Evaluating Leaf Water Potential of Maize Through Multi-Cultivar Dehydration Experiments and Segmentation Thresholding

Abstract

Estimating leaf water potential (Ψ_leaf) is essential for understanding plant physiological processes’ response to drought. The estimation of Ψ_leaf based on different regression analysis methods with hyperspectral vegetation indices (VIs) has been proven to be a simple and efficient technique. However, models constructed by existing methods and VIs still face challenges regarding the generalizability and limited ranges of field experiment datasets. In this study, leaf dehydration experiments of three maize cultivars were applied to provide a dataset covering a wide range of Ψ_leaf variations, which is often challenging to obtain in field trials. The analysis screened published VIs highly correlated with Ψ_leaf and constructed a model for Ψ_leaf estimation based on three algorithms—partial least squares regression (PLSR), random forest (RF), and multiple linear stepwise regression (MLR)—for each cultivar and all three cultivars. Models were constructed using PLSR and MLR for each cultivar and PLSR, MLR, and RF for the samples from all three cultivars. The performance of the models developed for each cultivar was compared with the performance of the cross-cultivar model. Simultaneously, the normalized ratio (ND) and double-difference (DDn) were applied to determine the VIs and models. Finally, the relationship between the optimal VIs and Ψ_leaf was analyzed using discontinuous linear segmental fitting. The results showed that leaf spectral reflectance variations in the 350~700 nm bands and 1450~2500 nm bands were significantly sensitive to Ψ_leaf. The RF method achieved the highest prediction accuracy when all three cultivars’ data were used, with a normalized root mean square error (NRMSE) of 9.02%. In contrast, there was little difference in the predictive effectiveness of the models constructed for each cultivar and all three cultivars. Moreover, the simple linear regression model built based on the DDn(2030,45) outperformed the RF method regarding prediction accuracy, with an NRMSE of 7.94%. Ψ_leaf at the breakpoint obtained by discontinuous linear segment fitting was about −1.20 MPa, consistent with the published range of the turgor loss point (Ψ_TLP). This study provides an effective methodology for Ψ_leaf monitoring with significant practical value, particularly in irrigation decision-making and drought prediction.

1. Introduction

Under the global warming context, agricultural drought frequency has significantly increased [1]. Reduced rainfall and its uneven spatiotemporal distribution have led to more severe drought threats to crops [2]. Simultaneously, with the background of water resource scarcity, greater emphasis is placed on the precision and efficiency of agricultural irrigation management. Leaf water potential (Ψ_leaf) is a crucial indicator for estimating the crop water status as it takes into account the leaf osmotic potential, the apoplastic water content, and the plant cell elasticity [3].

The turgor loss point (Ψ_TLP) is the negative water potential at which the cell membrane separates from the cell wall, resulting in the cell turgor dropping to zero [4,5]. When Ψ_leaf is lower than Ψ_TLP, the plant will experience severe water stress. Ψ_TLP is widely regarded as a key surrogate for drought tolerance, reflecting a crop’s adaptive capacity under water stress [6,7,8]. Research shows that the lower the Ψ_TLP, the more drought-tolerant the plant [9]. In this context, Ψ_leaf and Ψ_TLP, as key physiological indicators, play a crucial role in assessing crop water deficits [4]. These indicators can directly reflect the dynamic water status of plant tissues and are considered to be more sensitive than soil water content monitoring [10]. The precise and efficient estimation of Ψ_leaf and Ψ_TLP provides practical value and serves as a foundation for optimizing irrigation decisions and breeding drought-resistant crop cultivars.

Conventional methods of Ψ_leaf determination (pressure chamber) provide accurate data but are laborious to operate and cause irreversible damage to the samples, which limits their application in large-scale field studies. Maize leaves are delicate and lack petioles, which makes measuring the pressure–volume (PV) curve particularly challenging. Consequently, the repeated use of the pressure chamber method for Ψ_TLP determination using the traditional PV curve approach may be impractical [11]. Therefore, achieving the efficient and non-destructive monitoring of Ψ_leaf and Ψ_TLP has become a challenge in agricultural water management.

