[Significance] Against the backdrop of consumption upgrading and growing health awareness, consumer concerns about agricultural products are shifting from mere safety to freshness, taste, and brand reputation, making quality a core target of modern agricultural upgrading and accelerating the transition of production modes toward precise and intelligent control. However, the traditional centralized cloud‑computing architecture dominated by a central cloud platform is increasingly constrained in weak‑network environments, low‑latency control, and local autonomy, and thus can hardly support the closed‑loop requirements of "timely detection-rapid response-continuous optimization" in quality control. Consequently, cloud-edge-device integrated collaborative architectures are regarded as an important development direction, and it is necessary to systematically sort out their key technologies and application foundations in the agricultural domain so as to provide unified theoretical and technical support for the design of integrated middleware oriented to agricultural product quality control. [Progress] Starting from the current application status of cloud-edge-device integration in agriculture, the paper focuses on three typical scenarios—field planting, facility horticulture, and livestock farming—and summarizes the functional division of perception, computation, and control among cloud, edge, and device, as well as the differences in business requirements under various geographical conditions and production modes. On this basis, it identifies multiple technical challenges in practical deployments, including heterogeneous adaptation difficulties caused by the coexistence of numerous industrial and IoT protocols, poor interoperability due to inconsistent data definitions and encodings, resource imbalance arising from the mismatch between edge computing capacity and task loads, and coarse‑grained, inefficient cloud-edge collaborative scheduling. To address these challenges, the paper synthesizes relevant research on artificial intelligence, edge computing, federated learning, and blockchain, and proposes improvement ideas for cloud-edge-device collaboration from the perspectives of resource‑allocation optimization, edge‑intelligent inference, privacy protection, and trusted sharing. Furthermore, targeting the concrete requirements of intelligent quality control for agricultural products, it constructs a cloud-edge-device integrated technical framework: At the data layer, it integrates standardized mapping of quality data and a multi‑source data dictionary system; at the access layer, it realizes heterogeneous device‑protocol conversion and unified access; at the computing layer, it achieves dynamic matching between computing resources and services through intelligent task scheduling and load balancing, multi‑level caching, model hot‑updating, and federated deployment. At the collaboration level, it proposes a "device-crop-model" trinity abstraction that brings physical devices, agronomic objects, and algorithmic models under unified middleware management, thereby masking underlying hardware differences while supporting elastic adaptation and capability reuse across diverse scenarios. [Conclusions and Prospects] The cloud-edge-device integrated collaboration and its adaptive middleware constitute the key infrastructure and connective hub of intelligent quality‑control systems for agricultural products, providing reusable architectural concepts and technical references for the development and engineering implementation of systems for precise quality perception, automatic control, and intelligent decision‑making. Looking ahead, further in‑depth research is needed on cross‑scenario data and protocol standard systems oriented to quality indicators and control actions, on the integrated application of edge intelligence and large models, on the enhancement of privacy‑protection and trusted‑computing mechanisms, and on large‑scale engineering demonstrations and validation, so as to continuously improve the intelligence level, operational reliability, and scalability of agricultural product quality‑control systems and to promote the transition of related technologies from pilot applications to large‑scale deployment.

[Significance] Livestock body-size parameters are important phenotypic indicators for evaluating growth and development, body condition, production performance, genetic evaluation, and breeding selection. Traditional manual measurement is time-consuming, labor-intensive, operator-dependent, and may induce stress and safety risks during animal restraint. Although two-dimensional imaging enables non-contact measurement, its lack of depth information limits the accurate estimation of girth-related traits and local body morphology. By directly describing three-dimensional body-surface structures, point clouds provide a promising basis for automated and objective livestock phenotyping. The aim of this review is to systematically summarize advances in three-dimensional point cloud-based livestock body measurement, to clarify the main technical workflow and methodological routes, to compare their characteristics and application scenarios, and to identify key challenges and research priorities for intelligent three-dimensional phenotyping and smart livestock farming. [Progress] Three-dimensional point cloud technology can reconstruct livestock body structures while retaining the advantages of non-contact measurement. It can be used for measuring linear traits, such as body length, body height, and body width; girth-related traits, including chest girth, abdominal girth, rump girth, and cannon circumference; and local morphological features such as rump shape, hindquarter fullness, and limb and hoof structure. With pigs, cattle, and sheep as the main research objects, this review examines the technical workflow of point cloud acquisition, preprocessing, completion, segmentation, keypoint localization, and body-size calculation. Livestock point cloud acquisition sensors and fixed and mobile acquisition schemes are summarized. Preprocessing methods, including multi-view registration, filtering and denoising, posture normalization, and point cloud completion, are reviewed. These procedures can reduce background interference, motion distortion, occlusion, data loss, and posture variation, thereby improving measurement consistency. Two technical routes for body-size measurement are discussed. In the first route, anatomical keypoints are directly localized from RGB images, depth images, or point clouds using geometric features, projection analysis, curvature, skeleton models, or deep learning-based detectors. This route is suitable for linear measurements under standardized postures. In the second route, the animal is first segmented into anatomically relevant regions, after which keypoints or measurement sections are identified within the corresponding regions. By narrowing the search space, segmentation-based methods can reduce interference from unrelated body parts and