Identification of hybrid tuliptree and parental species by near infrared hyperspectral imaging （NIR-HSI） spectroscopy

doi:10.21203/rs.3.rs-1993392/v1

Download PDF

Article

Identification of hybrid tuliptree and parental species by near infrared hyperspectral imaging （NIR-HSI） spectroscopy

https://doi.org/10.21203/rs.3.rs-1993392/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Hybrid tuliptree (Liriodendron sion-americanum P.C. Yieh ex Shang et Z.R. Wang) has good wood quality and been widely used in furniture and other fields. Compared with parental species, hybrid tuliptree displayed faster growth rate, better wood quality and wider adaptability. But the morphological properties of hybrid tuliptree and parental species are highly similar, its identification has become an urgent problem. This study aims to evaluate the capabilities of near infrared hyperspectral imaging (NIR-HSI) technique to identify hybrid tuliptree, Chinese tuliptree, and tuliptree. For this purpose, the building models were carried out using Partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) after different data treatment. The results demonstrated that NIR-HSI spectroscopy combined with successive projections algorithm (SPA) treatment and SVM model can achieve the highest prediction accuracy (100%). This study shows that NIR-HSI technique has great potential in the genus Liriodendron identification and is suitable for classifying highly similar woods.

NIR-HSI

PLA-DA

SVM

wood identification

Hybrid tuliptree (Liriodendron sion-americanum P.C. Yieh ex Shang et Z.R. Wang) is a hybrid species obtained by crossbreeding Chinese tuliptree (Liriodendron chinese Sarg.) and tuliptree (Liriodendron tulipifera L.). Compared with parental species, hybrid tuliptree displayed faster growth rate and wider adaptability, and it can be used to make furniture, toys, plywood, and pulp. According to LY/T 2054 − 2012 “Methods for evaluating machining properties of lumber”, the hybrid tuliptree wood had good performances in sanding, planning, milling, lathing and tenoning(Ge et al., 2019a). In addition to the compressive strength, the other physical properties of hybrid tuliptree was superior to its parents(Ge et al., 2019b).

However, the morphological properties of hybrid tuliptree, Chinese tuliptree, and tuliptree are highly similar, it is difficult to distinguish the three woods by structural characteristics. Since hybrid tuliptree offers obvious advantages for wood quality and growth speed, timber traders often use Chinese tuliptree as hybrid tuliptree for profiteering(Ge et al., 2019b). There was no regulatory authority in China can provide reliable methods for hybrid tuliptree wood identification, which may lead to consumer rights abuses.

To maintain the order of the timber market and the interests of consumers, it is necessary to establish a fast and reliable method for hybrid tuliptree identification. Traditional wood identification methods were mainly carried out by the perspective of physics and anatomy. Despite the effectiveness of these methods, it has intrinsic disadvantages of low accuracy, high cost, and lack of speed(Ravindran et al., 2020). By contrast, optical and chemotaxonomical methods had proven useful in wood identification(Brunswick et al., 2021), extremely similar wood species can be distinguished by analyze the spectral characteristics.

To our knowledge, near-infrared spectroscopy (NIR) is an efficient, high-speed, non-destructive technology and have been widely used in analyzing woody materials(Dalmolin Ciarnoschi et al., 2022). Near infrared hyperspectral imaging (NIR-HSI) combines spectroscopy and imaging, enabling the acquisition of the entire spectrum at various locations in the image plane. Compared to NIR, NIR-HSI can be better detect the optical characteristic of samples, which enables both morphological and compositional inspection. Therefore, this technology is more suitable for the analysis of heterogeneous samples and allows more reliable qualitative identification using both spectral and spatial information(Xue et al., 2022).

NIR-HSI has proved to be a reliable method to analyze soils, water, food, seeds, and other samples(Tigabu and Odén, 2003, Tigabu and Odén, 2004, McLean et al., 2014, da Conceicao et al., 2021, Piarulli et al., 2021, Zhang et al., 2021, Zhao et al., 2021). Little research has focus on the analysis of highly similar materials, such as wood products taken from relative species. It is worthwhile to establish an effective method to identify highly similar woods, considering that NIR-HSI has more advantages than NIR spectroscopy.

