A calibration transfer optimized single kernel near-infrared spectroscopic method

Deployment strategy of multiple miniaturized near-infrared spectrometers based on spectral transfer for characterizing soil organic matter and nitrogen

Developing timely, convenient, and low-cost methods for high-frequency characterization of soil nutrients is necessary for implementing precise soil nutrient management. With the current availability of numerous calibration models of laboratory benchtop near-infrared (NIR) spectrometers for rapid soil nutrient characterization and the appearance of low-cost, convenient miniaturized NIR spectrometers, this study proposes an efficient deployment strategy to address model failure due to inter-device variation based on spectral transfer. The strategy involves using Direct Standardization (DS) to migrate the spectra from multiple miniaturized NIR spectrometers with a laboratory benchtop NIR spectrometer and then directly applying the existing calibration models of the laboratory benchtop instrument to the transferred spectra for soil nutrient analysis. The results indicated that the DS method successfully transferred the spectra of miniaturized devices to be consistent with the spectra of the laboratory benchtop instrument. The soil organic matter (SOM) predictions using the transferred spectra and the calibration models of the laboratory benchtop instrument were even more accurate than those using the respective models developed for each miniaturized devices, with root mean square error (RMSE) of 0.177 %, 0.177 %, and 0.150 %, respectively, while the performances of total nitrogen (TN) predictions were comparable to those using the respective models, with RMSE of 0.013 %, 0.012 %, and 0.010 %, respectively. Bland-Altman plots demonstrated good consistency between the strategy proposed in this study and the strategy of developing respective models for each miniaturized device, with no difference in predictions for the independent validation set compared to the laboratory benchtop instrument. This study proved the feasibility of deployment strategy of multiple miniaturized NIR spectrometers based on spectral transfer, offering a new solution for high-frequency on-site soil nutrient characterization.

Sequentially orthogonalized canonical partial least squares for improved multiple responses modeling in multiblock data sets

Multiblock data sets and modeling techniques are widely encountered in the chemometric community. Although the currently available techniques, such as sequential orthogonalized partial least squares (SO-PLS) regression are mainly focused on the prediction of a single response and deal with the multiple response(s) case using PLS2 type approach. Recently, a new approach called canonical PLS (CPLS) was proposed for extracting the subspaces efficiently for multiple response(s) cases, supporting both regression and classification. 'Efficiently' here means more information in fewer latent variables. This work suggests a combination of SO-PLS and CPLS, sequential orthogonalized canonical partial least squares (SO-CPLS), to model multiple response(s) for multiblock data sets. The cases of SO-CPLS for modeling multiple response(s) regression and classification were demonstrated on several data sets. Also, the capability of SO-CPLS to incorporate meta-information related to samples for efficient subspace extraction is demonstrated. Furthermore, a comparison with the commonly used sequential modeling technique, called sequential orthogonalized partial least squares (SO-PLS), is also presented. The SO-CPLS approach can benefit both the multiple response(s) regression and classification modeling and can be of high importance when meta-information such as experimental design or sample classes is available.

Feasibility of an NIR spectral calibration transfer algorithm based on optimized feature variables to predict tobacco samples in different states

The prediction accuracy of calibration models for near-infrared (NIR) spectroscopy typically relies on the morphology and homogeneity of the samples. To achieve non-homogeneous tobacco samples for non-destructive and rapid analysis, a method that can predict tobacco filament samples using reliable models based on the corresponding tobacco powder is proposed here. First, as it is necessary to establish a simple and robust calibrated model with excellent performance, based on full-wavelength PLSR (Full-PLSR), the key feature variables were screened by three methods, namely competitive adaptive reweighted sampling (CARS), variable combination population analysis-iteratively retaining informative variables (VCPA-IRIV), and variable combination population analysis-genetic algorithm (VCPA-GA). The partial least squares regression (PLSR) models for predicting the total sugar content in tobacco were established based on three optimal wavelength sets and named CARS-PLSR, VCPA-IRIV-PLSR and VCPA-GA-PLSR, respectively. Subsequently, they were combined with different calibration transfer algorithms, including calibration transfer based on canonical correlation analysis (CTCCA), slope/bias correction (S/B) and non-supervised parameter-free framework for calibration enhancement (NS-PFCE), to evaluate the best prediction model for the tobacco filament samples. Compared with the previous two transfer algorithms, NS-PFCE performed the best under various wavelength conditions. The prediction results indicated that the most successful approach for predicting the tobacco filament samples was achieved by VCPA-IRIV-PLSR when coupled with the NS-PFCE method, which obtained the highest determination coefficient (R_p² = 0.9340) and the lowest root mean square error of the prediction set (RMSEP = 0.8425). VCPA-IRIV simplifies the calibration model and improves the efficiency of model transfer (31 variables). Furthermore, it pledges the prediction accuracy of the tobacco filament samples when combined with NS-PFCE. In summary, calibration transfer based on optimized feature variables can eliminate prediction errors caused by sample morphological differences and proves to be a more beneficial method for online application in the tobacco industry.

