Non-destructive evaluation of soluble solid content in fruits with various skin thicknesses using visible–shortwave near-infrared spectroscopy

2.1 Sample and spectra measurement

Dragon fruit (Hylocereus polyrhizus), tomato (Solanum lycopersicum cultivar Servo F1), sapodilla (Manilkara zapota), guava (Psidium guajava L.), and banana (Musa Acuminata B.) were used as fruit samples. Each cultivar has 100 intact samples that were purchased from local markets in Yogyakarta Indonesia. After samples were transported to the laboratory, the reflectance spectral data of each fruit were collected by a Vis–SWNIR spectrometer (Flame-T-VIS-NIR Ocean Optics, Dunedin, FL, USA; 350–1,000 nm with a resolution of 0.22 nm) equipped with a tungsten halogen lamp (360–2,400 nm, HL-2000-HP-FHSA Ocean Optics, nominal bulb power 20 W, typical output power 8.4 mW) and a reflectance probe (QR400-7 VIS-NIR Ocean Optics). Spectra acquisition was set as shown in Figure 1. Spectra capture settings are carried out before the spectra acquisition process using the OceanView 1.6.7 with a scan to the average of 100, the integration time of 100 ms, and the boxcar width of 1. Spectra acquisitions were done inside a black box to minimize environmental interference.

Figure 1

Spectra acquisition set up using Vis–SWNIR spectroscopy.

2.2 Chemical and physical analysis

2.3 Data analysis and PLSR model development

The individual-cultivar studies were established to quantitatively determine the SSC in fruit. The whole spectra of Vis–NIR spectrometer covers a wavelength range of 350–1,100 nm. However, due to its heavy noise, the initial spectra at the Vis region were cut to which the remaining spectra in the region of 400–1,000 nm were used to build the PLSR model. Three spectra of fruits obtained from different surface positions in the equator area were used for analysis without being averaged. The total spectra of each sample fruit were 300 each sample fruit

All reflectance spectra data were then analyzed using the Unscrambler^® X software (CAMO, Oslo, Norway). Data were randomly divided into two-thirds of the data serving as the calibration set and one-third data serving as the prediction set. Several pre-processing techniques were applied to the spectra, such as the Savitzky–Golay second derivative (SGD2), Savitzky–Golay smoothing (SGS), area normalization (AN), standard normal variate (SNV), and multiple scatter correction (MSC). SGD and SGS are used to remove vertical offsets and linearly sloping baselines [19]. Normalization pre-processing implies allocating the same weight to all absorbencies. MSC and SNV are related due to light scattering and particle size [20]. The optimal spectra transformations that indicated the best pre-processing spectra were chosen based on the highest values of coefficient of determination (R ²) and the lowest of root mean square of error (RMSE) of PLSR models.

3.1 SSC analysis

Table 1 shows the statistical data of SSC for all fruit samples. A range of SSC fruits used in this research were 7.90–13.30, 3.40–5.90, 18.86–31.23, 9.43–25.93, and 5.35–9.85°Brix for dragon fruit, tomato, banana, sapodilla, and guava, respectively. SSC of tomato and guava was in a narrow interval, compared to dragon fruit, banana, and sapodilla. High contents of SSC were recorded for banana and sapodilla, while low contents were recorded for dragon fruit, tomato, and guava. The highest standard deviation (SD) was observed for banana and sapodilla then followed by dragon fruit and guava, while tomato had the lowest SD. A low SD implies a small variation in SSC content. Although might generate a high RMSE, a large SD is required to obtain a good regression model [21].

3.2 Spectra profiles

Figure 2 is reflectance spectra of dragon fruit, tomato, banana, guava, and sapodilla along the Vis and shortwave near-infrared (SWNIR) region at the wavelength of 450–970 nm derived from original spectra (Figure 2a) and SGD2 spectra (Figure 2b). The figure shows considerably different profiles for each fruit. From both figures, peaks or valleys of spectra were noticeable at the Vis region of 450–700 nm as a response to the presence of pigments [22], differences in reflectance spectra in the Vis region can be attributed to pigment differences in each fruit. While at the SWNIR region of 700–970 nm, the spectra were relatively flat. By applying the SGD2 method, the peaks around 960–980 nm (Figure 2b) were Vis due to the absorption of water and carbohydrate [23] and around 910–930 nm due to CH and CH2 stretching [24].

