Spectral Pre-processing Techniques Research Articles

Improving the quantification of sediment source contributions using different mathematical models and spectral preprocessing techniques for individual or combined spectra of ultraviolet–visible, near- and middle-infrared spectroscopy

In recent years, several sediment fingerprinting studies have used ultraviolet–visible (UV–Vis), near-infrared (NIR) and middle-infrared (MIR) spectroscopy as a low cost, non-destructive and fast alternative to obtain tracer properties to estimate sediment source contributions. For this purpose, partial last square regression (PLSR) has often been used to build predictive parametric models. However, spectra preprocessing and more robust and non-parametric models such as support vector machines (SVM) has gained little attention in these studies. Accordingly, the objectives of the current research were to evaluate (i) the accuracy of two multivariate methods (PLSR and SVM), (ii) the effect of eight spectra preprocessing techniques, and (iii) the effect of using the information contained in the UV–Vis, NIR and MIR regions considered either separately or in combination on sediment source apportionment. The estimated source contribution was then compared with contributions obtained by the conventional fingerprinting approach based on geochemical tracers. This study was carried out in the Arvorezinha catchment (1.23 km2) located in southern Brazil. Forty soil samples were collected in three main potential source (cropland surface, unpaved roads and stream channels) and twenty-nine suspended sediment samples collected at the catchment outlet during nine rainfall-runoff events were used in this study. Both PLSR and SVM models showed a higher accuracy when calibrated and validated with the spectra submitted to spectral processing when compared to the direct use of the raw spectra. The best model results were obtained with PLSR and SVM mathematical models associated with the spectral preprocessing techniques 1st derivative Savitzky-Golay (SGD1), normalization (NOR) and combining NOR + SGD1 in the UV–Vis + NIR + MIR. The lowest errors were observed when the UV–Vis + NIR + MIR bands were combined due to the gain in information and, consequently, the increase in discriminant power achieved by the models. Despite the good accuracy of the models calibrated and validated with the artificial mixtures, significant errors remain when results of source contributions are compared to those obtained with the conventional sediment fingerprinting technique based on geochemical tracers. Nevertheless, the magnitude of the contributions calculated by the spectroscopy and geochemical approaches remains very similar for all sources, especially when using the SVM-UV–Vis + NIR + MIR model. Therefore, spectroscopy proved to be a fast, cheap and accurate technique, offering an alternative to the conventional geochemical approach for discriminating sediment source contributions in agricultural catchments located in subtropical regions.

Application of VIS-NIR spectroscopy for estimation of soil organic carbon using different spectral preprocessing techniques and multivariate methods in the middle Indo-Gangetic plains of India

Soil organic carbon (SOC) is the main source of soil nutrients, which are essential for the growth and development of agricultural crops. Precise and quick quantification of SOC is of utmost importance in crop husbandry and soil health/carbon sequestration quantification. In order to evaluate visible and near-infrared diffuse reflectance spectroscopy (VIS-NIR) as an alternative to precise and quick method of quantification of SOC in the Indo-Gangetic plains, 280 soil samples were collected covering Inceptisols, Entisols and Alfisols and their spectra recorded. Six preprocessing techniques ((reflectance, absorbance, multiplicative scatter correction (MSC), standard normal variate (SNV), Savitzky–Golay smoothing first derivative (SG-FD) and Savitzky–Golay smoothing second derivative (SG-SD)) and four multivariate methods (partial least-squares regression (PLSR), random forest (RF), support vector regression (SVR) and multivariate adaptive regression splines (MARS)) were evaluated to predict SOC from VIS-NIR spectra. The considerable prediction accuracy and robustness were achieved using the PLSR model (RV2 = 0.73, RMSEV = 0.07, and RPDV = 1.90), RF model (RV2 = 0.69, RMSEV = 0.07, and RPDV = 1.74), SVR model (RV2 = 0.57, RMSEV = 0.08), and RPDV = 1.50), and MARS model (RV2 = 0.63, RMSEV = 0.10, and RPDV = 1.05). Findings from this study identified the reliability of SOC determinations by examining how preprocessing techniques and multivariate methods affect spectral analyses.

