Predicting Soil Organic Matter, Available Nitrogen, Available Phosphorus and Available Potassium in a Black Soil Using a Nearby Hyperspectral Sensor System.

Black soils, which play an important role in agricultural production and food security, are well known for their relatively high content of soil organic matter (SOM). SOM has a significant impact on the sustainability of farmland and provides nutrients for plants. Hyperspectral imaging (HSI) in the visible and near-infrared region has shown the potential to detect soil nutrient levels in the laboratory. However, using portable spectrometers directly in the field remains challenging due to variations in soil moisture (SM). The current study used spectral data captured by a handheld spectrometer outdoors to predict SOM, available nitrogen (AN), available phosphorus (AP) and available potassium (AK) with different SM levels. Partial least squares regression (PLSR) models were established to compare the predictive performance of air-dried soil samples with SMs around 20%, 30% and 40%. The results showed that the model established using dry sample data had the best performance (RMSE = 4.47 g/kg) for the prediction of SOM, followed by AN (RMSE = 20.92 mg/kg) and AK (RMSE = 22.67 mg/kg). The AP was better predicted by the model based on 30% SM (RMSE = 8.04 mg/kg). In general, model performance deteriorated with an increase in SM, except for the case of AP. Feature wavelengths for predicting four kinds of soil properties were recommended based on variable importance in the projection (VIP), which offered useful guidance for the development of portable hyperspectral sensors based on discrete wavebands to reduce cost and save time for on-site data collection.

Discrimination of Deoxynivalenol Levels of Barley Kernels Using Hyperspectral Imaging in Tandem with Optimized Convolutional Neural Network.

Deoxynivalenol (DON) in raw and processed grain poses significant risks to human and animal health. In this study, the feasibility of classifying DON levels in different genetic lines of barley kernels was evaluated using hyperspectral imaging (HSI) (382-1030 nm) in tandem with an optimized convolutional neural network (CNN). Machine learning methods including logistic regression, support vector machine, stochastic gradient descent, K nearest neighbors, random forest, and CNN were respectively used to develop the classification models. Spectral preprocessing methods including wavelet transform and max-min normalization helped to enhance the performance of different models. A simplified CNN model showed better performance than other machine learning models. Competitive adaptive reweighted sampling (CARS) in combination with successive projections algorithm (SPA) was applied to select the best set of characteristic wavelengths. Based on seven wavelengths selected, the optimized CARS-SPA-CNN model distinguished barley grains with low levels of DON (<5 mg/kg) from those with higher levels (5 mg/kg < DON ≤ 14 mg/kg) with an accuracy of 89.41%. The lower levels of DON class I (0.19 mg/kg ≤ DON ≤ 1.25 mg/kg) and class II (1.25 mg/kg < DON ≤ 5 mg/kg) were successfully distinguished based on the optimized CNN model, yielding a precision of 89.81%. The results suggest that HSI in tandem with CNN has great potential for discrimination of DON levels of barley kernels.

Rapid and Low-Cost Quantification of Adulteration Content in Camellia Oil Utilizing UV-Vis-NIR Spectroscopy Combined with Feature Selection Methods.

This study aims to explore the potential use of low-cost ultraviolet-visible-near infrared (UV-Vis-NIR) spectroscopy to quantify adulteration content of soybean, rapeseed, corn and peanut oils in Camellia oil. To attain this aim, test oil samples were firstly prepared with different adulterant ratios ranging from 1% to 90% at varying intervals, and their spectra were collected by an in-house built experimental platform. Next, the spectra were preprocessed using Savitzky-Golay (SG)-Continuous Wavelet Transform (CWT) and the feature wavelengths were extracted using four different algorithms. Finally, Support Vector Regression (SVR) and Random Forest (RF) models were developed to rapidly predict adulteration content. The results indicated that SG-CWT with decomposition scale of 25 and the Iterative Variable Subset Optimization (IVSO) algorithm can effectively improve the accuracy of the models. Furthermore, the SVR model performed best for predicting adulteration of camellia oil with soybean oil, while the RF models were optimal for camellia oil adulterated with rapeseed, corn, or peanut oil. Additionally, we verified the models' robustness by examining the correlation between the absorbance and adulteration content at certain feature wavelengths screened by IVSO. This study demonstrates the feasibility of using low-cost UV-Vis-NIR spectroscopy for the authentication of Camellia oil.

Non-destructive classification of apple bruising time based on visible and near-infrared hyperspectral imaging.

BACKGROUND: Bruising time of apple is one of the most important factors for internal quality assessment. The present study aimed to establish a non-destructive method for the classification of apple bruising time using visible and near-infrared (VNIR) hyperspectral imaging. In this study, VNIR hyperspectral images were obtained and analyzed at seven bruising periods. Moreover, regions of interest (ROIs) were chosen to construct the bruised region classification model, and spectra of bruised regions were collected and resampled based on four different methods. Subsequently, machine learning algorithms were employed and used for dealing with the time classification model of apples. In order to reduce data redundancy and improve the accuracy of the classification model, a tree-based assembling learning model was used to select feature wavelengths, and linear discriminant analysis (LDA) was used to improve the discernibility of data. RESULTS: The results revealed that the random forest (RF) model can precisely locate bruised regions, while the gradient boosting decision tree (GBDT) model can validly classify apple bruising times with 70.59% accuracy. Data of 128 wavebands were compressed to 13 wavebands, providing a high accuracy of 92.86%. CONCLUSION: The results prove that the hyperspectral technique can be used for predicting apple bruising time, which will help to assess the internal quality and safety of apples. © 2018 Society of Chemical Industry.

