Enhancing Grape Brix Prediction in Precision Viticulture: A Benchmarking Study of Predictive Models Using Hyperspectral Proximal Sensors ^†

Abstract

Sustainable and efficient agricultural production is a growing priority in modern society. Viticulture, an important agricultural and food sector, also faces this challenge. Precision Viticulture (PV) has gained prominence as it aims to foster high-quality, efficient, and environmentally sustainable practices. The Soluble Solids Content (SSC) is essential for assessing grape ripeness and quality in the winemaking process. Conventional methods for determining SSC values (expressed in °Brix) are invasive, expensive, and labour-intensive, necessitating sample preparation, making large-scale analysis impractical. In response to these limitations, this study presents an innovative approach within the field of Precision Viticulture. It focuses on the non-invasive prediction of SSC using low-cost proximal hyperspectral optical sensors. These sensors rely on spectral reflectance measurements in the range of 340–850 nm. This study was conducted in a commercial vineyard in the Demarcated Douro Region, Cima-Corgo sub-region, Portugal, over six weeks during ripening. In total, 169 grape berries from Touriga Nacional vines were analysed under three irrigation regimes (no irrigation, 30% ETc, and 60% ETc). After organising and preprocessing the data, machine learning algorithms, namely Partial Least Squares Regression (PLS), Random Forest (RF), and the Generalised Linear Model (GLM), were applied to predict SSC values. These models’ performance was thoroughly evaluated using cross-validation techniques. The performance of different models was evaluated, showing significant differences according to the metrics used (R2, RMSE, and MAPE). The RF model demonstrated effectiveness and precision. A high R² value of 0.9312, coupled with low RMSE (0.9199 °Brix) and MAPE (3.88%), signifies a strong fit to the data and accurate predictive capabilities. The results of this benchmarking study on predictive models of SSC provide valuable insights into the performance of various models, aiding winegrowers and winemakers in decision making.

1. Introduction

The grapevine (Vitis vinifera L.) is a traditionally non-irrigated crop, but given the need to adapt to climate change on viticultural activity, several studies have shown that changes in the water status of the vine at critical phenological stages have a direct effect on the composition and qualitative attributes of the grape, affecting vegetative growth, yield, the microclimate of the canopy and the metabolism of the fruit [1,2].

In the vineyard, careful management is essential to determine the harvest date and select grapes at optimum ripeness according to the desired characteristics. Some methods traditionally used to measure physiological parameters in the vineyard, such as the Soluble Solids Content (SSC), expressed in °Brix, are destructive, expensive, and laborious as they require the preparation of samples and do not allow them to be spatialised from a PV perspective, making them less useful [3]. The real-time determination of processes related to abiotic stress and physiological processes related to fruit ripening (SSC, anthocyanins, carotenoids, and organic acids) provides support for precision practices of great relevance to vineyards and wine [4]. The environmental conditions that most strongly influence °Brix include sunlight, temperature, and humidity. Irrigation timing also affects °Brix due to reduced water availability during fruit development [5].

The common procedure used today by wine producers to evaluate the maturity of grapes in a vineyard is using a refractometer [6]. Proximity hyperspectral optical technologies, such as hyperspectral optical sensors (HOSs), offer potential in the non-destructive and cost-effective assessment of grape ripeness in wine. However, high costs are still a challenge for most farmers. INESC TEC, in partnership with the Faculdade de Ciências da Universidade do Porto (FCUP), has been developing low-cost HOS and thus is able to obtain a cause-effect physiological adherence to the spectral data collected in vineyards. After collecting the spectral data, they were processed and analysed. Machine learning (ML), an area of AI, develops models to learn from data and improves performance by identifying complex patterns and using them in predictions. This strategy has wide applications, both in agronomic decisions regarding crop performance in a given environment and in supporting cultural practices [7]. Despite these promising technological advances, there are still factors limiting the full adoption of PV systems, particularly in validating this proximal (HOS) under field conditions. These include shortcomings in terms of data acquisition, processing, and modelling to obtain useful information. There is a lack of a high-throughput system for mapping spacio-temporally in vineyards to fill the quality gap in the context of PV.