Figure 2

Reflectance spectra of dragon fruit, tomato, banana, guava, and sapodilla developed using (a) original spectra and (b) SGD2 spectra.

The fruit’s skin and flesh condition are related to colorimetric measurements and pigment content. Colors in fruits are created by a mixture of pigments, including the green chlorophylls, red orange, yellow carotenoids, and red, blue, and violet anthocyanins, as well as other pigments [25]. Chlorophylls, carotenoids, and anthocyanins are the most common pigments found in fruits, which influence the coloration of the whole Vis reflectance spectra of the fruit at 400–700 nm. From Figure 2a, several peaks and valleys can be observed. Region 450–500 nm refers to the reflectance of carotenoid pigment which exhibits yellow, orange, and bright red colors. However, at slightly overlapped wavelength, chlorophylls a and b are also present at 400–500 nm [24]. In general, lower light intensity reflected by tomato and sapodilla, which may be caused by higher absorption of carotenoid, compared to banana, guava, and dragon fruit. At 500–600 nm, fruit reflects intensity, which is due to its anthocyanin pigment. Anthocyanins are responsible for red, purple, and blue colors of fruits [25]. Higher reflectance at the anthocyanin region was detected for banana and guava, while lower reflectance was detected for dragon fruit, tomato, and sapodilla. Reflectance values observed at the wavelength of 680 nm are associated with chlorophyll in which dragon fruit and banana reflect higher intensity compared to guava, tomato, and sapodilla. High reflectance at 680 nm means low absorbance of chlorophyll at 680 nm. Dragon fruit having red color reflects the highest light intensity at 680 nm compared to other fruits (Figure 2a), which can be assumed that dragon fruit contains the lowest chlorophyll content.

Chlorophyll is the primary pigment in photosynthesis while carotenoids and anthocyanins are categorized as accessory pigments taking a role in secondary metabolites [26]. Chlorophyll is found in chloroplast, which adsorbs light to drive the photosynthesis process into which the chlorophyll converts CO₂ and H₂O into simple sugars with the help of sunlight. In addition, anthocyanins are derived from non-sugar components (aglycone or anthocyanidins), which are bound with sugar fraction [27]. Those pigments appear in the Vis wavelength; therefore, the absorption of pigments recorded by Vis–NIR spectroscopy can be used to determine SSC in fruits.

The reflectance level between fruits in the Vis and the SWNIR region differs based on the spectral appearance in Figure 2. The thickness of the fruit skin in Table 2 is 4.10 mm for dragon fruit and 3.26 mm for banana. Meanwhile, tomatoes, sapodilla, and guavas are thin-skinned fruit (0.16–0.29 mm). At Vis wavelengths of 600–700 nm, the reflectance from the lowest is tomatoes, sapodilla, guava, banana, and dragon fruit. The order of low to high reflectance is determined by the thickness of the fruit skin. However, at the SWNIR wavelength, guava has the highest reflectance, even though the skin is thin. Despite the fact that the dragon fruit skin is thicker, the reflectance spectrum is under guava. The guava’s thin skin should allow light to penetrate into the flesh of the fruit.

The absorbance in the NIR region is 10–100 times lower than the fundamental absorption band in the IR electromagnetic spectrum regions [28]. That is what may cause the reflectance of dragon fruit to be high. In addition, the SWNIR area is mostly affected by water absorption features [29]. According to Table 2, the fruit skins with the highest water content are dragon fruit, tomato, guava, banana, and sapodilla. On the other hand, dragon fruit and guava have higher reflectance spectra than banana and sapodilla. Wavelength drift may be due to changes in the optical path [30]. Furthermore, [22] in the reflection optical geometry, the detector can receive the illuminated part of the fruit in the form of specular and diffuse reflectance. The shiny surface of the fruit causes specular reflectance, while the rough surface of the object causes diffuse reflectance [3].

