Maize is susceptible to rapid quality deterioration and fungal infection due to complex environmental fluctuations during "North-to-South Grain Transfer" strategies. However, conventional detection is often time-consuming, destructive, and labor-intensive under different storage and transportation environments. It is an urgent need to non-destructively and rapidly identify the mold ratio, and then continuously predict the quality index. In this study, a synchronous prediction was proposed for maize mold ratio and storage/transportation quality using advanced image processing and deep learning technologies. Maize samples with a controlled gradient mold ratio ranging from 0 to 12% were selected as the research objects. A systematic simulation was conducted on typical temperature and humidity environments of both waterway and overland transportation routes. Key quality indices of the maize were measured to quantify the deterioration rates in the simulated storage and transportation periods, including moisture content, fatty acid value, and electrical conductivity. Simultaneously, maize images were collected using a standard smartphone. Digital image processing was also integrated with the Vision Transformer (ViT) deep learning model and statistical modeling. The mold ratio was then detected to precisely predict the quality indices. The results indicated that the high-humidity environment of waterway transportation accelerated the deterioration of maize kernel quality (P<0.05). Specifically, the moisture content of the waterway samples rapidly exceeded the threshold of 14% in the national safe storage standard when the mold ratio reached 2%. In contrast, the moisture content of the overland transportation samples remained stable in the safe range of 12.207% to 12.772%. Furthermore, the fatty acid value and electrical conductivity increased by 57.070% and 38.357%, respectively, under waterway conditions, as the maize mold ratio increased progressively. These deterioration rates were higher than those under overland conditions, indicating the lower increases of 29.035% and 27.714%, respectively. In terms of the deep learning algorithms, the ViT architecture achieved exceptionally high precision in identifying moldy maize kernels, reaching an impressive overall accuracy of 99.00%. Subsequently, a Mean Absolute Error (MAE) of only 0.52% was achieved, indicating the accurate and reliable prediction of the overall maize mold ratio. Visual features were extracted and further screened to construct Multiple Linear Regression (MLR) models for quality evaluation. In the waterway samples, the coefficients of determination (R²) of the prediction models reached 0.859, 0.955, and 0.942, respectively, for moisture content, fatty acid value, and electrical conductivity. In the overland samples, the R² values of prediction models were 0.930 and 0.937, respectively, for the fatty acid value and electrical conductivity, indicating accurate prediction for the quality of maize during storage and transportation. In conclusion, the dynamic quality deterioration of moldy maize can provide a low-cost, easy-to-operate, and entirely non-destructive pathway for maize quality detection. This finding can also offer an effective and practical analytical tool to dynamically monitor quality and safety for risk early warning during the complex grain circulation.

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.

Fruit-tree canopies are typically characterized by complex branching architecture and dense foliage, leading to severe self-occlusion, uneven light distribution, and low light-use efficiency. However, experience-driven decisions cannot fully meet the large-scale pruning of the standard canopy shapes and optimal regulation of tree vigor in the orchard. In this study, an intelligent pruning model was proposed to optimize the canopy structure reconstruction and light-efficiency evaluation using Neural Radiance Fields (NeRF). A reproducible, quantitative, and visualization-friendly workflow was also provided for canopy analysis and pruning under a real orchard. Qingcuili plum (Prunus salicina cv. ‘Qingcuili’) was selected as the target species. Multi-view videos were captured around each tree from multiple angles for sufficient coverage of the canopy under field lighting and background. A NeRF reconstruction was used to learn volumetric radiance and density fields from the video frames. A high-fidelity 3D representation of the tree was generated from the reconstructed structure. Branch topology and geometric descriptors (e.g., branch order, orientation, length, and spatial distribution) were extracted for decision-making. The Monte Carlo Ray Tracing (MCRT) module was integrated to simulate ray–canopy interactions. Both direct and diffuse radiation pathways were estimated to quantify light conditions in the 3D canopy space. Two indicators were computed: light interception ratio (LIR) to characterize the proportion of incident light intercepted by the canopy, and energy interception ratio (EIR) to reflect the effective energy under the simulated radiation field. A pruning recommendation model was constructed to fuse: (i) branch-classification and geometric features, and (ii) light-efficiency weights derived from the MCRT outputs. Candidate-branch suggestions were given for the improved canopy illumination with structural feasibility. In addition, a virtual interactive pruning system was developed to support human-in-the-loop validation, intuitive visualization of light distribution, and pruning effects before field implementation. The framework was validated on a dataset of 102 Qingcuili plum trees with diverse canopies and growth. The NeRF-based reconstruction was achieved with high accuracy, with an average reconstruction error of 4.3%, indicating reliable recovery of canopy structure under complex occlusion. The pruning recommendation performance reached 93.2% accuracy, compared with expert labelling. Pruning recommendations were consistent with practical knowledge. The optimal pruning strategy was applied to substantially improve the canopy light conditions. LIR and EIR increased by approximately 15.2% and 18.9%, respectively, indicating enhanced light interception and more effective energy capture. From a deployment perspective, the end-to-end system response time remained stable in 2.3 min using cloud GPU acceleration (NVIDIA V100), indicating favorable real-time applicability for decision support and interactive analysis. A canopy structure–light-efficiency optimization was integrated with NeRF 3D reconstruction, MCRT light simulation, and feature–weight fusion for intelligent pruning. The 3D canopy reconstruction and quantitative pruning improved the light-use efficiency and the precision of tree vigor regulation in a real orchard. The finding can provide a feasible technical pathway toward digital twins and smart pruning for fruit-tree production.