In this study, cube wood samples of hybrid tuliptree, Chinese tuliptree and tuliptree were analyzed by NIR-HSI spectroscopy technique with SG (Savitzky-Golay) smoothing, SNV (Standard Normal Variate), Normalize, MSC (Multiple scattering correction), and successive projections algorithm (SPA) pre-processing methods, aims to establish a reliable approach for the classification of the species from the genus Liriodendron. Meanwhile, this study also evaluated the potential application of this technique to the analysis of highly similar materials.

2.1 Samples

Wood samples were taken from Xiashu Forest Farm in Jurong, Zhenjiang. The plantation was established in 1979 and each plot was assigned to Chinese tuliptree (three seed sources tested), tuliptree (two seed sources tested), and hybrid tuliptree (six combinations). Twelve trees of each species were randomly selected, and cube wood samples were collected at tree height of 1.3 m (totaling 60 trees). Five air-dried samples (moisture content of 11.5%-12%) were prepared for each tree with the dimensions of 100 mm x 100 mm x 100 mm (sap wood). According to our previous research, transverse section of wood is more suitable for wood identification(Xue et al., 2022). In this study, spectral data was scanned in transverse section of the samples.

We declare that this study has the official permission to collect the plant sample and complies with the Chinese legislation. Samples were stored in Nanjing Forestry University.

2.2 NIR-HSI spectra collecting

NIR-HSI spectra were collected using NIR-HSI spectrophotometer (ImSpectorV 10E), camera (R aptor EM285CL, UK) and 350 W halogen light source (Illumination Technologies, USA). The system was operated by IR CP0076 Software (Isuzn, Taiwan). The NIR-HSI analyses were performed in the spectral range from 982–2562 nm with a wavelength interval of 6.2 nm and a resolution of 10 nm; the distance between the light source and the camera lens were 20.5 cm and 30 cm respectively. In order to reduce the shadow generation and achieve high quality hyperspectral images, we aimed the light source at an angle of 45° to the samples and set the exposure time to 2.5 ms. Meanwhile, the delivery speed was set to 17.38 mm·s^− 1. After achieving the hyperspectral images of the wood samples, region of interest (ROI) was selected and calculated by ENVI 5.3 (ENVI Inc., USA; URL: https://www.envi.com) as the average reflection spectrum to use for model construction.

We performed hyperspectral image correction every 25 minutes to reduce the interference signals, which requires scanned white and black image. At the same wood sample image acquisition conditions, the white plate (99.9% reflectance Teflon plate) was scanned to get a white image, and the black image was obtained by covering the camera lens. All the collected sample spectral images were then converted to relative reflectance values according to the following equation:

$$R=\frac{{R}_{0}-B}{W-B}$$

Where R is the standardized light reflectance value, R₀ donate sample reference images, W donate white reference images and B donate dark image.

2.3 Data analysis

The datasets were processed by MATLAB R2018b (MathWorks Inc., Natick, MA, USA; URL: https://www.mathworks.com), with PLS_toolbox 802 (Eigenvector Research, Inc., Manson, WA, USA; URL: https://eigenvector.com).

NIR-HSI provides 256 spectral images at wavelengths from 982 to 2562 nm, but some bands may have noise that affects the modeling accuracy. In this study, wavelengths from 982 to 2186 nm were less suffered from interference, so they were selected for modelling. In this study, 200 samples were included in the training set, and the testing set included 100 samples.

Spectral data pretreatment had proved to be an effective method to improve modeling accuracy. In this study, the spectral data was subjected to different pre-processing, i.e., SPA, SG smoothing, MSC, Normalize, and SNV. PLS-DA and SVM methodology employed in this study allow for wood identification and cross-validated by Venetian blind-cross validation (5 segments). The principal components analysis (PCA) was used as an exploratory approach to examine outliers and to observe trends in clustering or separation of samples.

The value of sensitivity, specificity, and misclassification rate have been used to determine the performance of models. Sensitivity allows assessing how well the model can identify samples that belong to a particular class, and specificity measures the capacity of the model to reject nonbelonging samples. The misclassification rate is the ratio of false positives to the total number of samples. In this study, three statistical parameters were considered to evaluate the models.

3.1 Spectroscopic characterization

It is apparent from Fig. 1 that NIR-HSI spectra of tuliptree, Chinese tuilptree and hybrid tuliptree have several differences, presenting systematic variations of the baseline. The spectral region above 2025 nm did not present any important information and may affect the performance of models. However, the reflectance values have large difference within range of 982 ~ 1152 nm and 1435 ~ 1823 nm. It can be seen that all samples have important visible absorption peaks around 1111, 1305, 1670 and 1880 nm. The spectra of Chinese tuliptree and hybrid tuliptree were similar among the three wood, which may lead to the low classification accuracy.

