Uncertainty assessment for firmness and total soluble solids of sweet cherries using hyperspectral imaging and multivariate statistics

https://doi.org/10.1016/j.jfoodeng.2020.110177Get rights and content

Highlights

  • Hyperspectral imaging was used to predict internal quality attributes of sweet cherries using PLSR and GPR.

  • The prediction uncertainty of PLS and GPR models was discussed.

  • Accurate predictions can be achieved by GPR approach.

Abstract

Quantifying cherry fruit quality parameters is essential to maintaining high quality produce throughout the supply chain as it influences consumer confidence in the product. Hyperspectral imaging offers high potential as a non-destructive and fast analytical tool for estimating various quality parameters in different food products. The objective of the study is to investigate the potential of hyperspectral imaging for quality (total soluble solids concentration, TSS and flesh firmness, FF) assessment in fresh cherry fruits. Partial least squares regression (PLSR) and Gaussian process regression (GPR) was used to evaluate the prediction performance and predictive uncertainty. Test dataset results highlight that GPR can be used to predict TSS (RPDT = 3.04; R2T = 0.88; RMSET = 0.43%) and firmness (RPDT = 2.54; R2T = 0.60; RMSE = 0.38 N) of cherry fruits with high accuracy. In addition, GPR models showed lower uncertainty with a prediction interval coverage probability (PICP) of 0.90–0.97. Overall, hyperspectral imaging combined with multivariate data analysis using GPR can be used as a robust and reliable tool to estimate cherry fruit quality parameters.

Introduction

The sweet cherry (Prunus avium L.) is a widely appreciated stone fruit for its taste, flavour, colour, texture, myriad of nutrients and their health benefits. Sweet cherries are a highly perishable fruit and contain significant amount of nutrients, phenolic antioxidants which reduces the risk of cancer, dietary fibre and anthocyanin (Habib et al., 2017). In recent years, the demand for high quality cherries has been growing rapidly due to unprecedented demand from countries such as China, Vietnam and Thailand (Kuprienko, 2017; Anonymous, 2018; Carson and East, 2018; Anonymous, 2020). Despite high consumer demand for fresh and high-quality cherries, the current grading system faces several challenges including maintenance of high-quality produce and assurance of food safety. Traditionally, assessment of quality and safety of cherry fruits involves human visual inspection and chemical experiments, which are tedious, destructive and expensive. Although there are a few automatic vision-grading systems used in the industry, the system characterises the fruits purely based on external visual quality characteristics such as colour, shape, size and appearance. Hence these vision systems are not sufficient to provide information of internal quality parameters and some internal defects such as bruises and chilling injury. In this modern world, market and consumer's perspective, the internal quality parameters such as dry matter concentration, total soluble solids content (TSS), firmness and acidity are more important and yet cannot be determined by vision systems. This necessitates the demand for technologies that enable assessment of both internal and external quality parameters, which could be of significant benefit to the cherry growers/industry.

Visible/near infrared (Vis-NIR) spectroscopy is recognised as an efficient and effective method for various food quality assessments, with its capability of extracting the internal bio-chemical information of food products such as to quantify firmness and the total soluble solids (TSS) concentration of kiwifruit (Li et al., 2017). This technique studies the spectral property of an object when subject to electromagnetic radiation between 780 and 2500 nm. Spectral absorptions caused by the chemical compositions present in the object. Chemical bonds such as the CH, NH, OH and CO groups are subject to vibrational energy change in forms of stretching or bending when irradiated by NIR light (Marten et al., 1985). These changes correspond to specific absorption bands in the NIR region and are affected by NIR-active chemical groups which are widely present in food media such as a fruit tissue. However, diffusional reflectance spectra of fruit are non-specific because of multiple overlapping absorption features. Therefore, multivariate statistical techniques are required to extract relevant information from a spectrum (Nicolai et al., 2007).

With the advent of optical technology, hyperspectral imaging (HSI) has emerged as a new analytical tool for food quality inspection (Caporaso et al., 2018; Gowen et al., 2007; Reis et al., 2018; Somaratne et al., 2019). HSI is a state-of-the-art technology that combines vision technology and Vis-NIR spectroscopy, which allows the acquisition of spectral and spatial information simultaneously. Vision systems have been shown to be useful for evaluation of various external characteristics of fruits. Recent studies have also shown the feasibility of HSI in spatial mapping of internal quality of fruits and vegetables such as quantifying SSC and firmness for blueberries (Leiva-Valenzuela et al., 2013) and plums (Li et al., 2018a). Although HSI can be implemented in different modes (reflectance, transmittance and interactance), reflectance is most common mode used for analysing food quality (Gowen et al., 2007; Ravikanth et al., 2017) due to its ease of use. Manufacturers of hyperspectral sensors offer flexible and customisable options for the number and resolution of spectral wavebands (Su and Sun, 2018), meaning that lower cost sensors with a selection of useful wavebands built into the system could be developed, making this technology more affordable for industrial applications.