Self-supervised maize kernel classification and segmentation for embryo identification

Introduction

Computer vision and deep learning (DL) techniques have succeeded in a wide range of diverse fields. Recently, these techniques have been successfully deployed in plant science applications to address food security, productivity, and environmental sustainability problems for a growing global population. However, training these DL models often necessitates the large-scale manual annotation of data which frequently becomes a tedious and time-and-resource- intensive process. Recent advances in self-supervised learning (SSL) methods have proven instrumental in overcoming these obstacles, using purely unlabeled datasets to pre-train DL models.

Methods

Here, we implement the popular self-supervised contrastive learning methods of NNCLR Nearest neighbor Contrastive Learning of visual Representations) and SimCLR (Simple framework for Contrastive Learning of visual Representations) for the classification of spatial orientation and segmentation of embryos of maize kernels. Maize kernels are imaged using a commercial high-throughput imaging system. This image data is often used in multiple downstream applications across both production and breeding applications, for instance, sorting for oil content based on segmenting and quantifying the scutellum's size and for classifying haploid and diploid kernels.

Results and discussion

We show that in both classification and segmentation problems, SSL techniques outperform their purely supervised transfer learning-based counterparts and are significantly more annotation efficient. Additionally, we show that a single SSL pre-trained model can be efficiently finetuned for both classification and segmentation, indicating good transferability across multiple downstream applications. Segmentation models with SSL-pretrained backbones produce DICE similarity coefficients of 0.81, higher than the 0.78 and 0.73 of those with ImageNet-pretrained and randomly initialized backbones, respectively. We observe that finetuning classification and segmentation models on as little as 1% annotation produces competitive results. These results show SSL provides a meaningful step forward in data efficiency with agricultural deep learning and computer vision.

Effects of Variations in the Chemical Composition of Individual Rice Grains on the Eating Quality of Hybrid Indica Rice Based on Near-Infrared Spectroscopy

The chemical composition of individual hybrid rice (F2) varieties varies owing to genetic differences between parental lines, and the effects of these differences on eating quality are unclear. In this study, based on a self-developed near-infrared spectroscopy platform, we explored these effects among a set of 143 hybrid indica rice varieties with different eating qualities. The single-grain amylose content (SGAC) and single-grain protein content (SGPC) models were established with coefficients of determination (R²) of 0.9064 and 0.8847, respectively, and the dispersion indicators (standard deviation, variance, extreme deviation, quartile deviation, and coefficient of variation) were proposed to analyze the variations in the SGAC and SGPC based on the predicted results. Our correlation analysis found that the higher the variation in the SGAC and SGPC, the lower the eating quality of the hybrid indica rice. Moreover, the addition of the dispersion indicators of the SGAC and SGPC improved the R² of the eating quality model constructed using the composition-related physicochemical indicators (amylose content, protein content, alkali-spreading value, and gel consistency) from 0.657 to 0.850. Therefore, this new method proved to be useful for identifying high-eating-quality hybrid indica rice based on single near-infrared spectroscopy prior to processing and cooking.

Detection of Pesticide Residue Level in Grape Using Hyperspectral Imaging with Machine Learning

Rapid and accurate detection of pesticide residue levels can help to prevent the harm of pesticide residue. This study used visible/near-infrared (Vis-NIR) (376–1044 nm) and near-infrared (NIR) (915–1699 nm) hyperspectral imaging systems (HISs) to detect the level of pesticide residues. Three different varieties of grapes were sprayed with four levels of pesticides. Logistic regression (LR), support vector machine (SVM), random forest (RF), convolutional neural network (CNN), and residual neural network (ResNet) models were used to build classification models for pesticide residue levels. The saliency maps of CNN and ResNet were conducted to visualize the contribution of wavelengths. Overall, the results of NIR spectra performed better than those of Vis-NIR spectra. For Vis-NIR spectra, the best model was ResNet, with the accuracy of over 93%. For NIR spectra, LR was the best, with the accuracy of over 97%, but SVM, CNN, and ResNet also showed closed and fine results. The saliency map of CNN and ResNet presented similar and closed ranges of crucial wavelengths. Overall results indicated deep learning performed better than conventional machine learning. The study showed that the use of hyperspectral imaging technology combined with machine learning can effectively detect the level of pesticide residues in grapes.

