Prediction mapping of physicochemical properties in mango by hyperspectral imaging

Publication Type: Journal contribution (peer reviewed)
Authors: Rungpichayapichet, P; Nagle, M; Yuwanbun, P; Khuwijitjaru, P; Mahayothee, B; Müller, J
Year of publication: 2017
Published in: Biosystems Engineering
Editor: Elsevier
Band/Volume: 159/
DOI: 10.1016/j.biosystemseng.2017.04.006
Page (from - to): 109-120
Keywords: Chemometrics, fruit quality, Hyperspectral imaging, Mangifera indica, vis-NIR spectroscopy

Abstract

Hyperspectral imaging (HSI) techniques using a newly-developed frame camera were applied to determine internal properties of mango fruits including firmness, total soluble solids (TSS) and titratable acidity (TA). Prediction models were developed using spectral data from relative surface reflectance of 160 fruits in the visible and near infrared (vis/NIR) region of 450–998 nm analysed by PLS regression. For data reduction, MLR analysis showed 16 significant factors for firmness, 17 for TA, and 20 for TSS. The results of MLR did not substantially affect the prediction performance as compared to PLS. An original approach with combined chemometric and HSI data analyses was applied using R programming. Significant correlations were found between HSI data and firmness (R2 = 0.81 and RMSE = 2.83 N) followed by TA (R2 = 0.81 and RMSE = 0.24%) and TSS (R2 = 0.5 and RMSE = 2.0%). Prediction maps of physicochemical qualities were achieved by applying the prediction models to each pixel of HSI to visualise their spatial distribution. The variation of firmness, TSS, and TA within the fruit indicated fruit ripening started from shoulder toward to tip part. From these results, HSI can be used as a non-destructive technique for determining the quality of fruits which could potentially enhance grading capabilities in the industrial handling and processing of mango.Graphical abstract Figure optionsKeywords<ul><li>Mangifera indica;</li><li>Fruit quality;</li><li>Hyperspectral imaging;</li><li>Chemometrics;</li><li>Spectral mapping;</li><li>vis-NIR spectroscopy</li></ul>1. IntroductionMango quality plays an important role on market price and consumer satisfaction. Superior quality induces a customer's willingness to pay premium prices for a product. Overall, physical appearances such as colour and firmness are considered the primary criteria customers use to evaluate fruit quality (Delwiche, Mekwatanakarn, & Wand, 2008). Sivakumar, Jiang, and Yahia (2011) have meanwhile reported that consumer preference and order decisions are more related to fruit taste. However, in commercial operations, heterogeneous fruit quality is encountered due to the variation in degree of ripeness of harvested fruits, which causes non-uniform ripening behaviour and requires different postharvest handling (Kienzle et al., 2012). This problem leads to reduced shelf-life and postharvest losses during handling and transport. Therefore, inspection and controlling quality of mango at packing is an imperative process in the mango industry to assure that only superior fruits are distributed to high-value markets.The potential of non-destructive techniques as tools for sorting and grading have been demonstrated in numerous studies. Near infrared spectroscopy (NIRS) is widely implemented to identify the quality of many fruits, including mango. Studies have shown the potential of NIRS to determine the physicochemical properties of mango such as firmness, total soluble solids content (TSS), titratable acidity (TA) and bioactive compounds (Jha et al., 2010; Nagle et al., 2010; Rungpichayapichet et al., 2015; Rungpichayapichet et al., 2016 ;  Subedi and Walsh, 2011; Theanjumpol et al., 2014 ;  Valente et al., 2009; Watanawan, Wasusri, Srilaong, Wongs-Aree, & Kanlayanarat, 2014). More recently, hyperspectral imaging (HSI) is being regarded as an effective non-destructive method to determine the quality of agricultural products. The advantages of HSI over NIRS are the simultaneous procurement of spatial and spectral information from the sample and the flexibility of area selection for spectral extraction after data acquisition (Schmilovitch et al., 2014 ;  Zhang et al., 2013). Overall, HSI can provide information on the spatial distribution of physicochemical parameters, enhancing the perception of quality changes within samples (ElMasry, Sun, & Allen, 2013). The ability of HSI to evaluate the quality of fruits was recently reviewed by Magwaza and Opara (2015). For mango, HSI has been applied to detect fruit fly infestation (Haff et al., 2013 ;  Saranwong et al., 2011), mechanical damage (Vélez Rivera et al., 2014), moisture distribution in dried mango slices (Pu & Sun, 2015) and postharvest qualities including colour, firmness and TSS (Makino et al., 2013 ;  Sivakumar, 2006). However, little documentation currently exists describing the variability of physicochemical qualities within mango fruits and studies applying HSI to monitor these spatial quality changes are lacking. Issues also remain with software for HSI analysis, especially with respect to real-time application (ElMasry & Nakauchi, 2016) and high computational requirements for HSI data. Therefore, this study aims to demonstrate the feasibility of HSI as a tool to observe the spatial changes of quality of mango during ripening by (i) testing hyperspectral frame camera technology for rapid image acquisition, (ii) developing reduced-factor HSI prediction models in freely-available, programmable