Assessment of Smoke Contamination in Grapevine Berries and Taint in Wines Due to Bushfires Using a Low-Cost E-Nose and an Artificial Intelligence Approach

Bushfires are increasing in number and intensity due to climate change. A newly developed low-cost electronic nose (e-nose) was tested on wines made from grapevines exposed to smoke in field trials. E-nose readings were obtained from wines from five experimental treatments: (i) low-density smoke exposure (LS), (ii) high-density smoke exposure (HS), (iii) high-density smoke exposure with in-canopy misting (HSM), and two controls: (iv) control (C; no smoke treatment) and (v) control with in-canopy misting (CM; no smoke treatment). These e-nose readings were used as inputs for machine learning algorithms to obtain a classification model, with treatments as targets and seven neurons, with 97% accuracy in the classification of 300 samples into treatments as targets (Model 1). Models 2 to 4 used 10 neurons, with 20 glycoconjugates and 10 volatile phenols as targets, measured: in berries one hour after smoke (Model 2; R = 0.98; R2 = 0.95; b = 0.97); in berries at harvest (Model 3; R = 0.99; R2 = 0.97; b = 0.96); in wines (Model 4; R = 0.99; R2 = 0.98; b = 0.98). Model 5 was based on the intensity of 12 wine descriptors determined via a consumer sensory test (Model 5; R = 0.98; R2 = 0.96; b = 0.97). These models could be used by winemakers to assess near real-time smoke contamination levels and to implement amelioration strategies to minimize smoke taint in wines following bushfires.


Introduction
When bushfires occur within the grape growing season, vineyards can be affected at critical stages (véraison to harvest) [1], which could result in different levels of smoke contamination in berries and smoke taint in wines [2,3]. The intensity, number, and severity of bushfires are increasing due to climate change as well as the window of opportunity [4]. from treatments (Model 4), and (v) consumer sensory analysis using 12 wine descriptors (Model 5; Figure 1). The models obtained were of high accuracy, which could allow the implementation of this artificial intelligence (AI) technology in the winemaking process to assess the effect of ameliorating management techniques in the field (Model 1) through micro-vinifications, to assess the best timing for skin contact during fermentation for red wines, the addition of activated carbon to adsorb smoke-related compounds, wine filtration using membranes, reverse osmosis, and other commercial fining agents, among others [34,35].
Not only could the implementation of this technique help winemakers evaluate the different amelioration techniques mentioned above, but it could also monitor almost real-time changes in the aroma profiles of wine and assess which technology could best maintain a certain quality or style target.
This paper described how the e-nose was implemented for the different treatments and wine samples used and the specific machine learning algorithms used to develop five machine learning models with their respective analyses for accuracy and performance. A discussion on potential applications of the e-nose and models was also described for the wine industry to monitor and reduce smoke taint in wines.

Description of Treatments and Wine Samples
Field trials involving the application of smoke and/or in-canopy misting to Cabernet Sauvignon grapevines have been reported previously [3]. Briefly, three different smoke treatments were applied to vines (at approximately 7 days post-véraison): (i) low-density smoke exposure (LS), (ii) high-density smoke exposure (HS), and (iii) high-density smoke exposure, with in-canopy misting (HSM). Two controls were also included: (iv) a control without misting (C; no smoke treatment) and (v) a control with misting (CM; no smoke treatment). Treatments were applied to six adjacent vines, except for HSM, which was applied to five adjacent vines (i.e., one vine was missing). Smoke treatments involved exposure of grapevines to straw-derived smoke using a purpose-built tent for 1 h. At least one buffer vine separated treatments. The wine was subsequently produced on a small scale (i.e.,~5 kg per fermentation, performed in triplicate for each treatment), as described previously [3].

