Maturity Prediction in Yellow Peach (Prunus persica L.) Cultivars Using a Fluorescence Spectrometer

Technology for rapid, non-invasive and accurate determination of fruit maturity is increasingly sought after in horticultural industries. This study investigated the ability to predict fruit maturity of yellow peach cultivars using a prototype non-destructive fluorescence spectrometer. Collected spectra were analysed to predict flesh firmness (FF), soluble solids concentration (SSC), index of absorbance difference (IAD), skin and flesh colour attributes (i.e., a* and H°) and maturity classes (immature, harvest-ready and mature) in four yellow peach cultivars—‘August Flame’, ‘O’Henry’, ‘Redhaven’ and ‘September Sun’. The cultivars provided a diverse range of maturity indices. The fluorescence spectrometer consistently predicted IAD and skin colour in all the cultivars under study with high accuracy (Lin’s concordance correlation coefficient > 0.85), whereas flesh colour’s estimation was always accurate apart from ‘Redhaven’. Except for ‘September Sun’, good prediction of FF and SSC was observed. Fruit maturity classes were reliably predicted with a high likelihood (F1-score = 0.85) when samples from the four cultivars were pooled together. Further studies are needed to assess the performance of the fluorescence spectrometer on other fruit crops. Work is underway to develop a handheld version of the fluorescence spectrometer to improve the utility and adoption by fruit growers, packhouses and supply chain managers.


Introduction
Fruit maturity indices are used to inform harvest logistics and supply chain management decisions for the delivery of fruit with optimal quality to consumers. Soluble solids concentration (SSC), flesh firmness (FF), starch concentration, titratable acidity, skin and flesh colour, fruit size and shape, ethylene production and respiration rate are useful indices used for stone and pome fruit maturity assessment [1,2]. However, until more recent times, the determination of these parameters has been mostly carried out destructively on small sample sizes, leading to time-consuming operations, expensive labour and often subjective data influenced by operators' skills. Only in recent years, the introduction of spectrometers has led to the increasing adoption of non-destructive devices for food quality estimation (e.g., near-infrared, fluorescence meters, mid-infrared and multispectral/hyperspectral imagery) [3][4][5][6]. Based on the maturity index of interest, one technology can be more reliable than others. The handheld, non-destructive Delta Absorbance (DA)meter was introduced by Ziosi et al. in 2008 [7] to determine the index of absorbance difference between 670 and 720 nm (I AD ) and has often been used for maturity

Experimental Site, Plant Material and Fruit Sampling
The experiment was carried out at the Tatura SmartFarm, Agriculture Victoria, Australia (36 • 26 7 S and 145 • 16 8 E, 113 m a.s.l.) between January and March 2020 on 5-7 years old trees of four yellow peach cultivars (Prunus persica L. Batsch, 'August Flame', late 'O'Henry', 'Redhaven' and 'September Sun'). The experimental orchard's soil had clay-loam texture and trees were irrigated, fertigated, pruned and pest/disease managed following established local commercial practices. Fruit of the four cultivars were harvested based on the I AD maturity classes reported by the Victorian Horticulture Industry Network [38]. A DA-meter (TR Turoni, Forlì, Italy) was used to determine the I AD and fruit were harvested at the harvest-ready (onset of ethylene) maturity stage [38], with the exception of 'Redhaven' fruit that were harvested late (i.e., mature-peak of ethylene emission) due to temporary unavailability of the DA-meter. Two hundred fruit per cultivar were picked in mid-January 2020 ('Redhaven'), mid-February 2020 ('O'Henry' and 'August Flame') and early-March 2020 ('September Sun'). Fruit were hand-picked early in the morning (i.e., before 0800 h, AEST) from different orchard rows and trees, and with a sampling strategy of selecting a wide range of fruit size, shape and skin red colour coverage within each cultivar. After harvest, fruit were brought to the laboratory and left to adjust to a constant temperature of 25 • C for approximately two hours. Meanwhile, one cheek per fruit was marked and specimens were numbered progressively from 1 to 200. All the fruit non-destructive and destructive measurements were carried out within 5-8 h of harvest.

