A Machine Learning Approach to Monitor the Physiological and Water Status of an Irrigated Peach Orchard under Semi-Arid Conditions by Using Multispectral Satellite Data

Abstract

Climate change is making water management increasingly difficult due to rising temperatures and unpredictable rainfall patterns, impacting crop water availability and irrigation needs. This study investigated the ability of machine learning and satellite remote sensing to monitor water status and physiology. The research focused on predicting different eco-physiological parameters in an irrigated peach orchard under Mediterranean conditions, utilizing multispectral reflectance data and machine learning algorithms (extreme gradient boosting, random forest, support vector regressor); ground data were acquired from 2021 to 2023 in the south of Italy. The random forest model outperformed in predicting net assimilation (R² = 0.61), while the support vector machine performed best in predicting electron transport rate (R² = 0.57), Fv/Fm ratio (R² = 0.66) and stomatal conductance (R² = 0.56). Random forest also proved to be the most effective in predicting stem water potential (R² = 0.62). These findings highlighted the potential of integrating machine learning techniques with high-resolution satellite imagery to assist farmers in monitoring crop health and optimizing irrigation practices, thereby addressing the challenges determined by climate change.

1. Introduction

Irrigation is an agronomic practice necessary to achieve sufficient yield and quality in fruit tree systems. It is needed, especially in semi-arid environments like the Mediterranean basin, which are increasingly affected by water scarcity [1]. Determining plant water status and physiology is a starting point for scheduling irrigation efficiently, avoiding both plant stresses and water waste. Usually, in irrigated tree-cropping systems, water status and physiology are monitored by using pressure chambers and leaf gas exchange systems [2]. These monitoring methods require time-consuming fieldwork and acquiring a representative number of samples within the field [3]. The main plant parameters that are commonly measured to investigate the water status and physiology of trees are leaf and stem water potential, net photosynthetic rate, stomatal conductance, and other photosynthesis-related parameters, such as electron transport rate and the ratio Fv/Fm [2,4,5,6]. Drought and water stress significantly affect plant physiology, causing a series of changes in morphology and physiology aimed at improving water efficiency and survival during water scarcity. Morphologically, plants often experience diminished growth in roots, stems, leaves, and fruits [7,8]. Physiologically, drought stress lowers plant water potential and turgor pressure, resulting in reduced cell expansion [7]. Photosynthetic efficiency is influenced by reduced chlorophyll production and modified electron transport mechanisms, in combination with stomatal closure to minimize transpiration and CO₂ uptake [9,10]. Generally, good production can be achieved when soil and climatic conditions are not limiting the photosynthetic activity of plants. Climate change is making water management problems worse. The Intergovernmental Panel on Climate Change (IPCC) predicts a temperature increase of 1.5 °C by 2040 [11]. Higher temperatures affect rainfall patterns and the occurrence of extreme weather events, which, as a result, influence crop demand and water availability [12,13]. As a result, irrigation systems need to be more resilient and adaptable to guarantee food security in the world [14]. In addition, rising temperatures accelerate evaporation, reducing the amount of water available for agriculture, and areas of the Mediterranean, which used to receive regular rainfall in the past, are experiencing prolonged periods of drought alternating with sudden violent floods. Using remote sensing, it becomes possible to obtain key information on the temporal and spatial assessment of the water status of crops over vast areas without acquiring them destructively [15]. GIS (Geographic Information Systems) also allows users to visualize the variability by mapping different variables at field, regional, and global scales [16,17,18]. Remote sensing is applied in agriculture using different sources of reflectance data, like unmanned aerial vehicles (UAV), aircraft, and satellite platforms. However, the use of satellite imagery has additional benefits, such as the ability to monitor crops more easily and to conduct time series analyses. The use of machine learning techniques is constantly growing in every area of agriculture [19]. Machine learning approaches are used to predict crop variables by using reflectance data from remote sensing [18]. The most common algorithms used for this purpose, generally with good results, are random forest, extreme gradient boosting, and support vector regressor [20,21,22,23,24]. Few studies have investigated the integration of machine learning and remote sensing in predicting tree crop water status and physiology [25,26]. This study aimed to compare the performance of different machine learning algorithms for the prediction of the water status and physiology of peach trees under irrigated conditions, which is a tree crop with an important economic value in the Mediterranean basin and in particular in the Apulia region [27].

