Detection of internal qualities of hami melons using hyperspectral imaging technology based on variable selection algorithms

Hyperspectral reflectance imaging technology in near-infrared regions (900–1,700 nm) was used to evaluate soluble solids content (SSC), firmness, moisture content (MC), and pH values of ‘Fuji’ apples during a 13-week storage period. Totally, 167 apples were divided into calibration set (125) and prediction set (42) based on the joint x-y distance sample set partitioning method. Mean spectrum of the regions of interest in the hyperspectral image of each apple was used for analysis. Two typical variable selection methods, i.e., successive projection algorithm (SPA) and uninformative variable elimination (UVE), were applied to extract the characteristic variables from full spectra (FS). The partial least squares (PLS) regression, least squares support vector machine (LSSVM), and backpropagation (BP) network modeling methods were used to establish models to predict SSC, firmness, MC, and pH of apples, respectively. The results showed that the SSC and MC could be predicted exactly by all developed models, and SPA-LSSVM and FS-BP could be used to predict pH value roughly. All models failed to predict firmness. The SPA-LSSVM model had better comprehensive ability in determining SSC, MC, and pH than others, with the correlation coefficient of prediction of 0.961, 0.984, and 0.882 and residual predictive deviation of 3.49, 5.51, and 2.06, respectively. The results demonstrated the feasibility of using near-infrared hyperspectral reflectance imaging technology as a non-invasive method for predicting SSC, MC, and pH of apples simultaneously.

Detection of mango soluble solid content using hyperspectral imaging technology

To meet rapid and non-destructive identification of selenium-enriched agricultural products selenium-enriched millet and ordinary millet were taken as objects. Image regions of interest (ROI) were selected to extract the spectral average value based on hyperspectral imaging technology. Reducing noise by the Savitzky-Golay (SG) smoothing algorithm, variables were used as inputs that were screened by successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), uninformative variable elimination (UVE), CARS-SPA, UVE-SPA, and UVE-CARS, while sample variables were used as outputs to build support vector machine (SVM) models. The results showed that the accuracy of CARS-SPA-SVM was 100% in the training set and 99.58% in the test set equivalent to that of CARS-SVM and UVE-CARS-SVM, which was higher than that of SPA-SVM, UVE-SPA-SVM, and UVE-SVM. Therefore, the method of CARS-SPA had superiority, and CARS-SPA-SVM was suitable to identify selenium-enriched millet. Finally, 454.57 nm, 484.98 nm, 885.34 nm, and 937.1 nm, which were obtained by wavelength extraction algorithms, were considered as the sensitive wavelengths of selenium information. This study provided a reference for the identification of selenium-enriched agricultural products.

Can light penetrate through pomelos and carry information for the non-destructive prediction of soluble solid content using Vis-NIRS?

IntroductionThe soluble solids content (SSC), titratable acidity (TA), and pH are key indicators for evaluating the quality of table grapes during storage. Conventional detection methods are typically destructive and time-consuming. To address this limitation, visible-near infrared (Vis-NIR) spectroscopy was employed in this study to enable rapid and non-destructive quality assessment of fresh table grapes throughout the storage period.MethodsSeedless White table grape samples were analyzed within the 200–1100 nm spectral range, and calibration models for key quality parameters (SSC, TA, and pH) were established. Three machine learning algorithms, partial least squares regression (PLSR), support vector machine (SVM), and extreme learning machine (ELM), were employed to develop spectral prediction models based on characteristic wavelengths selected using different feature extraction strategies, including the successive projection algorithm (SPA), uninformative variable elimination (UVE), and competitive adaptive reweighted sampling (CARS).ResultsThe results demonstrate that the SNV-CARS-SVM models achieved excellent performance in predicting SSC with a root mean square errors (RMSEP) of 0.673, a coefficient of determination for the prediction data set (Rp) of 0.928 and an RPD of 3.311. Similarly, the SNV-SPA-SVM models exhibited excellent predictive accuracy for TA, yielding an RMSEP of 0.553, an Rp of 0.873, and an RPD of 2.662. Good performances were achieved with Rp of 0.758 and RMSEP of 0.113 with the SNV-CARS-PLSR model for pH.DiscussionThis study, for the first time, utilized Vis-NIR spectroscopy to achieve the simultaneous and rapid determination of multiple quality attributes in table grapes, providing a novel and efficient strategy for real-time and non-destructive quality evaluation during storage. The proposed approach showed considerable potential for rapid quality assessment and postharvest management of grapes. Future research will focus on expanding the diversity of grape cultivars and investigating various storage conditions to improve the robustness and transferability of the predictive model, thereby promoting the industrial validation and practical application of Vis-NIR spectroscopy in fruit quality monitoring.

