The textural formation mechanism of goji berry (Lycium barbarum L.) during pulsed vacuum drying (PVD) was investigated using multiscale approach integrating low-field nuclear magnetic resonance (LF-NMR), transmission electron microscopy (TEM), X-ray diffraction (XRD), rheological analyses, etc. Results indicated non-uniform moisture migration and transition in water transport pathways during PVD. Textural and rheological alterations were associated with pectin depolymerization and subsequent matrix densification. The progressive loss of cell membrane integrity was observed, as evidenced by ten-fold increase in malondialdehyde content. Concurrent modifications in cell wall polysaccharides (cellulose, hemicellulose, and pectin) revealed the crucial roles of pectin methyl esterase (PME) and polygalacturonase (PG) in facilitating the accumulation and transformation of chelator soluble pectin (CSP) from both water-soluble pectin (WSP) and Na2CO3-soluble pectin (NSP). These findings provide fundamental insights into the structural and mechanical evolution of berry during PVD, offering theoretical foundation for optimizing industrial dehydration processes to improve product quality.
Background Cold chain logistics (CCL) is essential for ensuring the quality and safety of fresh agricultural products and reducing spoilage. However, traditional CCL systems show poor environmental sensing, delayed responses, and inefficient coordination. These gaps raise the risk of chain breaks, product loss, and higher emissions. Smart transformation is thus urgently needed to address the dual challenges of high spoilage and high emissions. Scope and approach This review first outlines how the smart CCL concept evolved and then assesses recent advances in multidimensional sensing, adaptive control, and collaborative management. The synthesis draws on a focused set of seminal and illustrative studies to map the current status, core challenges, and emerging trends within an integrated framework. Key findings and conclusions Next-generation information technologies (e.g., Internet of Things, artificial intelligence, and blockchain) enable closed-loop, smart agricultural CCL, which significantly enhances the precision of environmental control, energy efficiency, and carbon emission management through multi-dimensional sensing, adaptive regulation, and collaborative control. Various key challenges remaining include sensors drifting under harsh conditions, as well as the difficulty of fusing data from many sources and performing real-time analyses. Models also adapt poorly to change, and blockchain can be costly and raise privacy issues. Overcoming these challenges will require robust multi-modal sensing, effective collaborative frameworks to drive the green transition, and digital twin platforms with standardized execution interfaces to enhance resilience, minimize product loss, and optimize energy use.
Meeting global food demands by 2050 requires a 45–60
The practical application of spectroscopic models developed with large soil spectral libraries (SSLs) should be to "think globally but fit locally," particularly for in-situ soil spectral measurements under field conditions. Our study explored deep transfer learning (DTL) as a robust calibration transfer strategy to enable SSL-based models to effectively predict soil organic matter (SOM) from in-situ spectra. A mid-infrared (MIR) SSL comprising 3,259 surface soil samples collected from 11 provinces across China was used to develop calibration models with partial least squares regression (PLSR) and one-dimensional convolutional neural network (1D-CNN). The SSL-PLSR and SSL-1D-CNN models demonstrated high predictive accuracy (R-2 > 0.95, RPD > 4.66, and RMSE <= 2.23 g/kg), while the direct application to in-situ spectra resulted in substantial performance degradation. Furthermore, the SOM predictions derived from in-situ MIR measurements remained less accurate, and even after applying direct standardization (DS) and DS combined with adaptive iteratively reweighted penalized least squares (DS-airPLS), the transferred in-situ spectra still failed to achieve the accuracy of lab-based MIR measurements. To address this issue, four transfer learning approaches were evaluated: DS, DS-airPLS, semi-supervised parameter-free calibration enhancement (SS-PFCE), and DTL. Compared with DS, DS-airPLS (RPD < 2.0) and SS-PFCE (RPD < 1.4), DTL achieved higher prediction accuracy for SOM (R-2 = 0.84, RPD = 2.46, and RMSE = 1.46 g/kg). The results demonstrated the superior transferability of DTL compared to conventional calibration transfer methods. Furthermore, SHapley Additive exPlanations (SHAP) analysis identified key spectral regions (1577, 2941, and 2501 cm (1)) that contributed most to SOM prediction, indicating that the DTL model retained critical spectral features when adapting to in-situ spectral data. Overall, this study highlights the potential of deep transfer learning for the practical and cost-effective application of SSL calibration models to in-situ SOM estimation, which minimized the reliance on traditional lab-based calibration and further enhanced efficiency of SOM monitoring.
Accurate and rapid detection of early decay remains a significant challenge in the postharvest grading of citrus fruit, particularly for varieties with weak or no fluorescence. While ultraviolet-induced fluorescence imaging performs well for strongly fluorescent fruit such as oranges, it is ineffective for fluorescence-deficient varieties. To address this limitation, a novel approach combining structured-illumination reflectance imaging (SIRI) with deep learning was proposed for early decay detection in dekopon fruit. Structured light images were captured at four spatial frequencies (0.10, 0.15, 0.20, and 0.25 cycles mm-1), demodulated, and evaluated through visual assessment and contrast index (CI) analysis. The optimal frequency (0.25 cycles mm- 1) was selected for further analysis. Texture features (56 per image) were extracted from the direct component (DC) images, amplitude component (AC) images and the ratio (RT) images of AC and DC, and the Random Frog (RF) algorithm was employed to optimize feature selection. Three machine learning models (i.e., KNN, LS-SVM, and XGBoost) were constructed using both the full and the selected feature sets. XGBoost combined with 12 optimal features extracted from RT images achieved the highest classification accuracy of 95.0 % on the test set. To further enhance detection robustness and deployment efficiency, a lightweight deep learning model, YOLOv11-SSConv, was developed by introducing a spatial and channel reconstruction convolution (SCConv) module. Using RT images as input, this model achieved 98.4 % detection accuracy, while reducing parameters by 4.8 % and increasing inference speed by 25 % compared to the original YOLOv11. These results demonstrate that combining SIRI with deep learning enables accurate and efficient early decay detection in dekopon fruit, offering strong potential for industrial applications, especially where fluorescence imaging is ineffective.