Overall, only spectral alterations information can be provided by raw spectra curve since the overlapping and combination of bands, but a high prediction accuracy of wood species can be achieved by modeling method.

3.2 Principal components analysis

PCA was performed on NIR-HSI spectra dataset to visualize any variation among hybrid tuliptree, Chinese tuliptree, and tuliptree. PCA reduces the dimension of spectral data by projecting the spectral data onto a set of latent variables. In order to get a better separation, SNV pre-processing was used and showed better separation than other pre-processing methods. PCA scores of raw spectra and SNV pre-processed spectra are illustrated in Fig. 2.

In view of Fig. 2a, the most two principal components explained 89.3% of the total variance of the wood samples, with PC1 accounting for 71.0%, and PC2 accounting for 18.3% of the variance. After SNV pre-processing, the sum of PC1 and PC2 in the 3-dimensional scatter plot was 86.4% (72.0% for PC1 and 14.4% for PC2, Fig. 2b). This variability observed on the scores could be due to chemical bonds or the larger particle size of the samples, which could affect the spectral signal(Manley, 2014). Meanwhile, it is possible to choose wavelengths that suitable for wood classification by observing the PC loadings. The clusters of hybrid tuliptree, Chinese tuliptree, and tuliptree were partially overlapping. PCA was able to observe trends in clustering or separation of samples but could not accurately classify the three woods.

3.2 Sample classification

In this study, for the sake of suppressing the bad influence brought by noise, SG smoothing, SNV, Normalize, and MSC pre-processing were used to analyze the NIR-HSI spectra data.

SG smoothing is an effective spectral pre-processing method with a wide range of application and a variety of different modes (Steinier et al., 1972). The number of smoothing points is an important parameter. Setting too few smoothing points will cause errors, but if too many points are set, the spectral information will be lost. So unsuitable smoothing points number would decrease the accuracy of model, select reasonable number of points is necessary (Chen et al., 2011). In this study, 15 smoothing points were been selected for spectra data pre-processing. MSC is widely used as a spectral pre-processing method to remove imperfect data from the matrix before model construction. This method has two steps, including correction coefficients estimation and recorded spectrum correcting. Normalize and SNV are classic pre-processing method for scatter correction of NIR data, both methods do not involve a least-square fitting in their parameter estimation, they can be sensitive to noisy entries in the spectrum (Rinnan et al., 2009).

Table 1 summarized the data for model results and compared different pre-processing algorithm. It can be seen that the PLS-DA and SVM model performed well in hybrid tuliptree, Chinese tuliptree and tuliptree identification. PLS-DA model as a commonly used classification model has the advantage of less computational cost, it can find the best functional match for a set of data by minimizing the sum of squares of the errors(Lestander et al., 2008). It is apparent from Table 1 that PLS-DA combined with MSC and SNV pre-processing methods can achieve the highest prediction accuracy, while SG smoothing and Normalize treatment preformed relatively poor.

SVM as a flexible and powerful machine learning tool, can provide solutions for regression and classification problems. Compared with PLS-DA model, SVM model takes longer modeling time but has higher accuracy. Similar with PLS-DA model, SVM combined with MSC and SNV have high prediction accuracy. What is interesting about the SVM model results was Normalize pre-processing will significantly reduce the prediction accuracy.

NIR-HSI can provide spectral information over a large number of wavelengths for each sample. But in many cases, NIR-HSI instrument have information redundancy when getting data, which lead the increased workload and even models unreliable. Variable selection techniques can used to improve the parsimony and prediction ability of multivariate calibration models.

SPA as an effective data optimization algorithm that can extract characteristic bands from complex spectral data and reduce modeling calculation time. This method has attracted increasing interest in the analytical-chemistry community, this method was originally be used in Multiple Linear Regression (MLR) models (Soares et al., 2013). As shown in Fig. 3, fifteen optimal wavelengths were selected for wood samples. The value of RMSE was related to the number of optimal wavelengths (Fig. 4). A comparison of the results reveals that the SPA has high prediction accuracy while improving modeling speed. When processing the data after SPA, it can be observed that the other four pre-processing methods are no longer important. Overall, the accuracy of the SVM model is higher than PLS-DA model, regardless of whether it is processed by SPA.