2. Material and Methods

2.1. Experimental Site and Field Data

The study was carried out in a peach orchard within the experimental farm of the Council for Agricultural Research and Economics, Research Center for Agriculture and Environment, located in Rutigliano (southern Italy) (latitude: 40°59′29.0″ N; longitude: 17°01′53.1″ E; 148 m above sea level) during three consecutive years (2021, 2022, and 2023) (Figure 1). The climate of the study area is Mediterranean, classified as CSa, according to Köppen and Geiger [28], with hot and dry summers and moderately cold winters; the average annual rainfall is about 545 mm, concentrated mainly during autumn and winter. Maximum annual temperatures range between 32 °C and 43 °C, with a minimum between 0 °C and 5 °C. Generally, August is the hottest month, July is the driest, and January is the coldest month. Climate data during the three years of the experiment were acquired from the agro-meteorological station installed within the experimental farm. The soil is classified as clay loam (USDA classification); soil properties are reported in

Table 1. Physical and chemical parameters of the soil of the peach orchard used for the study detected at 0.30 m (E.C., electrical conductivity; SOC, soil organic carbon).

Parameter
Sand (g·100 g−1)	21
Silt (g·100 g−1)	37
Clay (g·100 g−1)	42
E.C. (dS·m−1)	0.6
SOC (g·kg−1)	14

.

The peach orchard was planted with Prunus persica ((L.) Batsch), variety Redcal, grafted on GF677, and spaced at 5 m × 5 m. The orchard was irrigated using a drip irrigation system with two drippers per tree (flow rate of 16 L/h per dripper); the irrigation was scheduled each year following the methodology reported in Allen et al. [29].

Field data were collected between May (shoot development) and September (ripening) each year on 18 sample trees. Peach water status was determined at midday by measuring the stem water potential (SWP; MPa) on mature, fully expanded, and sun-exposed leaves using a pressure chamber from Soil Moisture Equipment Corp. (Model 3000; Santa Barbara, CA, USA). SWP is a reliable and critical indicator of plant water status, as it integrates the effects of soil, plant, and atmospheric conditions; the SWP is often preferred to leaf water potential since it provides a more robust and reliable measure of water stress [30,31]. Before SWP measurement, the samples were placed in aluminum foil for 60 min. Therefore, each leaf was positioned in the chamber, and gas nitrogen was pumped in until equilibrium, and then the value was read on the console. Physiological parameters were acquired on the same tree samples at midday by using a leaf gas exchange system (LI-6400XT, LI-COR Inc., Lincoln, NE, USA); measurements were acquired at light saturation (PAR ≥ 1600 mol photons·m⁻² s⁻¹). The physiological data acquired and used in this study were net assimilation (An, μmol CO₂ m⁻² s⁻¹), which is the amount of CO₂ used by the leaf; stomatal conductance (gs, mol H₂O m⁻² s⁻¹), the water vapor losses of the leaf; Fv/Fm (relative unit), which is the ratio between the variable fluorescence and the maximum fluorescence, and it represents the maximum quantum efficiency of the photosystem II (PSII); and electron transport rate (ETR, µmol CO₂ m⁻² s⁻¹), the measure of the rate at which electrons are transported through the electron transport chain during photosynthesis. ETR is often used in association with other parameters, such as Fv/Fm, to assess the overall photosynthetic performance of plants.