The potential nature of non-destructive, cost-effective and rapid sensing procedures is widely accepted in the fresh produce industry for quality assessment and detection of storage and quality issues. For sweet cherries, most applications focused on developing correlations between spectral data and quality attributes associated with chemical components within the fruit. For instance, DMC and SSC were successfully estimated in sweet cherries using handheld NIR devices (Escribano et al., 2017). Li et al. (2018b) also demonstrated the feasibility of NIR hyperspectral imaging for estimation of TSS and pH of cherry fruit and classifying fruit of different maturity stages. Predictions of textural properties have been more challenging as the physique-chemical relationship between input spectra and textural quality data (e.g. firmness) is complex and obscure and usually requires blackbox model approach (Li, 2017). For example, Lu (2001) used NIR spectroscopy combined with chemometrics to estimate TSS and firmness of sweet cherries and highlighted the difficulty in identifying single wavelengths that strongly correlated with the measured quality attributes. There is potential to improve the interpretability of the model by utilising more advanced data analytical tools. Therefore, in the current study, an attempt was carried out to develop quantitative regression models for both chemical (TSS) and textural (firmness) properties of sweet cherries by employing advanced chemometrics and machine learning algorithms to provide more insight to the underlying relationship between spectral data and quality outcome.

In recent years, a wide range of competent regression algorithms have been developed for analysing hyperspectral images. For instance, partial least squares (PLS) is a widely used regression method in spectroscopic studies because it is computationally fast, accurate and effectively extracts useful information from high dimensional data (Wold et al., 2001). Moreover, nonlinear chemometric methods such as support vector regression (Li et al., 2009), random forest regression (Fawagreh et al., 2014), kernel ridge regression, artificial neural networks (Blanco et al., 2000; Dębska and Guzowska-Świder, 2011), have shown advantages over linear ones because of their high predictive performance. However, these methods require optimization of several model parameters to achieve robust performance. In comparison, gaussian process regression (GPR) is a nonparametric Bayesian approach flexible in exploring wide variety of relations between input and output by optimising several parameters and capable of producing better estimates (Rasmussen, 2006; Verrelst et al., 2011). It also provides uncertainty intervals for prediction estimates which is essential for determining model quality. A study by Verrelst et al. (2012) compared the performance of GPR model with support vector regression (SVR), kernel ridge regression (KRR) and artificial neural network (ANN) and achieved improved results with GPR. Chen et al. (2007) have tested the capability of GPR for linear and non-linear spectral data, where GPR yielded improved accuracy over PLS and ANN with satisfactory results. Consequently, we have used both PLS and Gaussian process regression in this paper to evaluate the performance of HSI for quality prediction of cherries. Generally, predictive models are associated with some extent of uncertainty which influence the performance of the model to carry out predictions on unknown samples. It is essential to describe the level of uncertainty in order to indicate model robustness and reliability. Obtaining uncertainty estimates using PLS and GPR is achievable due to their inherent algorithm architecture. To the best of the author's knowledge, there is no research dealing with regression uncertainty for predicting cherry fruit quality.

This study aims to investigate (i) the potential application of HSI combined with advanced data analytical tools for predicting TSS and firmness of commercial sweet cherries, (ii) quantifying prediction uncertainties using PLS and GPR models.

Section snippets

Materials and methods

A total of 350 fresh cherries were harvested on February 25, 2019 from a commercial orchard located on the South Island, New Zealand. After harvest, fruits were transported by air freight to Massey University, Palmerston North and stored in a refrigerated cool room at 4 °C with 95% relative humidity for 2 days. Prior to hyperspectral image acquisition, the cherries were left at room temperature for a period of 20 h to allow equilibrium of fruit temperature and to avoid condensation on the

Results

The descriptive statistics of training and test datasets (Table 1) showed that the size of the training population was three times of test dataset and both datasets shared similar amounts of variation for both quality parameters of cherry fruits. As we attempted to maintain a typical range of export quality cherry fruits, the variation (1.31 ≤ std ≤ 1.40; 6.51≤ CV ≤ 6.91) in TSS is lower than the variation of firmness values (0.92 ≤ std ≤ 1.00; 20% ≤ CV ≤ 23%). The average first derivative

Discussion

As illustrated in Fig. 2, the spectral changes relevant to quality parameters were visually not clear but the quality relevant spectral features might be masked by other dominant parameters such as water (Osborne, 2000), structure and colour (Cen and He, 2007). Therefore, the use of multivariate data analysis techniques is necessary.