In this research, cadmium concentrations contained in Beche-de-mere (BDM) due to pollution of heavy metals in the sea were determined using hyperspectral imaging (HSI) technology and deep neural networks regression method. This detection system was important for maintaining the quality and safety of BDM which was one of the most important export food commodities in Indonesia. Several BDM samples of sandfish type were obtained from local markets in Indonesia. BDM samples were put in the glass jars that contained solutions of CdCl2 with various concentrations from 0 ~ 5 mg/L for three days. After three days, samples were dried using drying oven for 5 days. Next, the samples were scanned three times for each sample by hyperspectral imaging camera to obtain their hyperspectral data. Finally, after the spectral data acquired, the cadmium concentrations of the samples were ready to be determined by using Absorption Atomic Spectroscopy (AAS). Four regions of interests (ROI) were selected from each sample, and with three times scanning for each sample, resulted in total of 96 ROIs. Spectral profile data of the ROIs showed that there were variations in spectral reflectance of the BDM related to the different concentrations of cadmium used in the treatment of samples. Savitzky-Golay (SG) filter was used to clean noises from the spectral data. The model prediction was developed using deep neural networks. The results were also compared with shallow neural networks, partial least squared regression (PLSR), and support vector machine regression (SVR) methods. The best-established model which was deep neural networks with SG filtering had coefficient of correlation (R<inf>p</inf>) of 0.93, and root-mean-square-error prediction (RMSEP) of 9.21.

Accurate determination of Soil Organic Carbon (SOC), which is the foundation of soil health and safeguards ecological and food security, is crucial in local agricultural production. We aimed to investigate the influence of soil texture on hyperspectral models for predicting SOC content and to evaluate the role of different preprocessing methods and feature band selection algorithms in improving modeling efficiency. Laboratory-determined SOC content and hyperspectral reflectance data were obtained using soil samples from daylily cultivation areas in Yunzhou District, Datong City. Mathematical transformations, including Savitzky–Golay smoothing (SG), First Derivative (FD), Second Derivative (SD), Multiplicative Scatter Correction (MSC), and Standard Normal Variate (SNV), were applied to the spectral reflectance data. Feature bands extracted based on the successive projection algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS) were used to establish SOC content inversion models employing four algorithms: partial least-squares regression (PLSR), Random Forest (RF), Backpropagation Neural Network (BP), and Convolutional Neural Network (CNN). The results indicate the following: (1) Preprocessing can effectively increase the correlation between the soil spectral reflectance process and SOC content. (2) SPA and CARS effectively screened the characteristic bands of SOC in daylily cultivated soil from the spectral curves. The SPA algorithm and CARS selected 4–11 and 9–122 bands, respectively, and both algorithms facilitated model construction. (3) Among all the constructed models, the FD-CARS-PLSR performed most prominently, with coefficients of determination (R2) for the training and validation sets reaching 0.93 and 0.83, respectively, demonstrating high model stability and reliability. (4) Incorporating soil texture as an auxiliary variable into the PLSR inversion model improved the inversion accuracy, with accuracy gains ranging between 0.01 and 0.05.