Table 1

Summary of the prediction results of PLS-DA and SVM model.
	Pre-processing	Prediction accuracy
		PLS-DA	PLS-DA(SPA)	SVM	SVM(SPA)
Hybrid tuliptree	raw	98.67%	98.67%	100.00%	100.00%
	SG smoothing	98.67%	92.00%	100.00%	77.33%
	MSC	100.00%	97.33%	100.00%	98.67%
	Normalize	98.67%	98.67%	24.00%	54.67%
	SNV	100.00%	97.33%	100.00%	100.00%
Chinese tuliptree	raw	93.33%	94.67%	98.67%	100.00%
	SG smoothing	92.00%	90.67%	98.67%	96.00%
	MSC	98.67%	93.33%	100.00%	96.00%
	Normalize	92.00%	92.00%	46.67%	73.33%
	SNV	98.67%	93.33%	100.00%	98.67%
Tuliptree	raw	96.00%	100.00%	98.67%	100.00%
	SG smoothing	94.67%	96.00%	98.67%	69.33%
	MSC	96.00%	98.67%	100.00%	100.00%
	Normalize	96.00%	100.00%	37.33%	85.33%
	SNV	96.00%	98.67%	100.00%	100.00%

3.3 Mixing matrix

Mixing matrix can clearly reflect the prediction accuracy of the model. Figure 5 shows the detailed classification of hybrid tuliptree and parental species. It can be observed that SVM model has higher classification accuracy, especially for hybrid tuliptree. Overall, wood samples of Chinese tuliptree and hybrid tuliptree were easily confused, while tuliptree was easily identified, it may relate to the spectral characteristics of the samples. As shown in Fig. 5, SVM performs better in three tree species classifying, and PLS-DA performed relatively worse. Overall, it can be concluded that NIR-HSI is a robust and fast technique in separating the 3 woods in the focus of this study.

3.3 Statistical parameters of models

The sensitivity, specificity and misclassification rate given in Table 2 and Table 3 were used as a measure of the classification performance of the PLS-DA and SVM model. The calculation results shown that two models have good predictive ability. In particular, the sensitivity values demonstrated that PLS-DA model after SPA pre-processing was able to correctly identify the samples (97.80% for the training set, 97.33% for the testing set), and the specificity shows that the model does not misclassify multiple times (97.77% for the training set, 94.67% for the testing set).

Table 3 shows that SVM model allows to get better results in classification: 100% sensitivity, 100 specificity and 0% misclassification rate. The same statistical parameters indicated that SPA could simplify the model by reducing the number of the wavelengths but did not reduce the prediction accuracy.

Table 2

Summary of the quality statistical parameters of PLS-DA model.
	Pre-processing	Tests	Sensitivity (%)	Specificity (%)	Misclassification Rate (%)
PLS-DA	raw	Training	95.10	97.77	4.00
	raw	Testing	97.33	98.00	2.67
	SG smoothing	Training	95.10	96.67	4.44
	SG smoothing	Testing	97.33	97.33	4.00
	MSC	Training	98.23	99.77	1.78
	MSC	Testing	97.80	99.33	0.00
	Normalize	Training	95.10	97.80	4.00
	Normalize	Testing	94.67	97.13	4.00
	SNV	Training	98.23	99.77	1.78
	SNV	Testing	97.80	99.33	0.00
PLS-DA (SPA)	raw	Training	97.80	97.77	2.22
	raw	Testing	97.33	94.67	2.67
	SG smoothing	Training	94.23	94.00	6.22
	SG smoothing	Testing	93.33	94.67	6.67
	MSC	Training	97.33	97.33	2.67
	MSC	Testing	100.00	98.00	1.33
	Normalize	Training	98.23	98.00	3.11
	Normalize	Testing	98.67	96.00	2.67
	SNV	Training	97.33	97.33	2.67
	SNV	Testing	100.00	98.00	1.33

Table 3

Summary of the quality statistical parameters of SVM model.
	Pre-processing	Tests	Sensitivity (%)	Specificity (%)	Misclassification Rate (%)
SVM	raw	Training	100.00	100.00	0.00
	raw	Testing	100.00	100.00	0.00
	SG smoothing	Training	100.00	100.00	0.00
	SG smoothing	Testing	100.00	100.00	0.00
	MSC	Training	100.00	100.00	0.00
	MSC	Testing	100.00	100.00	0.00
	Normalize	Training	33.33	66.67	66.67
	Normalize	Testing	33.33	66.67	66.67
	SNV	Training	100.00	100.00	0.00
	SNV	Testing	100.00	100.00	0.00
SVM (SPA)	raw	Training	100.00	100.00	0.00
	raw	Testing	100.00	100.00	0.00
	SG smoothing	Training	83.57	91.77	16.44
	SG smoothing	Testing	84.00	92.00	16.00
	MSC	Training	98.23	99.10	1.78
	MSC	Testing	100.00	100.00	0.00
	Normalize	Training	78.70	89.33	21.33
	Normalize	Testing	77.33	88.67	22.67
	SNV	Training	100.00	100.00	0.00
	SNV	Testing	100.00	100.00	0.00