Microplastics can enter agro-food systems via multiple pathways, including agricultural production, environmental transport, post-harvest handling, processing, packaging, and distribution. Major sources can be attributed to the residue, weathering, fragmentation, and secondary breakdown of agricultural plastics, such as mulching films, greenhouse covers, irrigation tapes, pipes, nets, and food-contact materials. Particles can be redistributed through irrigation water, surface runoff, soil dust resuspension, atmospheric deposition, and the agricultural application of compost or sewage sludge. Additional contamination can also cause from abrasion, shedding, and migration of processing equipment, filtration media, packaging materials, and storage interfaces. Together, these pathways can pose a great challenge to the multi-source, multi-stage, and cross-media exposure pattern, leading to difficult-to-source identification, risk interpretation, and food safety. This review aims to focus on microplastic monitoring, assessment, and process control in the agro-food chain. Major source categories and contamination pathways were summarized in crop production, animal production, and processing environments, and then examined the reported occurrence in plant-derived foods, animal-derived foods, and processed products. Available studies showed that microplastics were widely detected in vegetables, fruits, cereals, shellfish, fish, meat products, bottled water, salt, sugar, honey, beer, and liquid milk. But reported concentrations varied substantially over studies, leading to orders of magnitude. Such heterogeneity was strongly influenced by commodity differences and regional context, method factors, including particle-size thresholds, reporting metrics with particle number or mass, pretreatment intensity, recovery correction, blank control, confirmation criteria, and sampling contexts, such as washing, peeling, retail cutting, cooking, and packaging. Direct comparison was limited for the quantitative interpretation, especially when trophic transfer coexisted with process-related contamination. Therefore, current occurrence data were more useful to identify high-concern commodities, high-concern links, and major uncertainty sources, compared with the cross-study ranking under heterogeneous analytical conditions. Analytical challenges were further highlighted under complex agro-food matrices rich in lipids, proteins, polysaccharides, pigments, and inorganic particulates, where microplastics often occurred at trace levels over broad size ranges. Practical bottlenecks were closely related to matrix removal, polymer preservation, particle loss control, and procedural contamination prevention. Pretreatment strategies, including chemical digestion, enzymatic digestion, density separation, and membrane filtration, were compared from the perspective of matrix applicability, polymer compatibility, recovery performance, and quality-control requirements. Chemical digestion provided effective organic matter removal, but caused damage to polymers under strong acidic or oxidative conditions. Enzymatic digestion offered milder treatment and better polymer preservation but remained constrained by cost, duration, and reagent background. Density separation was used for enrichment efficiency, but it was required for the selection of separation media and recovery validation, especially for small particles and high-density polymers. Membrane filtration functioned as a concentration step as a critical interface affecting optical imaging, verification, and background interference. Verification workflows were reviewed over microscopy and fluorescence imaging, vibrational spectroscopy, thermal analysis, and emerging high-throughput platforms. Among them, microscopy and fluorescence staining supported rapid morphological screening, but visual identification alone remained vulnerable to false positives and operator subjectivity. Micro-FTIR and micro-Raman spectroscopy provided for chemical identification central to polymer at the particle level, yet they were constrained by throughput, diffraction limits, fluorescence interference, and spectral-library dependence. Thermal evaluations, such as Py-GC/MS and TED-GC/MS, were mass-based quantification of visually undetectable particles, but their performance in food matrices remained sensitive to marker interference, calibration strategy, and background decomposition products. Emerging techniques, including laser direct infrared imaging, optical photothermal infrared spectroscopy, atomic force microscopy infrared spectroscopy, hyperspectral imaging, surface-enhanced Raman scattering, and portable sensing systems, were expected to improve throughput, particle-size coverage, and field-oriented monitoring, although their broader application still depended on scenario adaptation and quality. Artificial intelligence was reviewed as an auxiliary tool for high-throughput screening, spectral interpretation, and multimodal data integration. Image-based models were used to improve