Soluble solid content (SSC), firmness, and color (L*, a*, and b*) are important physicochemical indices for assessing the quality and maturity of kiwifruits. Therefore, this research aimed to realize the nondestructive detection and visualization map for the physicochemical indices of kiwifruits at different maturity stages by hyperspectral imaging coupled with the chemometrics. To further improve the detection accuracy and working efficiency of the models, competitive adaptive reweighted sampling (CARS) and successive projection algorithm were employed to choose feature wavelengths for predicting the physicochemical indices of kiwifruits. Multiple linear regression (MLR) was designed to develop simplified detection models based on feature wavelengths for determining the physicochemical indices of kiwifruits. The results showed that 32, 18, 26, 29, and 32 feature wavelengths were extracted from 256 full wavelengths to predict the SSC, firmness, L*, a*, and b*, respectively, with the CARS algorithm. Not only was the working efficiency of the CARS-MLR model improved, but the prediction accuracy of the CARS-MLR model for determining the physicochemical indices was also at its relative best. The residual predictive deviations of the CARS-MLR model for determining the SSC, firmness, L*, a*, and b* were 3.09, 2.90, 2.32, 2.74, and 2.91, respectively, which were all above 2.3. Compared with the model based on the full spectra, the CARS-MLR model could be used to predict the physicochemical indices of kiwifruits. Finally, the visualization map for the physicochemical indices of kiwifruits at different maturity stages was generated by calculating the spectral response of each pixel on the kiwifruit samples with the CARS-MLR model. This made the detection for the physicochemical indices of kiwifruits more intuitive. This study demonstrates that hyperspectral imaging coupled with the chemometrics is promising for the nondestructive detection and visualization map for the physicochemical indices of kiwifruits, and also provides a novel theoretical basis for the nondestructive detection of kiwifruit quality.

The moisture content (MC) and acidity content (AC) of the fermented grains used in liquor production directly affect the liquor quality and yield; as such, they are important indicators used to evaluate the quality of fermented grains. In this study, extreme gradient enhancement algorithm (XGBoost), partial least square regression (PLSR), and extreme learning machine (ELM) models were developed based on spectral data collected by near-infrared (NIR) hyperspectral imaging (HSI) technology. First, PLSR models were established after SNV and MSC algorithms preprocessed the HSI data, and the best preprocessing method was determined (MC: SNV; AC: MSC). Then, the competitive adaptive reweighting sampling (CARS) algorithm and principal component analysis (PCA), both combined with the successive projection algorithm (SPA), were used to extract the characteristic wavelengths from the full-band spectral data. Ultimately, the XGBoost model developed using the characteristic wavelengths extracted by CARS-SPA most accurately predicted the MC (RPD = 6.4167, RP 2 = 0.9757, RMSEP = 0.0442 g·100 g−1) and AC (RPD = 13.0308, RP 2 = 0.9941, RMSEP = 0.0216 mmol·10 g−1). The results showed that the XGBoost model could more accurately predict the MC and AC of the fermented grains from hyperspectral images of the grains, providing an effective method for the rapid analysis of raw materials used in the fermentation of liquor.

Soil nutrients, including soil available potassium (SAK), soil available phosphorous (SAP), and soil organic matter (SOM), play an important role in farmland soil productivity, food security, and agricultural management. Spectroscopic analysis has proven to be a rapid, nondestructive, and effective technique for predicting soil properties in general and potassium, phosphorous, and organic matter in particular. However, the successful estimation of soil nutrient content by visible and near-infrared (Vis-NIR) reflectance spectroscopy depends on proper calibration methods (including preprocessing transformation methods and multivariate methods for regression analysis) and the selection of appropriate variable selection techniques. In this study, raw spectrum and 13 preprocessing transformations combined with 2 variable selection methods (competitive adaptive reweighted sampling (CARS) and the successive projections algorithm (SPA)) and 2 regression algorithms (support vector machine (SVM) and partial least squares regression (PLSR)), for a total of 56 calibration methods, were investigated for modeling and predicting the above three soil nutrients using hyperspectral Vis-NIR data (400–2450 nm). The results show that first-order derivatives based on logarithmic and inverse transformations (FD-LGRs) can provide better predictions of soil available potassium and phosphorous, and the best form of soil organic matter transformation is SG+MSC. CARS was superior to the SPA in selecting effective variables, and the PLSR model outperformed the SVM models. The best estimation accuracies (R2, RMSE) for soil available potassium, phosphorous, and organic matter were 0.7532, 32.3090 mg/kg; 0.7440, 6.6910 mg/kg; and 0.9009, 3.2103 g/kg, respectively, and their corresponding calibration methods were (FD-LGR)/SPA/PLSR, (FD-LGR)/SPA/PLSR, and SG+MSC/CARS/SVM, respectively. Overall, for the prediction of the soil nutrient content, organic matter was superior to available phosphorous, followed by available potassium. It was concluded that the application of hyperspectral images (Vis-NIR data) was an efficient method for mapping and monitoring soil nutrients at the regional scale, thus contributing to the development of precision agriculture.