Discrimination of common defects in loquat fruit cv. ‘Algerie’ using hyperspectral imaging and machine learning techniques

Abstract Loquat (Eriobotrya japonica L.) is an important fruit for the economy of some regions of Spain that is very susceptible to mechanical damage and physiological disorders. These problems depreciate its value and prevent it from being exported. Visible (VIS) and near infrared (NIR) hyperspectral imaging was used to discriminate between external and internal common defects of loquat cv. ‘Algerie’. Two classifiers, random forest (RF) and extreme gradient boost (XGBoost), and different spectral pre-processing techniques were evaluated in terms of their capacity to distinguish between sound and defective features according to three approaches. In the first approach the fruit pixels were classified into two classes, sound or defect, with a 97.5% rate of success; in the second the defective features were considered internal or external defects, achieving a 96.7% rate of success; and in the third approach each type of defect, i.e. purple spot, bruising, scars and flesh browning, were considered separately with a correct classification rate of 95.9%. The results indicated that the XGBoost classifier was the best method in all cases.

Deep Spectral-Spatial Features of Near Infrared Hyperspectral Images for Pixel-Wise Classification of Food Products.

Hyperspectral imaging (HSI) emerges as a non-destructive and rapid analytical tool for assessing food quality, safety, and authenticity. This work aims to investigate the potential of combining the spectral and spatial features of HSI data with the aid of deep learning approach for the pixel-wise classification of food products. We applied two strategies for extracting spatial-spectral features: (1) directly applying three-dimensional convolution neural network (3-D CNN) model; (2) first performing principal component analysis (PCA) and then developing 2-D CNN model from the first few PCs. These two methods were compared in terms of efficiency and accuracy, exemplified through two case studies, i.e., classification of four sweet products and differentiation between white stripe (“myocommata”) and red muscle (“myotome”) pixels on salmon fillets. Results showed that combining spectral-spatial features significantly enhanced the overall accuracy for sweet dataset, compared to partial least square discriminant analysis (PLSDA) and support vector machine (SVM). Results also demonstrated that spectral pre-processing techniques prior to CNN model development can enhance the classification performance. This work will open the door for more research in the area of practical applications in food industry.

Quantification of phenolic acids by partial least squares Fourier-transform infrared (PLS-FTIR) in extracts of medicinal plants.

Phenolic compounds are ubiquitous compounds found in all plants as their secondary metabolites. Phenols are becoming increasingly important particularly because of their beneficial effects on health. To provide a faithful calibration model for the simultaneous determination and quantification of phenolic acids, as salicylic, vanillic, p-hydroxybenzoic acids, eugenol and thymol in different extracts of medicinal plants, a comparative study was made between two methods of infrared measurements based on attenuated total reflectance (ATR) and transmission. Characteristic absorbance peak heights of mid-infrared spectra of individual phenolic acids were measured for the compounds. For partial least squares regression (PLS-R) calibration mixtures of phenolic acids, wavenumber ranges, spectra pretreatment and number of latent variables, were assayed to improve the prediction capability of models using different spectral preprocessing techniques after mean centring of infrared data. Plant extracts were prepared by using water/methanol and ethanolic extraction solvents followed by Fourier-transform infrared (FTIR)-spectrometry analysis. The concentrations of phenolic compounds contained in the extracts were obtained by using the best models selected of the PLS calibration. PLS-ATR-mid-infrared (MIR) measurement provided the most accurate results and offers a good methodology for the determination of phenolic acids. The analysis showed that the rate of phenolic acids and monoterpenic phenols in extracts of medicinal plants is in the same range obtained with the Folin-Ciocalteu method, which confirm that the developed method using PLS is therefore, highly specific and selective. The simultaneous direct quantification of various phenolic acids in different plant extracts was possible with a fast and simple methodology based on PLS-ATR-FTIR analysis.