particle localization, counting, and morphological characterization in fluorescence and microscopic datasets. While machine learning and deep-learning were supported spectral denoising, feature extraction, data matching, and multimodal fusion over imaging, FTIR, and Raman data. Their practical value depended on transparent training datasets, reproducible preprocessing pipelines, external validation over matrices and devices, drift monitoring, reject-option strategies, and alignment with conventional analytical indicators, such as recovery, blank correction, repeatability, detection limits, and particle-size-specific performance. As such, artificial intelligence was better supported by standardized monitoring and data interpretation rather than a substitute for chemical verification. Furthermore, a flux accounting perspective was introduced to link endpoint measurements with process control. Input and output fluxes of the individual unit were defined to quantify the removal efficiency and net introduction. This framework was used to identify critical control points and then evaluate mitigation options. Source reduction, process interception, and terminal were selected as complementary controls. Priority measures included the life cycle of agricultural plastics, fragmentation-prone inputs, water purification, control of sludge and compost application, high-friction processing interfaces, and standardized sampling, pretreatment, reporting, and quality-control requirements. Safety thresholds were required for microplastics in agro-food products, covering precautionary and prioritizing comparable data generation, critical control points, and controllable exposure reduction over the full chain. Overall, the microplastics in agro-food systems can be expected to move from isolated endpoints toward integrated monitoring, process evaluation, and process-oriented detection.

Xanthoceras sorbifolium Bunge, belonging to the Sapindaceae family, has been one of the most potential promising woody oil species in northern China. X. sorbifolium can play the important ecological roles, including desert greening, windbreak, and sand fixation, also providing for the edible and medicinal value among plant resources. Furthermore, X. sorbifolia can be used as nature food for health protection and disease prophylaxis in recent years, such as for tea, due to its high concentration of unsaturated fatty acids, especially neuroprotective nervonic acid. The leaves and buds of X. sorbifolium also share the nourishing ingredients and bioactive substances, including amino acids, proteins, soluble sugars, polyphenols, flavonoids, and saponins. In this study, a full-factor design was adopted with five levels of steeping temperature and seven levels of steeping time. Nutritional quality of the tea infusions was evaluated under 35 combinations. A systematic investigation was conducted to fully clarify the influence of brewing conditions on the quality of the tea infusions with the Xanthoceras sorbifolium bud green tea (XBT). Fuzzy A fuzzy membership function was used to screen the optimal brewing parameters. The aroma and taste of the tea infusions with XBT were characterized by gas chromatography-ion mobility spectrometry, electronic nose, and electronic tongue. The results showed that the content of tea polyphenols in the tea infusions of XBT first increased, then decreased, and finally tended to be stable, with the extension of steeping time. While the contents of total free amino acids, flavonoids, and caffeine generally showed a trend of first increasing and then decreasing. At the same time, a higher brewing temperature was conducive to the dissolution of soluble sugar components. Fuzzy The fuzzy membership function also showed that the long-term brewing (360, 720 min) was not conducive to the overall quality of the tea infusions, while short brewing time (5 min) together with low temperature (50, 60 ℃) failed to fully dissolve the nutrients in the tea infusions of XBT. Three optimal combinations of steeping parameters were obtained: 70 ℃ for 60 min, 90 ℃ for 30 min, and 80 ℃ for 30 min. A total of 60 volatile organic compounds and 13 key aroma substances were detected in the tea infusions. 2-methylbutyraldehyde and 3-methylbutyraldehyde jointly contributed to the roast aroma of tea infusions, while both octanal and hexanal contributed to the fresh fruit aroma of the tea infusions, and pentanal contributed to the grassy aroma of the tea infusions. The brewing conditions of 70 ℃ for 60 min and 80 ℃ for 30 min were more benefits beneficial to the retention of aroma substances in the tea infusions, thereby reducing the bitterness and astringency. According to the nutritional quality, taste, and aroma, the optimal combination of 80 ℃ for 30 min was recommended as the daily drinking brewing for XBT. The finding can also provide a theoretical basis for the subsequent processing of tea beverages. In short, the bud can be expected to serve as a tea products for the high economic value of Xanthoceras sorbifolium.