are particularly suitable for variable postures, partial occlusion, and girth-related measurements. After keypoint localization or section determination, linear traits are generally calculated using Euclidean distances, whereas girth traits are estimated through contour extraction, curve fitting, and accumulated curve length. Overall, a development trend from manual annotation and rule-based geometric analysis toward automated methods integrating keypoint detection, point cloud segmentation, regional constraints, posture adaptation, and cross-species transfer is identified. [Conclusions and Prospects] Existing studies indicate that three-dimensional point cloud-based livestock body measurement has promising application prospects for automated, non-contact, and high-precision phenotype acquisition. However, several challenges remain, including unstable point cloud quality in complex farming environments; insufficient robustness, transferability, and edge-deployment capability of algorithms; limited multi-species and multi-posture datasets; and inconsistent definitions, annotations, and evaluation metrics. Future research should focus on improving acquisition accuracy, synchronization, and environmental adaptability and developing low-cost, lightweight, and reliable systems. Multimodal fusion of depth cameras, LiDAR, and RGB images, together with mobile platforms and multi-view acquisition, should be investigated for continuous measurement of group-housed animals. Robust, lightweight, and transferable models should be developed through self-supervised or unsupervised pretraining, semi-supervised learning, transfer learning, model pruning, knowledge distillation, and lightweight feature extraction. Open datasets should cover multiple species, breeds, growth stages, postures, and farming scenarios, accompanied by unified keypoint definitions, anatomical annotations, data partitions, and error metrics. In addition, body-size measurements should be integrated with body condition scoring, local morphological evaluation, growth and development monitoring, production performance analysis, genetic evaluation, and breeding selection. This will promote the transition from geometric parameter extraction to intelligent three-dimensional phenotypic analysis, ultimately achieving accurate, real-time, and high-throughput automatic livestock body measurement for precision feeding and smart livestock farming.

[Objective] The gas exchange method is the most widely applied and most direct quantitative measurement technique in plant photosynthesis research. Plant photosynthetic measurement systems based on the gas exchange method usually rely on the "steady-state paradigm" to determine the photosynthetic rate and other physiological parameters. During the measurement process, the system needs to wait until the CO₂ concentration reaches a steady state. However, this process is affected by factors such as plant physiology, leaf chamber volume, and sensor response characteristics, which severely restricts measurement efficiency. To this end, a new method is proposed that predicts the steady-state CO₂ concentration in advance based on the dynamic response sequence of the sensor and adaptively determines the measurement termination time using predictive uncertainty. [Methods] A dual-branch encoding network, DBE-TSNet (Dual-Branch Encoder for Time-Series Network), was constructed. The overall architecture of the network consisted of dual feature encoders and an aggregation encoder. The dual feature encoders adopted a structurally symmetric and parameter-independent branch design. They combined differential enhancement, cross-channel convolution-linear channel transformation, and global average pooling to model the dynamic response processes of the CO₂ concentration decrease and increase stages perceived by the system, respectively. After concatenating the features of the two branches, the aggregation encoder generated a four-dimensional output vector through a two-layer multilayer perceptron decoder: predicted decrease value (t̂₁), predicted decrease uncertainty (σ̂₁²), predicted increase value (t̂₂), and predicted increase uncertainty (σ̂₂²). During training, a Gaussian negative log-likelihood loss function was adopted to simultaneously learn the steady-state prediction together with its associated predictive uncertainty, while sample weights were adaptively adjusted accordingly. By setting uncertainty thresholds, adaptive truncation of the input sequence was realized, enabling determination of the prediction timing and the steady-state value, thereby improving measurement efficiency. A photosynthetic measurement system was built based on a single CO₂ sensor, using a time-shared differential measurement mode. Gas exchange data was collected from eight plant species (corn, potato, radish, bok choy, lettuce, loquat, orange tree, and paper mulberry) under three light intensity levels (low, medium, and high) for model training and validation. [Results and Discussions] By comparison with six typical time-series prediction models, including one-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN), the DBE-TSNet model achieved the best performance in the steady-state prediction task: The MAE of the decreasing stage was 1.22 μmol/mol, and that of the increasing stage was 2.07 μmol/mol; the R² values reached 0.996 and 0.984, respectively. To evaluate the relationship between model predictive reliability and output uncertainty, all test samples were divided into ten intervals according to the uncertainty output by the model, and the relationships between each interval and mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE), and the input segment ratio were counted. The results showed that as uncertainty gradually increased, the prediction error gradually increased, resulting in a significant decline in model predictive reliability. Based on this analysis, the uncertainty thresholds of the decreasing stage and increasing stage were determined to be 0.026 7 and 0.014 6, respectively. Early prediction experiments based on the set thresholds showed that under the premise of maintaining the error below 1 μmol/mol, the model saved about 128.87 s of measurement time on average, shortened the measurement cycle by about 51%, and significantly improved measurement efficiency. [Conclusions] DBE-TSNet integrated dual-branch encoding and uncertainty estimation mechanisms to achieve accurate and early prediction of steady-state CO₂ concentration in plant photosynthetic measurement, effectively solving the problem of long steady-state measurement time in the gas exchange method and transforming the traditional "waiting for steady state" process into a data-driven decision-making process of "dynamic steady-state prediction". The model showed stable performance under multiple plant species and multiple light conditions, with good generalization ability, and could provide key technical support for improving the efficiency of plant photosynthetic measurement systems based on the gas exchange method.