Monitoring grapevines over the ripening process in different hydric regimes, the main goal of this study was to develop a predictive model of SSC based on proximal detection data. The specific goals included (i) testing the performance of low-cost sensors developed at INESC TEC, (ii) benchmark ML models for SSC prediction, comparing the performance of each, and using appropriate metrics after applying pre-processing techniques to minimise undesirable effects, reducing data dimensionality and the matrix effect (spectral information on grape composition is characterised by multi-scale interference) and (iii) analyse the performance of predictive SSC models in different hydric conditions.

2. Materials and Methods

2.1. Grape Sampling and Data Acquisition

This research, carried out in the Douro region, was implemented at Quinta dos Aciprestes—latitude 41.21° N; and longitude 7.43° W. The farm is located next to the river, at altitudes of between 100 and 350 m, in the Douro Demarcated Region, a sub-region of Cima-Corgo. The farm benefits from a Mediterranean climate, with the following two distinct seasons: the wet season from October to April and the dry season from May to September. The experimental design used was randomised blocks, with each block including 6 plants subjected to different irrigation treatments: no irrigation (NR), 30% crop evapotranspiration (30% Etc), and 60% crop evapotranspiration (60% Etc). To obtain the most optimised model, it is crucial to evaluate the SCC of grapes from the initial ripeness stage until the ideal harvesting time. For each irrigation treatment, 4 grape berries were collected from each vine, in 2 rows and in 2 different locations in the row, over 6 weeks. Each week, one per vine was randomly selected from these berries, totalling 169 samples (28 July: n = 36, 4 August: n = 36, 11 August: n = 8, 18 August: n = 18, 25 August: n = 35, 1 September: n = 36).

The SSC expressed in the °Brix of the grapes was measured using an RHB-32ATC portable refractometer (Laxco Inc., Bothell, WD, USA)—a destructive method. The refractometer measures from 0 to 32 °Brix, with an accuracy of 0.20 °Brix and a resolution of ±0.2 °Brix. It was calibrated with a drop of distilled water and set to read 0 °Brix. The grapes previously measured using the spectroradiometer (next section) were carefully cut and pressed to use their juice for analysis with the refractometer. For the spectral acquisition, the equipment acquires spectra covering the ultraviolet, visible, and near-infrared zones, recording hyperspectral signatures between 340 nm and 850 nm of the electromagnetic spectrum [8]. This sensor has an LED light source (active sensor), which makes it possible to obtain spectra at night [9]. Hyperspectral point-of-measurement (HS-POM) measurements were taken by touching the berry to the light source; the power of the light source and the integration time were adjusted for the optimal recording of the spectra within the linear quantification region and, finally, the grape spectra were stored. Each spectrum was associated with the corresponding SSC reference measurements, resulting in the final dataset.

2.2. Modelation

The preprocessing data includes the filtering and reduction in spectral data. The hyperspectral data were filtered using the Savitzky–Golay digital filter to smooth the data, reduce noise, and preserve the important characteristics of the signals. Different window sizes were evaluated, assessing the impact on the result. Due to the dimensionality of the data, principal component analysis (PCA) was also conducted based on the standardised data set. This means that data was centred (by subtracting the mean of each variable) and scaled (divided by the corresponding standard deviations) before the PCA analysis, ensuring that variables with different scales did not dominate the principal components (PCs) due to their magnitudes and mitigating the matrix effect.