The TSS was predicted with high accuracy (RPD = 3.33) which was much better than the findings of Li et al. (2018b) where an RPD was 2.70. However, they used a

Conclusion

In this study, the use of hyperspectral imaging and multivariate data analysis to estimate internal quality parameters of cherries was investigated. The results showed that the TSS and firmness of cherry can be estimated with high accuracy using gaussian process regression combined with image analysis. Compared to PLSR approach, the uncertainty in GPR models was significantly lower. This indicates GPR has the potential to be effectively used for retrieving accurate estimates from hyperspectral

CRediT authorship contribution statement

Reddy R. Pullanagari: Conceptualization, Project administration, Funding acquisition, Methodology, Formal analysis, Investigation, Validation, Visualization, Writing - original draft. Mo Li: Methodology, Writing - review & editing, Writing - original draft.

Acknowledgement

This research was funded by Massey University Research Funding (MURF), New Zealand. We acknowledge “New Zealand Cherry Corporation” for providing export quality grade cherries. The authors would like to thank Sebastian Rivera, Abhimanyu Singh Garhwal and Praveen Veeregowda for their support in quality measurements.

References (57)

  • Y. Huang et al.

    Prediction of tomato firmness using spatially-resolved spectroscopy

    Postharvest Biol. Technol.

    (2018)
  • G.A. Leiva-Valenzuela et al.

    Prediction of firmness and soluble solids content of blueberries using hyperspectral reflectance imaging

    J. Food Eng.

    (2013)
  • B. Li et al.

    Application of hyperspectral imaging for nondestructive measurement of plum quality attributes

    Postharvest Biol. Technol.

    (2018)
  • H. Li et al.

    Support vector machines and its applications in chemistry

    Chemometr. Intell. Lab. Syst.

    (2009)
  • X. Li et al.

    SSC and pH for sweet assessment and maturity classification of harvested cherry fruit based on NIR hyperspectral imaging technology

    Postharvest Biol. Technol.

    (2018)
  • B.P. Malone et al.

    Comparing regression-based digital soil mapping and multiple-point geostatistics for the spatial extrapolation of soil data

    Geoderma

    (2016)
  • K. Ncama et al.

    Application of Vis/NIR spectroscopy for predicting sweetness and flavour parameters of ‘Valencia’orange (Citrus sinensis) and ‘Star Ruby’grapefruit (Citrus x paradisi Macfad)

    J. Food Eng.

    (2017)
  • B.M. Nicolai et al.

    Nondestructive measurement of fruit and vegetable quality by means of NIR spectroscopy: a review

    Postharvest Biol. Technol.

    (2007)
  • J. Qin et al.

    Measurement of the optical properties of fruits and vegetables using spatially resolved hyperspectral diffuse reflectance imaging technique

    Postharvest Biol. Technol.

    (2008)
  • O. Rahmati et al.

    Predicting uncertainty of machine learning models for modelling nitrate pollution of groundwater using quantile regression and UNEEC methods

    Sci. Total Environ.

    (2019)
  • M.M. Reis et al.

    Chemometrics and hyperspectral imaging applied to assessment of chemical, textural and structural characteristics of meat

    Meat Sci.

    (2018)
  • W. Saeys et al.

    Potential for onsite and online analysis of pig manure using visible and near infrared reflectance spectroscopy

    Biosyst. Eng.

    (2005)
  • M. Uddin et al.

    Nondestructive determination of water and protein in surimi by near-infrared spectroscopy

    Food Chem.

    (2006)
  • J. Verrelst et al.

    Machine learning regression algorithms for biophysical parameter retrieval: opportunities for Sentinel-2 and-3

    Remote Sens. Environ.

    (2012)
  • J. Verrelst et al.

    Gaussian processes uncertainty estimates in experimental Sentinel-2 LAI and leaf chlorophyll content retrieval

    ISPRS J. Photogramm.

    (2013)
  • R.A. Viscarra Rossel

    ParLeS: software for chemometric analysis of spectroscopic data

    Chemometr. Intell. Lab. Syst.

    (2008)
  • F. Westad et al.

    Validation of chemometric models – a tutorial

    Anal. Chim. Acta

    (2015)
  • S. Wold et al.

    PLS-regression: a basic tool of chemometrics

    Chemometr. Intell. Lab. Syst.

    (2001)
  • Cited by (46)

    View all citing articles on Scopus
    View full text