[Objective] Hyperspectral remote sensing enables accurate estimation of crop chlorophyll content, a key indicator of photosynthetic capacity and nitrogen nutrient status. However, models trained in one region (source domain) often suffer significant performance degradation when applied to another region (target domain) due to "domain shift" caused by differences in soil, climate, and management practices. This study proposed a robust adaptive transfer learning framework (RATL) to mitigate domain shift in cross-regional chlorophyll estimation of winter wheat and achieve accurate inversion. [Methods] A total of 1 491 paired samples of canopy hyperspectral reflectance and leaf soil and plant analyzer development (SPAD) values were collected during the late grain filling stage of winter wheat in the 2023－2024 growing season from Xinxiang and Zhoukou in Henan province. The degree of domain shift between the two regions was quantified using multiple metrics, including maximum mean discrepancy (MMD), Wasserstein distance, and correlation differences. The RATL framework consists of three synergistic modules: (1) Adaptive feature selection, which combines source-target correlation weighting and stability constraints to select 100 core features from the original 1 734-dimensional feature set, with preferential selection of red-edge bands (680－750 nm); (2) daptive feature weight calculation, which fuses random forest importance scores from both domains and feature stability metrics to assign weights to each feature, guiding the model to focus on domain-invariant features; and (3) domain adaptation training, which employs a two-stage XGBoost regressor (pre-training on source domain followed by fine-tuning with 30% of target domain samples) and incorporates a domain discriminator loss (balancing parameter λ=0.1) to encourage learning of domain-invariant representations. Four experimental scenarios were designed: source-only validation (Scenario A), direct transfer (Scenario B), transfer learning comparison with Transfer Component Analysis and Correlation Alignment(Scenario C), and the ideal upper bound using combined data from both regions (Scenario D). Shapley additive explanations (SHAP) analysis was employed to interpret model decision mechanisms, and quantile regression forests were used to generate 90% prediction intervals for uncertainty assessment. [Results and Discussions] Quantitative domain shift analysis revealed significant distribution differences between Xinxiang and Zhoukou maximum mean discrepancy was 0.20, P<0.001, confirming the challenge of domain shift in cross-regional modeling. Direct transfer exhibited substantial performance degradation on the target domain (R²=0.61). Traditional transfer learning methods TCA and CORAL achieved only marginal improvements (R²≈0.64－0.65, transfer gains<7%), proving insufficient for addressing complex domain shift. In contrast, RATL achieved optimal performance on the target domain (R²=0.75, root mean square error was 7.49), representing a 23.40% improvement over direct transfer and reaching 105.60% of the ideal performance achieved with combined data from both regions. SHAP analysis demonstrated that RATL successfully shifted the model's reliance from environmentally sensitive features to red-edge vegetation indices (modified Simple Ratio at 705 nm, normalized difference vegetation index,Green Index) that exhibit consistent responses across regions, enhancing both physiological rationality and regional adaptability of model decisions. Uncertainty quantification showed that RATL achieved higher prediction interval coverage (79.52%) compared to direct transfer (77.94%), while adaptively widening intervals for extreme SPAD values to provide more realistic and reliable uncertainty estimates. These results demonstrate that RATL's multi-module collaborative strategy effectively mitigates agricultural hyperspectral domain shift, significantly improving model interpretability and reliability while maintaining prediction accuracy, outperforming traditional linear alignment methods. [Conclusions] The proposed RATL framework achieves high-precision and high-reliability cross-regional chlorophyll inversion with only a small set of target samples through the synergistic effects of adaptive feature selection, feature weight adjustment, and domain adaptation training. The framework effectively focuses on physiologically meaningful domain-invariant features and provides reliable uncertainty estimates, offering a practical technical solution for regional-scale agricultural remote sensing monitoring. Future work will extend RATL to address temporal domain shift (e.g., across different growth stages) and explore lightweight model implementations for real-time applications.