Electronic Nose
Wine samples were measured (in triplicate) using a portable, user-friendly, and low-cost e-nose, comprising nine different sensors, which were sensitive to different gases, as mentioned in Table 1, plus a humidity and temperature sensor (AM2320; Guangzhou Aosong Electronics Co., Ltd., Guangzhou, China). Sensor details have already been reported [20]. A total of 100 mL of wine was poured into a 500 mL beaker, and the e-nose was placed on top of the container for 1 min to capture the gases present in the sample. The e-nose was calibrated for 20-30 s before and after measuring each sample to reset the readings to baseline. Values from all sensors were automatically recorded in a comma-separated values (.csv) file to facilitate analysis.
A range of volatile phenol glycoconjugates (Table 2) was measured using high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS/MS) according to stable isotope dilution analysis (SIDA) methods previously described [18,20]. The analysis was performed using an Agilent 1200 high-performance liquid chromatography (HPLC) equipped with a 1290 binary pump, coupled to an AB SCIEX Triple QuadTM 4500 tandem mass spectrometer, with a Turbo VTM ion source (Framingham, MA, USA). The preparation of the isotopically labeled internal standard d 3 -syringol gentiobioside has been previously reported [18,20]. The limit of quantitation for volatile phenol glycosides was 1 µg kg −1 .

Sensory Evaluation-Consumer Test
A consumer test was conducted with participants (N = 31; age range: 21-59 years; 77% female and 23% male) constituted of staff and students from The University of Melbourne (UoM; Ethics ID: 1545786.2) that had been recruited via e-mail. According to the power analysis conducted using the SAS ® Power and Sample Size v. 14.1 software (SAS Institute Inc., Cary, NC, USA), the number of participants was enough to find significant differences between samples (power: 1 − β > 0.99). The session was carried out in the sensory laboratory of the Faculty of Veterinary and Agricultural Sciences (FVAS) in individual booths with uniform white light-emitting diode (LED) lights. Each booth was equipped with a tablet PC in which the Bio-Sensory Application (The University of Melbourne, Parkville, VIC, Australia) was set up with the questionnaire to gather consumer responses. The appearance, overall aroma, smoke aroma, bitterness, sweetness, acidity, astringency, a warming sensation, and overall liking were assessed on a likeness scale (i.e., dislike extremely-neither like nor dislike-like extremely). The levels of smoke aroma and perceived quality were rated on an intensity scale (i.e., absent-intense). Both liking and intensity measures were presented on a 15 cm non-structured continuous scale. In addition, emotional responses were recorded, using a 0-100 FaceScale, where 0 = sad

Sensory Evaluation-Consumer Test
A consumer test was conducted with parti and 23% male) constituted of staff and students 1545786.2) that had been recruited via e-mail. A SAS ® Power and Sample Size v. 14.1 software participants was enough to find significant diff session was carried out in the sensory labora Sciences (FVAS) in individual booths with uni booth was equipped with a tablet PC in whi Melbourne, Parkville, VIC, Australia) was se responses. The appearance, overall aroma, smok a warming sensation, and overall liking were a neither like nor dislike-like extremely). The lev on an intensity scale (i.e., absent-intense). Both l cm non-structured continuous scale. In addition FaceScale, where 0 = sad ☹, 50 = neutral , an a 3-digit code, and 10 mL samples were served a Wine Tasting Glasses (Bormioli Luigi, Fidenza, I bias. Plain water and water crackers were used