A brief note on application of domain-invariant PLS for adapting near-infrared spectroscopy calibrations between different physical forms of samples

Near-infrared (NIR) calibration models are widely developed and routinely used for the prediction of physicochemical properties of samples. However, the main challenge with NIR models is that they are highly specific to the physical form of the samples. For example, a NIR calibration established for solid samples can usually not be used for the same samples in powdered form. Domain adaption (DA) techniques, such as domain invariant partial least-squares (di-PLS) regression, have recently appeared in the chemometric domain which allow adapting NIR calibrations for new sample-/instrument- or environment-associated conditions in a standard free manner. A practical use case of di-PLS can be assumed as the adaption of NIR calibration models to be used in different physical forms of samples. In this contribution we show, for the first time, application of di-PLS regression analysis for adapting a near-infrared (NIR) calibration for solid rice kernels to be used on powdered rice flour without the need for new reference measurements for the latter. di-PLS is a domain adaption technique that removes the differences between different but related data sources (i.e. domains) to reach generalized models. The study found that di-PLS allowed a direct adaption of calibration based on solid rice kernels to be used on powdered rice flour without requiring any reference protein measurements for the latter. Our results suggest that DA tools, such as di-PLS, can support a wider usage of chemometric calibrations especially when models need to be adapted to different physical forms of the same samples.

A correlation-analysis-based wavelength selection method for calibration transfer

Considering that the spectral signals vary among different instruments, calibration transfer is required for further popularization and application of the near-infrared spectroscopy (NIRS). To achieve good calibration transfer results, spectral variables with stable and consistent signals between instruments and containing the target component information should be selected. In this study, a correlation-analysis-based wavelength selection method (CAWS) is proposed for calibration transfer. This method relies on the selection of wavelengths at which the spectral responses of master and slave instruments are well correlated (high absolute values of Pearson's correlation coefficient (|R_i|)). The proposed CAWS method was applied to two available datasets, corn and rice bran, and its calibration transfer performances were compared with other wavelength selection methods. The effects of pretreatment methods and calibration transfer algorithms were also assessed. The CAWS optimized models obtained lower root mean square errors of prediction (RMSEP_trans) after calibration transfer, suggesting that the proposed method is capable of effectively improving the efficiency of calibration transfer. Combinations of this method with other wavelength selection methods and calibration transfer algorithms may further enhance the efficiency of calibration transfer, and thus should be thoroughly investigated.

The method of calibration model transfer by optimizing wavelength combinations based on consistent and stable spectral signals

Basing on the wavelengths with consistent and stable spectral signals between spectrometers, wavelength combinations were screened by different methods to obtain robust and simple near infrared spectra (NIR) calibration models that can be shared by slave spectrometers directly. Firstly, the wavelength set of Usc, at which the spectral signals between spectrometers are consistent and stable, was obtained by the method of screening the wavelengths with consistent and stable signals between spectrometers (SWCSS for short). Then, the wavelength set of Uscr whose spectral responses are correlated with dependent variables strongly was selected from Usc. Basing on Uscr, the methods of uninformative variable elimination (UVE), variable importance in projection (VIP) and selectivity ratio (SR) were applied to further screen optimal wavelength sets to obtain better NIR calibration models. These sets were recorded as Uscr^UVE, Uscr^VIP and Uscr^SR, respectively. The NIR partial least squares (PLS) models for predicting total alkaloids content of tobacco leaves were built on the three optimal wavelength sets, and named as Uscr^UVE-PLS, Uscr^VIP-PLS, Uscr^SR-PLS, respectively. Both Uscr^UVE-PLS and Uscr^VIP-PLS give satisfactory prediction errors for master and slave samples, and work better than the PLS model built on the whole wavelengths (WW-PLS) after piecewise direct standardization (PDS) calibration. The results show that further optimizing wavelength combinations based on consistent and stable spectral information cannot only simplify PLS models and improve the models' efficiency, but also ensure the models' accuracy when they are transferred to slave spectrometers. Wavelength selection based on the whole wavelengths without considering spectra consistency between spectrometers can improve the performance of the calibration models on the master spectrometer but cannot ensure the prediction accuracy of the slave samples.