Fluorescence Spectroscopy
Fluorescence spectra were collected on the marked cheek of each fruit using a custom-built fluorescence spectrometer prototype (Rubens Technologies Pty Ltd., Rowville, VIC, Australia) featuring excitation UV LED sources emitting at a wavelength of 400 nm. The spectral sensitivity of the device was in the range 350-950 nm. Since the fluorescence emission is orders of magnitude weaker than the excitation UV intensity, a high-pass filter with cut-off wavelength of 500 nm was placed at the spectrometer entrance to cut the primary UV light and retain only the fluorescent emission from the sample. Recorded spectra represented the average of five readings per fruit obtained in 5-10 s. Data were collected using a custom software running on the Windows 10 operating system and via a USB connection to a laptop.
Spectral measurements were carried out inside a black enclosure to remove influence from external light. The distance between the spectrometer entrance window and the surface of the fruit was kept constant at a value of 80 mm. The measured spectra were normalised to the total emission.

Determination of Maturity Indices and Maturity Classes
Fruit equatorial diameter from cheek to cheek (FD, mm) and fruit weight (FW, g) were obtained using a digital calliper and a digital scale with two decimal places, respectively. Subsequently, the marked cheek of each individual fruit previously exposed to the fluorescence spectrometer was scanned with the DA-meter to obtain the I AD and data were downloaded into a PC using an internal micro-SD card. Fruit were classified into three maturity classes based on cultivar-specific I AD thresholds obtained from relationships with fruit ethylene emission [38], as shown in Table 1.
Skin colour was determined on a single point per fruit-in the centre of the marked cheek-using a portable spectrophotometer (Nix Pro TM , Nix Sensor Ltd., Hamilton, ON, Canada) with a 14 mm aperture, a D50 illuminant and a 2 • observer angle. The Nix Pro TM spectrophotometer was operated using the Nix Pro TM Color Sensor application on an Android smartphone via Bluetooth connectivity. Data were logged in the smartphone memory and downloaded at the end of each measurement session. Colour data was stored in the RGB, XYZ, CMYK, HEX and CIELAB formats, but only the latter was used in this study (i.e., L*, a*, b*, C* and H • ). Next, fruit skin was removed from the marked cheek Sensors 2020, 20, 6555 4 of 17 using a potato peeler and the flesh was immediately scanned for colour assessment using the Nix Pro TM spectrophotometer with the same methodology adopted for skin colour determination.
The portion of fruit flesh that was scanned with the Nix Pro TM was then exposed to a penetrometer (FT327, FACCHINI srl, Alfonsine, Italy) with an 8 mm tip and FF was measured on a scale from 0 to 15 kgf. Lastly, a few drops of juice (~2 mL) were extracted from the same area and measured with a digital refractometer (PR-1; ATAGO CO., LTD., Saitama, Japan) to obtain SSC ( • Brix). Table 1. Fruit maturity classes for yellow-flesh peach based on cultivar-specific thresholds of index of absorbance difference (I AD ) according to the Victorian Horticulture Industry Network [38].

Cultivar
Maturity Classes (I AD ) Immature 1 Harvest- FD, FW, FF, SSC, I AD , skin and flesh colour attributes (i.e., L*, a*, b*, C* and H • ) were compared among the four cultivars using one-way analysis of variance (ANOVA), followed by the Games-Howell post hoc test for comparison of groups with unequal variances.
The correlations between pairs of the FD, FW, FF, SSC, I AD , skin and flesh colour attributes was tested using Pearson's correlation coefficients (r) and presented in a correlation heatmap. From this point onwards, a* and H • were the only colour attributes considered for further data analyses, as b*, C* and L* are less important for yellow peach fruit characteristics.
Linear regression models of flesh a* and H • against I AD were calculated to verify the validity of these two flesh parameters for maturity assessment. In addition, these models were used to estimate a* and H • maturity thresholds corresponding to I AD maturity classes in the four peach cultivars (Table 1) in order to provide a clear range of a* and H • values for harvest-ready fruit, as these two parameters are not established in the stone fruit industry. ANOVA, post hoc tests and Pearson's correlations were carried out using R (v. 4.0.2) [39] and the "Userfriendliscience" package [40]. Graphs were generated using SigmaPlot 12.5 (Systat software Inc., Chicago, IL, USA).