[Objective] In recent years, global warming and increasingly frequent extreme precipitation events have intensified waterlogging stress in agriculture, establishing it as a major abiotic threat to food security.To achieve scientific and precise control of wheat waterlogging disaster and increase yield, a rapid, accurate, and non-destructive method for identifying wheat waterlogging stress and regulation effect was proposed. [Methods] Field trials of control (CK), waterlogging stress, silicon-fertilizer regulation, and amino-acid regulation were conducted during the jointing-booting and flowering-filling stages of wheat. During the experiments, unmanned aerial vehicle (UAV) multispectral images of the wheat canopy were collected using DJI P4 Multispectral. The multispectral image had five bands, namely blue (B), green (G), red (R), red-edge (RE) and near-infrared (NIR), with reflectivity values of (450±16), (560±16), (650±16), (730±16), and (840±26) nm, respectively. At the end of the experiments, grain yield was measured for each treatment and used to evaluate the effectiveness of the regulation measures. The multispectral images were processed, which included radiometric correction, geometric correction, image alignment, segmentation, etc. Then a dataset was established that includes four categories: CK, waterlogging stress, silicon fertilizer regulation, and amino acid regulation. The stratified sampling was used for dataset partitioning. According to the ratio of 7:1.5:1.5, all original multispectral images of each category were randomly divided into the training, validation and testing set, respectively. Then, the original images were segmented into data blocks of 48×48 pixel size for model training and testing. A wheat waterlogging stress and regulation effect recognition model using visual geometry group (VGG) 21 was established by adding convolutional layers, adjusting the convolutional layers of convolutional block, replacing activation functions, and optimizing model hyperparameters based on VGG19. The VGG21 model included 4 convolutional blocks and the convolutional layers were 3, 3, 6, and 6, respectively. To verify the performance of the VGG21 model, the accuracy, precision, recall and F₁-Score of the ResNet50, Swin-Transformer, VGG19, VGG21, and VGG23 models were compared under the same conditions, and the confusion matrix and Grad-CAM (Gradient-Weighted Class Activation Mapping) were employed to evaluate the CK, waterlogging stress, and silicon fertilizer regulation model. [Results and Discussions] The performance comparison results showed that the VGG model performed better on multispectral image data recognition than the ResNet50 and Swin-Transformer model, and the VGG21 model demonstrated the most favorable overall performance. The comprehensive analysis of model recognition performance showed that the accuracy of the VGG21 model was more than 91% for silicon fertilizer regulation and waterlogging stress. And the precision, recall, and F₁-Score for silicon fertilizer regulation samples reached 96.42%, 89.53%, and 92.85%, respectively. When the silicon fertilizer regulation and stress recognition model was used for amino acid regulation sample identification, the accuracy, precision, recall, F₁-Score were comparable to the amino acid regulation and stress recognition model. A recognition model for CK, waterlogging stress, and silicon fertilizer regulation was constructed using VGG21, and the precision, recall, and F₁-Score for silicon fertilizer regulated samples achieved 95.13%, 88.71%, and 91.80%, respectively. The confusion matrix of the CK, waterlogging stress, and silicon fertilizer regulation model indicated that there was a high proportion of misclassifications between CK and waterlogging stress samples, resulting in a low accuracy of model, which only reached 77.45%. The heatmap analysis results of the CK, waterlogging stress, and silicon fertilizer regulation model indicated that after passing through the fourth convolutional block, the concentrated regions of the output features for misclassified CK and waterlogging stress samples became more dispersed, and even disappear. The CK and waterlogging stress samples with unfocused region heatmap accounted for a large proportion in the dataset, which led to poor recognition of these two types of samples by the regulation model. The comparison between recognition results of VGG21 model and the yield of regulation experiment showed that the silicon fertilizer had a better regulation effect, and the wheat regulated by silicon fertilizer had significant spectral and image features with high recognition accuracy. [Conclusions] The results demonstrated that UAV multispectral imaging technology and VGG21 model were feasible and useful for recognizing waterlogging stress and the regulatory effect of wheat. This technique could provide theory and technical support for integrated waterlogging management of wheat and for refined, intelligent wheat production management. However, the quantitative relationship between the identification results of regulation effect and yield was still unclear. In future research, it will be necessary to quantify the increased yield of regulation measures, and then use statistical methods to quantify the quantitative relationship between the yield-increasing regulation effect and identification indicators of recognition model.

