In-Field Hyperspectral Proximal Sensing for Estimating Grapevine Water Status to Support Smart Precision Viticulture ^†

Abstract

Predawn leaf water potential (Ψ_pd) is the main parameter to determine plant water status, and it has been broadly used to support irrigation management. However, the Scholander pressure chamber methodology is laborious, time-consuming and invasive. This study examined a low-cost hyperspectral proximal sensor to estimate the Ψ_pd in grapevine (Vitis vinifera L.). For this, both the Ψ_pd and spectral reflectance (340–850 nm) were accessed in grapevines in a commercial vineyard located in the Douro Wine Region, northeast Portugal. A machine-learning algorithm was tested and validated to assess grapevine’s water status. The experiment was performed in a randomized design with 12 grapevines (cv. Touriga Nacional) per irrigation treatment: non-irrigated, 30% crop evapotranspiration (Etc), and 60% Etc. The dataset was analyzed using Principal Component Analysis (PCA), and the machine-learning regression algorithm applied was Extreme Gradient Boosting (Xgboost). Results from the validation dataset (n = 108) for the Xgboost tested exhibited a root mean square error (RMSE) of 0.23 MPa, a mean absolute error (MAPE) of 16.57% and an R² value of 0.95. These results demonstrate that the hyperspectral sensor and Xgboost algorithm show potential for predicting the Ψ_pd in vineyards, regardless of a plant’s water status.

1. Introduction

Viticulture is an activity of major economic importance in Mediterranean regions. However, climatic changes have affected wine production and quality over the years in these regions [1]. Because of that, irrigated vineyards have been increasing due to the necessity to reduce climatic changes’ effects on viticulture [2]. Simultaneously, human habits demand water resources and put pressure on agriculture to reduce water waste and improve irrigation management [2,3].

Predawn leaf water potential (Ψ_pd) is an eco-physiological indicator used to assess grapevine’s water status and support irrigation management in vineyards [4]. Commonly, it is measured using a Scholander pressure chamber. Nonetheless, this method is destructive, time-consuming, laborious, and depends on set of measurement points to represent large planting areas [5].

Technological advances such as optical sensors are a fast, cost-effective, and non-invasive method to estimate Ψ_pd with high accuracy [6]. Specifically, hyperspectral proximal sensors and machine-learning algorithms showed good results in predicting Ψ_pd [7]. However, more studies are necessary to develop this technology, focusing on making it economical and readily available to wine producers [8]. This study examined a low-cost hyperspectral proximal sensor to estimate the Ψ_pd in grapevine (Vitis vinifera L.).

2. Materials and Methods

2.1. Test Site

The research was carried out in Quinta dos Aciprestes (41.21° N; 7.43° W; 145 m), a commercial vineyard (wine company Real Companhia Velha) in the Douro Wine Region, northeast of Portugal (Figure 1). The vines in the test site follow a Bilateral Royat System, with canopy height around 1.5 m and plant spacing of 2.2 m × 1 m. The cultivar studied was Touriga Nacional, in a randomized design in 2 plots, with two replicate areas. Each plot included three irrigation treatments: non-irrigated, irrigation to replace 30% of evapotranspiration (Etc) water volume, and 60% Etc, managed by the wine company. For each irrigation treatment, we considered six grapevines, totaling 36 grapevines being assessed.

2.2. Predawn Leaf Water Potential and Spectroscopy Methodology

The predawn leaf water potential (Ψ_pd) was measured using a Scholander pressure chamber [9] (PMS600, Albany, OR, USA) at predawn on one leaf from each vine per block sampled. The hyperspectral data were measured using a prototype directed vertically close to the leaf, which recorded reflectance signatures between 340 nm and 850 nm of the electromagnetic spectrum [10]. The Ψ_pd and hyperspectral data were measured in the same leaves during six consecutive weeks in 2022 after veraison (25 July) on evaluation dates: 28 July—3 days after veraison (DAV), 4 August (10 DAV), 11 August (17 DAV), 19 August (25 DAV), 25 August (31 DAV) and 1 September (38 DAV), totaling 216 observations.

2.3. Statistical and Principal Component Analysis

A one-way analysis of variance (ANOVA) with a p-value associated with Fisher’s exact test (p ≤ 0.05) was performed to compare the means of Ψ_pd between the irrigation treatments and the evaluation dates.

A Principal Component Analysis (PCA) was performed to discover patterns that could explain the variance of spectral data on evaluation dates. The PCA considered three evaluation dates to represent the different stages of grape ripening: 28 July (3 DAV), 11 August (17 DAV) and 1 September (38 DAV).

2.4. Data Processing and Modelling

A multiplicative scattering correction logarithm was applied in the hyperspectral reflectance to correct baseline and scattering artefacts in spectra data before algorithm processing [7].

The machine-learning algorithm tested was Extreme Gradient Boosting (Xgboost). For Xgboost’s application, the dataset was randomly split into two sub-datasets: 50% to calibration (n = 108 observations) and 50% to validation (n = 108 observations) [11]. The statistical analysis and machine-learning algorithm were computed in R (version 4.2.3) [12].