3.3 PLSR analysis

Calibration and prediction datasets used for developing PLSR predicting SSC of dragon fruit, tomato, banana, sapodilla, and guava are shown in Table 3. The table shows that the mean values of SSC from the highest are banana, sapodilla, dragon fruit, guava, and tomato were 27.37, 19.86, 10.78, 7.25, and 4.25°Brix, respectively, for calibration datasets. Moreover, for prediction datasets, the mean values of SSC from the highest are banana, sapodilla, dragon fruit, guava, and tomato were 27.33, 19.59, 10.63, 7.14, and 4.27°Brix, respectively. The values of SD from the highest to the lowest are sapodilla, banana, dragon fruit, guava, and tomato with the SD of 2.20, 2.11, 1.16, 0.92, and 0.39°Brix, respectively, for calibration datasets. Meanwhile, for prediction datasets, SD from the highest to the lowest is sapodilla, banana, dragon fruit, guava, and tomato with the SD of 2.20, 2.15, 1.22, 0.88, and 0.35°Brix, respectively. In general, selection data for developing calibration and prediction model are appropriate since prediction datasets are within the range of calibration datasets.

To evaluate PLSR performance in this research, it is important to compare the sample characteristics as shown in Table 3 with other research reports. SSC of banana samples in this research was narrower compared to the study by ref. [5] that used banana in the SSC of 5.5–28.4°Brix and the SD of 5.48°Brix which resulted in the coefficient of determination of calibration (R ² _C) of more than 0.90. SSC of tomatoes was smaller compared to samples used by ref. [31], which were in the SSC of 3.5–7.1°Brix and the SD of 0.64°Brix, resulting in the correlation coefficient for calibration (R _C) of 0.85 and the root mean square error of prediction of 0.33°Brix. Samples of sapodilla in this research were within the sample range of previous research conducted by ref. [32] that used samples with the TSS of 11.7–32.20°Brix and the SD of 2.21°Brix, which yielded the R ² _C of 0.82. Guava samples analyzed in this study had wider values compared to ref. [33] that measured TSS in guava using samples with the TSS of 7.82–8.80°Brix, which resulted in the R ² _C of 0.85. Dragon fruits in this study were in a narrower range compared to samples used by ref. [34] with the TSS of 9.20–17.60°Brix, which yielded the coefficient correlation (r) of 0.93 and the standard error of prediction of 0.66°Brix.

Table 4 shows the PLSR performance of calibration and cross-validation models developed using original and pre-processed spectra. The best calibration SSC model for each fruit based on the highest R ² _C and the lowest root mean square error of calibration (RMSEC) was obtained from SGS spectra for dragon fruit with the R ² _C and RMSEC of 0.70 and 0.52°Brix using six latent variables (LVs) and tomato with the R ² _C and RMSEC of 0.80 and 0.15°Brix using seven LVs. Moreover, calibration models were developed based on SGD2 using seven LVs, which resulted in the R ² _C of 0.88, 0.92, and 0.77, the RMSEC of 0.36, 0.36, and 0.34°Brix, for banana, sapodilla, and guava, respectively.