Spatial assessment of soluble solid contents on apple slices using hyperspectral imaging

In-field and non-destructive monitoring of grapes maturity by hyperspectral imaging

Wheat protein and fat contents directly affect the flavor and yield of liquor. This study utilized hyperspectral imaging (HSI) technology along with an integrated learning model to rapidly and non-destructively detect protein and fat contents in wheat. The hyperspectral data of the wheat grains was preprocessed using various methods. The competitive adaptive reweighted sampling (CARS) algorithm, the successive projections algorithm (SPA), and the combination of the two algorithms (CARS-SPA) were used to extract the characteristic wavelengths. Three support vector regression (SVR) models were created to forecast wheat protein and fat contents using full-band spectral data and characteristic wavelengths. These models are gray wolf optimized support vector regression (GWO-SVR), particle swarm optimized support vector regression (PSO-SVR), and support vector regression-based integrated learning (AdaBoost-SVR). Comparing the models showed that the AdaBoost-SVR model, utilizing the Eigen wavelengths extracted through the CARS-SPA and CARS algorithms, stood out as the most accurate in predicting wheat protein and fat contents, with Rp 2 values of 0.9789 and 0.9651, respectively. The study findings indicate that combining HSI with AdaBoost-SVR integrated learning modeling enables rapid, non-destructive determination of wheat protein and fat contents with greater accuracy than using a single machine learning (ML) model.

Skin defects are one of the main problems that occur during postharvest grading and processing of loquat, leading to deterioration of loquat, reducing the economic value of the commodity, and causing food quality and safety problems. Improving the identification rate of skin defects in loquat can reduce the economic loss caused by transportation and storage. In this paper, hyperspectral imaging technology was used to collect the reflectance (R), absorbance (A), and Kubelka–Munk (KM) spectra of loquat with skin defects for classification of defects types. Principal component analysis (PCA) was used to obtain the characteristic wavelength images. Competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), uninformative variables elimination (UVE), and Monte Carlo combined with uninformative variables elimination (MCUVE) were used to reduce the dimension of spectral data to obtain the characteristic wavelength. The spectral data and grayscale features were used to establish the spectral model (SPEC) and grayscale features combined with spectral model (MIX). Extreme learning machine (ELM), least squares support vector machine (LS‐SVM), and k‐nearest neighbors (KNN) algorithm were applied to establish a classification model for skin defects in loquat. Comparing the model classification results of the three spectral parameters combined image features, it was found that the A‐CARS‐MIX‐ELM model had the highest accuracy, with a classification accuracy of 98.18%. The number of selected characteristic spectra was 37, accounting for 21.02% of the total spectral number of the whole band. In the online detection process of a large number of fruits, we usually need to improve the detection speed on the premise of high‐precision detection. In this case, the R‐SPA‐MIX‐ELM model can be selected, and the classification accuracy is 94.55%. The number of selected spectra is 10, accounting for 5.68% of the number of wavelengths in the whole band. Consequently, it also provides a theoretical reference for the rapid, nondestructive, and high‐precision fruit online detection technology in the future.

Strawberry anthracnose, caused by Colletotrichum spp., is a major disease that causes tremendous damage to cultivated strawberry plants (Fragaria × ananassa Duch.). Examining and distinguishing plants potentially carrying the pathogen is one of the most effective ways to prevent and control strawberry anthracnose disease. Herein, we used this method on Colletotrichum gloeosporioides at the crown site on indoor strawberry plants and established a classification and distinguishing model based on measurement of the spectral and textural characteristics of the disease-free zone near the disease center. The results, based on the successive projection algorithm (SPA), competitive adaptive reweighted sampling (CARS), and interval random frog (IRF), extracted 5, 14, and 11 characteristic wavelengths, respectively. The SPA extracted fewer effective characteristic wavelengths, while IRF covered more information. A total of 12 dimensional texture features (TFs) were extracted from the first three minimum noise fraction (MNF) images using a grayscale co-occurrence matrix (GLCM). The combined dataset modeling of spectral and TFs performed better than single-feature modeling. The accuracy rates of the IRF + TF + BP model test set for healthy, asymptomatic, and symptomatic samples were 99.1%, 93.5%, and 94.5%, the recall rates were 100%, 94%, and 93%, and the F1 scores were 0.9955, 0.9375, and 0.9374, respectively. The total modeling time was 10.9 s, meaning that this model demonstrated the best comprehensive performance of all the constructed models. The model lays a technical foundation for the early, non-destructive detection of strawberry anthracnose.