SPORT pre-processing can improve near-infrared quality prediction models for fresh fruits and agro-materials

Abstract Near-infrared spectroscopy (NIRS) is a key non-destructive technique for rapid assessment of the chemical properties of food materials. However, a major challenge with NIRS is the mixed physicochemical phenomena captured by the interaction of the light with the matter. The interaction often results in both absorption and scattering of the light. The overall NIRS signal therefore contains information related to the two phenomena mixed. To predict chemical properties such as dry matter, Brix and lipids, light refelction/absorption is used. Therefore, when the aim of the data analysis is to predict chemical components, it is necessary to remove as much as possible the scattering effects from the spectra. Several pre-processing techniques are available to do this, but it is often difficult to decide which one to choose. In this article we present the use of a recently developed pre-processing approach, sequential pre-processing through orthogonalization (SPORT), to improve the predictive power of multivariate models based on NIR spectra of food materials. The SPORT approach utilizes sequential orthogonalized partial least square regression (SOPLS) for the fusion of data blocks corresponding to several spectral preprocessing techniques. The results were compared with commonly used pre-processing techniques in the analysis of food materials by NIRS. The comparison was made by analyzing 5 different datasets comprised of apples, apricots, olive oils and grapes associated with chemical properties such as dry matter (DM), Brix, lipids and citric acid. The datasets were from both reflection and transmission measurements. The results showed that the fusion-based pre-processing methodology is an ideal choice for pre-processing of NIRS data. For four out of five datasets, the prediction accuracies (high R2pred and low RMSEP) were improved. The improvement led to as much as a 20 % increase in R2pred and a 25 % decrease in RMSEP compared to the standard 2nd derivative pre-processing. The pre-processing fusion was more effective for the reflection mode compared to the transmission mode. Multiple pre-processing techniques provided complementary information, and therefore, their fusion using the SPORT approach improved the model performance. The methodology is not only applicable to food materials but can in fact be used as a general pre-processing approach for all types of modeling of spectral data.

Novel Application of Near-infrared Spectroscopy and Chemometrics Approach for Detection of Lime Juice Adulteration.

The aim of this study is to investigate the novel application of a ‎handheld near infra-red spectrophotometer coupled with classification methodologies as a screening approach in detection of adulterated lime juices. For this purpose, a miniaturized near infra-red spectrophotometer (Tellspec®) in the spectral range of 900–1700 nm was used. Three diffuse reflectance spectra of 31 pure lime juices were collected from Jahrom, Iran and 25 adulterated juices were acquired. Principal component analysis was almost able to generate two clusters. Partial least square discriminant analysis and k-nearest neighbors algorithms with different spectral preprocessing techniques were applied as predictive models. In the partial least squares discriminant analysis, the most accurate prediction was obtained with SNV transforming. The generated model was able to classify juices with an accuracy of 88% and the Matthew’s correlation ‎coefficient ‎value of 0.75 in the external validation set. In the k-NN model, the highest accuracy and Matthew’s correlation ‎coefficient in the test set (88% and 0.76, respectively) was obtained with multiplicative signal correction followed by 2nd-order derivative and 5th nearest neighbor. The results of this preliminary study provided promising evidence of the potential of the handheld near infra-red spectrometer and machine learning methods for rapid detection of lime juice adulteration. Since a limited number of the samples were used in the current study, more lime juice samples from a wider range of variability need to be analyzed in order to increase the robustness of the generated models and to confirm the promising results achieved in this study.

Measuring Internal Maturity Parameters Contactless on Intact Table Grape Bunches Using NIR Spectroscopy.