The reflectance in fruit samples is dominated by specular reflectance [3] and influenced by fruit skin characteristics [35]. Based on Table 4, low R ² _C belonged to dragon fruit and guava, which guava has the thinner skin, and higher R ² _C belonged to banana, sapodilla, and tomato, which banana has thinner skin. Fruit skin influences the amount of light, which enters the fruit [36]. Subedi and Walsh [37] reported the thickness of the banana skin affected the dry matter model. Likewise for the determination of dry matter in avocado, by using various Vis–NIR instruments, Subedi and Walsh [16] reported the better R ² for PLSR in skin-removed compared to intact fruits. Arendse et al. [38] reviewed several studies of SSC fruits with thick rinds, which were measured using Vis–SWNIR (300–1,100 nm) and IR (up to 2,400 nm) regions. Some of results were satisfactory by using the Vis–SWNIR region such as watermelon, melon, passion fruit, and pomegranate. This is due to the higher energy at Vis, which allows deeper light penetration. The NIR wavelength is 700–900 nm, and the maximum light penetration is 4 mm. At higher wavelengths, 900–1,900 nm, the maximum light penetration is only about 2–3 mm [14]. Moreover, Guthrie et al. [39,40] obtained unsatisfactory results for predicting SSC in pineapple and melon using NIR spectroscopy. Furthermore, not just does the peel thickness obstruct NIRS capability, yet also the non-uniformity of internal fruit pulp as well as macrostructures including hard black seeds [3]. Dragon fruit and guava were observed to have large amounts of small grains in the pulp, resulting in the low performance of the calibration model. In short, all the best-selected calibration models have R ² _C between 0.70 and 0.92, which is applicable for screening even for quality assurance [41].

Apart from skin thickness, concentrations of targeted analyte also affect the performance of the PLSR calibration model. Sapodilla and banana have higher SSC compared to dragon fruit, guava, and tomato. According to Table 4, sapodilla and banana have higher R ² _C and lower RMSEC compared to dragon fruit, guava, and tomato. Sim and Kimura [42] found that prediction of moisture in transformer oil and lubricating oil was poorer in the low moisture concentration samples compared to the high ones. A similar finding was reported by ref. [37] in which bananas containing higher SSC had higher R ² compared to the lower ones. There is a high correlation between SSC concentration and firmness, so that firmness decreases as the concentration of sugars and/or by-products increases, causing cell wall disassembly and degradation, which induces increased absorption above 900 nm [43].

Prediction of the SSC content is also affected by the light absorption of fruit skin and internal fruit structure. Skin or peel of fruit is critical in determining internal fruit quality since the IR light is scattered or absorbed by peel before it reaches the pulp [44]. Chlorophyll and other color pigments in fruits are mostly present in the peel [45]. Chlorophylls, which are the majority pigments present in some fruits, are inversely related to SSC during ripening [24] and, thus, can be used to predict SSC. Wang et al. [45] reported that the PLS model yielded a higher correlation (r) for navel orange intact fruits compared to the peeled fruits. The inclusion of the Vis region in reflectance mode produced better results compared to the only SWNIR used in the PLSR model. Although the SSC of peel of navel orange was higher compared to its pulp, Wang et al. [46] confirmed that the spectra of peels contained valuable information about SSC of pulp. However, in contrast, Wang et al. [47] achieved better results for banana pulp compared to intact banana by considering the mesocarp TSS content, which was highly correlated (R ² > 0.85) with Hunter a and a/b skin color. Based on Table 4, fruits that contain higher chlorophyll (perceived from the peel color) such as tomato, banana, sapodilla, and guava, show higher R ² and lower RMSE compared to red dragon fruit. If the surface layer of the sample is the same as the inner layer of the fruit flesh, the content of the fruit skin can be used to provide information about dry matter, dyes, sugars, and other constituents [8].

Not only peel, but also internal food structure affects the light reflectance of fruit. Optical absorption and scattering properties are induced by structural properties and chemical changes of the product [18,43]. Tomato is heterogeneous internal structure fruit with watery and seeds. Although it has a thin skin, it is different from sapodilla, which is included in the homogeneous pulp thin skin with similar anatomical features such as the homogeneous pericarp in apricot. Based on the results of the reflectance spectra, the internal structure of tomato, which contains high water, has lower reflectance intensity than sapodilla at C–H and O–H absorption wavelengths. Although guava also has a heterogeneous internal structure, the pericarp structure layer has a harder texture than the pulp region, which also contains a lot of seeds, resulting in a high reflectance spectra [47]. The experimental results show that the SSC model’s prediction accuracy and the accuracy of SSC online sorting of peach are higher than that of the navel orange. Therefore, it confirms that under the same detection device, when the light passes through the two kinds of fruits, the peach reflects more light. The soft flesh leads to a lower opacity and greater light penetration [43]. The navel orange is a citrus fruit with no or small grains inside, which does not affect the optical path [47]. As in bananas with a thick skin but small seeds in the flesh compared to guava and sapodilla, at a wavelength of 700–1,000 nm, there is lower reflectance than guava, sapodilla, and dragon fruit, but when compared to tomatoes, the reflectance of bananas is much higher. However, dragon fruit allows thick skin to act as a barrier and prevents to predict the composition of internal pulp [18]; the model results (Figure 3) show the dominant color pigment information at Vis than SWNIR wavelengths.

