Hao Tian’s research while affiliated with Zhejiang University and other places

Visible/Near-infrared (Vis/NIR) spectroscopy is widely used in the detection of dry matter content (DMC) of potatoes. However, biological variability (e.g., cultivar and season) will affect the potato DMC and spectral features, and will further cause the DMC prediction model ineffective. This study aimed to develop robust Vis/NIR models for predicting potato DMC with influence of cultivar and season. The local and global models were developed to explore the influence of cultivar and season. The Mahalanobis distance and concentration gradient (MD-CG) method was developed to select representative samples, and the combinations of different variable selection methods (CARS, SPA and CSMW) and model updating methods (SBC and recalibration) were investigated for model enhancement. The results indicated that 10 new samples selected by MD-CG method, combined with variable selection and model updating, were sufficient to improve the performance of the local (RPDp>1.7) and global (RPDp>2) models. In the local models, for the datasets with different cultivars (EG-2021, XS-2021 and AT-2021), the optimal results were obtained using CSMW combined with recalibration, and the RMSEp was decreased from 4.18%, 1.14%, 2.54–1.05%, 0.72%, 0.79%, respectively. For the datasets with different seasons (FA-2022), the optimal result was obtained by using SPA combined with recalibration, and the RMSEp was decreased from 3.70% to 0.91%. For the global model, CSMW combined with recalibration and SPA combined with SBC obtained better results, with RMSEp decreasing from 0.83% to 0.52% and 0.51%, respectively. The MD-CG method and the combinations of variable selection and model updating proposed in this study are important to reduce the influence of external conditions and enhance the model robustness to biological variability.

Potatoes are generally consumed directly as a staple food or used for processing, depending on the quality of raw materials. Dry matter content (DMC) is the most critical characteristic of potatoes, as it determines the processing and the final product quality. This study aimed to investigate the potential of different optical sensing systems in predicting the DMC of intact potato tubers, and the efficacy of classifying potatoes based on dry matter levels. The whole tubers were scanned using three optical sensing modes (transmittance spectra, interactance spectra and hyperspectral imaging). PLSR and different classifiers (PLSDA, SVM and ANN) were utilized to build the prediction and classification models, respectively. To extract the most influential wavelengths related to the prediction of DMC, the CARS and CSMW techniques were applied. The results indicated that the DMC of two sections on the equator of the potato tuber belly can well represent the DMC of the intact potato, and together with the spectral detection at the equatorial position, it provided good performance. The CARS-PLSR prediction model in transmittance mode showed stronger correlations than other systems, with Rp and RMSEP values of 0.968 and 0.413%, respectively. The CARS-SVM-Linear classification model exhibited the best performance with classification rates of 100% and 97.62% in the training and testing sets, respectively. Moreover, the spectra preferred by CARS and CSMW variable selection methods in transmittance mode overlapped near the absorption peak at 980 nm, indicating the importance of this band for predicting DMC. This study presented the feasible application of using spectroscopy to evaluate the DMC of intact potatoes and classify potatoes based on thresholds that are crucial to consumers and food processors. Graphical abstract

Flesh content of pomelos sometimes does not match the price due to thick peel, which reduces their commercial value and consumer satisfaction. This study demonstrated the potential use of X-ray imaging method for non-destructive evaluation of the edible rate and the flesh content of ‘Hongrou’ pomelos and ‘Guanxi’ pomelos. An adaptive threshold segmentation method was used to segment the X-ray image into background, flesh region and peel region. Then, 2D edible rate and 2D flesh content were defined based on region area ratio and gray level logarithmic sum, respectively, and the multiple linear regression (MLR) models of edible rate and flesh content were established for quantitative analysis. The results showed that the residual predictive deviation (RPD) value of edible rate of ‘Hongrou’ and ‘Guanxi’ pomelos in prediction set were up to 2.78 and 2.82, respectively. The hybrid model based on both two cultivars pomelos also achieved good prediction accuracy (RPD = 2.83). In terms of flesh content prediction, the model prediction performance of ‘Hongrou’ pomelo (RPD = 2.92) was obviously better than that of ‘Guanxi’ pomelo (RPD = 2.05), and the RPD value of hybrid model in prediction set was 2.62. Further, both two-grade and three-grade linear discriminant analysis (LDA) classifiers were trained to explore the feasibility of using X-ray images to classify the edible rate of pomelos, and the classification accuracy of hybrid samples were 96.7% and 90.0%, respectively. Overall, the features extracted from X-ray image of pomelo could allow the non-destructive evaluation of the edible rate for pomelo.