Statistical Analysis and Machine Learning Mod
Analysis of variance (ANOVA) was cond Addinsoft Inc., New York, NY, USA), and Tuke was used to assess significant differences betwe Machine learning modeling was performed pattern recognition and regression models, usin Natick, MA, USA) developed to test 17 differ developed using 20 data points from the peak o 1 (pattern recognition) used the scaled conjuga samples into the five different treatments: (i) regression models were developed using the Le the use of the 20 glycoconjugates and 10 volati being exposed to smoke as targets. In compariso volatile phenols in berries but measured at glycoconjugates and seven volatile phenols an hand, Model 5 was developed to predict 12 sens overall aroma, (iii) smoke aroma, (iv) bitterness warming sensation, (ix) overall liking, and (x) th and (iii) the FaceScale emotional response as tar All inputs and targets were normalized fro models, with 60% of the data being used for the t Model 1 used a cross-entropy loss to test per onsumer test was conducted with participants (N = 31; age range: 21-59 years; 77% female male) constituted of staff and students from The University of Melbourne (UoM; Ethics ID: .2) that had been recruited via e-mail. According to the power analysis conducted using the wer and Sample Size v. 14.1 software (SAS Institute Inc., Cary, NC, USA), the number of ants was enough to find significant differences between samples (power: 1 − β > 0.99). The was carried out in the sensory laboratory of the Faculty of Veterinary and Agricultural (FVAS) in individual booths with uniform white light-emitting diode (LED) lights. Each as equipped with a tablet PC in which the Bio-Sensory Application (The University of rne, Parkville, VIC, Australia) was set up with the questionnaire to gather consumer es. The appearance, overall aroma, smoke aroma, bitterness, sweetness, acidity, astringency, ing sensation, and overall liking were assessed on a likeness scale (i.e., dislike extremelyike nor dislike-like extremely). The levels of smoke aroma and perceived quality were rated tensity scale (i.e., absent-intense). Both liking and intensity measures were presented on a 15 structured continuous scale. In addition, emotional responses were recorded, using a 0-100 le, where 0 = sad ☹, 50 = neutral , and 100 = happy . Samples were randomly assigned code, and 10 mL samples were served at room temperature (20 °C) in International Standard sting Glasses (Bormioli Luigi, Fidenza, Italy). Samples were served in random order to avoid in water and water crackers were used as palate cleansers between samples.
istical Analysis and Machine Learning Modeling alysis of variance (ANOVA) was conducted on e-nose data using XLSTAT (ver. 19.3.2, ft Inc., New York, NY, USA), and Tukey's honest significant difference test (HSD; α = 0.05) d to assess significant differences between treatments. chine learning modeling was performed based on artificial neural networks (ANN) for both recognition and regression models, using codes written in Matlab ® R2019b (Mathworks, Inc., MA, USA) developed to test 17 different training algorithms. Five distinct models were ed using 20 data points from the peak of the e-nose outputs (nine sensors) as inputs. Model rn recognition) used the scaled conjugate gradient training algorithm to classify the wine into the five different treatments: (i) LS, (ii) HS, (iii) HSM, (iv) C, and (v) CM. All four on models were developed using the Levenberg Marquardt algorithm. Model 2 consisted of of the 20 glycoconjugates and 10 volatile phenols (Table 2) found in berries one hour after posed to smoke as targets. In comparison, Model 3 used the same 20 glycoconjugates and 10 phenols in berries but measured at harvest. The targets used for Model 4 were 17 njugates and seven volatile phenols analyzed in the wine samples (Table 2). On the other odel 5 was developed to predict 12 sensory responses, using the liking of (i) appearance, (ii) aroma, (iii) smoke aroma, (iv) bitterness, (v) sweetness, (vi) acidity, (vii) astringency, (viii) a g sensation, (ix) overall liking, and (x) the intensity of (i) smoke aroma, (ii) perceived quality, the FaceScale emotional response as targets. inputs and targets were normalized from −1 to 1. Data were divided randomly for all ANN with 60% of the data being used for the training stage, 20% for validation, and 20% for testing.
, and 100 = happy s 2020, 20 2) that had been recruited via e-mail. According to the power analysis conducted using the Power and Sample Size v. 14.1 software (SAS Institute Inc., Cary, NC, USA), the number of cipants was enough to find significant differences between samples (power: 1 − β > 0.99). The on was carried out in the sensory laboratory of the Faculty of Veterinary and Agricultural ces (FVAS) in individual booths with uniform white light-emitting diode (LED) lights. Each was equipped with a tablet PC in which the Bio-Sensory Application (The University of ourne, Parkville, VIC, Australia) was set up with the questionnaire to gather consumer nses. The appearance, overall aroma, smoke aroma, bitterness, sweetness, acidity, astringency, rming sensation, and overall liking were assessed on a likeness scale (i.e., dislike extremelyer like nor dislike-like extremely). The levels of smoke aroma and perceived quality were rated intensity scale (i.e., absent-intense). Both liking and intensity measures were presented on a 15 on-structured continuous scale. In addition, emotional responses were recorded, using a 0-100 cale, where 0 = sad ☹, 50 = neutral , and 100 = happy . Samples were randomly assigned igit code, and 10 mL samples were served at room temperature (20 °C) in International Standard Tasting Glasses (Bormioli Luigi, Fidenza, Italy). Samples were served in random order to avoid Plain water and water crackers were used as palate cleansers between samples.