[Objective] Winter wheat yield formation is jointly affected by multiple factors, including canopy structure, growth status, photosynthetic function, and meteorological conditions. The contributions of remote sensing, meteorological data, and photosynthetic information to yield prediction differ among phenological stages. However, the influence of information accumulation at different phenological stages on county-level winter wheat yield estimation remains unclear, which restricts further improvement in early prediction capability and agricultural interpretability of yield prediction models. Therefore, a phenological-stage progressive multi-source data fusion method was proposed for county-level winter wheat yield prediction, to improve the prediction accuracy of county-level winter wheat yield, clarify the influence of information accumulation at different phenological stages on yield estimation performance, and provide methodological support for dynamic regional yield prediction and monitoring of key growth stages. [Methods] 100 major winter wheat-producing counties of Henan province were selected as the study area. Remote sensing variables, meteorological variables, and solar-induced chlorophyll fluorescence (SIF)-based photosynthetic variables from 2013 to 2022 were integrated to construct a county-level multi-source time-series feature dataset covering the period from the tillering stage to the maturity stage of winter wheat. Among these variables, remote sensing variables were used to characterize canopy structure and crop growth status, meteorological variables were used to reflect heat and water conditions during the growth period, and SIF variables were used to characterize changes in crop photosynthetic function. Based on the winter wheat planting distribution mask and county-level administrative boundaries, county-scale time-series features of each variable were extracted. According to the developmental sequence of tillering, overwintering, greening, jointing, heading, anthesis, and maturity stages, the model input window was progressively expanded to form input sequences under different phenological-stage accumulation conditions. For model construction, a convolutional neural network-long short-term memory-transformer (CNN-LSTM-Transformer) model was adopted. The CNN was used to extract local variation features between adjacent time steps, the LSTM was used to capture continuous temporal dependencies among stages, and the Transformer Encoder was used to model global associations among time steps in the input sequence. To evaluate model performance, the CNN-LSTM-Transformer model was compared with random forest (RF), Extreme Gradient Boosting (XGBoost), and CNN-LSTM models. In addition, a leave-one-year-out validation strategy was used to analyze the interannual generalization ability of the CNN-LSTM-Transformer model under different year conditions. [Results and Discussions] In the test set based on county-level sample partitioning, the CNN-LSTM-Transformer model outperformed the RF, XGBoost, and CNN-LSTM models and achieved the best yield estimation performance. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute percentage error (MAPE) were 0.827, 588.25 kg/hm², and 8.02%, respectively. Compared with the CNN-LSTM model, the R² of the CNN-LSTM-Transformer model increased by 0.054, while RMSE and MAPE decreased by 85.69 kg/hm² and 1.54 percentage points, respectively. These results indicated that introducing the Transformer Encoder helped enhance the model's ability to represent global multi-source time-series information. The phenological-stage accumulation analysis showed that model prediction accuracy continuously improved as growth-stage information gradually increased. Prediction accuracy was relatively low at the tillering and overwintering stages, whereas model performance improved most obviously from the greening stage to the heading stage. This indicated that multi-source information during this period could more fully reflect rapid canopy development, enhanced photosynthetic capacity, and the yield formation process of winter wheat. Therefore, the period from greening to heading was identified as the key stage for the rapid improvement of county-level winter wheat yield prediction ability. The residual analysis across yield zones showed that the CNN-LSTM-Transformer model alleviated, to some extent, the overestimation in low-yield areas and the underestimation in high-yield areas. The leave-one-year-out validation results showed that the proposed method maintained good prediction stability across different test years. The spatial distribution results further showed that the predicted yield could effectively reflect the county-level spatial pattern of winter wheat yield in Henan province, with lower values in the western region and higher values in the central and eastern regions. [Conclusions] Overall, the proposed phenological-stage progressive multi-source data fusion method effectively integrated remote sensing, meteorological, and SIF-based photosynthetic information, improved the prediction accuracy of county-level winter wheat yield, and revealed the influence of information accumulation at different phenological stages on yield estimation performance. The CNN-LSTM-Transformer model collaboratively represented local variation features, stage-wise temporal dependencies, and global long-term associations, thereby effectively characterizing multi-source time-series dynamic changes during winter wheat yield formation. The results showed that the period from the greening stage to the heading stage was the key period for the rapid improvement of county-level winter wheat yield prediction ability. This study could provide a reference for early winter wheat yield estimation, key growth-stage monitoring, and regional grain production management.