Prediction Modelling
To achieve maturity prediction this work focused on analysing the fluorescence spectra using two different machine learning approaches. The first approach was used for the prediction of continuous variables such as FF, SSC, I AD , skin and flesh a* and H • . The second approach was put in place to predict fruit maturity classes (as shown in Table 1) that represent simple categories that are widely understood by growers and industry stakeholders.
Prediction models for FF, SSC, I AD and the most relevant colour attributes in both skin and flesh (a* and H • ) were carried out using partial least square (PLS) regression with spectral band optimisation. Fluorescence spectra were not used for the prediction of FD and FW, as no relationship between fluorescence and fruit size was expected. All models were optimised using a k-fold cross-validation procedure using k = 10 splits. The basic idea was to select the most informative wavelength bands out of the entire spectral response. The selection was done via a random optimisation procedure called simulated annealing (SA) [41]. This algorithm begins with a randomly selected set of bands and, at each iteration, randomly changes a subset of these bands. A PLS regression model was developed at each iteration and the Akaike information criterion (AIC) [42] was used as the cost function. The aim Sensors 2020, 20, 6555 5 of 17 of the algorithm was to decrease the cost functions, hence, to improve the model. The coefficients of determination for the cross-validation procedure (R 2 CV ) and Lin's concordance correlation coefficients (r c ) [43] were calculated to assess models' robustness, with the latter being a reliable measure of the agreement between two variables. The r c is particularly useful when comparing two measures of the same variable, such as when x (observed values) and y (predicted values) are in the same unit. The best models had the highest agreement between x and y and generated a r c closer to 1 in a 0-1 range. The root mean square error in cross-validation (RMSE CV ) was also calculated to provide a measure of error in the same unit of the measured variable.
The SA optimisation procedure is a variant of a greedy optimisation strategy whereby the algorithm seeks to decrease the cost function, but not monotonically. SA allows for the occasional increase of the cost function to reduce the likelihood of the process getting stuck in a local minimum of the parameter space. In order for the SA process to be effective, the increase of the cost function must be occasional; in other words, it must happen with a small probability, governed by a hyper-parameter, which we chose to be equal to 0.1% of the current value of the AIC at each iteration.
The algorithm workflow was as follows: 1.
Raw data were smoothed using convolution with a Gaussian kernel with σ = 1.5 units (corresponding to a physical sigma of approximately 3 nm at the centre of the spectrum).
The resulting spectra were then mean-centred and scaled to unit variance.

4.
After these pre-processing steps the spectra were fed in the SA algorithm. The algorithm starts with a random draw of 20 bands (out of the total 72). At each iteration, a subset of two bands were randomly swapped, a PLS model was developed and cross-validated on the collection of bands selected by the SA. The optimum PLS model at each step was obtained by minimising the AIC. The algorithm was run for 5000 iterations.
The number of selected bands (20 in our case) was fixed for all models and initially chosen in accordance with the empirical optimal value of the latent variables selected by the cross-validation process. As such value was of the order of 10, we chose the number of bands to be double that, to provide some redundancy for the dimensionality reduction of the PLS algorithm. In other words, if the number of bands was to be further decreased, the number of latent variables tended to become similar to the number of bands, hence the model became over-constrained. R 2 CV , r c , RMSE CV , AIC and number of latent variables (LV) in the models were presented for the preferred PLS models.
Lastly, the selected 20 wavelength bands were also used to predict the fruit maturity classes ( Table 1) in fruit of the four peach cultivars pooled together (n = 800) using a linear discriminant analysis (LDA). A three-class confusion matrix was used to visualise the performance of the model. The overall model's accuracy was based on the likelihood to predict the correct maturity class and expressed in F1-scores (0-1 range, where 1 is perfect prediction and 0 is no correct predictions) for each maturity class and for the whole model. The F1-score was calculated as the harmonic mean of precision and recall. Precision represents the number of true positive results divided by the sum of true and false positive results. Recall represents the number of true positive results divided by the sum of true positive results and false negative results. The model's error was calculated as 1 − F1.
The algorithms were implemented in Python TM 3.7.6 using the PLS regression and LDA routines available in the Scikit-learn package (v. 0.22.2) [44]. Sample scripts of the algorithms are freely available as a Project Jupyter Notebook [45].