Satellite remote sensing can identify the irrigation information, because of its rapid and wide-area observation. However, only a single source is often extracted from remote sensing data. Spatial and temporal resolution cannot fully meet the requirement for the high-accuracy and dynamic identification of irrigation information at the regional scale, especially for the strong spatial heterogeneity of agricultural activities in the complex terrain. In this study, a remote sensing framework was developed to identify the irrigation information using drought index analysis and spatiotemporal fusion. The Guanzhong Region was also taken as the study area. The temperature vegetation dryness index (TVDI) was selected as the identification index after correlation analysis between drought indices and soil moisture. Elevation correction with fusion optimization was introduced to characterize the variation in the soil moisture. Its spatiotemporal fusion accuracy was also enhanced under complex terrain conditions. Subsequently, the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM) was used to fuse high-spatial-resolution Landsat imagery and high-temporal-resolution MODIS data for the high-spatiotemporal-resolution TVDI time series. Spring irrigation information in the Guanzhong Region in 2024 was identified using the threshold method with precipitation data. The results showed that the correlations between the remote sensing drought indices and soil moisture at the 10-20 cm depth were generally higher than those with soil moisture at the 0-10 cm depth. Elevation topographic correction effectively reduced the influence of terrain on land surface temperature. There was a strong correlation between TVDI and soil moisture. Furthermore, the elevation-corrected TVDI showed a strong negative correlation with soil moisture at the 10-20 cm depth, with the maximum correlation coefficient of −0.77 during the spring crop growth period. Normalized difference vegetation index (NDVI) and land surface temperature (LST) were fused for the higher accuracy of TVDI than the strategy of first calculation and then fusion. R² and RMSE values of 0.76 and 0.07 for the former, whereas 0.44 and 0.13 for the latter, respectively. The validation showed that the overall accuracy was 90.8% for the identification in the Donglei Phase II irrigation district, with a Kappa coefficient of 0.80. The mean error was 15.1% and 14.3%, respectively, for accumulated and actual irrigated areas in the irrigation districts. Regional identification results indicated that the spring irrigation was mainly concentrated from March to April, with the irrigation frequency ranging from one to two times. Irrigated areas were distributed in the relatively flat Weihe Plain, with a spatial pattern characterized by broader irrigation extent and higher irrigation frequency in the eastern and western parts. While the central part exhibited relatively lower irrigation intensity. The spatial distribution and irrigation frequency of spring irrigation were dominated by regional topography, water supply, and cropping structure. The finding can provide a strong reference to identify the regional-scale irrigation information and water resources under complex terrains.

Sinohyriopsis cumingii is one of the economically important freshwater mussels in the pearl aquaculture industry. Phenotypic traits of S. cumingii can be expected to evaluate the individual growth performance. Germplasm resources are identified to implement precise genetic breeding. However, conventional manual measurements cannot fully meet the scalability and applicability of the large-scale production in intelligent aquaculture, due to their labor-intensive, time-consuming, and highly susceptible to subjective errors. In this study, an improved measurement was proposed for the non-destructive, rapid, and accurate acquisition of phenotypic parameters using YOLOv8n, termed YOLOv8n-CBM. 1) An integrated phenotypic measurement for S. cumingii was constructed to combine the dynamic transmission, machine vision, and digital image processing. The system comprised a conveyor device, a high-precision industrial camera, and an image processing module. The mussel samples were automatically transported into the imaging area, thus enabling standardized image acquisition and high-throughput phenotypic measurement under continuous dynamic conditions. 2) Three targeted improvements were implemented in the original YOLOv8n network architecture, according to the characteristics of mussel images. In the backbone network, four convolutional block attention modules (CBAM) were embedded after each C2f block to enhance the extraction of contour edges and local features of mussel samples, while effectively suppressing irrelevant background interference. In the neck network, the bidirectional feature pyramid network (BiFPN) was introduced to strengthen bidirectional fusion of multi-scale features for the targets of different sizes and postures. Meanwhile, the original C2f module was replaced with a multi-scale dilated attention (MSDA) module to expand the network’s receptive field for the local fine-grained and global contextual information. Finally, the key phenotypic parameters were extracted, including shell