2.2. Satellite Images

For this study, satellite images provided by Planet Labs PBC (hereinafter Planet) have been used. Planet is a U.S. company involved in Earth observation, capturing high-resolution, high-time-frequency images. Particularly the images of the Planet satellites “SuperDove” (Imagery© 2021, 2022, and 2023; Planet Labs PBC, San Francisco, CA, USA [32]) were acquired (time of crossing the equator from 7.30 a.m. to 11.30 a.m. local standard time). The images had eight spectral bands (SBs) with a spatial resolution of 3 m; the SBs consisted of coastal blue (443 nm), blue (490 nm), green I (531 nm), green (565 nm), yellow (610 nm), red (665 nm), red edge (705 nm), and NIR (865 nm) [32]. The satellite images were downloaded from the tool “Planet Explorer” as orthorectified and radiometrically corrected TIFs [32]. The images used were those temporally matching the date of the eco-physiological measurements taken on the field.

2.3. Statistical Analysis and Machine Learning

In this research, the overall data of the three years of physiological variables and water status were independently considered as target variables, using the reflectance of the eight SBs as predictive variables (features). For the machine learning analyses, per target, the dataset (n = 315) was randomly split into a calibration dataset (80%) and a testing dataset (20%). The models were fitted using the calibration dataset and then tested on the testing dataset to assess their performance. In this work, three machine-learning algorithms were evaluated and compared with the classical linear model (LM) in RStudio [33]. The machine-learning algorithms were extreme Gradient Boosting (XGB), random forest (RF), and support vector regressor (SVR). XGB is a gradient-boosting algorithm that builds an ensemble of decision trees sequentially. This process continues iteratively, with each tree focusing on samples that were incorrectly classified or poorly predicted by previous trees. XGB is widely used in agriculture, e.g., for the prediction of crop yield or integrated into recommendation systems for farmers [22,23,24,25,26,27,28,29,30,31,32,33,34]. RF is another tree-based ensemble algorithm that improves regression accuracy by combining the predictions of multiple decision trees trained on random subsets of the dataset. The final prediction is made by averaging the outputs of all the individual trees in the forest; it is used in many fields of agronomy, including remote sensing applications [3,35,36]. SVR has been used to predict several crop traits, e.g., corn canopy nitrogen content from multispectral images [20]; it finds the best-fitting line (or hyperplane) that maximizes the margins between data points in a transformed feature space; once the optimal hyperplane is found, new data points are predicted by determining their position relative to the hyperplane and margins, making predictions based on the learned relationship between the input features and the target variable [21,37]. A systematic grid search approach, utilizing the expand.grid function, was implemented to fine-tune the hyperparameters for all trained models. To avoid the risk of overfitting, the 10-fold cross-validation was implemented using the trainControl function of the ‘caret’ package [38].

The coefficient of determination (R²; Equation (1)), root mean square error (RMSE; Equation (2)), and normalized root mean square error (nRMSE; Equation (3)) were calculated to evaluate and compare the performance of the models as follows:

$$R^2=1-\left({\displaystyle \frac{S{S}_{res}}{S{S}_{tot}}}\right)$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(S_i-O_i\right)}^2}$$

(2)

$$nRMSE=100{\displaystyle \frac{\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(S_i-O_i\right)}^2}}{nval}}$$

(3)

where SS_res is the sum of squares of the residuals, SS_tot is the total sum of squares, S is the simulated value, O is the observed value, and n is the number of observations.

The selection of the best model was based on the value of testing R² for each target. In addition, once the best model for each parameter has been individuated by using the raster package [39], the best model has been applied to further SuperDove images to model the trend of each parameter approximately every 2 weeks for the 3-irrigation seasons considered.

The analysis of variance (ANOVA), followed by Tukey’s test, has been used to detect statistical differences among years for each ground truth parameter acquired (significance level set at 0.05). The software RStudio (RStudio 2024.04.1 + 748 “Chocolate Cosmos” for Windows) and SigmaPlot (SigmaPlot, Systat Software Inc. (San José, CA, USA), Version 14 for Windows) have been used for machine learning analyses and plotting.