The determination of internal maturity parameters of table grape is usually done destructively using manual methods that are time-consuming. The possibility was investigated to determine whether key fruit attributes, namely, total soluble solids (TSS); titratable acidity (TA), TSS/TA, pH, and BrimA (TSS – k x TA) could be determined on intact table grape bunches using Fourier transform near-infrared (FT-NIR) spectroscopy and a contactless measurement mode. Partial Least Squares (PLS) regression models were developed for the maturity and sensory quality parameters using grapes obtained from two consecutive harvest seasons. Statistical indicators used to evaluate the models were the number of latent variables (LVs) used to build the model, the prediction correlation coefficient (R2p) and root mean square error of prediction (RMSEP). For the respective parameters TSS, TA, TSS/TA, pH, and BrimA, the LVs were 21, 23, 5, 7, and 24, the R2p = 0.71, 0.33, 0.57, 0.28, and 0.77, and the RMSEP = 1.52, 1.09, 7.83, 0.14, and 1.80. TSS performed best when moving smoothing windows (MSW) + multiplicative scatter correction (MSC) was used as spectral pre-processing technique, TA with standard normal variate (SNV), TSS/TA with Savitzky-Golay first derivative (SG1d), pH with SG1d, and BrimA with MSC. This study provides the first steps towards a completely nondestructive and contactless determination of internal maturity parameters of intact table grape bunches.

Improvement of proximate data and calorific value assessment of bamboo through near infrared wood chips acquisition

In a previous study, a near infrared (NIR) spectroscopy model was developed using a spectra of ground bamboo samples. Although the previous report on ground bamboo described a good performance, operation in the power plant was found to be inconvenient due to preparation costs and labour required for the necessary preparation of ground samples. Thus, this study presents the comparison of the performance of an NIR model that was developed by direct scanning of bamboo chips to the previously developed model for ground samples. Special emphasis is put on the comparison of the spectral reproducibility. Bamboo chip models were developed based on PLS regression with variable selection methods in order to achieve the optimal model. The moisture content (MC) and ash content (A) of the developed bamboo chip models could be applied toward quality assurance. The volatile matter (VM) and fixed carbon (FC) models could be used for approximating predictions. The gross and net calorific value (GCV and NCV) models could be used for most applications. The root mean square (RMS) value of pre-treated spectra of different particle size had no statistically significant differences. The study’s findings indicate that the model developed using NIR spectroscopy protocol with wood chips spectra can be used as a classification tool and is an effective method for estimating bamboo chip energy quality. The big particle size of wood chips affect negatively the prediction model, however, it could be solved through spectral pre-processing technique, thus eliminating the need for grinding feedstock samples.

Determination of Moisture Content and Basic Density of Poplar Wood Chips under Various Moisture Conditions by Near-Infrared Spectroscopy

AbstractThe potential of near-infrared (NIR) spectroscopy coupled with partial least-squares (PLS) regression was used to determine the moisture content and basic density of poplar wood chips. NIR spectra collected from the surface of wood chips were used to develop calibration models for moisture content and basic density predication, and various spectral preprocessing techniques were applied to improve the accuracy and robustness of the prediction models. The models were tested using totally independent sample sets and exhibited acceptable predictive performance for moisture content (coefficient of determination for prediction [R2p] = 0.98 and standard error of prediction [SEP] = 2.51 percent) and basic density (R2p = 0.87 and SEP = 17.61 kg m–3). In addition, the effect of moisture variations on prediction of basic density was investigated based on NIR spectra from wood chips under various moisture levels. The results demonstrated that broad absorption bands from water molecules, especially when free water exists in the cell lumen, overlap with informative signals related to wood properties and weaken the calibration relation between spectral features and basic density. Thus, maintaining wood chips in a low and even moisture state would help achieve reliable estimates of wood density by NIR analysis models.