The on-line detection of paddy moisture content (MC) during harvest has gained increasing interest recently due to its unique role for the control of combine harvester, yield evaluation and post-harvest grain handling operations. However, it is very difficult to achieve good performance under the complex and changeable situation during field harvest. In this study, paddy varieties, paddy flow, feeding types and algorithms were comprehensively considered to optimize the MC detection method. Firstly, an on-line near-infrared sensing system supplemented for grain tank of combine harvester was designed, and spectra were collected under the most common and essential detecting conditions, which including paddy varieties, feeding types and straw effect. Then, ensemble preprocessing, parameter optimization and accuracy test were performed. The best result of all conditions was extreme learning machine (ELM) coupled with the ensemble preprocessing of orthogonal signal correction with savitzky-golay (OSC + SG). The root mean standard error of prediction (RMSEPV) of this method after validation on unknown sample was as low as 1.0791% w.b, and the residual predictive deviation (RPDV) was higher than 3.5646. Stability tests were carried out under conditions of varying feeding types and straw quantities. The results showed that ELM had enough robustness to cope with complex detecting conditions and maintain proper accuracy as the mean value of repeatability, conditions and reproducibility were calculated as 0.0213%, 0.4471% and 0.6868% w.b, respectively. Despite the preliminary feasibility for on-line MC measurement of paddy, the on-line near-infrared sensing system needs to be verified on combine harvester during harvest.

The purpose of this study was to compare the accuracy and robustness of the detection models based on bulk optical properties (BOP) with that based on conventional spectroscopy in kiwifruit quality evaluation. 81 kiwifruit were selected as experimental samples in this study. A single integrating sphere system was built to estimate the bulk absorption coefficient (μa) and bulk reduced scattering coefficient (μs′) of samples and a self-designed online system was used to obtain transmission spectra. The relationship of μa and μs′ with SSC and flesh firmness was analyzed, and detection models were established using partial least squares regression (PLSR). Competitive adaptive reweighted sampling (CARS) was also used to eliminate the variables in the original spectra that do not contribute to the improvement of the model performance. Results showed that μa at 670 nm decreased with the increase of SSC, μa at 720–900 nm and 950–1000 nm increased with the increase of SSC, and spectra of μs′ decreased with the decreasing firmness. CARS-PLSR models were developed, based on μa, μs′, μa×μs′, μeff, μt′, and transmission spectra. The accuracy of the model based on BOP in predicting internal quality was better than that based on transmission spectra. The model based on μa was the best for SSC (Rp2 = 0.97, RMSEP = 0.25%), and the model based on μa×μs′ was the best for flesh firmness (Rp2 = 0.97, RMSEP = 0.02 N). 20 kiwifruit that differed from the experimental samples in planting orchard and harvest time were used to compare the robustness and portability of the models. Results showed that all SSC models and the firmness model based on μa×μs′ had good robustness and portability. However, the model based on transmission spectra had a poor performance in predicting the firmness of samples from the external validation set. This study provides an effective reference for the prediction of firmness and SSC based on BOP and transmission spectra of kiwifruit.

The penetration depth of light and the distribution of components in fruit affect the accuracy of models in visible-near infrared spectroscopy (Vis-NIRS). These issues are particularly prominent for predicting the soluble solids content (SSC) of large fruit with thick rinds, such as pomelo. To describe the light penetration and distribution inside pomelos, three different modes of puncture measurement were performed by an original system with a puncture optical fiber. The asymmetrical changes and distribution of light intensity were observed in tissues. The semi-transmittance mode was adopted combined with the limited penetration depth, photon aggregation near the light source, and the distribution characteristics of SSC in pomelo. Subsequently, the multi-point spectra at six points in semi-transmittance mode were employed to evaluate the influence of the asymmetrical light and SSC distribution using cross-validation of partial least squares regression (PLSR). The models established by the mean spectra got good performances by eliminating the difference in light distribution. In addition, the probability distribution of SCC could affect the model performance more than the spatial distribution. Finally, the effective wavelengths were selected by competitive adaptive reweighted sampling (CARS) to establish the calibration model. The global model of CARS-PLSR using mean spectra and mean SSC of six points got the best performance in terms of root mean square error of prediction (RMSEP) and residual predictive deviation (RPD), with values of 0.25% and 2.62, respectively. Overall, multi-point detection in semi-transmittance mode could weaken the distribution difference caused by biological variability and allow the non-destructive prediction of SSC in pomelo.