software to quantitatively determine physicochemical qualities of mango, and (iii) establishing prediction maps showing the distribution pattern of quality parameters within fruits for potential grading applications.2. Materials and methods2.1. MaterialsA total of 160 mango fruits (cv. Nam Dokmai, subcv. Si Thong) from two production areas (Phitsanulok and Phetchabun provinces, Thailand) were obtained at commercial export maturity, corresponding to 120 days after full bloom. Eighty fruits from each area were harvested during the peak season in April 2013. Mature fruits were preselected in the field for fruit mass about 310 g and density greater than 1000 kg m−3. Fruits then were transported to Nakhon Pathom province within one day, where the experiments were carried out. Fruits free from visual defects were washed, air-dried, and stored in baskets covered with paper at ambient temperature (32.1 ± 0.6 °C and 73.0 ± 7.9% relative humidity) for up to 8 days of ripening. Ten fruits were randomly selected each day and subjected to HSI analysis and quality reference measurements.2.2. Reference analyses of quality attributesFruits were cut along the sagittal plane from the proximal to distal end into two slabs on either side of the endocarp. Slabs were divided into three positions, i.e. shoulder, cheek, and tip parts (Fig. 1a). Firmness was determined only at the cheek position according to Jha, Kingsly, and Chopra (2006). A puncture test was performed with a needle probe (2 mm diameter) on the peeled mesocarp at a speed of 0.5 mm s−1 and a compression depth of 10 mm using a texture analyser (TA-xT2i, Stable Micro Systems, Godalming, UK). Maximum force (N) was used to describe the firmness. For chemical analyses, each mesocarp section was homogenised separately (Ultra-Turrax T25 basic, IKA Labortechnik, Germany). Five grams of the homogenate was squeezed through filter cloth and used to determine TSS by a digital refractometer (PAL-1, Atago, Tokyo, Japan). TA was measured by alkaline titration of 100 mL dilution prepared from 5 g of homogenate with 0.1 N NaOH until a pH of 8.1 and expressed as mass percentage (%) of citric acid (AOAC, 1999). Determinations of TSS and TA were done in duplicate.Fig. 1. Fruit positions for (a) reference quality analyses sampling and (b) region of interest (ROI) selection for hyperspectral measurements; ROIs are not to scale.Figure options2.3. Hyperspectral system and image acquisitionA HSI measurement system was setup for spectral data acquisition in this study (Fig. 2), which was similar to the equipment and methods used by Jung, Vohland, and Thiele-Bruhn (2015). The system consisted of (i) a sample holder (ii) a 50 W tungsten halogen lamp (Pro Lamp, ASD Inc. CO, USA) and (iii) a hyperspectral frame camera UHD 285 (Cubert GmbH, Ulm, Germany), which has been developed for the advantage of real-time data acquisition (Hagen, Kester, Gao, & Tkaczyk, 2012). The distance between the camera and the sample was fixed at 700 mm. The hyperspectral imaging system was controlled by a PC with the Cubert GUI (Cubert GmbH, Ulm, Germany). Prior to image acquisition, the light source was switched on for 15 min to stabilise illumination. In addition, two light-diffusers (150 × 250 mm) were placed nearby the sample holder to provide uniform light distribution. Images were acquired at a spatial dimension of 910 × 900 pixels, with a respective pixel size of 0.32 × 0.32 mm representing a pixel area of 0.10 mm2. The integration time required for capturing the complete hyperspectral image was 10 ms. Spectral data included 138 bands covering the spectral range 450–998 nm in 4 nm increments with a spectral resolution of 8 nm. Dark reference was measured with the shutter closed while covering of the camera lens with black paperboard. A professional light reflector disc made from Teflon-coated nylon fabric was used as white reference. Image correction was done automatically by Cubert software according to Eq. (1).equation(1)%RI=IS(x,y)|ID(x,y)IR(x,y)−ID(x,y)×100Turn MathJax onwhere RI is the relative reflectance at pixel location (x, y) in percentage, IS is the sample image value, ID is the dark frame image value and IR is the white reference image value ( Leiva-Valenzuela, Lu, & Aguilera, 2014).Fig. 2. Hyperspectral imaging system: (a) tungsten halogen lamp, (b) hyperspectral frame camera with CCD detector, (c) mirrors, (d) sample holder and (e) computer with data acquisition software.Figure options2.4. Data analysis2.4.1. Spectral analysisThe spectral information was extracted from hyperspectral image file using HyperSpec interface (Beleites & Sergo, 2012) in the R environment (version 3.1.0, R Core Team, 2014). Ten region of interest (ROI) areas of 10 ✕ 10 pixels were defined in each of the fruit segments (Fig. 1b). The reflectance spectrum of each ROI was averaged from the corresponding 100 pixels. The HSI data from the three fruit sections (shoulder, cheek, and tip), representing the respective reference measurements, were used to develop the calibration models. In addition, RGB and monochromatic images at frequencies corresponding to chlorophyll (680 nm) and sucrose (913 nm) were generated from the Cubert software provided with the camera. Images were converted to greyscale in ImageJ software (Rasband, 2014) where the fruit area was automatically identified and subsequent light intensity (IL) of the fruit surface, as a mean of pixel values ranging from 0 (black) to 255 (white) was calculated by the