Single kernel wheat hardness estimation using near infrared hyperspectral imaging

Wheat is a major component of human diets across the world; hence, it is important to produce superior quality wheat for consumption. Hardness of wheat is one of its major quality and end-use determinants and it is desirable that this parameter could be rapidly and non-destructively measured along with its other functional attributes. Recent research has established that a number of quality parameters could be accurately measured using near infrared (NIR) hyperspectral imaging. In order to fully characterize wheat kernels using NIR hyperspectral imaging, this study aims to provide an optimal methodology to predict hardness of individual kernels. The predicted hardness of wheat kernels using this non-destructive imaging technique was referenced against the industry adopted hardness index (HI). Principal component analysis (PCA) was used as an exploratory method for classification of wheat kernels based on their hardness index into three categories of soft, medium, and hard.Various spectral pre-processing techniques were used and the best combination (Savitzky–Golay 1st derivative with 11 smoothing points followed by orthogonal signal correction and mean centering) was selected for partial least squares (PLS) regression and artificial neural network (ANN) models. The PLS model yielded moderate prediction of wheat hardness for calibration (R2 = 0.81 and RMSECV = 10.85) and prediction (R2 = 0.80 and RMSEP = 12.90). The ANN superseded the PLS model for calibration (R2 = 0.94 and RMSECV = 5.38) and prediction (R2 = 0.90 and RMSEP = 6.59). Although, these preliminary results have clearly established the potential of hyperspectral imaging for non-destructive and on-line pre-screening of individual wheat kernels for hardness determination, further studies on larger datasets are needed to scale-up the technique to commercial throughput levels.

Study of the scattering effects on NIR data for the prediction of ash content using EMSC correction factors

AbstractScattering effect is a really common physical phenomenon during near‐infrared analysis. It is an undesired variation in the spectral data due to a deviation of light from a straight trajectory into different paths. The nonlinearities introduced can be handled by using spectral preprocessing techniques. The situation is completely different when the parameter of interest is physical by nature, such as ash content, in this case removing the physical artifacts of scattering would be negative for the final model. In this study, we have decided to investigate the ash content parameter trying to figure out if the information useful for its prediction is related to the scattering effects, the chemical features, or a mixture of them. To this aim, two near‐infrared spectral datasets were taken into consideration: woodchip for energy sector and pellet samples for feed sector. A new regression model (CORR‐PLS) was developed by including principal components analysis scores and extended multiplicative scatter correction (EMSC) factors as physical parameters into the partial least squares (PLS) regression model. The prediction performance of the regular PLS models (PLS on the raw data and MSC pre‐treated data) were compared with that of the CORR‐PLS model both with regard to prediction uncertainty and model complexity in order to evaluate which is the relevant information for prediction of the ash content.

A memory-based learning approach utilizing combined spectral sources and geographical proximity for improved VIS-NIR-SWIR soil properties estimation

The current study was driven by the need to derive improved soil information whereby the strengths of memory-based learning and soil spectroscopy are exploited towards addressing regional challenges and supporting sustainable development across the Balkan, North Africa, and Middle East regions. In this study we focused on a novel derivation of the Spectrum Based Learner (SBL) algorithm by i) taking into account both the geographical proximity and the spectral similarity in the computation of the distance between samples, and ii) using an optimal pair of spectral pre-treatments with the first used for the determination of neighbours while the latter for the estimation of the target property. The proposed approach was tested on a recently developed standardized Soil Spectral Library (SSL) across the VIS-NIR-SWIR spectral region range (350–2500 nm) which comprises 1760 soil samples from 9 different countries and is considered the largest and most diverse soil database infrastructure in the region compatible with other SSLs. Here, we found that our approach presents a great potential for predicting soil texture contents (Clay, Sand, and Silt), organic carbon (OC), Calcium Carbonate (CaCO3), pH, and Electrical Conductivity (EC). These results outperform the predictive performance of other state of the art global and local approaches that have been applied to similar large and complex soil datasets. The use of geographical coordinates in the computation of the sample similarities enhanced the predictions of soil properties, since it allowed the generation of local subsets that present similar soil compositions. In addition, we conclude that the various spectral sources as derived by a set of predefined spectral pre-processing techniques contain complementary information which should be combined instead of relying solely on the best spectral pre-processing technique. This approach could be effectively utilized to enhance the predictions of soil properties in large and complex SSLs, since it decreased the Root Mean Square Errors (RMSEs) by a relative mean of 6.47% (average value across the properties – decrease ranging from 2.90% to 9.09%) for the various soil properties, compared to other global and local algorithms. We conclude that national SSLs that were measured under a standardization process could further contribute to the global initiative to address challenges and support a data-centric approach for informed decision making with regards to environmental and agricultural issues.