length, full height, shell height, and radial rib length of the buttock angle, according to the geometric relationship between rotated bounding boxes and biological key points. A series of experiments was conducted on a dataset of 50 manually annotated S. cumingii samples with diverse sizes and postures. The results show that the mean average precision (mAP50-95) of the YOLOv8n-CBM model reached 98.2%, indicating the rotated object detection performance over the original YOLOv8n model. The average localization deviation of biological key points was less than 2.0 mm, indicating the high precision in feature detection. The mean absolute errors (MAE) of shell length, full height, shell height, and radial rib length of the buttock angle were 1.51, 1.08, 1.019, and 1.998 mm, respectively. In all groups stratified by different shell lengths and full heights, the measurement errors of YOLOv8n-CBM were consistently lower than those of the original YOLOv8n model, with the maximum absolute error within 2.879 mm. Measurement accuracy and robustness were effectively improved with diverse morphologies and postures. In conclusion, the reliable technical approach was used to realize the rapid, accurate, and non-destructive acquisition of phenotypic traits in S. cumingii. Shellfish growth evaluation and genetic breeding can be expected to support the transition of the pearl industry from empirical farming to data-driven and intelligent aquaculture. The findings can also offer a valuable reference for phenotypic measurement in the molluscan species.

Kitchen waste (KW) composting often suffers from prolonged processing time and strong odor emissions due to the high moisture content and complex organic composition of the substrate. This study aimed to elucidate how inoculation with an immobilized bacterial consortium (IBC) regulates the microbial community, co-occurrence network structure, and metabolic functions in a KW composting system, thereby improving composting efficiency and mitigating odor generation. A composting system inoculated with an IBC composed of six functional bacterial strains was established, with a non-inoculated treatment serving as control. The physicochemical parameters of the compost, including temperature, moisture content, pH, and germination index (GI), were continuously monitored throughout the 15-day process. Bacterial community composition and succession were analyzed via 16S rRNA gene sequencing. Co-occurrence networks were constructed for different composting phases to reveal changes in microbial interactions. Functional Annotation of Prokaryotic Taxa (FAPROTAX) was applied to predict metabolic pathways related to carbon, nitrogen, and sulfur cycling. Partial Least Squares Path Modeling (PLS-PM) was used to explore causal relationships among physicochemical conditions, microbial community structure, network complexity, metabolic functions, and composting efficiency. The IBC treatment sustained a longer and more stable thermophilic phase than the control, accelerating compost maturity, with the GI reaching 88.89% on day 15 compared to 58.89% in the control. Inoculation significantly reshaped the bacterial community structure and enhanced deterministic assembly processes, guiding microbial succession toward functional guilds specialized in organic degradation and nutrient transformation. The inoculated compost exhibited greater network complexity, characterized by increased node and edge numbers, higher average degree, and reduced path length and network diameter, indicating stronger microbial connectivity and synergistic metabolic cooperation. Functional prediction showed that carbon cycling was dominated by chemoheterotrophy and aerobic chemoheterotrophy, both increasing over time, while fermentation functions gradually declined. In the nitrogen cycle, nitrite respiration and dissimilatory ammonification were most active during the early phase, but nitrogen fixation became dominant in the later cooling and maturation stages. Sulfur respiration pathways were markedly suppressed in the inoculated group, implying the inhibition of reductive sulfur metabolism and reduced potential for odor emission. PLS-PM analysis further demonstrated that microbial inoculation reversed the relationship between physicochemical properties and bacterial community from negative to positive, promoting the enrichment of core functional taxa. The relationship between community structure and metabolic function shifted from diversity-driven to functional taxa-driven patterns. Although the direct effect of network complexity on composting efficiency declined, it indirectly enhanced system functionality through improved robustness and cooperative stability. The immobilized bacterial consortium effectively optimized the composting physicochemical environment, reconstructed microbial interaction networks, and reinforced functional coupling among key taxa. These integrated effects accelerated organic matter degradation, shortened the composting period, and reduced odor emissions. The study provides new ecological insights into the microbial regulatory mechanisms of KW composting and supports the development of efficient, low-emission, and sustainable biotechnological strategies for organic waste recycling.