Chinese chestnuts, members of the Fagaceae family, provide ecological and economic benefits and have been cultivated in China for over three millennia (Barakat et al., 2009). Because of their distinctive taste, chestnuts are highly sought after in international markets, which contain lower levels of fat and protein but higher amounts of water and carbohydrates compared to other nuts (Desmaison et al., 1984; Ümran et al., 2006). This unique property allows chestnuts to be consumed either raw or cooked, as well as processed into various forms such as canned foods, cakes, and confectioneries (Senem et al., 2021).

Sweetness is a crucial component of fruit flavor and plays an integral role in fruit quality formation (Wang et al., 2022). Soluble sugar content and composition are the main determinants of fruit sweetness, which not only effectively modulates the human taste system but also provides beneficial carbohydrates to humans (Vânia et al., 2021; Alessandra et al., 2022). Therefore, the content and composition of soluble sugars are essential for the evaluation of fruit quality (Li M. et al., 2018). Chestnuts are rich in soluble sugars, mainly consisting of sucrose, fructose and glucose, and their sweetness also depends on the content of these three sugars. However, their content is influenced by factors such as variety, growing location, and other environmental conditions (Freinkel, 2009; Pereira-Lorenzo et al., 2010). Genetic selection is necessary to obtain good quality chestnuts and to reduce the impact of environmental factors on soluble sugars. However genetic selection usually depends on extensive experiments and large sample sizes (Couture et al., 2016). While traditional methods for determining soluble sugar content in chestnuts yield accurate measurements, they require kernel peeling and drying, rendering them time-consuming, laborious, and destructive, thereby rendering them unsuitable for large-scale genetic selection. Therefore, there is a pressing need to develop rapid, non-destructive, and straightforward methods for determining soluble sugars in chestnuts to expedite the progress of genetic selection.

With the advancements in chemometric methods and spectroscopic instrument hardware technology, spectroscopic analysis has become a mainstream technique for non-destructive detection of the internal quality of fruits (Maria et al., 2021). Hyperspectral imaging technology is a rapid, eco-friendly, non-destructive, and efficient detection method (Yu et al., 2009; Yang et al., 2020). It reflects the absorption information of molecules, such as hydrogen-containing groups like C-H, O-H, and N-H, in the ensemble and multiplicity frequencies. When combined with chemometric methods, spectroscopy can perform qualitative and quantitative analysis of relevant chemical components (Luypaert et al., 2007; Shi and Yu, 2017), it has been widely used in various fields.

However, the spectral information obtained directly using spectral analysis techniques often contains noise and other irrelevant information that can affect model building. To filter out useful spectral information and ensure model stability, preprocessing of the spectra is necessary (Du et al., 2020; Albanell et al., 2021). Commonly used spectral preprocessing methods include smoothing, first-order derivative, second-order derivative, standard normal variable transformation, batch normalization and so on (Gai et al., 2022; Xiao et al., 2022).

Furthermore, spectral data often exhibit problems such as wide spectral bands, overlapping absorption peaks, serious co-linearity between adjacent bands, and contain a large amount of redundant information (Webb et al., 2020). To reduce the modeling wavenumber, simplify the model, and improve model prediction accuracy, it is necessary to select spectral variables and remove noise and interference variables that are independent of the target attribute before establishing the model (Li and Zhao, 2019). Successive projections algorithm (SPA), Monte Carlo uninformative variable elimination (MCUVE), uninformative variable elimination (UVE) and randomized tests (RT) are common variable selection algorithms (Li P. et al., 2018; Cheng et al., 2020).

The establishment of a stable and accurate quantitative analysis model is of utmost importance in the application of spectral analysis techniques. Commonly used methods for model establishment include PLS, principal component regression (PCR), multiple linear regression (MLR), and artificial neural networks (ANN) (Hein, 2010; Yang et al., 2018). Among these methods, PLS is widely used and is particularly effective in handling regression relationships between multiple variables when variables are highly correlated. It is able to effectively solve the problem of many variables and small sample size (Keshav, 2021). The PLS method combines the advantages of principal component regression and multiple linear regression methods, with the smallest sum of squares of errors and a relatively simple model with high predictive accuracy (Mehmood et al., 2020; Beyaztas and Lin, 2022).

In recent years, hyperspectral imaging techniques have been widely used to detect quality parameters in fruits and vegetables, such as apples (Zhang Y. et al., 2021), peaches (Jiang et al., 2021), carrots and tomatoes (Roberto et al., 2018). However, no hyperspectral techniques have been reported for the detection of quality parameters in chestnuts to date.