Combination of fractional order derivative and memory-based learning algorithm to improve the estimation accuracy of soil organic matter by visible and near-infrared spectroscopy

Visible and near-infrared (Vis–NIR) spectroscopy is used to estimate soil organic matter (SOM). Spectral preprocessing techniques and multivariate modeling methods play important roles in the quantitative analysis of SOM. First and second derivatives (i.e., the conventional integer order derivatives) are commonly used spectral derivatives, which, however, may ignore some detailed spectral information regarding SOM. Here, we presented a fractional order derivative (FOD) method to preprocess the reflectance spectra. Robust modeling methods are still required for accurate estimation of SOM. Local modeling technique (memory-based learning, MBL) was introduced to compare with two global modeling approaches, namely, partial least square (PLS) and random forest (RF). A total of 535 topsoil samples were gathered from Hubei Province, Central China, with their reflectance spectra and SOM contents measured in the laboratory. FOD was allowed to vary from 0 to 2 with an increment of 0.25 at each step. Coefficient of determination (R2) and ratio of the performance to deviation (RPD) were employed as performance statistics during validation. Results showed that with the increase of derivative order, the baseline drifts and overlapping peaks were gradually removed but the spectral strength decreased concurrently. Higher derivative order reflectance (i.e., 1.5-order, 1.75-order, and 2-order reflectance) were more susceptible to spectral noise interferences. The correlation coefficient of SOM with FOD processed spectra at some specific wavelengths was larger than that with the original reflectance. MBL performed better than PLS and RF, regardless of FOD transformation. Calibration with 0.25-order reflectance and MBL provided the most accurate estimation of SOM, with an RPD of 2.23. Our results confirm the effectiveness of FOD and local modeling (MBL) in the development of Vis–NIR models for SOM estimation.

Pattern recognition-based optical technique for non-destructive detection of Ectomyelois ceratoniae infestation in pomegranates during hidden activity of the larvae.

In this research, the feasibility of utilizing visible/near-infrared (Vis/NIR) spectroscopy as an optical non-destructive technique combined with both supervised and unsupervised pattern recognition methods was assessed for detection of Ectomyelois ceratoniae, carob moth, infestation in pomegranates during hidden activity of the larvae. To this end, some fruits were artificially contaminated to the carob moth larvae. Vis/NIR spectra of the blank samples and the contaminated pomegranates without and with external visual symptoms of larvae infestation were analyzed one and two weeks after contaminating the samples as three groups of "Healthy", "Unhealthy-A" and "Unhealthy-B", respectively. Principal component analysis (PCA) as unsupervised pattern recognition method was used to verify the possibility of clustering of the pomegranate samples into the three groups. Discriminant analysis (DA) based on PCA was also used as a powerful supervised pattern recognition method to classify the samples. The calibration models of linear, quadratic and Mahalanobis discriminant analyses were developed based on different spectral pre-processing techniques. The best PCA-DA model was obtained using Mahalanobis distance method and first derivative (D1) pre-processing. The total percentage of correctly classified samples with the best calibration model was 97.9%. The developed model could also predict unknown samples with total percentage of correctly classified samples of 90.6%. It was concluded that Vis/NIR spectroscopy combined with pattern recognition method of PCA-DA can be an appropriate and rapid technology for non-destructively screening the pomegranates for detection of carob moth infestation during hidden activity of the larvae.