software.2.4.2. Regression model developmentA total of total of 4800 reflectance spectra were obtained from 160 fruits (10 spectra from each section resulting in 30 spectra per fruit). To establish calibration and validation sets, fruit samples were sorted according to their physicochemical values. The two highest and two lowest values were assigned to the calibration set. After that, two-thirds of the samples were fixed as calibration data and one-third of the samples were defined as validation data in a 2:1 leave-one-out procedure. Thus, calibration data was established from 3240 reflectance spectra (67.5% of the measured spectra) using full-cross validation and the remaining 1560 reflectance spectra (32.5%) were set as external validation samples.Different mathematical pre-processing of the spectra were performed during building of prediction models to improve the performance, including Savitzky–Golay smoothing (SG), Savitzky–Golay second derivative (SG″) using 15 points with second order polynomial regression (Savitzky & Golay, 1964), standard normal variate (SNV) transformation and multiplicative scatter correction (MSC) using the ‘prospectr’ package for R (Stevens & Ramirez-Lopez, 2013). Models were formulated to relative reflectance spectra and reference values for each fruit part by applying partial least squares (PLS) regression using the full cross-validation method (Mevik & Wehrens, 2007). Multifactorial analysis (backward-elimination-stepwise method) was used for selecting significant wavelengths. Candidate variables were chosen from wavelengths which were found to be related with quality by PLS regression coefficient plots. Multiple linear regression (MLR) was used to develop the reduced-factor models. Performance of the prediction models was evaluated using coefficient of determination (R2), root mean square error (RMSE) and bias. The model fit was assessed based on high R2, low RMSE and low bias. These parameters are defined as follows, according to Esbensen (2004, chap. 9):equation(2)RMSE=1n∑i=1n(y&circ;i−yi)2Turn MathJax onequation(3)bias=1n∑i=1n(y&circ;i−yi)Turn MathJax onwhere y&circ;i is the predicted values calculated by the calibration equation; yiyi is the measured values obtained from the reference analysis; n is the number of samples used in the calibration (RMSEcal) and validation (RMSEval), respectively. The number of PLS components was defined based on combination of high R2 together with low RMSEcal and RMSEvalrepresenting goodness-of-fit for the prediction models.2.4.3. Hyperspectral classification of physicochemical propertiesPrediction maps of firmness, TA and TSS were generated using graphics functions programmed in R language by applying the PLS prediction model to each pixel with its own corresponding spectrum. The prediction results of each pixel have been displayed and plotted according to the constituent of the sample based on their spectral signature to provide the distribution gradient in the sample (ElMasry et al., 2013). Prediction images of mango quality during ripening were established by applying the best prediction model to all pixels in segmented HSI images (background removed). Spatial distributions of physicochemical attributes were illustrated by colour gradients which were used to depict respective values in the fruit samples. Prediction maps for the three predicted parameters required 9.5 s to generate using the programmed functions. Discriminant analysis (DA) was applied to HSI spectral data using SPSS software (ver. 16.0, SPSS, Inc., Chicago, IL) to classify fruits according to ripeness. Mean spectra from three positions were subjected to DA and the classification model was validated using cross-validation method.2.4.4. Statistical analysesStatistical analyses were performed using SAS program (version 9.3, SAS Inst., New Jersey, USA). Data was assessed for correlations between the individual quality parameters. Relationships were evaluated using Pearson correlation coefficient and level of significance (p value). General linear model program (GLM) and least significant difference (LSD) were applied to determine significant differences between chemical compositions for each fruit position using ripening day as blocks. The results were assessed at the significance level p ≤ 0.05. The difference between laboratory reference methods (measured) and HSI-predicted quality parameters were analysed by conducting paired t-tests.3. Results3.1. Physicochemical changes during ripeningPhysicochemical changes of mango during ripening (also respective to fruit position) are shown in Fig. 3. Firmness was found to clearly decrease in the initial phase of ripening and decline more gradually in the final phase. Significant differences of firmness were found from the first to fifth day of ripening, decreasing from 25.04 to 8.45 N and then further declining to 6.32 N (Fig. 3a). Ketsa, Chidtragool, Klein, and Lurie (1998) reported that softening of mango pulp takes place by breakdown of the structurally-bound pectin in the cell wall into water-soluble pectin by β-galactosidase and polygalacturonase. In addition, firmness decline can be augmented by starch deterioration during ripening (Yashoda, Prabha, & Tharanathan, 2006). TA sharply declined in the first phase of ripening and thereafter reduced slightly over the end phase (Fig. 3a and d). Significant reductions in TA (85%) were found over the first four days of ripening, from 1.65 to 0.25% (Fig. 3b). Vélez Rivera et