Comparisons between Cultivars
Maturity indices were found to be significantly different between the four yellow peach cultivars ( Table 2). 'September Sun' had the largest FD and FW, whereas 'Redhaven' fruit were the smallest. Since 'Redhaven' fruit were harvested late, when already mature, their FF and I AD were significantly lower than in other cultivars ( Table 2). In 'August Flame', 'O'Henry' and 'September Sun', the median I AD was within the respective harvest-ready I AD range reported in Table 1. 'O'Henry' fruit had the highest sugars (median SSC of 14.8 • Brix), whilst 'Redhaven' fruit were the least sweet (11.9 • Brix).  Skin colour was significantly different between the cultivars under study (Table 2). 'Redhaven' fruit had the significantly highest skin L* (i.e., lighter colouration), a* (i.e., redder than other cultivars in the green-red scale), b* (i.e., more yellow than other cultivars in the blue-yellow scale), C* (i.e., higher colour vividness than other cultivars) and H • (i.e., more orange-red than other cultivars in the red to yellow scale, i.e., 0 to 90 • ).
Flesh colour was also found to be significantly different between 'August Flame', 'O'Henry', 'Redhaven' and 'September Sun' ( Table 2). Even in this case, the flesh of 'Redhaven' fruit had the highest L*, a*, b* and C*, but the lowest H • when compared to the other cultivars. Figure 1 highlights a colour-coded correlation strength between pairs of parameters, with darkest colours associated to strong correlations (blue for r ≥ 0.75 and red for r ≤ −0.75, p < 0.001) and white colour representing absent correlation (−0.10 < r < 0.10, p > 0.05). A strong correlation was observed between FW and FD in all the cultivars (Figure 1), as these two fruit size parameters are logically directly associated. In addition, strong correlation within skin colour attributes or within flesh colour attributes were not considered, as they are a consequence of collinearity among variables that are derived from their interdependency due to the fact that some of these variables are calculated from others (e.g., when

Correlations between Maturity Indices
Sensors 2020, 20, 6555 7 of 17 a* increases, H • decreases, when C* increases, b* increases). SSC was not strongly correlated with other parameters (r < 0.50 or >−0.50), whereas FF and I AD had an association with r > 0.50 in all the cultivars. FF and I AD were in turn highly correlated (r > 0.50) with the flesh colour attributes a* and H • . Flesh a* was negatively correlated with I AD , whereas flesh H • had a positive relationship.
The scatter plots with linear fits in Figure 2 highlight the relationships of flesh a* and H° with IAD in the four yellow peach cultivars. Table 3 shows linear regression models in which the IAD maturity threshold values reported in Table 1 were set as x to predict a* and H° thresholds between maturity classes in the four cultivars under study. 'Redhaven's' linear regression model generated relatively low R 2 (<0.50) for both a* and H°, as expected from correlation heatmaps (Figure 1), whereas in the other cultivars the models' R 2 were higher than 0.50. The range of a* and H° values in harvestready fruit was relatively small in each of the four cultivars-from 2.6 ('September Sun') to 6.0 ('Redhaven') for a* and from 2.6 ('September Sun') to 5.8 ('Redhaven') for H°. Maturity thresholds appeared to be cultivar-specific, although 'August Flame' and 'O'Henry' fruit had almost identical a* and H° thresholds. The linear regression slopes were similar but inverse for a* and H°.  Although 'Redhaven' fruit were picked at a late maturity stage, correlations between parameters had in most cases similar strengths to those observed in 'August Flame', 'O'Henry' and 'September Sun' (Figure 1).
The scatter plots with linear fits in Figure 2 highlight the relationships of flesh a* and H • with I AD in the four yellow peach cultivars. Table 3 shows linear regression models in which the I AD maturity threshold values reported in Table 1 were set as x to predict a* and H • thresholds between maturity classes in the four cultivars under study. 'Redhaven's' linear regression model generated relatively low R 2 (<0.50) for both a* and H • , as expected from correlation heatmaps (Figure 1), whereas in the other cultivars the models' R 2 were higher than 0.50. The range of a* and H • values in harvest-ready fruit was relatively small in each of the four cultivars-from 2.6 ('September Sun') to 6.0 ('Redhaven') for a* and from 2.6 ('September Sun') to 5.8 ('Redhaven') for H • . Maturity thresholds appeared to be cultivar-specific, although 'August Flame' and 'O'Henry' fruit had almost identical a* and H • thresholds. The linear regression slopes were similar but inverse for a* and H • .   Table 3 for regression coefficients). The three colours indicate IAD maturity classes reported in Table 1 Table 3. Linear regression models and predicted thresholds (y) of flesh a* and H° for the maturity classes "harvest-ready" and "mature" based on the indices of absorbance difference (x) reported in Table 1. Data reported for fruit of four yellow peach cultivars at harvest (n = 200). Intercept, slope and  Table 3 for regression coefficients). The three colours indicate I AD maturity classes reported in Table 1