3. Results

3.1. Field Data

The 2021 growing season was the driest of the three years of the experiment, with only 67.27 mm of rainfall occurring from May to September. The average daily maximum temperature was 28.50 °C, and the average daily minimum temperature was 16.75 °C. The estimated seasonal reference evapotranspiration (ETo) was 768 mm. The 2022 season was rainier than the previous year, with a cumulated seasonal rainfall of 126.2 mm. The average daily maximum and minimum temperatures were comparable with those of the previous year (25.86 °C and 16.71 °C, respectively). For 2022, the estimated seasonal ETo was 748 mm. The 2023 season was the rainiest of the three years, with 189 mm of precipitation. The average daily maximum temperature was 27.89 °C, and the average daily minimum temperature was 16.92 °C. The 2023 cumulated ETo was 694 mm (Figure 2). The 2021 irrigation season started on DOY 133 and ended on DOY 236, while in 2022 it started on DOY 144 and terminated on DOY 242. During 2023, the irrigation season began later than the previous years (mid-June, DOY 172), and it finished in the middle of September (on DOY 263). The irrigation water applied during the 2021, 2022, and 2023 growing seasons was 131.57 mm, 133.31 mm, and 129.13 mm, respectively (Figure 3).

Table 2. Descriptive statistics of the eco-physiological parameters acquired during the three years of the experiment and the overall data. An, net assimilation; gs, stomatal conductance; ETR, electron transport rate; Fv/Fm, the ratio between the variable fluorescence and the maximum fluorescence; SWP, stem water potential. Letters indicate differences among years (p < 0.05).

Parameter	Year	Mean	s.d.	Median	Min	Max
An	2021	11.09	4.92	10.69	4.12	25.15	c
	2022	18.58	4.56	18.13	9.30	30.21	a
	2023	14.42	4.64	14.57	4.53	26.11	b
ETR	2021	116.86	54.88	126.96	22.54	219.20	b
	2022	151.20	38.40	153.61	47.65	253.23	a
	2023	140.01	47.31	145.63	47.75	221.56	a
Fv/Fm	2021	0.59	0.13	0.63	0.39	0.97	a
	2022	0.52	0.05	0.52	0.36	0.70	c
	2023	0.55	0.09	0.53	0.39	0.75	b
gs	2021	0.12	0.07	0.10	0.03	0.38	c
	2022	0.20	0.07	0.19	0.07	0.47	a
	2023	0.18	0.04	0.18	0.08	0.31	b
SWP	2021	−1.33	0.26	−1.30	−2.00	−0.8	b
	2022	−1.14	0.27	−1.1	−2.00	−0.70	a
	2023	−1.16	0.39	−1.02	−2.40	−0.56	a
An	overall	15.62	5.42	15.47	4.12	30.21
ETR		140.57	46.75	144.07	22.54	253.23
Fv/Fm		0.18	0.07	0.17	0.03	0.47
gs		0.54	0.09	0.53	0.36	0.97
SWP		−1.13	0.34	−1.10	−2.40	−0.42

reports the results of the descriptive statistics of the eco-physiological parameters acquired during the three irrigation seasons considered for this study.

3.2. Models’ Performance

The following sub-sections report the results of the performance of the developed machine learning models to estimate the physiology and water status of the peach orchard. Accurate performance evaluation is fundamental to determining the reliability and usefulness of the algorithm in predicting peach orchard conditions. Models’ metrics are used to quantify the accuracy and robustness of the models, allowing a direct comparison between the different machine learning approaches.