al. (2014) reported that acid degradation of mango is mainly caused by metabolism of citric acid. Figure 3c illustrates the increase of TSS during ripening. Significant differences were found on first three days of ripening, when TSS rose sharply from 13.5 to 18.8%, then continued to moderately increase to a maximum of 19.5% on the sixth day of ripening (Fig. 3c). Increasing sweetness during ripening is a result of starch hydrolysis by amylase, whereas slight decreases in the final phase may be due to conversion by respiration or fermentation during senescence (Jha et al., 2006; Medlicott and Thompson, 1985 ;  Palafox-Carlos et al., 2012). With respect to fruit position, variations of TA and TSS in individual fruits were clearly observed (Fig. 3d and e). Significant differences of TA at different fruit parts were found during the initial phase of ripening (Fig. 3d). The highest acidity was observed at the tip part, while the shoulder part was significantly lower in concentration of organic acids. Shoulder parts had the highest TSS contents, but did not significantly differ from the tip parts. The results indicated that overall ripening process of mango cv ‘Nam Dokmai’ progressed from the shoulder toward the tip part. Results also showed that all three quality parameters were significantly correlated with one another at p < 0.001. The clearest relationship was TA:F with r = 0.93, while correlations with TSS were weaker (r = −0.66 for TSS:TA and r = −0.62 for TSS:F).Fig. 3. Mean values of (a) firmness, (b) TA, and (c) TSS as well as differences in (d) TA and (e) TSS within fruit samples during ripening.Figure options3.2. Reflectance imagingRGB images, monochromatic images at 680 and 913 nm and the average original reflectance spectra of mango are shown in Fig. 4. A characteristic of the Nom Dokmai cultivar is the pale yellow (in the web version) peel colour at the unripe stage that changes to intense yellow during ripening, which is evident in the RGB images. The images at 680 and 913 nm are shown for monitoring the changes of chlorophyll and sugar content during ripening, respectively, since these wavelengths are responsive to the corresponding compounds (ElMasry et al., 2007 ;  McGlone and Kawano, 1998). Differences between the images at 680 and 913 nm were not clearly seen at the same ripening day, while noticeable changes were found between ripening days. A relative reflectance was observed in fruits at 680 nm on the first day of ripening. Mango at the unripe stage showed higher reflectance as seen by the brighter image as compared to ripened samples. For example, IL at 680 nm changed from 121 to 101, whereas IL at 913 nm decreased from 111 to 80 during ripening. A higher relative reflectance spectrum was acquired from the cheek section, while the lowest was found on tip section. The result corresponded with reference measurements, whereby the cheek part contained the lowest TSS content. In addition, correlations between IL and physicochemical properties were examined. Significant correlations (p < 0.05) were found between IL at 680 nm and TSS, TA and firmness with correlation coefficients (R) of −0.3, 0.53 and 0.68, respectively. Unexpected non-significant correlation was found between IL at 913 nm and TSS (p ≥ 0.05), while TA and firmness showed weak correlations with IL with R of 0.34 and 0.57, respectively. Therefore, using the changes of IL at these defined wavelengths was found to be inadequate for predicting the physicochemical properties of mango.Fig. 4. RGB images, monochromatic images at 680 and 913 nm and reflectance spectra at 1, 4, 6 and 8 days of ripening.Figure options3.3. Prediction models for physicochemical propertiesTable 1 shows the statistics of physicochemical properties in the calibration and validation sets for PLS analysis and regression results for mango qualities are summarised in Table 2. Scatter plots showing the predicted TSS, TA and firmness versus measured values are illustrated in Fig. 5. As shown in Table 2, prediction of fruit firmness by HSI was possible with R2 of 0.77–0.81 and RMSEval of 2.83–3.15 N, and a slightly positive bias of 0.18–0.23 N. Calibration models for firmness developed from SG pre-processed spectra required a relatively higher number of PLS components to achieve comparable accuracy to other modelling techniques. Best prediction ability for firmness was found for the model from spectra pre-processed with MSC at R2 of 0.81 and RMSEval of 2.85 N, with a bias of 0.20 N. TA was most accurately predicted by using the model developed from SNV spectra which gave R2, RMSEval, and a bias of 0.81, 0.24%, and −0.00%, respectively. Compared to firmness and TA, HSI showed less predictive ability for TSS attributes. The highest R2 for TSS prediction model was found to be 0.5 with RMSEval of 2.0% and a bias of −0.0% using SNV pre-processed spectra with 13 PLS components (Table 2). The results showing pre-processing spectra with SNV technique gave the most accurate prediction model for mango quality attributes. MLR analysis showed 16 significant factors for firmness, 17 for TA, and 20 for TSS, with the top weighted coefficients at 510 and 574 nm for firmness, 506 and 530 nm for TA, and 506, 586, and 730 nm for TSS (Table 3). The results of MLR models were not substantially different in terms of prediction performance as compared to PLS. This would allow the elimination of unnecessary channels for prediction and would decrease time-consuming computations for quicker handling during industrial application.Table 1. Descriptive statistics of firmness, TA and TSS from reference analyses used for calibration and validation sets in PLS modelling.