Near-infrared Hyperspectral Imaging Data Research Articles

Cognitive spectroscopy for the classification of rice varieties: A comparison of machine learning and deep learning approaches in analysing long-wave near-infrared hyperspectral images of brown and milled samples

Rapid and non-destructive detection of genuine Thai Jasmine rice (Khao Dawk Mali 105 (KDML105)) from Pathum Thani1 (PTT1) and Phitsanulok2 (PSL2) under either milled or brown conditions is required to disrupt fraudulent. This study aimed to resolve this real issue using long-wave near infrared hyperspectral imaging (NIR-HSI) coupled with machine learning and deep learning approaches. The best classification accuracy for the milled rice was achieved using the spectral imaging-based analysis on the NIR-HSI data with selected wavelength, approximately 95% for the test set either by convolutional neural network or support vector machine (SVM), whereas for the brown rice, the SVM model based on the averaged NIR spectra could achieve the best classification accuracy of 95.4%. It suggests the chemical component difference and its spatial distribution in the milled rice could contribute higher classification accuracy. Additionally, the surface bran effects of brown rice could be reduced by using averaged spectral data coupled with the SVM method.

Wavelength Bands Reduction Method in Near-Infrared Hyperspectral Image based on Deep Neural Network for Tumor Lesion Classification

In this paper, we propose a method for wavelength bands reduction of near-infrared (NIR) hyperspectral imaging data to extract the cancer region with minimum input data. NIR hyperspectral imaging data has a spectrum data of each pixel and is suitable for distinguishing tumors region of the body rather than RGB imaging data. However, it is difficult to applicate to the medical field because of processing time consumption and hardware size limitation. Therefore, it is necessary to remove the redundant wavelength bands which are not (or little) contributed to tumor region extraction. Although several previous studies for wavelength bands reduction have been conducted, these approaches focused on the characteristics of the wavelength itself. In this research, the proposed wavelength bands reduction method is focused on the node weights of the post-training deep neural network which is an indicator directly related to classification. The experimental results using GIST specimen demonstrated that four wavelength bands selected from all wavelengths bands by using the proposed method are effective for tumor distinguishing as well as all wavelength bands.

Rapid and non-destructive seed viability prediction using near-infrared hyperspectral imaging coupled with a deep learning approach

Seeds are the basis of the agricultural food industry, greater insights into seed viability before sowing could improve storage management and field performance. In the present study, we aimed to address this issue by using highly cost-efficient near-infrared hyperspectral imaging (NIR-HSI) and a convolutional neural network (CNN) deep learning approach. An NIR-HSI camera was used because it can recognize both molecular vibration information (i.e. chemical component differences) and its spatial distribution in each seed sample; this camera is much more informative than a regular RGB digital camera. Using this technology, the emphasis of this study was firstly to provide a methodology for enhancing the interpretability of viable and non-viable seeds via principal component analysis (PCA) and support vector machine (SVM) viability classification analysis of NIR-HSI data. A CNN was then constructed to“cognize” the differences in viable and non-inviable seeds and classify them automatically. Experimental results indicate that the methodology produces a ~90% classification accuracy for both a five-fold cross-validation set and a test set of naturally aged Japanese mustard spinach seeds. Therefore, this study provides a new strategy for effective and practical seed viability prediction.

Evaluation of near infrared hyperspectral imaging for detection of tuna powder contaminated with shrimp powder

Near-infrared hyperspectral imaging (NIR-HSI) was investigated to detect the contamination of shrimp powder (SP) in tuna powder (TP) with partial least squares regression (PLSR) model. The principal component analysis was performed with NIR-HSI data for classification of tuna and shrimp powder. Samples for NIR-HSI data analysis were prepared using tuna powder contaminated with shrimp powder in concentration of 0%, 0.01%, 0.05%, 0.1%, 0.5%, 1%, 5%, 10%, 25%, 50%, 75% and 100% (w/w). The NIR-HIS in a wavelength range 864.5 to 1695.1 nm of the samples were used to create a prediction model using a partial least squares regression (PLSR) model. The result showed that the best model was based on spectra pretreated wit second derivative combined with standard normal variate pretreatments, The performance of the prediction was expressed with the following values; factor = 3, Rc2 = 0.989, RMSEC = 3.48%, Rcv2 = 0.984, RMSECV = 4.218%, Rp2 = 0.991, RMSEP = 3.110%. The regression coefficients of the PLSR model from 2D+SNV spectral pre-treatments were used to identify functional groups from the chemical composition of each sample. The study demonstrated that the NIR-HSI can be used for quantitative analysis of TP contaminated with SP which rapid nondestruction technique.

Study of modeling optimization for hyperspectral imaging quantitative determination of naringin content in pomelo peel

With the development of modern computational science and data metrology, hyperspectral imaging technology has been utilized in the field of remote sensing for the application in precision agriculture and crop quality testing. In this paper, the near infrared hyperspectral imaging (NIRHSI) technique was applied to quantitatively determine the naringin content in pomelo peel samples. Model optimization was studied by investigating the influence of system measuring conditions and parameters of the built-up HSI instrumental system. The HSI data acquisition was built up using a prevalent pushbroom scanner in the near infrared region. Calibration models were established using the partial least squares (PLS) regression in the mode of cross validation, in combined use of the Savitzky-Golay smoother (SGS) for data pretreatment. These multivariate analytical models were optimized in comparison about the region of interest (ROI) for NIRHSI model optimization. In the process of NIRHSI data acquisition, the rational values of some system measurement parameters were also tuned and tested, such as the use of different watts for light intensity, different lenses and different materials as the scanning backgrounds. Results showed that the cross-validation PLS regression methods performed well in the calibration and prediction processes, working well together with the parameter tuning of the SGS pretreatment. In addition to the fact that different materials as the scanning backgrounds obviously affected the quantitative result, there is no apparent difference in the comparing cases of different light intensities and different lenses. This work validates the capability of applying NIRHSI technique to quantitative determine the content of naringin in pomelo peel samples. The test for model optimization by comparing the measurement parameters and the system properties has prospective application ability in fields of other spectral/hyperspectral data analysis. It is an important lab simulation of remote sensing. It contributes significant theoretical reference to the design of the large-scale online hyperspectral data acquisition systems.