Nondestructive detection of saponin content in Panax notoginseng powder based on hyperspectral imaging.

This study investigated the feasibility of using hyperspectral imaging (HSI) technique to detect the saponin content in Panax notoginseng (PN) powder. The reflectance hyperspectral images of PN powder samples were collected in the spectral range of 400.6-999.9 nm. Savitzky-golay (SG) smoothing combined with detrending correction was utilized to preprocess the original spectral data. Two model population analysis (MPA) based methods, namely bootstrapping soft shrinkage (BOSS) and iteratively retains informative variables (IRIV) were employed to extract feature wavelengths from the full spectra. A generalized normal distribution optimization based extreme learning machine (GNDO-ELM) model was proposed to establish calibration model between spectra and saponin content, and compared with existing methods (GA-ELM, PSO-ELM and SSA-ELM). The result showed that the IRIV-GNDO-ELM model gave the best performance, with coefficient of determination for prediction (R2P) of 0.953 and root mean square error for prediction (RMSEP) of 0.115%. Therefore, it is possible to determine the saponin content of PN powder by using HSI technique.

Coupled Vis-NIR spectroscopy with chemometrics strategy for soil organic carbon prediction in the Agro-pastoral Transitional zone of northwest China.

Rapidly and accurately grasp the change of soil organic carbon content in farmland, which is of great significance in guiding the timely and effective mastery of farmland soil fertility and improvement of soil physical properties. In this study, an ASD FieldSpec 4 spectrometer was used to collect spectral reflectance data on 128 agricultural soil samples taken from Jingbian County, Yulin City, Shaanxi Province, China. Firstly, descriptive statistics of the SOC in the study area were performed, and secondly, after 10 spectral transformations were performed, the correlation analysis and the Boruta algorithm were used to extract the characteristic wavebands of soil organic carbon, respectively, in order to reduce the redundancy of the data. Finally, by comparing the accuracies of different strategies, we constructed a spectral prediction model of soil organic carbon in farmland of the Northwest Agricultural and Animal Husbandry Intertwined Zone that integrates the optimal preprocessing, feature selection strategy and modelling method. The results indicate that: 1) The mean SOC content of the farmland in the study area was low and at the nutrient deficient level, with the standard errors and coefficients of variation for the modelling and validation sets were 1.596 g kg-1, 1.457 g kg-1, 54 % and 52 %, respectively; 2) The shape and trend of spectral special curves with different SOC contents show consistency, and the SOC content is negatively correlated with spectral reflectance; 3) CA selects more feature bands, but the feature bands are more homogeneous, while the Boruta algorithm can effectively remove irrelevant variables and improve the SOC feature selection effect; 4) The SOC prediction model based on Boruta-FD-RF can be better for soil organic carbon estimation, with R2 of 0.899 and 0.748 for the training set and validation set, respectively, RMSE of 1.432 g kg-1 and 1.967 g kg-1, and RPD of 2.557 and 1.647, respectively. The results show that the SOC model established by integrating optimal spectral pre-processing, feature selection strategy and chemometrics strategy has obvious improvement in prediction accuracy and stability, and this study provides an important reference for the fast and accurate estimation of SOC content in farmland of Agro-pastoral Transitional zone in northwest China.

Identification and quantitative detection of illegal additives in wheat flour based on near-infrared spectroscopy combined with chemometrics.

As a common food raw material in daily life, the quality and safety of wheat flour are directly related to people's health. In this study, a model was developed for the rapid identification and detection of three illegal additives in flour, namely azodicarbonamide (ADA), talcum powder, and gypsum powder. This model utilized a combination of near-infrared spectroscopy with chemometric methods. A one-dimensional convolutional neural network was used to reduce data dimensionality, while a support vector machine was applied for non-linear classification to identify illegal additives in flour. The model achieved a calibration set F1 score of 99.38% and accuracy of 99.63%, with a validation set F1 score of 98.81% and accuracy of 98.89%. Two cascaded wavelength selection methods were introduced: The first method involved backward interval partial least squares (BiPLS) combined with an improved binary particle swarm optimization algorithm (IBPSO). The second method utilized the CARS-IBPSO algorithm, which integrated competitive adaptive reweighted sampling (CARS) with IBPSO. The two cascade wavelength selection methods were used to select feature wavelengths associated with additives and construct partial least squares quantitative detection models. The models constructed using CARS-IBPSO selected feature wavelengths for detecting ADA, talcum powder, and gypsum powder exhibited the highest overall performance. The model achieved validation set determination coefficients of 0.9786, 0.9102, and 0.9226, with corresponding to root mean square errors of 0.0024%, 1.3693%, and 1.6506% and residual predictive deviations of 6.8368, 3.5852, and 3.9253, respectively. Near-infrared spectroscopy in combination with convolutional neural network dimensionality reduction and support vector machine classification enabled rapid identification of various illegal additives. The combination of CARS-IBPSO feature wavelength selection and partial least squares regression models facilitated rapid quantitative detection of these additives. This study introduces a new approach for rapidly and accurately identifying and detecting illegal additives in flour.

