather. While the relative humidity was about 5% higher. The active ventilation also shared significant thermal insulation in rainy weather, where the indoor temperature increased by about 2-5 ℃. The dehumidification reduced the relative humidity by about 9 percentage points, due to the high humidity of the outdoor climate. Therefore, active ventilation was achieved in the dehumidification and thermal insulation under different weather conditions. Finally, the average annual investment and expected returns from active ventilation were calculated to significantly improve the crop yield and quality. The return on investment (ROI) of the active ventilation strategy was 2.62 for the economic benefits with relatively low investment, indicating the economic feasibility. The active ventilation with the spatial distribution of temperature and humidity was efficiently achieved in the dehumidification and insulation, leading to higher returns under indoor temperatures. The findings can provide theoretical and practical implications for dehumidification and insulation in winter greenhouses.