Throughout this study, different ML methods were evaluated to predict the SSC. The ML methods considered were as follows: Random Forest (RF), the Generalised Linear Model (GLM), and Partial Least Squares (PLS) [10]. RF is a model that is tolerant of data noise; its performance is high in determining spectral reflectance measurements due to its low sensitivity to outliers [11]. For this reason, two tests were conducted using RF, with spectral data filtered without PCA selection and with PCA selection, adjusting hyperparameters via Random Search. The PLS model was evaluated using Leave-One-Out Cross-Validation (LOOCV) and selecting the number of components (ncomp) based on the criterion of the lowest RMSE value. Finally, the GLM model was used to model a variety of distributions for the dependent variable, such as in the situation where data do not follow a normal distribution. The model was trained with a Gaussian distribution with cross-validation for the selection of PCs, and different hyperparameters were evaluated (distribution family, link functions) as well as other linear regressions, including Ridge and LASSO regressions.

To assess the generalisation capacity of the different models evaluated, the data were divided into two different sets via random sampling without replacement: the training set (70% of the data) and the validation set (30%). To assess the performance of each model, we evaluated and selected the most robust, and the following metrics were used: the coefficient of determination R²; root mean square error (RMSE); and mean absolute percentage error (MAPE). The residual analysis involved testing and analysing the distribution of residuals in the irrigation treatments.

3. Results

Analysing the average spectral curves per irrigation treatment (Figure 1), according to the absorption of the photosynthetic pigments, the reflectance of the irrigation treatments indicates that the NR treatment had the highest concentration of chlorophyll a (428 and 453 nm), chlorophyll b (642 and 661 nm) and carotenoids (400 and 500 nm) [12]. On the other hand, the 60% Etc treatment showed a lower reflectance in the chlorophyll and xanthophyll range (540–580 nm) [13] and anthocyanins (550 nm) [14].

The SSC predictive capacities of the models evaluated were assessed on the training and validation sets shown in

Table 1. Performance of the different ML models obtained in the training and validation sets for determining the SSC.

Models	Training Set			Validation Set
	R2	RMSE (°Brix)	MAPE (%)	R2	RMSE (°Brix)	MAPE (%)
RF	0.9895	0.3988	1.49	0.9312	0.9199	3.88
PCA+RF	0.9427	0.9072	3.66	0.7134	1.8585	8.35
PCA+PLS	0.6427	2.0696	0.08	0.6382	2.0414	0.09
PCA+GLM	0.9991	0.1009	0.00	0.9990	0.1076	0.01

. Comparing the two results obtained for each model evaluated allowed the most robust to be selected and overfitting to be analysed.

Preprocessing: Although 5 components explained 99% of the total variance, 3 to 15 PCs were considered for training and validating the different ML models, according to the results of the cross-validation applied to assess their performance.

The PCA+GLM model shows excellent results in the training and validation sets, with extremely high R² values and low RMSE and MAPE. However, these extremely good results may show overfitting due to high complexity [15]. Cross-validation stratified in relation to the different lambda values returned (data not showed) an RMSE of 3.45 °Brix and a MAPE of 17.72%, with relatively high cross-validation errors (MSE between 9.5 and 12.5). The RF model showed a slight drop in R² in the validation set compared to the training set. The PCA+RF model showed a significant drop in R² and an increase in RMSE and MAPE in the validation set compared to the RF model. The PCA+PLS model had a moderate performance in terms of R² on the validation set but an extremely low RMSE and MAPE. Given these results, the RF model was the most effective.

The SSC was estimated using the training and test sets of TN samples collected over the period under analysis. Figure 2 shows the results of the SSC estimation on both sets, allowing us to compare the predicted SSC values with the observed ones and conclude on the model’s performance in terms of its generalisation capacity. According to the metrics obtained in the validation set, this model explains 93.1% of the variance in the data (R² = 93.1%), with the average error of the predictions compared to the observed values of 0.920 (RMSE = 0.920 °Brix) and an average accuracy of the predictions deviating at 3.9% from the real values (MAPE = 3.88%).

Figure 3 presents the model’s residuals plotted against the observed °Brix.