[Objective] Carnation (Dianthus caryophyllus L.) is one of the most economically valuable cut flower crops worldwide. Postharvest openness is a key quality indicator influencing pricing, logistics tolerance, and shelf life. However, manual grading is inefficient and subjective due to dense petal overlap and complex edge structures. With the shift toward large-scale production and rising labor costs, accurate automated grading has become essential. Existing object detection models face a trade-off between computational efficiency and feature fidelity: High-precision architectures are computationally expensive for edge deployment, while lightweight models often lack sufficient feature representation. Additionally, conventional spatial downsampling introduces spectral aliasing, leading to the loss of high-frequency petal texture information and limiting the separability of adjacent openness grades. Therefore, a lightweight yet detail-preserving detection framework is required. To address this need, Flor-YOLO (Flower openness recognition You Only Look Once) was proposed by integrating frequency-domain perception with structural re-parameterization for efficient and accurate carnation openness grading. [Methods] Based on the YOLO11n baseline, the Flor-YOLO architecture was proposed with targeted improvements to the backbone, downsampling mechanism, and detection head. Backbone reconstruction: A lightweight LiteChimeraNet was constructed to enhance feature expression under limited computing power. A RepStem re-parameterization module was introduced at the input stage to establish an anti-aliasing mechanism via multi-branch training and single-path inference. Simultaneously, the C3k2_PConv module, utilizing partial convolution, was integrated to reduce memory access cost (MAC) and focus computation on petal foregrounds. Additionally, a RepNCSPELAN4_CAA module embedded with context anchor attention was incorporated in deep layers to capture long-range dependencies of the global flower topology. Frequency-domain downsampling: To mitigate texture aliasing and detail loss caused by spatial downsampling, a WaveletPool module was introduced. Utilizing the 2D discrete wavelet transform (2D-DWT), this module orthogonally decomposed feature maps into low- and high-frequency sub-bands, explicitly preserving high-frequency information in horizontal, vertical, and diagonal directions to alleviate spectral aliasing. Detection head optimization: A lightweight shared detail-enhanced detection head (SDL-Head) was designed. It reduced parameter redundancy through cross-scale weight sharing and incorporated detail-enhanced convolution (DEConv), fusing central and angular difference operators, to boost sensitivity to the geometric morphology of petal edges. Furthermore, a scale-adaptive layer combined with Intersection over Union (IoU)-aware soft labels was applied to improve multi-scale feature alignment. A dataset comprising 1 748 original images of "Hongkang" carnations was collected and expanded to 6 580 samples via hybrid data augmentation. The model was trained on an NVIDIA RTX 4060 GPU for 250 epochs using SGD optimization, and comparative evaluations were conducted against the YOLO series, NanoDet-m, and Hyper-YOLO-t. [Results and Discussions] Ablation studies and comparative experiments on the self-constructed dataset revealed significant performance gains. Ablation analysis: Reconstructing the backbone to LiteChimeraNet reduced floating point operations from 6.3 GFLOPs (baseline) to 1.5 GFLOPs, a decrease of 76.2%, while maintaining stable mean Average Precision (mAP@50), verifying its efficiency in removing background redundancy. Introducing WaveletPool significantly improved mAP@50 by 1.79 percentage points, confirming the critical role of explicitly preserving high-frequency components for serrated texture representation. Integrating SDL-Head further optimized feature alignment, increasing the recall rate to 94.47%. Overall performance: Flor-YOLO achieved a precision of 93.04%, recall of 94.47%, and mAP@50 of 96.10%. Compared to the YOLO11n baseline, these metrics improved by 3.52, 1.34, and 3.25 percentage points, respectively. Meanwhile, parameters and floating point operations were reduced by 51.2% to 1.26 M and 1.1 GFLOPs (82.5% reduction). Flor-YOLO exhibited distinct advantages over YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, and YOLOv12n in accuracy, mAP, and inference speed. Mechanism analysis: Spectral energy statistics showed that high-frequency energy intensified with increasing openness grades, aligning with the visual characteristics of petal expansion and wrinkle formation, thus validating the discriminative value of high-frequency information. Grad-CAM++ visualizations further validated that the improved model stably focused on petal edges and flower centers, demonstrating superior robustness over the baseline in complex backgrounds. [Conclusions] By constructing the LiteChimeraNet backbone, incorporating frequency-domain downsampling, and designing a detail-enhanced head, the proposed model effectively enhances the representation of critical details such as petal edges and flower centers while maintaining extremely low computational costs. Comprehensively, Flor-YOLO achieves an optimal balance between accuracy, model size, and real-time performance, demonstrating strong potential for deployment on low-power mobile terminals and embedded sorting equipment. Furthermore, the proposed frequency-aware lightweight design paradigm provides a valuable reference for other agricultural vision tasks relying on subtle textural differences.