In the present study, a model for the prediction of soluble sugars in chestnuts was developed using hyperspectral imaging combined with chemometrics. The aims of the study were to (1) investigate the potential of spectroscopic techniques in quantifying the soluble sugar content of chestnuts and establish an optimal prediction model; (2) identify the best pre-treatment method for soluble sugars during model building; and (3) determine the most important wavelengths associated with the soluble sugar content of chestnuts for non-destructive detection during model calibration.

As an emerging technology, hyperspectral imaging can acquire both image and spectral information of samples, which is efficient, fast and non-destructive, and is widely used in medicine, agriculture and food quality (Halicek et al., 2019; Lu et al., 2020; Zhu et al., 2020). In agriculture, hyperspectral imaging technology can be used as a non-destructive inspection tool, which has a broad application in fruit and vegetable quality monitoring and variety selection (Faqeerzada et al., 2020). Hyperspectral imaging has already been used for the detection and grading of apple bruises (Tan et al., 2018), the assessment of kiwi ripeness (Serranti et al., 2018), and the analytical determination of soluble solids content in oranges (Zhang et al., 2020a). In this study, a hyperspectral camera was used to obtain spectral and image information of chestnuts. The aim was to explore the feasibility of hyperspectral image techniques in detecting the soluble sugar content of chestnuts. Soluble sugars play a crucial role in plant growth and development, serving not only as signaling molecules but also as key indicators of fruit quality due to their content and fraction within the fruit (León and Sheen, 2003; Tang et al., 2021). Consequently, developing a rapid, convenient, and non-destructive method for determining soluble sugar content is essential for fruit quality assessment and new variety selection.

To this end, this study examined the relationship between soluble sugars and hyperspectral images by integrating BS preprocessing methods, original spectra, and three commonly used machine learning techniques. The results indicated that the optimal prediction model combined the BS preprocessing method with the rep variable selection method, achieving R² and RMSE values of 0.59 and 1.02, respectively. These values were lower than the R² and RMSE values obtained when predicting the soluble sugar content of sweet maize (R² = 0.8431, RMSE = 5.8292) (Yang et al., 2020). The discrepancy may be attributed to the freshness of the chestnut used for spectral data collection, their high water content, and the complex texture and composition of the chestnut kernels. The strong absorption of water in the NIR spectral band, in conjunction with other compounds present in the fruit, renders the collected spectra highly intricate (He et al., 2021). Moreover, the collected spectra encompass not only absorption information related to the chemical composition but also scattering information pertaining to the internal fruit structure, complicating the accurate assignment of specific absorption bands to particular chemical components (Yuan et al., 2022). This resembles the prediction of soluble solids content by PLSR models based on visible/near-infrared spectroscopy of lychee, which may also be due to the non-uniform thickness and roughness of the lychee peel, as well as the complex composition of the flesh interior, resulting in limited model prediction accuracy (Pu et al., 2016).

It has been demonstrated that sugars are not uniformly distributed in fruits; therefore, collecting spectral information from multiple regions of the fruit surface can enhance the accuracy of predictive models (Jiang et al., 2021; Tan et al., 2022). Hyperspectral imaging, a novel method that merges spectroscopy and image analysis, yields both image and spectral information for a sample, providing a comprehensive depiction of its shape, texture, and intrinsic characteristics (Quan et al., 2014; Zhang et al., 2020b). However, hyperspectral imaging is substantially influenced by the sample’s curvature, and the acquired hyperspectral image data often contain numerous non-linearities. These factors necessitate hyperspectral data processing for reliable extraction of sample properties (Luka et al., 2021). Various treatments possess distinct characteristics. For example, SNV primarily eliminates scattering phenomena and filters the impact of optical range changes on the spectral signal during the experiment (Bi et al., 2016). The derivative removes baseline drift and background interference, enhances spectral differences, and improves resolution (Xu et al., 2008). Detrend eliminates baseline drift and curvature interference from the spectral signal (Luypaert et al., 2003). In practice, the appropriate processing method should be chosen based on the spectral data (Luo et al., 2020).