3.2.1. Physiology

Comparative analysis of predictive models showed that the RF model outperformed other approaches in predicting An (R² = 0.61). SVR and XGB models demonstrated comparable performance (SVR R² = 0.58; XGB R² = 0.60), suggesting both algorithms effectively capture the variance in An. However, XGB exhibited the highest NRMSE among machine learning algorithms (15.30), indicating a larger discrepancy between predicted and observed values. The LM model yielded the weakest predictive results overall (Figure 4), suggesting an LM-based model may not adequately capture the relationship between An and the predictor variables. In predicting ETR, the SVR model demonstrated higher performance (R² = 0.57), followed by RF (R² = 0.51) and LM (R² = 0.39). Conversely, the XGB model performed poorly (R² = 0.26), suggesting XGB may not be the most suitable algorithm for modeling the ETR (Figure 5). SVR also exhibited higher performance in predicting the Fv/Fm ratio (R² = 0.66). RF performed similarly well (R² = 0.62) but with a slightly higher NRMSE than SVR (RF: 15.30; SVR: 15.50). The XGB model demonstrated weaker performance (R² = 0.57; NRMSE = 17.70), while the LM approach yielded the poorest results (R² = 0.33; NRMSE = 20.40) (Figure 6). For predicting gs, the SVR model proved most effective (R² = 0.56; NRMSE = 12.90), followed by RF (R² = 0.53), XGB (R² = 0.51), and LM (R² = 0.14) (Figure 7).

3.2.2. Stem Water Potential

The best machine learning algorithm for the prediction of the SWP was the RF-based model (R² = 0.62; NRMSE = 14.40); nevertheless, SVR had close performance in terms of R² (0.57) and an NRMSE slightly lower (13.60). The XGB-based approach produced weaker results (R² = 0.43; NRMSE = 17.90), while the LM yielded the poorest predictive accuracy among the tested algorithms (R² = 0.18; NRMSE = 21.10) (Figure 8).

3.3. Predicted Physiology and Water Status

The RF algorithm, found to be the best model to predict An, was used to model the trend of An during the irrigation seasons. During the three years of the experiment, the predicted physiological parameters showed different trends. During 2021, predicted An remained stable close to 20 µmol CO₂ m⁻² s⁻¹, and, in mid-July, it dropped to 13.06 ± 0.40 µmol m⁻² s⁻¹. Then, it slightly rose and dropped again in the first decade of September (9.73 ± 0.93 µmol m⁻² s⁻¹). During 2022, two decreasing peaks occurred between the end of June and the first half of July (DOY 180: 14.56 ± 0.40 µmol m⁻² s⁻¹; DOY 194: 14.68 ± 0.24 µmol m⁻² s⁻¹), and in August (DOY 232: 12.96 ± 0.38 µmol m⁻² s⁻¹). During the third year of the experiment, predicted An was steadily low between July and August (DOY 206: 12 ± 0.43 µmol m⁻² s⁻¹; DOY 213: 12.34 ± 0.57 µmol m⁻² s⁻¹; DOY 241: 11.95 ± 0.40 µmol m⁻² s⁻¹). At the end of each irrigation season, predicted An decreased in the second half of September, except for 2021, when it rose to a higher value compared to the first half of the month (Figure 9).

The SVR algorithm, being the best-performing model for predicting ETR, was applied to model its trend. Predicted ETR in 2021 had high values at the beginning of the irrigation season (DOY 151: 181.28 ± 3.16 µmol m⁻² s⁻¹), but it gradually decreased to lower values in the central part of the season (~145 µmol CO₂ m⁻² s⁻¹), then dropped in the first half of September (DOY 254: 88.80 ± 8.74 µmol m⁻² s⁻¹), and slightly rose in the second half of the month (DOY 264: 108.39 ± 7.08 µmol m⁻² s⁻¹). During 2022, two considerable downward peaks occurred: in mid-July (DOY 194: 114 ± 3.60 µmol m⁻² s⁻¹) and mid-August (DOY 232: 97.44 ± 4.98 µmol m⁻² s⁻¹). In 2023, as for the 2021 irrigation season, ETR was higher at the beginning of the season but had three decreasing peaks: in mid-June (DOY 164: 100.16 ± 3.38 µmol m⁻² s⁻¹), at the beginning of August (DOY 215: 102.59 ± 5.22 µmol m⁻² s⁻¹), and at the end of September (DOY 268: 72.12 ± 2.39 µmol m⁻² s⁻¹). During July, ETR showed high values (DOY 192: 174.16 ± 5.52 µmol m⁻² s⁻¹; DOY 206: 155.14 ± 6.89 µmol m⁻² s⁻¹), while during August and the first half of September, ETR ranged between 102.59 and 114.84 µmol m⁻² s⁻¹ (Figure 10).