[Objective] Cotton maturity assessment is a vital task in precision agriculture, playing a crucial role in supporting timely irrigation, fertilization, and harvesting decisions. However, traditional monitoring approaches are time-consuming and labor-intensive, and current deep learning-based models often struggle to effectively recognize cotton bolls at varying maturity stages, especially in complex field environments with dense foliage, occlusion, and illumination changes. To address these challenges, a high-accuracy and lightweight computer vision model was proposed for cotton maturity detection. [Methods] An enhanced detection framework named cotton maturity-detection transformer (CM-DETR) was proposed, based on an improved real-time detection transformer (RT-DETR) architecture. CM-DETR incorporated three core architectural innovations that significantly improved both detection accuracy and computational efficiency. First, to construct a lightweight and efficient backbone, a novel feature extraction module named Re-parameterized Group Convolution Spatial Enhancement Lightweight Attention Network (RGCSPELAN) was introduced. This module integrated Progressive Convolution, which captured hierarchical and local contextual features, with Re-parameterized Convolution (RepConv), which reduced computational complexity during inference by transforming multi-branch structures into a single-path representation. The combination effectively enhanced the model's feature representation capabilities and gradient propagation while minimizing the number of parameters and FLOPs. Furthermore, RGCSPELAN was designed with a scalable architecture, allowing its computational capacity to be adjusted via a scaling factor. This ensured compatibility with both small and large models, facilitating flexible deployment across resource-constrained edge devices and high-performance systems alike. Second, to address the issue of small target feature loss, a new module termed Deep Robust Feature Downsampling (DRFD) was proposed. DRFD employed a multi-scale feature fusion strategy by integrating multiple downsampling branches (e.g., convolutional, cut-based, and max-pooling pathways). This design enabled the model to retain fine-grained spatial details while expanding its receptive field. Third, the original loss function in RT-DETR was replaced with Focaler-CIoU, and an adaptive regression optimization strategy integrating sample reweighting and geometric constraints was implemented to improve bounding box localization under complex conditions. [Results and Discussions] Experimental results demonstrated that CM-DETR achieved mAP50 and mAP50~95 scores of 80.8% and 51.1%, respectively, outperforming the baseline model by 3.7 and 1.8 percentage points. Meanwhile, CM-DETR reduced the parameter count and computational cost by 31.7% and 22.8%, respectively, indicating a favorable trade-off between detection accuracy and model efficiency. The incorporation of the DRFD module enhanced the model's sensitivity to small and subtly distinct features related to cotton maturity, improved robustness under diverse field conditions, and enabled more precise detection of cotton bolls at different growth stages. Moreover, the optimized regression strategy contributed to more stable bounding box prediction performance in scenarios involving occlusion, scale variation, and dense foliage. Overall, the proposed architectural improvements effectively strengthened feature representation capability while maintaining lightweight characteristics, thereby demonstrating practical applicability in real-time agricultural environments. [Conclusions] The proposed CM-DETR model provides an efficient and scalable solution for automated cotton maturity detection. By enhancing multi-stage feature recognition, improving small-target sensitivity, and reducing the demand on computational resources, CM-DETR serves as a reliable tool for intelligent decision-making in precision agriculture. Its practical deployment can support more accurate timing for irrigation, fertilization, and harvesting, thereby contributing to improved crop management and yield optimization.

[Objective] Moisture content and kernel hardness are key indicators for evaluating the eating quality and harvest maturity of fresh corn. Accurate and simultaneous prediction of the two parameters is essential for quality grading and post-harvest management. However, conventional near-infrared spectroscopy (NIRS) methods are mostly based on single-point measurements, which are easily affected by husk shielding and kernel heterogeneity, resulting in limited prediction accuracy, especially for multi-attribute estimation. Spatially resolved spectroscopy, by acquiring spectral information from multiple measurement channels, can better reflect internal quality differences of corn kernels. Nevertheless, the high dimensionality and strong correlation of spatially resolved spectral data, together with the heterogeneity between different prediction tasks, pose significant challenges to traditional modeling approaches. Therefore, the purpose of this research was to develop an effective multi-task prediction model that can fully exploit spatially resolved spectral information while balancing feature sharing and task specificity for moisture content and kernel hardness prediction in husked fresh corn. [Methods] Husk-on fresh corn samples were collected across different maturity stages to ensure sufficient variability in moisture content and kernel hardness. A multi-channel visible-near infrared spatially resolved spectroscopy system was constructed to acquire spectral signals from four spatial measurement channels at different source-detector distances. Each corn ear was segmented into multiple positions, and spectral data were collected from different spatial locations to comprehensively capture internal quality information. In total, 500 valid spectral samples were obtained and randomly divided into training, validation, and test sets at a ratio of 7:1:2. Before modeling, the raw spectral data were preprocessed using z-score standardization to eliminate scale differences among channels and wavelengths