Figure 3

Regression coefficient from the best PLSR model for (a) dragon fruit, (b) sapodilla, (c) tomato, (d) banana, and (e) guava.

The sensitive wavelengths reflecting the spectral characteristics for SSC were obtained based on the regression coefficients (B), in each fruit calibration model shown in Figure 3 shows the peaks appear between 400–700 nm in Vis area and between 840 and 980 nm in SWNIR. The peak at 980 nm seems to be due to water stretching vibrations. Water binds to solutes, such as sugar groups, such that SSC information is still present on SWNIR [48]. The highest positive and negative peaks, which are associated to H₂O and carbon-hydrogen functional group in SWNIR wavelength, appeared at 969, 979, 985, and 990 nm in sapodilla, 992 and 994 nm in tomato, 960, 979, and 989 nm in banana, 979 nm in dragon fruit, and 938, 962, 979, and 986 nm in guava. All samples showed the same peak at 979 nm, except for tomato. According to the previous research, regions that included the sugar information are located below the wavelength of 1,000 nm, specifically 810–820, 868, 880–890, 910–940, and 980–995 nm [49]. The best result of the sapodilla, banana, and guava calibration models that were preprocessed by SGD2 seems to be more pronounced in the SWNIR area, which is the overlay of the 3rd overtones of C–H stretch and the 2nd and 3rd overtones of O–H [50]. Tomatoes and dragon fruit were built by SGS. Tomatoes are one of the fruits that contain high water content, so the absorption is very easily influenced by water (O–H bonds) in the SWNIR region [31]. The coefficient of regression of the tomato model is quite low due to the dominance of water absorption, which causes noise. The dragon fruit is fairly obvious in the Vis wavelength range but not in the SWNIR region. The preprocessing procedure is used to raise the spectra, but there is over-smoothing and polishing of spectral data, which results in the loss of valuable information [51].

The performance of the final predicted model was built by following the calibration model result. First, the highest coefficient correlation of calibration (R ² _C) and the lowest RMSEC. Prediction model shown in Figure 4, the PLS models from sapodilla and banana, showed satisfactory performance, providing better prediction results than others. According to the prediction model performance standards [41], the coefficient correlation of prediction (R ² _p) in the banana and sapodilla samples, namely 0.885 and 0.905, should be used with caution in various applications. The guava with the R ² _p score of 0.769 can still be utilized as a screening criterion and in some other “approximate” applications [41]. Tomatoes and dragon fruit R ² _p 0.646 and 0.596, respectively, are not recommended to be used and require additional research [41]. The predictive model for tomatoes was equally poor. Despite having the lowest RMSE, fruits with thinner skin and the lowest SD in the observed data (0.33%), led to a decrease in the model [37]. Variability in the chemical composition of the samples observed has the potential to provide a robust and reliable predictive model for several biomass species [52]. It is difficult to achieve an acceptable fruit quality prediction model for SSC assessment of rough and thick-skinned fruits such as dragon fruit compared to smooth and thin-skinned fruits using the reflectance mode due to the effects of light scattering and higher skin thickness [53]. Dragon fruit has the lowest prediction model for SSC with the highest RMSE value and at the same time has a high skin thickness.

Figure 4

Scatter plots between actual SSC versus Vis–NIR-predicted SSC values of validation sets: (a) dragon fruit, (b) sapodilla, (c) tomato, (d) banana, and (e) guava.