The residuals were spread on both sides of the “zero line”, with no tendency for the residuals. The Shapiro–Wilk test results returned the W-Statistic value = 0.96607 and the p-value = 0.1512 (p_value > 0.05). The ANOVA (F = 0.262; p > 0.05) confirmed no significant differences between the irrigation treatments.

4. Discussion

In the spectral averages per irrigation treatment, two peaks of reflectance were visible. Reflectance in the 420–460 nm range was related to the absorption of chlorophyll a [12], which are green pigments involved in photosynthesis. However, the significant presence of these pigments in ripe grapes is unusual, as the plant directs energy towards the production of ripening-related compounds rather than chlorophyll. Therefore, it may be more related to carotenoids. Carotenoids are common in many fruits and vegetables, including grapes, as they are responsible for colours ranging from yellow to red. Visible reflectance in the 500–600 nm range may be related to anthocyanins, which are the pigments responsible for the red and purple colours in red grapes [14].

The application of low-cost hyperspectral optical sensors with machine learning models for Precision Viticulture presents a promising alternative to destructive and expensive conventional techniques. The developed RF model seems to be the safest choice in terms of overfitting, as it performs well in the validation set and has a moderate difference between the training and validation sets. The model’s evaluation metrics show that approximately 98.95% of the variability in SSC values can be explained with high precision. The PCA+RF model showed a significant drop in R² and an increase in RMSE and MAPE in the validation set compared to the RF model. This suggests that dimensionality reduction with PCA led to a loss of valuable information and poorer performance, leading to overfitting due to the loss of information during dimensionality reduction. The PCA+PLS model has a moderate performance in terms of R² in the validation set but an extremely low RMSE and almost zero MAPE. This may be indicative of a model with a high bias. Finally, the PCA+GLM model shows results suggesting overfitting, which occurs when an ML model overfits the training data, including noise, reducing its performance on new data sets. This means that the model fits the training data so well that it cannot generalise effectively to independent data, thus affecting the model’s ability to predict accurately. Essentially, the model learns not only the actual structure of data but also random fluctuations, which reduces its usefulness in real situations [15].

The RF model is the most robust choice, as it performs well in the validation set, and there is a slight difference between the training and validation sets. The residual analysis confirmed the null hypothesis (H0) that the residuals follow a normal distribution, and their homoscedasticity is satisfied, i.e., the variance of the residuals does not vary significantly as the predicted values increase. The RF model proved to be effective and accurate. The high R² value (0.9312) and the relatively low RMSE (0.9199 °Brix) and MAPE (3.88%) indicate that the model is adequately adjusted to the data and can make accurate predictions. This model was tested on a dataset with high variability in SSC values. The results demonstrate that the irrigation treatments did not significantly impact the model’s performance, which indicates the potential generalisation of the model’s results. These results are in line with other studies carried out in the Douro region, with the R² value (0.959) and RMSE (1.026 °Brix) [16], as well as in other regions, namely the Mediterranean, R² value (0.83), and RMSE (1.99 °Brix) with an RF model applied to the Syrah grape variety [17].

5. Conclusions

This work allowed for the benchmarking of SSC predictive ML models. Differences in the performance of four models tested on TN grape berries collected throughout the ripening period were demonstrated. The RF model is the most robust, not only because it is one of the models with the highest rate of explanation for the variation in SSC by the independent variables (R² = 0.9312), but it is also the safest choice in terms of overfitting. The MAPE values suggest that the model can make good predictions on both the training data (1.49% MAPE) and the test data (3.9% MAPE). The different irrigation treatments and the ripening date did not have a significant impact on the predictive abilities of this model.

The potential demonstrated by some of the models justifies the investment in low-cost hyperspectral optical sensors, such as Metbots. Evaluating the generalisation capacity using different vintages could generate new studies related to the rapid and non-destructive assessment of the ripeness of wine grapes.