Prediction of Douglas-Fir Lumber Properties: Comparison between a Benchtop Near-Infrared Spectrometer and Hyperspectral Imaging System

Near-infrared (NIR) spectroscopy and NIR hyperspectral imaging (NIR-HSI) were compared for the rapid estimation of physical and mechanical properties of No. 2 visual grade 2 × 4 (38.1 mm by 88.9 mm) Douglas-fir structural lumber. In total, 390 lumber samples were acquired from four mills in North America and destructively tested through bending. From each piece of lumber, a 25-mm length block was cut to collect diffuse reflectance NIR spectra and hyperspectral images. Calibrations for the specific gravity (SG) of both the lumber (SGlumber) and 25-mm block (SGblock) and the lumber modulus of elasticity (MOE) and modulus of rupture (MOR) were created using partial least squares (PLS) regression and their performance checked with a prediction set. The strongest calibrations were based on NIR spectra; however, the NIR-HSI data provided stronger predictions for all properties. In terms of fit statistics, SGblock gave the best results, followed by SGlumber, MOE, and MOR. The NIR-HSI SGlumber, MOE, and MOR calibrations were used to predict these properties for each pixel across the transverse surface of the scanned samples, allowing SG, MOE, and MOR variation within and among rings to be observed.

Revealing the location of semen, vaginal fluid and urine in stained evidence through near infrared chemical imaging

Crime scene investigation (CSI) requires the ultimate available technology for a rapid, non-destructive, and accurate detection of a wide variety of evidence including invisible stains of bodily fluids. Particularly crucial is the discrimination of semen in stained evidence from sexual abuse cases. This is because those evidence have high odds of containing the DNA from the aggressor. To this aim, we demonstrated the potential of near infrared hyperspectral imaging (NIR-HSI) to make visible stains of semen, vaginal fluid, and urine on fabrics, which lays the bases to face the challenging visualization and discrimination of semen within bodily fluids mixtures. Combining the NIR-HSI data and simple chemometric techniques such as principal component analysis and classical least squares regression, we have revealed the location of semen, vaginal fluid and urine in bodily fluids stained evidence.

Near-infrared hyperspectral imaging combined with CARS algorithm to quantitatively determine soluble solids content in "Ya" pear

The present study proposed competitive adaptive reweighted sampling (CARS) algorithm to be used to select the key variables from near-infrared hyperspectral imaging data of "Ya" pear. The performance of the developed model was evaluated in terms of the coefficient of determination(r2), and the root mean square error of prediction (RMSEP) and the ratio (RPD) of standard deviation of the validation set to standard error of prediction were used to evaluate the performance of proposed model in the prediction process. The selected key variables were used to build the PLS model, called CARS-PLS model. Comparing results obtained from CARS-PLS model and results obtained from full spectra PLS, it was found that the better results (r(2)pre = 0. 908 2, RMSEP=0. 312 0 and RPD=3. 300 5) were obtained by CARS-PLS model based on only 15. 6% information of full spectra. Moreover, performance of CARS-PLS model was also compared with PLS models built by using variables got by Monte Carlo-uninformative variable elimination (MC-UVE) and genetic algorithms (GA) method. The result found that CARS variable selection algorithm not only can remove the uninformative variables in spectra, but also can reduce the collinear variables from informative variables. Therefore, this method can be used to select the key variables of near-infrared hyperspectral imaging data. This study showed that near-infrared hyperspectral imaging technology combined with CARS-PLS model can quantitatively predict the soluble solids content (SSC) in "Ya" pear. The results presented from this study can provide a reference for predicting other fruits quality by using the near-infrared hyperspectral imaging.

Influence of grain topography on near infrared hyperspectral images

Near infrared hyperspectral imaging (NIR-HSI) allows spatially resolved spectral information to be collected without sample destruction. Although NIR-HSI is suitable for a broad range of samples, sizes and shapes, topography of a sample affects the quality of near infrared (NIR) measurements. Single whole kernels of three cereals (barley, wheat and sorghum), with varying topographic complexity, were examined using NIR-HSI. The influence of topography (sample shape and texture) on spectral variation was examined using principal component analysis (PCA) and classification gradients. The greatest source of variation for all three grain types, despite spectral preprocessing with standard normal variate (SNV) transformation, was kernel curvature. Only 1.29% (PC5), 0.59% (PC6) and 1.36% (PC5) of the spectral variation within the respective barley, wheat and sorghum image datasets was explained within the principal component (PC) associated with the chemical change of interest (loss of kernel viability). The prior PCs explained an accumulated total of 91.18%, 89.43% and 84.39% of spectral variance, and all were influenced by kernel topography. Variation in sample shape and texture relative to the chemical change of interest is an important consideration prior to the analysis of NIR-HSI data for non-flat objects.