<table><thead><tr><th colspan="1" rowspan="2"></th><th colspan="3" rowspan="1">Calibration set (n = 108)a</th><th colspan="3" rowspan="1">Validation set (n = 52)</th></tr><tr><th colspan="1" rowspan="1">Mean</th><th colspan="1" rowspan="1">Range</th><th colspan="1" rowspan="1">SD</th><th colspan="1" rowspan="1">Mean</th><th colspan="1" rowspan="1">Range</th><th colspan="1" rowspan="1">SD</th></tr></thead><tbody><tr><th colspan="1" rowspan="1">Firmness (N)</th><td colspan="1" rowspan="1">12.43</td><td colspan="1" rowspan="1">4.94–29.09</td><td colspan="1" rowspan="1">6.87</td><td colspan="1" rowspan="1">12.27</td><td colspan="1" rowspan="1">5.18–28.16</td><td colspan="1" rowspan="1">6.59</td></tr><tr><th colspan="1" rowspan="1">TA (%)</th><td colspan="1" rowspan="1">0.53</td><td colspan="1" rowspan="1">0.10–2.87</td><td colspan="1" rowspan="1">0.61</td><td colspan="1" rowspan="1">0.51</td><td colspan="1" rowspan="1">0.11–2.51</td><td colspan="1" rowspan="1">0.56</td></tr><tr><th colspan="1" rowspan="1">TSS (%)</th><td colspan="1" rowspan="1">18.0</td><td colspan="1" rowspan="1">8.8–25.6</td><td colspan="1" rowspan="1">3.1</td><td colspan="1" rowspan="1">18.0</td><td colspan="1" rowspan="1">11.0–22.8</td><td colspan="1" rowspan="1">2.8</td></tr></tbody></table>an = number of samples.Table optionsTable 2. PLS components and accuracy of calibration and validation (RMSE & Bias) sets for firmness, TA and TSS using different pre-processing techniques for the data; bold treatments indicate best performance.<table><thead><tr><th colspan="1" rowspan="1">Quality attribute</th><th colspan="1" rowspan="1">Pre-processinga</th><th colspan="1" rowspan="1">PLS components</th><th colspan="1" rowspan="1">R2</th><th colspan="1" rowspan="1">RMSEcal</th><th colspan="1" rowspan="1">RMSEval</th><th colspan="1" rowspan="1">Bias</th></tr></thead><tbody><tr><td colspan="1" rowspan="5">Firmness (N)</td><td colspan="1" rowspan="1">None</td><td colspan="1" rowspan="1">15</td><td colspan="1" rowspan="1">0.79</td><td colspan="1" rowspan="1">2.68</td><td colspan="1" rowspan="1">3.01</td><td colspan="1" rowspan="1">0.18</td></tr><tr><td colspan="1" rowspan="1">SG</td><td colspan="1" rowspan="1">37</td><td colspan="1" rowspan="1">0.79</td><td colspan="1" rowspan="1">2.78</td><td colspan="1" rowspan="1">3.00</td><td colspan="1" rowspan="1">0.16</td></tr><tr><td colspan="1" rowspan="1">SG″</td><td colspan="1" rowspan="1">24</td><td colspan="1" rowspan="1">0.77</td><td colspan="1" rowspan="1">2.75</td><td colspan="1" rowspan="1">3.15</td><td colspan="1" rowspan="1">0.23</td></tr><tr><td colspan="1" rowspan="1">SNV</td><td colspan="1" rowspan="1">18</td><td colspan="1" rowspan="1">0.81</td><td colspan="1" rowspan="1">2.52</td><td colspan="1" rowspan="1">2.83</td><td colspan="1" rowspan="1">0.22</td></tr><tr><td colspan="1" rowspan="1">MSC</td><td colspan="1" rowspan="1">14</td><td colspan="1" rowspan="1">0.81</td><td colspan="1" rowspan="1">2.54</td><td colspan="1" rowspan="1">2.85</td><td colspan="1" rowspan="1">0.20</td></tr><tr><td colspan="1" rowspan="5">TA (%)</td><td colspan="1" rowspan="1">None</td><td colspan="1" rowspan="1">13</td><td colspan="1" rowspan="1">0.70</td><td colspan="1" rowspan="1">0.28</td><td colspan="1" rowspan="1">0.30</td><td colspan="1" rowspan="1">−0.00</td></tr><tr><td colspan="1" rowspan="1">SG</td><td colspan="1" rowspan="1">36</td><td colspan="1" rowspan="1">0.69</td><td colspan="1" rowspan="1">0.28</td><td colspan="1" rowspan="1">0.31</td><td colspan="1" rowspan="1">−0.01</td></tr><tr><td colspan="1" rowspan="1">SG″</td><td colspan="1" rowspan="1">23</td><td colspan="1" rowspan="1">0.73</td><td colspan="1" rowspan="1">0.29</td><td colspan="1" rowspan="1">0.29</td><td colspan="1" rowspan="1">−0.02</td></tr><tr><td colspan="1" rowspan="1">SNV</td><td colspan="1" rowspan="1">13</td><td colspan="1" rowspan="1">0.81</td><td colspan="1" rowspan="1">0.27</td><td colspan="1" rowspan="1">0.24</td><td colspan="1" rowspan="1">−0.00</td></tr><tr><td colspan="1" rowspan="1">MSC</td><td colspan="1" rowspan="1">14</td><td colspan="1" rowspan="1">0.79</td><td colspan="1" rowspan="1">0.27</td><td colspan="1" rowspan="1">0.25</td><td colspan="1" rowspan="1">0.00</td></tr><tr><td colspan="1" rowspan="5">TSS (%)</td><td colspan="1" rowspan="1">None</td><td colspan="1" rowspan="1">14</td><td colspan="1" rowspan="1">0.4</td><td colspan="1" rowspan="1">2.0</td><td colspan="1" rowspan="1">2.1</td><td colspan="1" rowspan="1">−0.0</td></tr><tr><td colspan="1" rowspan="1">SG</td><td colspan="1" rowspan="1">21</td><td colspan="1" rowspan="1">0.4</td><td colspan="1" rowspan="1">2.0</td><td colspan="1" rowspan="1">2.1</td><td colspan="1" rowspan="1">0.0</td></tr><tr><td colspan="1" rowspan="1">SG″</td><td colspan="1" rowspan="1">16</td><td colspan="1" rowspan="1">0.4</td><td colspan="1" rowspan="1">2.1</td><td colspan="1" rowspan="1">2.2</td><td colspan="1" rowspan="1">0.0</td></tr><tr><td colspan="1" rowspan="1">SNV</td><td colspan="1" rowspan="1">13</td><td colspan="1" rowspan="1">0.5</td><td colspan="1" rowspan="1">2.0</td><td colspan="1" rowspan="1">2.0</td><td colspan="1" rowspan="1">−0.0</td></tr><tr><td colspan="1" rowspan="1">MSC</td><td colspan="1" rowspan="1">12</td><td colspan="1" rowspan="1">0.5</td><td colspan="1" rowspan="1">2.0</td><td colspan="1" rowspan="1">2.1</td><td colspan="1" rowspan="1">0.0</td></tr></tbody></table>aSavitzky–Golay smoothing (SG), Savitzky–Golay second derivative (SG″), standard normal variate transformation (SNV) and multiplicative scatter correction (MSC).Table optionsFig. 