Spectra Characteristics
To highlight the response of fluorescence emission under different maturity classes (based on the classification in Table 2) Figure 3 shows spectra of three individual 'August Flame' fruits. The immature and harvest-ready fruit emitted fluorescence peaks at~680-690 nm, with the latter being sensibly lower than the former. However, the harvest-ready fruit showed higher emission between 450 and 650 nm compared to the immature fruit. The mature fruit had no clearly defined peaks, but instead showed a hump with maximum emission between 600 and 650 nm, a smaller but wider hump in the 480-550 nm range and a barely visible hump at 680-690 nm. Both the humps at 600-650 and 480-550 nm observed in the mature fruit were more pronounced than in the harvest-ready fruit.

Spectra Characteristics
To highlight the response of fluorescence emission under different maturity classes (based on the classification in Table 2) Figure 3 shows spectra of three individual 'August Flame' fruits. The immature and harvest-ready fruit emitted fluorescence peaks at ~680-690 nm, with the latter being sensibly lower than the former. However, the harvest-ready fruit showed higher emission between 450 and 650 nm compared to the immature fruit. The mature fruit had no clearly defined peaks, but instead showed a hump with maximum emission between 600 and 650 nm, a smaller but wider hump in the 480-550 nm range and a barely visible hump at 680-690 nm. Both the humps at 600-650 and 480-550 nm observed in the mature fruit were more pronounced than in the harvest-ready fruit.
Measured maturity indices corresponding to the fruit shown in Figure 3 are reported in Table 4. The immature fruit had the highest FF, SSC, IAD, skin and flesh H°, and the lowest skin a* and flesh a*. This trend was expectedly inverted in mature fruit, in line with previous results (Figure 1).  Table 1). Figure 3 are reported in Table 4. The immature fruit had the highest FF, SSC, I AD , skin and flesh H • , and the lowest skin a* and flesh a*. This trend was expectedly inverted in mature fruit, in line with previous results (Figure 1). Table 4. FF, SSC, I AD , skin and flesh a* and H • in three individual 'August Flame' fruit shown in Figure 3.

PLS Models
Once fluorescence emission spectra were processed with the PLS algorithm, predicted values of FF, SSC, I AD , skin a* and H • and flesh a* and H • were obtained and compared to measured data. Figure 4 shows cross-validation predicted I AD plotted against the observed I AD . A visual assessment of the concordance between the two variables suggested that the PLS models provided a good estimation of I AD , as model linear fits were relatively close to a y = x reference line. Linear fits also show that the PLS models in the four cultivars slightly overestimated I AD when its values were low-i.e., mature fruit-and underestimated them when they were high-i.e., immature fruit. However, overestimations at a minimum observed I AD = 0.0 (+0. 17 Table 1). Table 4. FF, SSC, IAD, skin and flesh a* and H° in three individual 'August Flame' fruit shown in Figure  3.