and to improve numerical stability during training. The proposed MaMoNet (Mamba-MMoE Network) model is composed of three main modules: a one-dimensional convolutional neural network (1D-CNN) for local feature extraction, a Mamba-based sequence modeling module for global dependency learning, and a Multi-gate Mixture-of-Experts (MMoE) module for task-adaptive feature allocation. First, the 1D-CNN module was employed to extract low-level local spectral patterns and reduce feature redundancy. The input spectral data were processed by two successive one-dimensional convolutional layers and down sampling was then applied to compress the spectral length, resulting in a more compact representation and reducing the computational burden for subsequent sequence modeling. Next, the compressed spectral features were fed into the Mamba module to capture long-range dependencies along the spectral dimension. Mamba is a selective state space model that enables efficient modeling of long spectral sequences by dynamically updating hidden states, allowing global spectral evolution patterns to be effectively learned with linear computational complexity. Finally, the output features of the Mamba module were input into an MMoE module to support multi-task learning of moisture content and kernel hardness. Multiple expert networks were shared across tasks, and task-specific gating networks were used to adaptively weight expert outputs, enabling flexible feature sharing and task-specific representation learning. The task-specific features were then passed to individual regression heads to generate the final predictions for moisture content and kernel hardness. To comprehensively evaluate the effectiveness of the proposed approach, MaMoNet was compared with several representative multi-task convolutional neural network models. In addition, ablation experiments were conducted by selectively removing the Mamba module or the MMoE module to analyze their individual contributions to overall performance. [Results and Discussions] Experimental results demonstrated that the proposed MaMoNet model consistently outperformed all comparison models on the test set for both prediction tasks. For moisture content prediction, MaMoNet achieved a coefficient of determination (R²) of 0.91, a root mean square error (RMSE) of 4.98%, and a residual predictive deviation (RPD) of 3.40, indicating excellent predictive accuracy and robustness. For kernel hardness prediction, the corresponding R², RMSE, and RPD values reached 0.89, 3.46 N, and 3.06, respectively, which also surpassed those of the benchmark models. The ablation study further verified the rationality of the proposed model design. Removing both the Mamba and MMoE modules resulted in the poorest performance, whereas introducing either module individually led to noticeable improvements. The best prediction results were achieved when both modules were jointly employed, indicating that their combination was critical for achieving optimal multi-task prediction performance. These results indicated that moisture content and kernel hardness, although correlated, emphasize different spectral characteristics. The flexible feature-sharing mechanism enabled by MMoE allows the model to balance information sharing and task specificity, while the Mamba module ensures effective utilization of long-range spectral information. Together, they contribute to improved generalization performance under limited sample conditions. [Conclusions] This study proposes a MaMoNet model that integrates Mamba-based state space modeling with an MMoE-based multi-task learning strategy for simultaneous prediction of moisture content and kernel hardness in husked fresh corn using spatially resolved spectroscopy. The proposed approach effectively overcomes the limitations of conventional single-point spectral analysis and rigid parameter-sharing multi-task models. Experimental comparisons and ablation analyses confirm that MaMoNet achieves improved accuracy, robustness, and generalization capability. The results demonstrate the potential of the proposed framework for rapid and non-destructive quality assessment of fresh corn and provide useful methodological insights for multi-attribute prediction based on high-dimensional spectral data.

[Objective] In the livestock industry, animal behavior serves as a critical indicator for evaluating physiological health and providing environmental early warning. To address the challenges of severe multi-object occlusion, complex illumination, and class imbalance in intensive farming, a detail-contextual attention-YOLO (DC-YOLO) model was developed for the 24/7 automatic detection of daily behaviors in housed Hu sheep, thereby facilitating precision management. [Methods] For severe multi-object occlusion, an adaptive detail-contextual attention (DCAttention) mechanism was introduced to construct the DCAC3K2 module, and bounding box similarity soft-NMS (BS-NMS) was adopted to prevent incorrect removal of occluded bounding boxes. For complex illumination, a Light-Encoder was pre-trained on illumination conditions and its parameters were transferred to DC-YOLO. For class imbalance, a comprehensive classification-quality focal loss was designed to adaptively increase loss weights for occluded and minority-class samples. The dataset, collected through 24/7 surveillance at Hu sheep farms, comprised 505 annotated images depicting 5 daily behaviors: drinking, eating, lying, licking, and standing. [Results and Discussions] DC-YOLO achieved mAP@50 of 91.4%, improving by 7.8 percentage points over YOLOv12. Moreover, DC-YOLO had 2.29 M parameters, an 8.74% reduction compared to YOLOv12. On CPU, the inference time of DC-YOLO was reduced to 115.5 ms and the frame rate increased to 8.50 f/s, corresponding to improvements of 33.2% and 48.9%, respectively. [Conclusions] Experimental results demonstrate that DC-YOLO effectively mitigates detection challenges caused by severe occlusion, complex illumination, and class imbalance while maintaining high inference efficiency. Consequently, it provides effective technical support for tracking and analyzing behaviors in intensive Hu sheep farming.