5. Prediction of (a) firmness, (b) TA and (c) TSS using SNV pre-processed spectra as compared to reference measurements; crossed and round symbols represent calibration and validation set, respectively.Figure optionsTable 3. Results of multiple linear regression (MLR) showing selected wavelengths and coefficients for firmness, TA and TSS.<table><thead><tr><th colspan="2" rowspan="1">Firmness</th><th colspan="2" rowspan="1">TA</th><th colspan="2" rowspan="1">TSS</th></tr><tr><th colspan="1" rowspan="1">Wavelength</th><th colspan="1" rowspan="1">Coefficient</th><th colspan="1" rowspan="1">Wavelength</th><th colspan="1" rowspan="1">Coefficient</th><th colspan="1" rowspan="1">Wavelength</th><th colspan="1" rowspan="1">Coefficient</th></tr></thead><tbody><tr><td colspan="1" rowspan="1">490</td><td colspan="1" rowspan="1">−14.93</td><td colspan="1" rowspan="1">490</td><td colspan="1" rowspan="1">−1.40</td><td colspan="1" rowspan="1">490</td><td colspan="1" rowspan="1">5.26</td></tr><tr><td colspan="1" rowspan="1">510</td><td colspan="1" rowspan="1">21.72</td><td colspan="1" rowspan="1">506</td><td colspan="1" rowspan="1">1.77</td><td colspan="1" rowspan="1">506</td><td colspan="1" rowspan="1">−13.51</td></tr><tr><td colspan="1" rowspan="1">514</td><td colspan="1" rowspan="1">13.25</td><td colspan="1" rowspan="1">510</td><td colspan="1" rowspan="1">1.41</td><td colspan="1" rowspan="1">522</td><td colspan="1" rowspan="1">5.34</td></tr><tr><td colspan="1" rowspan="1">538</td><td colspan="1" rowspan="1">−18.01</td><td colspan="1" rowspan="1">514</td><td colspan="1" rowspan="1">0.86</td><td colspan="1" rowspan="1">538</td><td colspan="1" rowspan="1">4.55</td></tr><tr><td colspan="1" rowspan="1">558</td><td colspan="1" rowspan="1">−3.90</td><td colspan="1" rowspan="1">530</td><td colspan="1" rowspan="1">−1.87</td><td colspan="1" rowspan="1">542</td><td colspan="1" rowspan="1">−4.28</td></tr><tr><td colspan="1" rowspan="1">562</td><td colspan="1" rowspan="1">−17.17</td><td colspan="1" rowspan="1">534</td><td colspan="1" rowspan="1">−1.32</td><td colspan="1" rowspan="1">546</td><td colspan="1" rowspan="1">2.69</td></tr><tr><td colspan="1" rowspan="1">574</td><td colspan="1" rowspan="1">21.95</td><td colspan="1" rowspan="1">538</td><td colspan="1" rowspan="1">−0.42</td><td colspan="1" rowspan="1">558</td><td colspan="1" rowspan="1">−5.45</td></tr><tr><td colspan="1" rowspan="1">602</td><td colspan="1" rowspan="1">−8.38</td><td colspan="1" rowspan="1">542</td><td colspan="1" rowspan="1">0.72</td><td colspan="1" rowspan="1">586</td><td colspan="1" rowspan="1">−12.72</td></tr><tr><td colspan="1" rowspan="1">618</td><td colspan="1" rowspan="1">7.48</td><td colspan="1" rowspan="1">546</td><td colspan="1" rowspan="1">−0.16</td><td colspan="1" rowspan="1">602</td><td colspan="1" rowspan="1">5.16</td></tr><tr><td colspan="1" rowspan="1">622</td><td colspan="1" rowspan="1">7.48</td><td colspan="1" rowspan="1">582</td><td colspan="1" rowspan="1">0.59</td><td colspan="1" rowspan="1">634</td><td colspan="1" rowspan="1">3.45</td></tr><tr><td colspan="1" rowspan="1">646</td><td colspan="1" rowspan="1">4.00</td><td colspan="1" rowspan="1">638</td><td colspan="1" rowspan="1">0.37</td><td colspan="1" rowspan="1">666</td><td colspan="1" rowspan="1">4.50</td></tr><tr><td colspan="1" rowspan="1">678</td><td colspan="1" rowspan="1">−8.63</td><td colspan="1" rowspan="1">670</td><td colspan="1" rowspan="1">−0.29</td><td colspan="1" rowspan="1">682</td><td colspan="1" rowspan="1">−3.40</td></tr><tr><td colspan="1" rowspan="1">710</td><td colspan="1" rowspan="1">5.35</td><td colspan="1" rowspan="1">682</td><td colspan="1" rowspan="1">−0.58</td><td colspan="1" rowspan="1">686</td><td colspan="1" rowspan="1">−2.10</td></tr><tr><td colspan="1" rowspan="1">758</td><td colspan="1" rowspan="1">−15.91</td><td colspan="1" rowspan="1">698</td><td colspan="1" rowspan="1">0.54</td><td colspan="1" rowspan="1">718</td><td colspan="1" rowspan="1">3.10</td></tr><tr><td colspan="1" rowspan="1">946</td><td colspan="1" rowspan="1">−6.31</td><td colspan="1" rowspan="1">706</td><td colspan="1" rowspan="1">0.25</td><td colspan="1" rowspan="1">730</td><td colspan="1" rowspan="1">−14.95</td></tr><tr><td colspan="1" rowspan="1">950</td><td colspan="1" rowspan="1">5.77</td><td colspan="1" rowspan="1">726</td><td colspan="1" rowspan="1">0.66</td><td colspan="1" rowspan="1">762</td><td colspan="1" rowspan="1">5.19</td></tr><tr><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1">754</td><td colspan="1" rowspan="1">−0.94</td><td colspan="1" rowspan="1">778</td><td colspan="1" rowspan="1">7.80</td></tr><tr><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1">794</td><td colspan="1" rowspan="1">−6.40</td></tr><tr><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1">954</td><td colspan="1" rowspan="1">−0.94</td></tr><tr><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1"></td><td colspan="1" rowspan="1">994</td><td colspan="1" rowspan="1">−2.09</td></tr></tbody></table>Table options3.4. Spatial evaluation of physicochemical propertiesResults of HSI prediction maps for firmness, TA and TSS are shown in Fig. 6. The intensity of the individual qualities is indicated by a blue-red (in the web version) colour