Model performance in estimating Ψ_pd was evaluated based on the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAPE).

3. Results

Table 1. Statistical results of predawn leaf water potential (Ψ_pd, MPa) for the irrigation treatments in different evaluation dates during the stages of grape ripening.

Evaluation Date	Non-Irrigated	30% Etc	60% Etc	ANOVA F
28 July	−1.263 ± 0.226 bA	−0.560 ± 0.072 dB	−0.469 ± 0.119 bB	5.72 × 10−14 ***
4 August	−1.294 ± 0.185 bA	−1.096 ± 0.231 aB	−0.575 ± 0.020 aC	7.12 × 10−11 ***
11 August	−1.475 ± 0.116 aA	−0.729 ± 0.090 cB	−0.650 ± 0.122 aB	<2 × 10−16 ***
19 August	−1.490 ± 0.123 aA	−0.717 ± 0.095 cB	−0.433 ± 0.087 bC	<2 × 10−16 ***
25 August	−1.558 ± 0.095 aA	−0.752 ± 0.155 cB	−0.446 ± 0.075 bC	<2 × 10−16 ***
1 September	−1.538 ± 0.124 aA	−0.958 ± 0.146 bB	−0.390 ± 0.120 bC	<2 × 10−16 ***
Mean	−1.436 A	−0.802 B	−0.494 C	<2 × 10−16 ***
ANOVA F	1.08 × 10−5 ***	4.04 × 10−12 ***	7.48 × 10−8 ***

ANOVA F is the p-value associated with Fisher’s exact test performed in the ANOVA. Means with a p-value less than 0.05 are considered statistically different. Means followed by the same lowercase letter in the same column (among dates) or means followed by the same uppercase letter in the same line (among irrigation treatments) are not significantly different (α = 5%). The asterisk in Table 1 indicates the p-value and represents the statistical significance of the Analysis of Variance (ANOVA). A value with *** signifies that the test was highly significant, with a probability value of p ≤ 0.001.

presents statistical results of the measured Ψ_pd and shows significant differences among all the evaluation dates and irrigation treatments. Considering the evaluation dates, non-irrigated grapevines presented a better statistical mean than irrigated grapevines. Due to a 1.6 mm precipitation on 10 August (16 DAV), the values of the Ψ_pd since 4 August showed a variability, mainly in 30% Etc treatment.

Figure 2 shows PCA that the spectral data in the evaluation dates indicate different cycles of leaf development. The first two principal components accounted for a variance > 75%. The figure represents the clustering according the dates of assessment, suggesting that the leaf’s ageing and abiotic stress (e.g., water stress) was be detectable through spectral data. The spectral data is grouped by evaluation date in a pattern related to the changes in the leaves’ composition caused by aging.

Table 2. Model performance metrics of training and validation datasets.

	Training Dataset	Validation Dataset	Total
R2	1	0.89	0.95
RMSE (MPa)	0.00	0.32	0.23
MAPE (%)	0.09	33.36	16.57

and Figure 3 show the performance of the Extreme Gradient Boosting regression model. This model performed a variance of 0.95% in Ψ_pd prediction, the RMSE of 0.23 MPa and the MAPE of 16.57%. The R² of 1 and RMSE of 0.00 in the training dataset indicate that there is a tendency of overfitting behavior observed in data features, probably due to the structure of the data. Figure 3 supports the tendency of overfitting, showing an effective metric for the training dataset and a lower prediction in the validation dataset.

4. Discussion

Low water regimes can induce water stress symptoms in grapevines and affect their phenology. Water stress causes stomata closure; in consequence, leaf temperatures increase through reducing transpiration rates and reducing CO₂ concentration, along with photodamage to photosystem II (PSII) reaction centers [13]. Spectral sensors provide information about plant physiology before water stress impacts on grapevine productivity.

The low-cost hyperspectral proximal sensor shows promising results in estimating the Ψ_pd in grapevine. The prototype sensor predicted the Ψ_pd considering a variance between −0.2 MPa and −1.7 MPa with an error of 0.23 MPa. These results allow for the use of the sensor by grapevine producers that conduct their vineyard under deficit irrigation. In these areas, the irrigation schedule starts when vineyards present Ψ_pd values below −0.4 MPa [14], inside the range detected by the sensor studied.

Even though the Scholander pressure chamber is a recognized method to access Ψ_pd, it is not widely applicable in large vineyards, besides being time and labor consuming. Jointly, the sensor and the Extreme Gradient Boosting model provide real-time and reliable detections of plants’ water status. They can be used to support irrigation decisions by winegrowers in an automated system that would allow adjustments in irrigation to the site-specific characteristics, according to yield and quality objectives [15].

5. Conclusions

This work shows the applicability of a low-cost hyperspectral proximal sensor and the Extreme Gradient Boosting regression model in predicting Ψ_pd by considering different evaluation dates and irrigation treatments. These results indicate the potential for this model and sensor to be applied in vineyards to support irrigation management.

There are other applications of this sensor to be studied, such as integrating this sensor in robots to optimize the measurements of Ψ_pd in vineyards, and quantifying berry quality parameters and leaves’ pigments.