Assessment of the optimal spectral bands for designing a sensor for vineyard disease detection: the case of \u2018Flavescence dor\xe9e\u2019

Disease detection and control is one of the main objectives of vineyard research in France. Manual monitoring of diseases is a time consuming operation especially when fields are large. Current research studies aim to develop automatic in-field systems to detect diseases. This study investigated the optimal spectral bands for the design of a dedicated high-resolution multispectral camera embedded on an unmanned aerial vehicle for identifying infected zones in a grapevine field. The target disease was Flavescence dorée, which is infectious, incurable and can result in considerable yield loss. An in-field spectrometry study was performed on four grapevine varieties in the Provence Alpes Côte d’Azur region in France. Two spectral analysis techniques were proposed for choosing the best spectral bands capable of discriminating healthy from diseased leaves. The first novel approach is a feature selection technique based on the successive projection algorithm (SPA), some spectral pre-processing techniques were jointly investigated. The second approach examines a set of traditional vegetation indices (VI). Support vector machine (SVM) and discriminant analysis (DA) are the two classifiers used in this paper and the accuracy of the results is compared for the two methods of analysis. The best models were computed as a function of the grapevine variety considered. The SPA technique performed better in general with respect to common VIs, the overall classification accuracy was more than 96%. Results demonstrated that employing a feature selection technique based on the SPA algorithm can provide a valid tool for determining the optimal bands that are sensitive to Flavescence dorée grapevine disease and assist in its identification. The benefit behind the presented procedure relies on the possibility of generalizing it for other infections and stresses or even for different crops.

Examining the Performance of PARACUDA-II Data-Mining Engine versus Selected Techniques to Model Soil Carbon from Reflectance Spectra

The monitoring and quantification of soil carbon provide a better understanding of soil and atmosphere dynamics. Visible-near-infrared-short-wave infrared (VIS-NIR-SWIR) reflectance spectroscopy can quantitatively estimate soil carbon content more rapidly and cost-effectively compared to traditional laboratory analysis. However, effective estimation of soil carbon using reflectance spectroscopy to a great extent depends on the selection of a suitable preprocessing sequence and data-mining algorithm. Many efforts have been dedicated to the comparison of conventional chemometric techniques and their optimization for soil properties prediction. Instead, the current study focuses on the potential of the new data-mining engine PARACUDA-II®, recently developed at Tel-Aviv University (TAU), by comparing its performance in predicting soil oxidizable carbon (Cox) against common data-mining algorithms including partial least squares regression (PLSR), random forests (RF), boosted regression trees (BRT), support vector machine regression (SVMR), and memory based learning (MBL). To this end, 103 soil samples from the Pokrok dumpsite in the Czech Republic were scanned with an ASD FieldSpec III Pro FR spectroradiometer in the laboratory under a strict protocol. Spectra preprocessing for conventional data-mining techniques was conducted using Savitzky-Golay smoothing and the first derivative method. PARACUDA-II®, on the other hand, operates based on the all possibilities approach (APA) concept, a conditional Latin hypercube sampling (cLHs) algorithm and parallel programming, to evaluate all of the potential combinations of eight different spectral preprocessing techniques against the original reflectance and chemical data prior to the model development. The comparison of results was made in terms of the coefficient of determination (R2) and root-mean-square error of prediction (RMSEp). Results showed that the PARACUDA-II® engine performed better than the other selected regular schemes with R2 value of 0.80 and RMSEp of 0.12; the PLSR was less predictive compared to other techniques with R2 = 0.63 and RMSEp = 0.29. This can be attributed to its capability to assess all the available options in an automatic way, which enables the hidden models to rise up and yield the best available model.

On-The-Go Hyperspectral Imaging Under Field Conditions and Machine Learning for the Classification of Grapevine Varieties.