scale representing low to high values, respectively. Mean values from three positions of the reference measurements and predicted values from HSI (in brackets) are located below each sample. For firmness, fruit softening was observed from the distribution map with the shoulder part exhibiting intense blue colour which implies that firmness declined from the shoulder toward to tip part of the fruit. Similar results were found for a hybrid mango cultivar, which confirmed that ripening process occurred from stem to tip positions of the fruit (Jha, Sethi et al., 2010). However, the uniform softened texture as visualised by even blue colour was observed in fruits later during ripening. The variation of TA was consistent with laboratory reference values indicating that the tip sections in unripe fruits were more acidic than shoulder sections in riper fruits. The variations of TSS content in individual fruits were clearly seen in distribution maps. The intense red colour over the shoulder position compared with other positions represents the higher TSS regions. Nevertheless, uniform distribution of TSS was found in fully ripened fruits due to the complete conversion of starches to sugars (Medlicott & Thompson, 1985). Firmness was the parameter most accurately predicted by HSI. Meanwhile, the results for TSS were uncertain and the capacity of HSI to predict TA was mostly likely a secondary effect due to correlation with firmness.Fig. 6. Prediction maps showing the distribution of firmness, TA and TSS in twelve mango samples during ripening with measured and predicted (in brackets) values indicating results of the HSI prediction model.Figure optionsResults for DA classification analysis for firmness are presented in Table 4. Samples were classified into three stages including unripe, ripe and overripe groups: fruits which had firmness value >14.0 N were grouped into unripe phase, while fruits having firmness value <8.0 N were classified as overripe and those with firmness value of 8.0–14.0 were defined as ripe phase. According to this classification, the major ripening stage of the fruit was ripe phase (60 fruits) followed by overripe (54 fruits) and unripe phase (46 fruits). Three-fourths of samples, consisting of all three ripening stages, were defined as the training set for development of the classification model and the remaining samples were used to validate the efficiency of classification. Overall, the DA was most successful at classifying unripe and overripe fruits, with ripe fruits being misclassified into either of the other two classes. Effective classification was realised in unripe fruits at values up to 89%.Table 4. Results from discriminate analysis (DA) of firmness with classification performance according to ripeness groups: unripe (1), ripe (2) and overripe (3).<table><thead><tr><th colspan="2" rowspan="3"></th><th colspan="4" rowspan="1">Calibration (na = 160)</th><th colspan="4" rowspan="1">Cross-validation (n = 120)</th></tr><tr><th colspan="4" rowspan="1">Predicted</th><th colspan="4" rowspan="1">Predicted</th></tr><tr><th colspan="1" rowspan="1">1</th><th colspan="1" rowspan="1">2</th><th colspan="1" rowspan="1">3</th><th colspan="1" rowspan="1">Total</th><th colspan="1" rowspan="1">1</th><th colspan="1" rowspan="1">2</th><th colspan="1" rowspan="1">3</th><th colspan="1" rowspan="1">Total</th></tr></thead><tbody><tr><th colspan="10" rowspan="1">Measured</th></tr><tr><td colspan="1" rowspan="3">n</td><td colspan="1" rowspan="1">1</td><td colspan="1" rowspan="1">41</td><td colspan="1" rowspan="1">5</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">46</td><td colspan="1" rowspan="1">26</td><td colspan="1" rowspan="1">8</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">34</td></tr><tr><td colspan="1" rowspan="1">2</td><td colspan="1" rowspan="1">4</td><td colspan="1" rowspan="1">43</td><td colspan="1" rowspan="1">13</td><td colspan="1" rowspan="1">60</td><td colspan="1" rowspan="1">6</td><td colspan="1" rowspan="1">22</td><td colspan="1" rowspan="1">17</td><td colspan="1" rowspan="1">45</td></tr><tr><td colspan="1" rowspan="1">3</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">6</td><td colspan="1" rowspan="1">48</td><td colspan="1" rowspan="1">54</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">4</td><td colspan="1" rowspan="1">37</td><td colspan="1" rowspan="1">41</td></tr><tr><td colspan="1" rowspan="3">%</td><td colspan="1" rowspan="1">1</td><td colspan="1" rowspan="1">89.1</td><td colspan="1" rowspan="1">10.9</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">100</td><td colspan="1" rowspan="1">76.5</td><td colspan="1" rowspan="1">23.5</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">100</td></tr><tr><td colspan="1" rowspan="1">2</td><td colspan="1" rowspan="1">6.7</td><td colspan="1" rowspan="1">71.7</td><td colspan="1" rowspan="1">21.7</td><td colspan="1" rowspan="1">100</td><td colspan="1" rowspan="1">13.3</td><td colspan="1" rowspan="1">48.9</td><td colspan="1" rowspan="1">37.8</td><td colspan="1" rowspan="1">100</td></tr><tr><td colspan="1" rowspan="1">3</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">11.1</td><td colspan="1" rowspan="1">88.9</td><td colspan="1" rowspan="1">100</td><td colspan="1" rowspan="1">0</td><td colspan="1" rowspan="1">9.8</td><td colspan="1" rowspan="1">90.2</td><td colspan="1" rowspan="1">100</td></tr></tbody></table>an = number of samples.Table options4. DiscussionA spatial gradient of physicochemical