tatistical Analysis and Machine Learning Modeling
Analysis of variance (ANOVA) was conducted on e-nose data using XLSTAT (ver. 19.3.2, insoft Inc., New York, NY, USA), and Tukey's honest significant difference test (HSD; α = 0.05) used to assess significant differences between treatments. Machine learning modeling was performed based on artificial neural networks (ANN) for both rn recognition and regression models, using codes written in Matlab ® R2019b (Mathworks, Inc., k, MA, USA) developed to test 17 different training algorithms. Five distinct models were loped using 20 data points from the peak of the e-nose outputs (nine sensors) as inputs. Model ttern recognition) used the scaled conjugate gradient training algorithm to classify the wine les into the five different treatments: (i) LS, (ii) HS, (iii) HSM, (iv) C, and (v) CM. All four ssion models were developed using the Levenberg Marquardt algorithm. Model 2 consisted of se of the 20 glycoconjugates and 10 volatile phenols (Table 2) found in berries one hour after exposed to smoke as targets. In comparison, Model 3 used the same 20 glycoconjugates and 10 ile phenols in berries but measured at harvest. The targets used for Model 4 were 17 conjugates and seven volatile phenols analyzed in the wine samples (Table 2). On the other , Model 5 was developed to predict 12 sensory responses, using the liking of (i) appearance, (ii) ll aroma, (iii) smoke aroma, (iv) bitterness, (v) sweetness, (vi) acidity, (vii) astringency, (viii) a ing sensation, (ix) overall liking, and (x) the intensity of (i) smoke aroma, (ii) perceived quality, iii) the FaceScale emotional response as targets. All inputs and targets were normalized from −1 to 1. Data were divided randomly for all ANN els, with 60% of the data being used for the training stage, 20% for validation, and 20% for testing.
. Samples were randomly assigned a 3-digit code, and 10 mL samples were served at room temperature (20 • C) in International Standard Wine Tasting Glasses (Bormioli Luigi, Fidenza, Italy). Samples were served in random order to avoid bias. Plain water and water crackers were used as palate cleansers between samples.