It is difficult to identify hybrid tree and parent tree by traditional methods because they have similar microstructure and chemical composition. Hybrid tuliptree and parental species have high similarity in cell structure, fiber morphology and volatile components, high professional knowledge is required in wood identification. In this study, the identification of hybrid tuliptree, Chinese tuliptree, and tuliptree was performed according to NIR-HSI method and obtained reliable results. NIR-HSI spectroscopy combined with SPA treatment and SVM models was confirmed to be a fast, robust, and non-destructive method for classify woods of genus Liriodendron according to the rate of sensitivity (100%), specificity (100%), and misclassification rate (0%). It is also noticeable that the SPA treatment has proved to be a reliable method to improve model performance. The real advantage of this technique is the possible to develop a dedicated spectrophotometer device or a faster portable device, due to the low number of optimal wavelengths. These devices can improve the accuracy of affinity trees identification and reduce the detection cost significantly.

Summarizing, NIR-HSI techniques can be used in official inspection institutions, supervision departments, and timber merchants for wood identification purposes. The main strength of this technique is the cheap, rapid, and non-destructive advantages in which the analysis can be performed to classify highly similar materials. Transfer from lab to large-scale application needs further research into the effects of moisture, temperature, and wood surface variations on identification accuracy. Meanwhile, since NIR-HSI requires stronger judgement to process optical information compared to NIR, more attention should be focused on how to efficiently extract information from samples.

Statement

On behalf of all authors, I declare that our research has no conflict of interest.

Additional information

The datasets generated and analyzed during the current study are available in the Figshare repository, [https://doi.org/10.6084/m9.figshare.20153645.v1]