PLS Models
Once fluorescence emission spectra were processed with the PLS algorithm, predicted values of FF, SSC, IAD, skin a* and H° and flesh a* and H° were obtained and compared to measured data. Figure 4 shows cross-validation predicted IAD plotted against the observed IAD. A visual assessment of the concordance between the two variables suggested that the PLS models provided a good estimation of IAD, as model linear fits were relatively close to a y = x reference line. Linear fits also show that the PLS models in the four cultivars slightly overestimated IAD when its values were lowi.e., mature fruit-and underestimated them when they were high-i.e., immature fruit. However, overestimations at a minimum observed IAD = 0.0 (+0. 17 Table 5.  2 Lin's concordance correlation coefficient, 3 Akaike information criterion, 4 root mean square error of the cross-validation and 5 number of latent variables. The calculated R 2 CV , r c and RMSE CV provided valuable information to better assess prediction power and error of the PLS models. A summary of these parameters is presented in Table 5 and highlight differences in models' robustness among cultivars. AIC values cannot be compared across cultivars and maturity indices, as they are only meant to be used to compare alternative models for the same dataset. The PLS model for FF and SSC predictions generated the overall lowest R 2 CV and r c among maturity indices. The fluorescence emission spectra could only poorly predict SSC and FF in 'September Sun' fruit (R 2 CV ≤ 0.50 and r c < 0.70). I AD was the only maturity index that was consistently and strongly predicted by the PLS models in all the cultivars (R 2 CV ≥ 0.80 and r c ≥ 0.89), regardless of maturity advancement, with RMSE CV < 0.25. R 2 CV and r c were also consistently in the 0.75-0.88 and 0.85-0.94 ranges, respectively, in models for the prediction of skin a* and H • , with the latter being always the one with highest precision between the two-colour attributes. Conversely, the accuracy of flesh a* and H • prediction differed in the four cultivars, ranging from the most precise predictions in 'O'Henry' (R 2 CV = 0.87 and r c = 0.93) to the poorest estimates in 'Redhaven' (R 2 CV < 0.50 and r c ≤ 0.65).

LDA Model
The LDA model for the pooled fruit dataset generated an overall F1 = 0.85; thus, the overall likelihood to estimate the correct maturity class was of approximately 85%, with an estimate error of 15%. The confusion matrix in Figure 5 shows that the model estimated 356, 185 and 135 true positives in mature, harvest-ready and immature fruit, when actual counts for the three classes were 387, 235 and 178, respectively. The single-class F1-scores were 0.93, 0.79 and 0.75 for mature, immature and harvest-ready fruit, respectively. The two opposite diagonal values in the black matrix cells show that only one mature fruit was classified as immature, and no immature fruit were classified as mature, providing additional strength to the model's results.

LDA Model
The LDA model for the pooled fruit dataset generated an overall F1 = 0.85; thus, the overall likelihood to estimate the correct maturity class was of approximately 85%, with an estimate error of 15%. The confusion matrix in Figure 5 shows that the model estimated 356, 185 and 135 true positives in mature, harvest-ready and immature fruit, when actual counts for the three classes were 387, 235 and 178, respectively. The single-class F1-scores were 0.93, 0.79 and 0.75 for mature, immature and harvest-ready fruit, respectively. The two opposite diagonal values in the black matrix cells show that only one mature fruit was classified as immature, and no immature fruit were classified as mature, providing additional strength to the model's results.

Discussion
The four cultivars selected for this study produced fruit with different characteristics, as highlighted in Table 2. The skin colour values reported for 'Redhaven' were partially influenced by advanced maturity due to the late sampling of fruit at harvest. However, skin colour was measured on the fruit cheek along the equatorial diameter, usually dominated by cover colour rather than ground colour, with the former being poorly influenced by maturity. Therefore, significantly different skin colouration among cultivars was most likely influenced by intrinsic genotypic characteristics. In flesh colour measurements, lower H° represented redder flesh, as also suggested by a* (Table 2), a typical consequence of fruit maturation. Therefore, flesh colouration differences in the four cultivars may be attributed to both genotypic characteristics and maturity stage.