parameters was observed for mango fruits showing non-uniform ripening process within individual samples. As reported by Crisosto, Slaughter, Garner, and Boyd (2001) and Peiris, Dull, Leffler, and Kays (1999), the ripening process and spatial distribution of the physicochemical constituents are dependent on fruit cultivar, production season, and fruit size. The pattern of TSS, higher in shoulder section then progressively decreasing toward the tip section, was also observed in mango cv. Hayden (Peiris et al., 1999). However, inconsistent results were found in mango cv. Cogshall showing mesocarp located in the tip part had significant higher TSS than the cheek and the shoulder of the fruit (Nordey, Joas, Léchaudel, Génard, & Joas, 2014). In this study, insignificant differences of TSS between shoulder and tip section could be explained by the variation of ethylene gradients in the mesocarp. Since a shorter distance exists between endocarp and pericarp at the shoulder, ethylene diffusion from the endocarp through the mesocarp and to the endocarp occurs quicker at the shoulder, thus causing faster ripening, e.g. higher TSS content (Nordey et al., 2014). From the results, changes of TA and TSS at each fruit position during ripening implied that ripening started from shoulder toward to tip part.Changes caused by mango ripening can be roughly observed from mono-spectral (greyscale) images. These findings agreed with the other studies which explained that high firmness and low soluble solid content or unripe samples provided higher reflectance than ripe samples (ElMasry et al., 2007; Leiva-Valenzuela et al., 2013; Rajkumar et al., 2012 ;  Wei et al., 2014). In this respect, SNV pre-processing exhibited the best prediction results since non-uniform scattering effects are reduced, i.e. baseline shifts and slope variations which are influenced by different physical property of fruit are minimised. Good prediction performance was achieved for firmness (RMSEval = 2.85 N, bias = 0.20 N) rather than for TA (RMSEval = 0.24%, bias = −0.00%) and TSS (RMSEval = 2.0%, bias = −0.0%). In other studies, firmness of persimmon could be related to HSI data with R2 and RMSE of 0.83 and 4.35 N, respectively (Wei et al., 2014). Comparing the results with other mango cultivars, firmness prediction models developed in this study obtained higher R2 and higher error (R2 of 0.81 and RMSE of 2.83 N) than another model for Tommy Atkins (R2 of 0.77 and SEP of 2.08 N) (Sivakumar, 2006). High RMSE found for the TA model greatly affected the accuracy as seen by the predicted negative values at low concentrations (Fig. 5b). However, the results from Baiano, Terracone, Peri, and Romaniello (2012) showed that HSI provided a higher accuracy for TA prediction in white grape and red grape with R2 of 0.95 and 0.82 and RMSE of 0.06 and 0.24 g L−1, respectively. Meanwhile, lower prediction ability was observed for TSS. Low R2 (0.5) with high prediction error (2.0%) was stated as inadequate to use for screening quality of fruit.Better performance of HSI for firmness and TA prediction rather than TSS prediction can be explained by the fact that HSI technique relies on scattering measurements which are more influenced by the structural features of the fruit tissue (Mendoza, Lu, Ariana, Cen, & Bailey, 2011). Moreover, the broader range of firmness and TA data as compared to TSS could have promoted stronger HSI correlations (Leiva-Valenzuela et al., 2014). The higher potential of HSI to predict firmness over TSS was also reported by Rajkumar et al. (2012) who found that HSI at 400–1000 nm showed a better prediction of firmness (R2 = 0.91) than TSS (R2 = 0.85) of banana. However, models established in this study could at least be used for rough screening due to excessive prediction error. The distribution of firmness within fruit was documented as an advantage of the HSI technique and analysis of the variation in hyperspectral maps is proposed as a possibility for grading systems in packing houses for quantitative and qualitative quality determination. Although considerable error was found for TA and TSS prediction models, the changes of these attributes within fruit during ripening were still possible to observe in prediction maps. False mapping occurred at the edge of the fruits resulting from the non-uniform distribution of light source on the curved fruit surface which caused stretching and shadowing (Gómez-Sanchis et al., 2008 ;  Vélez Rivera et al., 2014). In this study, error from insufficient illumination was minimised by placing light diffusers near the fruit. The challenge of using diffusers to optimise light distribution for HSI was also reported for a study on mandarin by Gómez-Sanchis et al. (2008). Meanwhile, cross-polarisation of light could be a factor to further enhance the proficiency of HSI measurements (Nguyen-Do-Trong, Keresztes, De Ketelaere, & Saeys, 2016). Overall, applying additional methods for correcting the effect of the illumination system may enhance the spatial reliability of the prediction models.The results of DA analysis for classification of fruit according to firmness as an indicator of ripeness were promising. Classification success rates greater than 90% are mostly desirable since losses of up to 10% in general are economically tolerable, but this level was

Involved persons

Prof. Dr. Joachim Müller

Prediction mapping of physicochemical properties in mango by hyperspectral imaging

Involved persons

Involved institutions