Grapevine varietal classification is an important plant phenotyping issue for grape growing and wine industry. This task has been achieved from destructive techniques like classic ampelography and DNA analysis under laboratory conditions. This work displays a new approach for the classification of a high number of grapevine (Vitis vinifera L.) varieties under field conditions using on-the-go hyperspectral imaging and different machine learning algorithms. On-the-go imaging was performed under natural illumination using a hyperspectral camera mounted on an all-terrain vehicle at 5 km/h. Spectra were acquired over two different leaf phenological stages on the canopy of 30 different varieties on a commercial vineyard located in La Rioja, Spain. A total of 1,200 spectral samples were generated. Support vector machines (SVM) and artificial neural networks (multilayer perceptrons, MLP) were used for the development of a large number of models, testing different algorithm parameters and spectral pre-processing techniques. Both classifiers yielded notable performance values and were able to train models with recall F1 scores and area under the receiver operating characteristic curve marks up to 0.99 for 5-fold cross validation. Statistical analyses supported that the best SVM kernel was linear and the best activation function for MLP was the hyperbolic tangent function. The prediction performance for individual varieties of MLP ranged from 0.94 to 0.99, displaying low levels of variability. In the case of SVM, slightly higher differences were obtained, ranging from 0.83 to 0.97 for individual varieties. These results support the possibility of deploying an on-the-go hyperspectral imaging system in the field capable of successfully classifying leaves from different grapevine varieties. This technology could thus be considered as a new useful non-destructive tool for plant phenotyping under field conditions.

Mapping the Pungency of Green Pepper Using Hyperspectral Imaging

The pungency level of green peppers is dependent on the amounts of capsaicin and dihydrocapsaicin they contain. This study was conducted to develop a non-destructive method for the prediction and mapping of the capsaicin and dihydrocapsaicin contents in green pepper. Hyperspectral images of 200 total green peppers of three varieties were acquired in the wavelength range of 1000–1600 nm, from which the mean spectra of each pepper variety were extracted. The reference capsaicin and dihydrocapsaicin contents of the samples were measured by high-performance liquid chromatography. Quantitative calibration models were built using partial least squares (PLS) regression with different spectral preprocessing techniques; the best performance was found by normalizing the preprocessed spectra with correlation coefficients (rpred) of 0.86 and 0.59, which showed the standard errors of prediction (SEPs) of 0.09 and 0.03 mg/g for capsaicin and dihydrocapsaicin, respectively. Seventeen and 16 optimal wavebands were selected using the successive projections algorithm; rpred of 0.88 and 0.68 and SEPs of 0.084 and 0.027 mg/g were obtained for capsaicin and dihydrocapsaicin, respectively, from the newly developed PLS calibration models using these optimal wavebands. The successive projections algorithm (SPA)-PLS model was used to map the capsaicin and dihydrocapsaicin contents of the green peppers. These maps provided detailed information on the pungency levels of the tested green peppers. The results of this study indicated that hyperspectral imaging is useful for the rapid and non-destructive evaluation of the pungency of green peppers.

Integration of Artificial Neural Network Modeling and Hyperspectral Data Preprocessing for Discrimination of Colla Corii Asini Adulteration

The study of hyperspectral imaging in tandem with spectral preprocessing and neural network techniques was conducted to realize Colla Corii Asini (CCA, E’jiao) adulteration discrimination. CCA was adulterated with pig skin gelatin (PSG) in the range of 5–95% (w/w) at 5% increments. Three methods were used to pretreat the original spectra, which are multiplicative scatter correction (MSC), Savitzky-Golay (SG) smoothing, and the combination of MSC and SG (MSC-SG). SPA was employed to select the characteristic wavelengths (CWs) to reduce the high dimension. Colour and texture features of CWs were extracted as input of prediction model. Two kinds of artificial neural network (ANN) with three spectral preprocessing methods were applied to establish the prediction models. The prediction model of generalized regression neural network (GRNN) in tandem with the MSC-SG preprocessed method presented satisfactory performance with the correct classification rate value of 92.5%. The results illustrated that the integration of preprocessing methods, hyperspectral imaging features, and ANN modeling had a great potential and feasibility for CCA adulteration discrimination.