Discussion
The four cultivars selected for this study produced fruit with different characteristics, as highlighted in Table 2. The skin colour values reported for 'Redhaven' were partially influenced by advanced maturity due to the late sampling of fruit at harvest. However, skin colour was measured on the fruit cheek along the equatorial diameter, usually dominated by cover colour rather than ground colour, with the former being poorly influenced by maturity. Therefore, significantly different skin colouration among cultivars was most likely influenced by intrinsic genotypic characteristics. In flesh colour measurements, lower H • represented redder flesh, as also suggested by a* (Table 2), a typical consequence of fruit maturation. Therefore, flesh colouration differences in the four cultivars may be attributed to both genotypic characteristics and maturity stage.
The fact that SSC was not strongly correlated with other parameters (Figure 1) suggests that sugars are a relatively independent fruit quality parameter that should be used with caution when attempting to assess maturity in yellow peaches. The strong correlations between FF and I AD indicate that there is a degree of interdependency between loss of firmness and chlorophyll degradation in yellow peaches. In addition, the strong correlation of chlorophyll degradation with the flesh colour attributes a* and H • was probably governed by the changes in flesh pigments, with an overall decrease of chlorophyll over the maturation period. An exception was highlighted in 'Redhaven', where the correlations between I AD and flesh a* and H • were poorer than in other cultivars, probably influenced by the fact that when fruit are more mature, I AD saturates towards its minimum value (i.e., 0.0) and fails to detect further physiological changes (e.g., cell wall and pectin modifications, internal darkening) that might instead be captured by variations in flesh colour attributes such as a* and H • . Our results support earlier work by Slaughter et al. [30], who observed a difference of approximately 5 • H of flesh hue between immature and mature peaches. The linear models shown in Table 3 highlight that the flesh colour attributes a* and H • may be used as maturity indices, although thresholds between maturity classes are cultivar specific. The similar but inverse slope in their relationship with I AD suggests that regression models based on one flesh colour attribute or the other are likely to have similar accuracy. This explains why literature shows the use of both a* and H • for fruit maturity assessments [30,46].
Typically, in immature fruit most of the fluorescence is emitted by chlorophyll pigments. Chlorophyll fluorescence has a primary peak at~683 nm and a secondary peak at~720 nm. The width of the peaks in stone fruit is usually comparable to their separation, so that the secondary peak appears as a shoulder between 700 and 750 nm [47] These features are clearly visible in immature and harvest-ready 'August Flame' fruit ( Figure 3). In mature fruit, since chlorophyll had mostly degraded, its fluorescence emission is greatly diminished, and the measurable signal at~683 nm is lower than the maximum emission emitted between 600 and 650 nm. The fluorescence emission observed at 600-650 nm is likely to be due to anthocyanins in the skin, as this range is compatible with fluorescence emission ranges of the 3-glucoside of malvidin (i.e., oenin) [48], an anthocyanin found in red grapes. Fluorescence emission at wavelength < 600 nm can be explained by a combination of flavonols and carotenoids. Flavonols such as quercetin, mostly found in yellow peach peels and scarcely present in flesh [49,50], emit green fluorescence with maxima near 520-530 nm [51]. Carotenoids have maximum fluorescence emission in the 450-550 nm range [52] and they are mainly present in the form of β-carotene and β-cryptoxanthin in yellow peach flesh [53]. On the one hand, the immature fruit showing a large chlorophyll peak at~683 nm had, expectedly, the lowest skin and flesh a* and the highest skin and flesh H • , i.e., greener (Table 4). On the other hand, the mature fruit showed the highest emission in the flavonol and carotenoid spectral ranges (Figure 3), meaning that these pigments were the most likely candidates to induce high a* and low H • in both skin and flesh (i.e., redder) ( Table 4). Changes of pigments concentrations in skin and flesh over fruit maturation are the basis for maturity prediction by fluorescence spectrometry.