In a study on the rapid prediction of sugar content in Dangshan pears, the influence of fruit shape caused significant fluctuations in the fruit surface’s reflectance, preventing accurate spectral information from being obtained across the entire fruit surface. Mean normalization of the spectra partially mitigated the impact of fruit shape on the obtained spectra, improved spectral intensity differences, and yielded a more uniform grayscale distribution compared to the original (Zhang et al., 2018). In this study, the spectral intensity of chestnuts was also affected by the curvature of the hemispherical surface. To minimize the impact of curvature and other interfering information, the spectra were processed using BS. BS can identify information in different blocks, extract complementary information from them, balance the influence of building modules, and prevent any block from dominating the model, thereby reducing the effects of interfering information such as scattering, fruit shape, etc. (Campos and Reis, 2020; Puneet et al., 2021).

As shown in Figure 5, it is evident that the BS-processed spectral data are much more tightly clustered, and the absorption peaks become more pronounced compared to the original average spectra, although the absorption peaks exhibit a broader spectral range. Spectra provide light intensity values at each increment of the wavelength range, resulting in a large number of variables that necessitate considerable computational power to combine (Esposito and Houser, 2019). Direct modeling of the original full spectrum not only replicates the structure and incurs significant costs, but also negatively impacts the model’s stability. Therefore, it is essential to extract valid wavelength information as needed (Qi and Fu, 2022). These selected effective wavelengths reduce data dimensionality, contain the most critical information related to the sample traits, and can replace the original full spectrum for modeling. This approach increases the model accuracy and robustness while accelerating data computational speed and reducing the computational cost of the generated model (Saputro and Handayani, 2017; Zhang J. et al., 2021). In this experiment, variable selection based on BS pretreatment using bve, rep, and sMC was combined with PLS to construct PLS models for predicting soluble sugar content in chestnuts. The results of each model are shown in Table 1. Models built with bve and sMC selected wavelengths exhibited lower accuracy than those constructed using the original full spectrum, which may be attributed to the elimination of some information variables related to soluble sugars during the variable selection covariance removal process. This is similar to the pear fruit sugar content prediction model building, where models using SPA-selected variables were built with lower accuracy than full-spectrum modeling (Zhang et al., 2018). The model built using the variables selected by the rep method demonstrated marginally better performance than the model built from the original full spectrum. This improvement is because the problems of covariance and overfitting are alleviated by using only the effective wavelengths, eliminating redundant wavelengths that do not carry much spectral information, and reducing the amount of operations (Aquinocruz et al., 2021). This has also been demonstrated in non-destructive testing of strawberry quality attributes, where models built using selected effective wavelengths have higher predictive performance than those built from the original full spectrum (ElMasry et al., 2007). In this study, the important variables selected closely related to soluble sugars in chestnuts were 464.3, 503.2, 539.3, 579, and 711.3 nm, similar to those selected in bananas, namely 440, 525, 633, 672, 709, 760, 925, and 984 nm (Rajkumar et al., 2012). Furthermore, the residual analysis of the best model and the original spectral model was performed in this experiment. As seen in Figures 6C, D, the residuals of the best model were more evenly distributed and had a narrower bandwidth in the −2 to 2 interval compared to the original spectrum. It has been found that if the residual values are evenly distributed in the horizontal band and have a narrower bandwidth, this indicates that the chosen model is more appropriate and has better fitting accuracy (Couture et al., 2016).

The presented findings demonstrate the feasibility of utilizing hyperspectral imaging for predicting soluble sugar content in chestnuts. However, further improvements are required to refine the model and achieve optimal performance. Therefore, a comprehensive investigation of factors affecting spectral analysis and mitigation of their negative effects is necessary for enhancing subsequent research. Additionally, the exploration of a wider range of combinations of spectral processing and modeling methods is recommended to augment the sample.

This study explored the feasibility of using hyperspectral techniques to predict the soluble sugar content in chestnut fruits. Soluble sugars play a crucial role in determining chestnut fruit quality, as they influence both nutrition and flavor. The study applied various preprocessing methods to the spectral data, including SNV, derivative, DET, and BS, and employed machine learning techniques to establish a prediction model for soluble sugar content. The results indicated that the BS pretreatment method combined with rep variable selection yielded the optimal prediction model, with an R² value of 0.59 and an RMSE value of 1.02. The selected model demonstrated higher accuracy in fitting the regression equation, suggesting that it was more appropriate and reliable for predicting soluble sugar content in chestnuts.

The original contributions presented in this study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

JY conducted the experiment and wrote the manuscript. XJ designed the study and revised the manuscript. BG revised and corrected the manuscript. All authors read and approved the final manuscript.