The SVR algorithm was also the best model for the prediction of Fv/Fm, and it was applied for the prediction of Fv/Fm over time. Predicted Fv/Fm ranged between 0.48 and 0.52 during the first irrigation season, until the first half of September, when a peak occurred (DOY 253: 0.65 ± 0.01). During the second irrigation season, Fv/Fm was stable too, ranging between 0.50 and 0.57. During the third year, Fv/Fm ranged between 0.52 and 0.56 until the second half of July, then in August (DOY 215) it rose to 0.62 ± 0.01. Another Fv/Fm peak occurred at the end of September (DOY 268: 0.67 ± 0.006) (Figure 11).

Also, in the case of gs, SVR, having shown the best performance, was used to model the seasonal trend of gs. During the 2021 irrigation season, gs had two decreasing peaks: in August (DOY 218: 0.15 ± 0.005 mol m⁻² s⁻¹), and in September (DOY 253: 0.11 ± 0.01 mol m⁻² s⁻¹). At the end of the season, gs rose to 0.16 ± 0.005 mol m⁻² s⁻¹. During the second irrigation season, gs had a notable peak in mid-June (DOY 165: 0.27 ± 0.008 mol m⁻² s⁻¹), then decreased to 0.16 ± 0.004 mol m⁻² s⁻¹ and ranged between 0.14 and 0.19 mol m⁻² s⁻¹ until the end of the season. In the third year of the experiment, the highest value of gs was detected at the beginning of the season (DOY 150: 0.26 ± 0.01 mol m⁻² s⁻¹), and then dropped to 0.18 ± 0.001 mol m⁻² s⁻¹ in DOY 164. During the second part of the irrigation season (from the end of July to the end of August) gs values were lower, ranging between 0.14 and 0.16. An increasing peak was recorded in mid-September (DOY 255: 0.17 ± 0.002 mol m⁻² s⁻¹), then dropped again to 0.14 ± 0.006 mol m⁻² s⁻¹ at the end of September (Figure 12).

In the case of the prediction of SWP, it was modeled using the RF-based algorithm since it showed the best predictive performance. In 2021, SWP predicted values lowered progressively from the beginning of the irrigation season (DOY 151: −1.07 ± 0.02 MPa) to the beginning of August (DOY 218: −1.34 ± 0.02 MPa), then rose to −1.14 ± 0.04 MPa in the second half of August and dropped again to −1.29 ± 0.04 MPa in the first half of September. During the second year, the predicted SWP ranged between −0.89 MPa and −1.01 MPa in June; thereafter, it markedly decreased to −1.49 ± 0.02 MPa. Predicted SWP gradually increased during the season: in July it ranged between −1.19 MPa and −1.22 MPa, in August it remained stable at ~−1.09 MPa, and in September predicted SWP increased to −0.91 ± 0.01 MPa in DOY 249 and lowered to −1.03 MPa in DOY 271. In the 2023 season, the predicted SWP ranged between −1.06 MPa and −0.94 MPa until the first half of July. SWP dropped to the lowest value of the season on DOY 206 (−1.46 ± 0.02 MPa). SWP showed another low peak at the end of August (DOY 241: −1.46 ± 0.03 MPa), then it rose in September, reaching values of −1.18 ± 0.01 MPa on DOY 255 and −1.06 MPa on DOY 268 (Figure 13).