Statistical Analysis and Machine Learning Modeling
Analysis of variance (ANOVA) was conducted on e-nose data using XLSTAT (ver. 19.3.2, Addinsoft Inc., New York, NY, USA), and Tukey's honest significant difference test (HSD; α = 0.05) was used to assess significant differences between treatments.
Machine learning modeling was performed based on artificial neural networks (ANN) for both pattern recognition and regression models, using codes written in Matlab ® R2019b (Mathworks, Inc., Natick, MA, USA) developed to test 17 different training algorithms. Five distinct models were developed using 20 data points from the peak of the e-nose outputs (nine sensors) as inputs. Model 1 (pattern recognition) used the scaled conjugate gradient training algorithm to classify the wine samples into the five different treatments: (i) LS, (ii) HS, (iii) HSM, (iv) C, and (v) CM. All four regression models were developed using the Levenberg Marquardt algorithm. Model 2 consisted of the use of the 20 glycoconjugates and 10 volatile phenols (Table 2) found in berries one hour after being exposed to smoke as targets. In comparison, Model 3 used the same 20 glycoconjugates and 10 volatile phenols in berries but measured at harvest. The targets used for Model 4 were 17 glycoconjugates and seven volatile phenols analyzed in the wine samples (Table 2). On the other hand, Model 5 was developed to predict 12 sensory responses, using the liking of (i) appearance, (ii) overall aroma, (iii) smoke aroma, (iv) bitterness, (v) sweetness, (vi) acidity, (vii) astringency, (viii) a warming sensation, (ix) overall liking, and (x) the intensity of (i) smoke aroma, (ii) perceived quality, and (iii) the FaceScale emotional response as targets.
All inputs and targets were normalized from −1 to 1. Data were divided randomly for all ANN models, with 60% of the data being used for the training stage, 20% for validation, and 20% for testing. Model 1 used a cross-entropy loss to test performance, while Models 2-5 were based on means squared error (MSE). Figure 1 shows the diagrams for Model A (Figure 1a), Models 2-4 (Figure 1b), and Model 5 ( Figure 1c); all models consisted of a two-layer feedforward network with the hidden layer using a tan-sigmoid function and the output layer using softmax neurons (Model 1) and a linear transfer function (Models 2-5). A trimming test (data not shown) was performed to find the optimal number of neurons (3,5,7,10) to get the best performance. Statistical data reported for regression models to assess under-or overfitting consist of the correlation coefficient (R), slope (b), MSE, and determination coefficient (R 2 ); the latter was calculated using the curve fitting tool found in Matlab ® .  Figure 2 shows the results from the ANOVA for the e-nose responses. It can be observed that there were significant differences (p < 0.05) between samples in the outputs from all nine sensors that integrated the e-nose. Ethanol gas (MQ3) presented the highest values for all wine samples with CM (mean = 4.07 V) being significantly different from HSM (mean = 3.85 V), HS (mean = 3.82 V), and C  Figure 2 shows the results from the ANOVA for the e-nose responses. It can be observed that there were significant differences (p < 0.05) between samples in the outputs from all nine sensors that integrated the e-nose. Ethanol gas (MQ3) presented the highest values for all wine samples with CM (mean = 4.07 V) being significantly different from HSM (mean = 3.85 V), HS (mean = 3.82 V), and C (mean = 3.92 V), and these from LS (mean = 3.66 V). Hydrogen sulfide (MQ136) was the lowest for all samples, and CM (mean = 0.34 V) was significantly different from all other samples (means = 0.23-0.27 V). The CO 2 sensor readings are inverse; therefore, higher Volts mean lower concentration; it can be observed that all the samples with smoke treatments (LS, HS, and HSM) had the lowest CO 2 and presented significant differences with control samples (CM and C).   Table 3 shows the minimum, maximum, and average values of the glycoconjugates and volatile phenols detected in berries one hour after smoking, in berries at harvest, and wine. It can be observed that there was a wide range of values for all of the compounds, which indicated these were adequate samples to be used for machine learning modeling and to detect smoke contamination.   Table 3 shows the minimum, maximum, and average values of the glycoconjugates and volatile phenols detected in berries one hour after smoking, in berries at harvest, and wine. It can be observed that there was a wide range of values for all of the compounds, which indicated these were adequate samples to be used for machine learning modeling and to detect smoke contamination.  Table 4 shows the minimum, maximum, and average values of the responses from the sensory session conducted with consumers when evaluating the wines. It can be observed that the results from all attributes were within the whole range of the scales used for liking and appearance (0-15) and FaceScale (0-100), which made the data suitable to be used for machine learning modeling.  Table 5 shows the statistical results from Model 1 for the classification of the samples into the five different treatments. It can be observed that there was a high accuracy for all stages (>90%) and 97% for the overall model. According to the performance values, there were no signs of overfitting, as the training stage had a cross-entropy value lower than the validation and testing, and these two had similar performance. In Figure 3, the results from the receiver operating characteristic (ROC) curve are shown. This graph depicted the sensitivity (true positive rate) and specificity (false positive rate) of the overall model, with optimal operating points of 98%, 100%, 93%, 93%, and 98% for C, CM, LS, HS, and HSM, respectively.  Table 5 shows the statistical results from Model 1 for the classification of the samples into the five different treatments. It can be observed that there was a high accuracy for all stages (>90%) and 97% for the overall model. According to the performance values, there were no signs of overfitting, as the training stage had a cross-entropy value lower than the validation and testing, and these two had similar performance. In Figure 3, the results from the receiver operating characteristic (ROC) curve are shown. This graph depicted the sensitivity (true positive rate) and specificity (false positive rate) of the overall model, with optimal operating points of 98%, 100%, 93%, 93%, and 98% for C, CM, LS, HS, and HSM, respectively.   Table 6 depicts the statistical data for the four regression models. Model 2 had very high overall correlation and determination coefficients (R = 0.98; Figure 4a; R 2 = 0.95). The close value of the validation and training correlation coefficients (R = 0.96 and R = 0.98, respectively), along with the fact that the performance of the training stage (MSE = 0.01) was lower than that of the validation and