Discrimination of wheat flour grade based on PSO-SVM of hyperspectral technique.

Rapid detection of wheat flour grade played an important role in the food industry. In this work, hyperspectral technology was used to detect five types of wheat flour. An analysis model was established based on the reflectance of samples at 968 ∼ 2576 nm. Moreover, multivariate scattering correction (MSC), standard normalized variate (SNV), and Savitzky-Golay (S-G) convolution smoothing were used for preprocessing, which was employed to reduce the influence of noise in the original spectrum. In order to simplify the model, competing adaptive reweighted sampling (CARS), successive projection algorithm (SPA), uninformative variable elimination (UVE) and the UVE-CARS algorithm were applied to extract feature wavelengths. Both partial least squares discriminant analysis (PLS-DA) model and support vector machine (SVM) model were established according to feature wavelengths. Furthermore, particle swarm optimization (PSO) algorithm was adopted to optimize the search of SVM model parameters, such as the penalty coefficient c and the regularization coefficient g. Experimental results suggested that the non-linear discriminant model for wheat flour grades was better than the linear discriminant model. It was considered that the MSC-UVE-CARS-PSO-SVM model achieved the best forecasting results for wheat flour grade discrimination, with 100% accuracy both in the calibration set and the validation set. It further shows that the classification of wheat flour grade can be effectively realized by using the hyperspectral and SVM discriminant analysis model, which proves the potential of hyperspectral reflectance technology in the qualitative analysis of wheat flour grade.

Research and Application Validation of a Feature Wavelength Selection Method Based on Acousto-Optic Tunable Filter (AOTF) and Automatic Machine Learning (AutoML).

Near-infrared spectroscopy has been widely applied in various fields such as food analysis and agricultural testing. However, the conventional method of scanning the full spectrum of the sample and then invoking the model to analyze and predict results has a large amount of collected data, redundant information, slow acquisition speed, and high model complexity. This paper proposes a feature wavelength selection approach based on acousto-optical tunable filter (AOTF) spectroscopy and automatic machine learning (AutoML). Based on the programmable selection of sub nm center wavelengths achieved by the AOTF, it is capable of rapid acquisition of combinations of feature wavelengths of samples selected using AutoML algorithms, enabling the rapid output of target substance detection results in the field. The experimental setup was designed and application validation experiments were carried out to verify that the method could significantly reduce the number of NIR sampling points, increase the sampling speed, and improve the accuracy and predictability of NIR data models while simplifying the modelling process and broadening the application scenarios.

Extensive evaluation of prediction performance for 15 pork quality traits using large scale VIS/NIRS data.

Visible and near-infrared spectroscopy (VIS/NIRS) has been extensively used in the livestock and food industries to quantify meat quality. Here, we collected VIS/NIRS data of 1206 pigs longissimus muscle, measured the corresponding 15 meat quality traits, and used seven models to predict these meat quality traits. The prediction performances of 7 models varied among predicted traits, with the Rcv2 of most traits above 0.9 in the best model. We have also established a new method, spectral-wide association analysis (SWAS), to select the feature wavelengths of measured traits. Results showed that the prediction performance is proportionate to the number of identified significant association wavelengths. We used the selected wavelengths to perform prediction again, and the prediction accuracy was similar to results with full wavelength using the best model, indicating effectiveness of feature wavelengths selection methods.

[Hyperspectral prediction model of soil nutrient content in the loess hilly-gully region, China.] / é»„åœŸä¸˜é™µæ²Ÿå£‘åŒºåœŸå£¤å »åˆ†å«é‡çš„é«˜å ‰è°±é¢„æµ‹æ¨¡åž‹.

Rapid and accurate estimation of soil nutrient content based on hyperspectral data is an optimal method for the monitoring of soil nutrient and inversion of soil physical and chemical characters. The relationship between soil nutrient content and spectral reflectance was analyzed with soil samples being collected from the loess hilly-gully region of northern Shaanxi Province. The prediction models of the content of soil organic matter, total nitrogen, total phosphorus and total potassium were constructed by the combination of three techniques, including partial least squares (PLS), multiple linear regression (MLR), and support vector machine (SVM). Then, the optimal model was selected by comparison analysis. The results showed good correlations between the content of soil nutrients and spectral reflectance in visible region (400-760 nm) and near infrared region (760-1100 nm). The maximum values of correlation coefficient located in both spectral regions. The SPA-SVM model had the best applicability and highest inversion accuracy for the contents of all soil nutrients, with simple and efficient modeling process. Our results provided a reference for applying machine learning algorithm in the construction of hyperspectral prediction model of soil nutrient content in the loess hilly-gully region.