Figure 14 reports the predicted An as a function of the predicted gs for the three years considered in this study. The relationship between the two parameters was significant for all three years; the highest R² (0.63) was found for 2023, and for 2021 and 2022, it was 0.56 and 0.39, respectively.

4. Discussion

This research presents a machine learning framework to predict different peach parameters useful for irrigation scheduling. Particularly, the model with the highest accuracy in the prediction, among all the eco-physiological parameters considered, was found to be SVR for the prediction of Fv/Fm. This result confirms the good performance of the SVR algorithm in the prediction of Fv/Fm. For instance, Chen et al. [40] successfully used SVR for Fv/Fm modeling in lettuce by using proximal data. On the contrary, Wu et al. [41] obtained low performance of Fv/Fm prediction in wheat when SVR was applied to remote sensing data; nonetheless, they used a different framework than the one presented in this work. Specifically, they used UAV-derived vegetation indices as predictors; furthermore, their analysis was conducted on a smaller dataset; instead, they obtained better results for the prediction of Fv/Fm with the RF model. For peach trees, Fv/Fm values under optimal conditions typically range from 0.80 to 0.83, indicating high photosynthetic efficiency of PSII, similar to other healthy plant species [42,43,44]. In the considered peach orchard, the average predicted Fv/Fm ranged between 0.65 and 0.48, 0.57 and 0.50, and 0.67 and 0.50 during the first, second, and third irrigation seasons. Values of Fv/Fm below 0.75 may indicate that the plant is experiencing stress and that the efficiency of PSII has decreased [44]. Nonetheless, in a study with different peach tree varieties, Fv/Fm values ranged from 0.70 in well-watered plants to 0.19–0.61 in plants under severe water stress. The decrease in Fv/Fm was observed at leaf water potentials from −0.7 to −1.13 MPa [45]. Both high and low temperatures can cause stress in plants and reduce Fv/Fm values. Thermal stress can damage the proteins and pigments involved in PSII [44]. For the prediction of An, RF showed the best results, confirming the findings of Garofalo et al. [46]. In their work, the authors found that RF outperformed LM in the prediction of carob tree An by using Planet spectral bands as predictors. Nonetheless, the RF performance in their work was better than the ones presented in the current study, suggesting that, even if the same modeling approach and source of data are used, the results could be strictly dependent on the crop considered and the specific field conditions. Other studies used machine learning techniques to predict An. For instance, Zhang et al. [47] compared the performance of SVR, XGB, RF, and a generalized additive model (GAM) by using some leaf parameters as features (e.g., leaf area, length, perimeter, and shape factor) for the prediction of Populus simonii photosynthesis; the authors found out that XGB outperformed the other algorithms, even if RF and SVR had quite similar performance. In a study on the prediction of Nicotiana tabacum photosynthesis, Fu et al. [48] obtained good results in the prediction of An by using proximal data (hyperspectral leaf reflectance) and the SVR algorithm. Wu et al. [49] used machine learning methods to retrieve the net photosynthetic rate of Oryza sativa from multispectral drone images; in their study, RF outperformed SVR and LM. As for An prediction, a similar discussion can be made for gs: in the present study, RF outperformed LM, but with generally lower accuracy, compared with the findings of Garofalo et al. [46], in which Planet spectral bands were used as predictors but SVR was not tested. Xie et al. [50] used SVR, RF, and K-nearest neighbors regression (KNR) to predict the gs of citrus under water stress conditions using drone images; KNR was found to be the best model for the prediction of gs. As found in this study, RF had a slightly worse performance than SVR. On the contrary, in a study by Zhou et al. [51], gs was predicted using proximal data (hyperspectral leaf reflectance), and their findings report better performance of RF compared to SVR. The relationship between predicted An and predicted gs of the peach orchard was stronger during the first and third irrigation seasons than in 2022; however, the constant significance across years indicates the importance of predicted gs in influencing predicted An, although other factors (e.g., climatic conditions, water availability) might also have a role to varying degrees each year [52,53]. Hyperspectral leaf reflectance was also used to predict maximum ETR in a study on grapes. Yang et al. [54] built an ETR predictive model using the Bayesian neural network (BNN). Under optimal conditions, a high ETR (electron transport rate) generally indicates greater photosynthetic efficiency and better plant performance [55,56,57,58]. However, an excessively high ETR may suggest that the plant is absorbing more light than it can use for carbon fixation, leading to the production of reactive oxygen species (ROS) and oxidative stress, potentially causing damage [56,58,59]. Therefore, ETR should be interpreted in conjunction with other physiological and environmental indicators to comprehensively evaluate the plant’s health and photosynthetic performance [55,56,60]. In this research, the predicted ETR had a decreasing peak in September during the first irrigation season, in August and September during the second, and at the end of September during the third, potentially indicating peach trees’ stress in those periods. A larger number of studies can be found in the literature for the prediction of plant water status using a machine-learning approach. In the present study, RF was found to be the outperforming algorithm for the prediction of peach water status. This result is confirmed by other findings reported in the literature. In a study on olive trees under Mediterranean conditions, Garofalo et al. [3] reported that RF outclassed SVR and LM in predicting olive tree SWP. Shi et al. [61] reported that RF had better performance in predicting the plant water status and canopy water status of winter wheat than SVR, using multi-source data. Tang et al. [62] tested different machine-learning algorithms to monitor vine water status from drone images and found that RF had better performance than SVR but tended to be similar to XGB. SWP is one of the most widely used plant water stress indicators due to its accuracy [63,64]. In this study, the predicted SWP of the peach orchard notably decreased at the end of June during the 2022 irrigation season and at the end of July and August during the 2023 irrigation season.