Discussion
Nowadays, the only alternative for grape growers is to apply potential amelioration techniques before the bushfires and hope for the best since there are limited tools that can be applied in the field or at the winemaking stage, which can render results in near real-time for proper decision-making [35,36]. Recently, non-invasive devices have been proposed using infrared thermal imaging to assess contaminated grapevine canopies in the field and smoke taint in berries and wines using near-

Discussion
Nowadays, the only alternative for grape growers is to apply potential amelioration techniques before the bushfires and hope for the best since there are limited tools that can be applied in the field or at the winemaking stage, which can render results in near real-time for proper decision-making [35,36]. Recently, non-invasive devices have been proposed using infrared thermal imaging to assess contaminated grapevine canopies in the field and smoke taint in berries and wines using near-infrared spectroscopy [5,33]. The research presented in this paper has contributed to the potential implementation of new and emerging sensor technologies and modeling strategies using machine learning in the viticulture and winemaking industry. These low-cost e-noses could become a game-changer for the management of smoke contamination and taint in berries and wines due to bushfires.
In general, previous applications of e-noses in the wine industry have been implemented mainly for the analysis of grapes and crushing methods [37], improvement of maceration and fermentation processes [38], to monitor the aging of wine in barrels [39][40][41], geographical classification [42], wine spoilage [28,43,44], and to assess correlations with human perception through sensory evaluation [27,29,45]. However, most of these studies have been based on multivariate data analysis and correlation analysis.
Low-cost sensors presented in this research, developed by integrating an array of gas sensors [20], could be used in the winery to assess the level of grapevine smoke exposure. In the present study, models were developed to evaluate the effects of different amelioration techniques (Model 1) for berries immediately after the bushfire event (Model 2), at harvest time (Model 3), and in the actual wines (Model 4). Since smoke-derived glycoconjugates in berries are difficult to detect using e-noses due to the binding of these compounds with sugars in the berries, these assessments need to be performed after the winemaking process, in which the compounds are released through the maceration and fermentation processes.
A further model (Model 5) developed to assess sensory characteristics of wines rapidly and objectively, which can be implemented in parallel with successful amelioration techniques to reduce smoke taint, such as the addition of activated carbon to wines or fining agents [2,34]. For the latter case, Model 5 will offer a near real-time assessment of the techniques used.
The advantages of implementing these models coupled with low-cost sensor technology are that grape growers and winemakers will not depend on random sampling, which may not render representative results, or external laboratory services, which may not deliver results in a timely manner due to being overwhelmed by large sample volumes that are delivered when concurrent bushfires occur. Knowing the levels of smoke-derived compounds and the effects on consumer appreciation in the winemaking process offer the following advantages: (i) rapid and user-friendly smoke taint determination; (ii) potential implementation of techniques to reduce smoke taint using activated carbon or fining agents on samples and re-test using the e-nose and models developed; (iii) sensory panel not required for assessments/modifications, minimizing the time for the commercial release of wines and economic impacts of smoke taint.
Further applications of these low-cost e-noses can be implemented to assess the maturity of grapes in the field, specifically through the alcohol-based sensors. The latest research has shown that ethanol is released from grape berries when they become oxygen stressed [46]. So, being able to assess when cell death begins would be a useful tool in monitoring berry health and fruit ripening potential. These processes of berry cell death assessment can be done non-destructively by near-infrared spectroscopy and machine learning modeling [47] or by tracking ethanol release from grapevine bunches through the implementation of low-cost e-noses in the field using sensor networks or as a payload of low altitude unmanned aerial vehicle (UAV) surveys [48,49].

Conclusions
Low-cost e-nose sensor technology coupled with machine learning offers the advantage of easy implementation in field conditions using sensor networks or in the winery. Machine learning models obtained could make available valuable information to winemakers and winegrowers for the decision-making process to produce commercial wines by minimizing smoke taint. An artificial intelligence system can be implemented based on sensor technology and machine learning developed here to obtain the least tainted wine or to target specific sensory aroma profiles to take advantage of the decontamination process to maximize the likability of wines.