BRUNSWICK, P., CUTHBERTSON, D., YAN, J., CHUA, C. C., DUCHESNE, I., ISABEL, N., EVANS, P. D., GASSON, P., KITE, G., BRUNO, J., VAN AGGELEN, G. & SHANG, D. 2021. A practical study of CITES wood species identification by untargeted DART/QTOF, GC/QTOF and LC/QTOF together with machine learning processes and statistical analysis. Environmental Advances, 5.10.1016/j.envadv.2021.100089
CHEN, H., PAN, T., CHEN, J. & LU, Q. 2011. Waveband selection for NIR spectroscopy analysis of soil organic matter based on SG smoothing and MWPLS methods. Chemometrics and Intelligent Laboratory Systems, 107, 139-146.10.1016/j.chemolab.2011.02.008
DA CONCEICAO, R. R. P., SIMEONE, M. L. F., QUEIROZ, V. A. V., DE MEDEIROS, E. P., DE ARAUJO, J. B., COUTINHO, W. M., DA SILVA, D. D., DE ARAUJO MIGUEL, R., DE PAULA LANA, U. G. & DE RESENDE STOIANOFF, M. A. 2021. Application of near-infrared hyperspectral (NIR) images combined with multivariate image analysis in the differentiation of two mycotoxicogenic Fusarium species associated with maize. Food Chem, 344, 128615.10.1016/j.foodchem.2020.128615
DALMOLIN CIARNOSCHI, L., CLAUDIO DE OLIVEIRA, L., LUCIA FERREIRA SIMEONE, M., DOS SANTOS PANERO, F., DOS SANTOS PANERO, P., RUIZ RODRIGUEZ, A., ALVES FILHO, E. G., GERVASIO PEREIRA, M. & MANOEL DA SILVA, L. 2022. Prediction of dry matter, carbon and ash contents and identification of Calycophyllum spruceanum (Benth) organs by Near-Infrared spectrophotometry. Microchemical Journal, 180.10.1016/j.microc.2022.107621
GE, X., W., WU, Z., H. & HUANG, Q., T. 2019a. Study on the use and promotion of Liriodendron sino-americanum and Liriodendron tulipifera Linn. Furniture, 40(02), 22-27.10.16610/j.cnki.jiaju.2019.02.005
GE, X. W., WU, Z. H. & HUANG, Q. T. 2019b. Study on Physical and Mechanical Properties of Liriodendron Sino-americanum and Its Adaptability for Solid Wood Furniture. China Forest Products Industry, 46(07), 27-29.10.19531/j.issn1001-5299.201907006
LESTANDER, T. A., LINDEBERG, J., ERIKSSON, D. & BERGSTEN, U. 2008. Prediction of Pinus sylvestris clear-wood properties using NIR spectroscopy and biorthogonal partial least squares regression. Canadian journal of forest research, 38, 2052–2062
MANLEY, M. 2014. Near-infrared spectroscopy and hyperspectral imaging: non-destructive analysis of biological materials. Chem Soc Rev, 43, 8200 – 14.10.1039/c4cs00062e
MCLEAN, J. P., JIN, G., BRENNAN, M., NIEUWOUDT, M. K. & HARRIS, P. J. 2014. Using NIR and ATR-FTIR spectroscopy to rapidly detect compression wood in Pinus radiata. Canadian Journal of Forest Research, 44, 820–830
PIARULLI, S., MALEGORI, C., GRASSELLI, F., AIROLDI, L., PRATI, S., MAZZEO, R., SCIUTTO, G. & OLIVERI, P. 2021. An effective strategy for the monitoring of microplastics in complex aquatic matrices: Exploiting the potential of near infrared hyperspectral imaging (NIR-HSI). Chemosphere, 286, 131861.10.1016/j.chemosphere.2021.131861
RAVINDRAN, P., THOMPSON, B. J., SOARES, R. K. & WIEDENHOEFT, A. C. 2020. The XyloTron: Flexible, Open-Source, Image-Based Macroscopic Field Identification of Wood Products. Front Plant Sci, 11, 1015.10.3389/fpls.2020.01015
RINNAN, Å., BERG, F. V. D. & ENGELSEN, S. B. 2009. Review of the most common pre-processing techniques for near-infrared spectra. TrAC Trends in Analytical Chemistry, 28, 1201-1222.10.1016/j.trac.2009.07.007
SOARES, S. F. C., GOMES, A. A., ARAUJO, M. C. U., FILHO, A. R. G. & GALVÃO, R. K. H. 2013. The successive projections algorithm. TrAC Trends in Analytical Chemistry, 42, 84-98.10.1016/j.trac.2012.09.006
STEINIER, J., TERMONIA, Y. & DELTOUR, J. 1972. Smoothing and differentiation of data by simplified least square procedure. Anal Chem, 44, 1906 – 9.10.1021/ac60319a045
TIGABU, M. & ODÉN, P. C. 2003. Discrimination of viable and empty seeds of Pinus patula Schiede & Deppe with near-infrared spectroscopy. New Forests, 25, 163-176.10.1023/a:1022916615477
TIGABU, M. & ODÉN, P. C. 2004. Simultaneous detection of filled, empty and insect-infested seeds of three Larix species with single seed near-infrared transmittance spectroscopy. New Forests, 27, 39-53.10.1023/A:1025037408922
XUE, X., CHEN, Z., WU, H. & GAO, H. 2022. Identification of Guiboutia species by NIR-HSI spectroscopy. Sci Rep, 12, 11507.10.1038/s41598-022-15719-0
ZHANG, L., SUN, H., LI, H., RAO, Z. & JI, H. 2021. Identification of rice-weevil (Sitophilus oryzae L.) damaged wheat kernels using multi-angle NIR hyperspectral data. Journal of Cereal Science, 101.10.1016/j.jcs.2021.103313
ZHAO, S., QIU, Z. & HE, Y. 2021. Transfer learning strategy for plastic pollution detection in soil: Calibration transfer from high-throughput HSI system to NIR sensor. Chemosphere, 272.10.1016/j.chemosphere.2021.129908

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of hybrid tuliptree and parental species by near infrared hyperspectral imaging （NIR-HSI） spectroscopy

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials And Methods

2.1 Samples

2.2 NIR-HSI spectra collecting

2.3 Data analysis

3. Results And Discussion

3.1 Spectroscopic characterization

3.2 Principal components analysis

3.2 Sample classification

3.3 Mixing matrix

3.3 Statistical parameters of models

4. Conclusions

Declaration

References

Additional Declarations

Status:

Version 1