Different satellite platforms have been used to assess the crop water status and water productivity of several crops [65,66,67]. Satellite data can also be used to monitor ecological responses to different environmental changes [68]. For example, Sentinel 2 (band resolution of 10 m to 60 m depending on the considered bands) has been used to predict the water status and stress in several crops, such as olive trees, vineyards, citrus, pomegranate, and cotton [17,69,70,71,72]. The advantage of using Planet images is surely related to the high resolution and high frequency of acquisition [32]. However, while using Planet images can lead to technical benefits for higher resolution, it should be taken into account that other platforms (e.g., Sentinel, Landsat 8 and 9) are free to use; this could certainly reduce the applicability of the presented framework when compared to free data sources. However, the results of this study, as for other agronomic variables, are related to the conditions of the field and need to be confirmed by further studies; particularly to validate the approach, long-term studies should be conducted in other regions, including other crops. Moreover, an irrigation approach based on remote sensing monitoring should be compared to the traditional approaches, e.g., soil sensor, tensiometer, or leaf water status-based approaches [2,73,74,75], which, although often accurate and helpful in irrigation scheduling, may not be sufficiently representative of the variability within the field [69,76].

5. Conclusions

In this study, a machine learning framework to predict some eco-physiological variables of peach has been presented. To the best of the authors’ knowledge, few studies have been conducted on the prediction of net assimilation, stomatal conductance, electron transport rate, and Fv/Fm of irrigated crops, especially when considering remote sensing data as predictive variables. This is the first study integrating machine learning algorithms and high-resolution satellite images for the prediction of the water status and physiology of an irrigated peach orchard under semi-arid conditions. Developing models for the prediction of the above-mentioned parameters could help farmers and technicians monitor the effect of climate conditions on peach trees and schedule irrigation more sustainably in the long term. The results of the model comparison, along with literature findings, suggest that the prediction of different parameters could need a different modeling approach, particularly when considering different crops under different agro-climatic conditions.