PLS models highlighted that fluorescence emission can estimate FF and SSC in yellow peaches at harvest, although prediction is not very accurate (Table 5). I AD and skin a* and H • were consistently predicted in all the cultivars with good accuracy (R 2 CV ≥ 0.75 and r c ≥ 0.85). The prediction power for flesh a* and H • diverged among cultivars, with models for 'Redhaven'-the most mature fruit-showing low accuracy (R 2 CV < 0.50 and r c ≤ 0.65, Table 5). The high median values of skin L* and C* in 'Redhaven' fruit ( Table 2) may have led to increased self-absorption of the fluorescence from the skin, hence, a decreased contribution of the pulp emission to the total measured fluorescence. This in turn, could have caused a reduction of the prediction ability of fluorescence spectrometry for flesh a* and H • in this cultivar.
The LDA model's F1-score was equal to 0.85, which represented a very good likelihood of predicting the correct maturity class. All the class predictions showed robust predictions (F1 ≥ 0.75). The lowest F1 (0.75) was observed in the intermediate maturity class (harvest-ready) as in this case there were more chances to have a false positive on both sides of the class, i.e., neighbouring mature and immature classes. The highest accuracy obtained from the prediction of mature fruit (F1 = 0.93) was likely to be influenced by the larger sample size for this class (n = 387) compared to the other classes. The PLS models obtained for I AD and the LDA models obtained for maturity classes satisfied the hypothesis of our study, as they generated maturity predictions with accuracies ≥ 0.85, in terms of r c and F1-score, respectively.
The fluorescence spectrometer used in this study captures the steady-state fluorescence from the fruit under constant excitation light illumination. This system is fundamentally different from the pulse-amplitude modulation (PAM) fluorometers [54], which employ modulated excitation and fast detection to decouple fluorescence yield information from the measured fluorescence intensity. The sensitivity of a PAM fluorometer is derived from its time resolution and it is usually restricted to chlorophyll fluorescence. The fluorescence spectrometer described here does not produce fluorescence yield information, but it enables rapid measurements of a large spectral band without requiring specific sample preparation, most notably without requiring dark adaption of the fruit samples.
This characteristic is especially promising in view of field applications using hand-held devices, as the spectra can be captured without picking the fruit. The development of a wireless, battery-operated, handheld device for field estimation of fruit maturity is presently underway and early prototypes are being tested in field pilots of stone fruits (nectarine, white peach, plum, apricot) and pome fruits (apple, pear). While the current system gives good results in maturity estimation, the prediction of commercial maturity parameters such as SSC or FF is somewhat limited and very dependent on the cultivar. This suggests that a combination of fluorescence and NIR is required for improved performances. At the same time, a better modelling of the self-absorption properties of the fruit skin may help compensate from the signal reduction from the pulp, hence improve the model accuracy.
The ability to correctly predict maturity classes, though amenable to improvements, is suggestive of a post-harvest application of the fluorescence principle in graders. This application is currently being explored, in order to take advantage of prediction models that have the potential to be, to a large extent, independent from cultivar information within the same crop (e.g., yellow peach), hence better suited to a commercial operation with constantly changing cultivars being processed over a season.

Conclusions
In conclusion, the fluorescence spectrometer used in this study accurately predicted fruit maturity in yellow peaches. The machine learning prediction algorithms have the potential to be easily implemented in different configurations of the spectrometer (e.g., fluorescence alone, fluorescence + NIR, fluorescence + RGB and NIR in grading machines, portable or benchtop) for both in situ and post-harvest fruit maturity estimations. The sensor's applicability for fruit maturity prediction and for the assessment of fruit quality parameters and storage disorders in stone fruits and apples is currently under investigation at the Tatura SmartFarm. Funding: This study was financially supported by the following projects: (1) 'Horticulture Development Plan Task 1: Yield and quality relationships with light interception' funded by the Victorian Government's Agriculture Infrastructure and Jobs Fund; (2) 'SF17006: Summerfruit orchard-phase II' funded by Horticulture Innovation Australia Limited using Summerfruit levy and the Australian Government with co-investment from Agriculture Victoria; and (3) 'Deploying real-time sensors to meet Summerfruit export requirements for China' funded by Food Agility CRC Ltd., under the Commonwealth Government CRC Program. The CRC Program supports industry-led collaborations between industry, researchers and the community.