With the development of hyperspectral remote sensing technology, it has become possible to monitor Ψ_leaf in a rapid, non-destructive, and real-time manner [12]. Cotrozzi et al. [13] established four irrigation treatments and successfully assessed the pre-dawn Ψ_LW and instantaneous Ψ_LW of Quercus oleoides using leaf spectral properties in the 950–2400 nm wavelength range. Building on this theoretical foundation, the present study focuses on vegetation indices (VIs) to develop a model for estimating leaf water potential across different maize cultivars. VIs are widely used to analyze vegetation biophysical properties because of their simplicity and high generalizability. At present, the main VIs used to estimate Ψ_leaf are based on hyperspectral information, and include the normalized vegetation index (NDVI), ratio vegetation index (RVI), double-difference vegetation index (DDn), and numerous other VIs calculated based on the conditions of raw spectral reflectance, first-order derivative of spectral reflectance, etc. [14,15,16]. Although the increasing use of VIs allows for a more comprehensive explanation of crop parameters, the redundancy between these indices can potentially affect the accuracy of estimations. Numerous studies have shown that VIs combined with multiple regression algorithms can accurately estimate the above-ground biomass of maize [17], above-ground biomass of potatoes [18], the canopy water content of rice [19], and biomass of wheat [20] and become an effective way to improve the efficiency and accuracy of modeling. Salah et al. [21] accurately estimated the relative water content (RWC), gravimetric water content (GWCF), and grain yield (GY) of two wheat cultivars under three irrigation regimes (100%, 75%, and 50% of crop evapotranspiration (ETc)) using RF, PLSR, and MLR. Although VIs combined with multiple regression algorithms have demonstrated effectiveness, their accuracy can be influenced by factors such as the plant species and water stress levels, potentially diminishing their generalizability under specific conditions. Furthermore, leaf water status may not accurately be predicted under extreme stress conditions due to experimental limitations, such as resource constraints, operational conditions, and insufficient irrigation gradients in field trials [22].

Previous studies have demonstrated that leaf dehydration experiments provide datasets covering a wide range of leaf water conditions and biological realities. These datasets may be more suitable for index identification than field measurements under water stress conditions as experiments not only efficiently track changes in leaf water status and physiological processes but also capture extreme leaf water conditions. The water indices dND(1415,1530) and dSR(1530,1895) have been used to estimate RWC and EWT based on progressive dehydration experiments conducted on four deciduous species [22]. The majority of existing maize Ψ_leaf estimation models primarily focus on studies involving a single cultivar, with limited systematic comparisons of model performance across multiple cultivars. Differences in leaf characteristics among maize cultivars, caused by artificial cultivation practices, may lead to significant variations in reflectance, which could affect the applicability and accuracy of these models, particularly when notable disparities in leaf traits are observed between cultivars. Furthermore, research connecting VIs with Ψ_TLP as an alternative to PV curve measurements remains limited, with insufficient exploration and validation of this approach.

To address the aforementioned challenges, a leaf dehydration trial was used to measure the hyperspectral reflectance and Ψ_leaf of three common maize cultivars widely planted in the North China Plain Region, followed by the development of a hyperspectral model for Ψ_leaf diagnosis. The specific research objectives were to (1) reveal the changes in hyperspectral reflectance and Ψ_leaf in maize leaves throughout the dehydration process; (2) validate the applicability of existing VIs in diagnosing maize Ψ_leaf and optimize the remote sensing estimation model for Ψ_leaf based on newly identified VIs; and (3) explore the feasibility of diagnosing Ψ_TLP in maize leaves with hyperspectral remote sensing techniques.

2. Materials and Methods

2.1. Leaf Sampling and Dehydration Measurements

The experiment was conducted in 2024 at the laboratory of the Experimental Station, National Center for Efficient Irrigation and Technology Research, in the North China Plain (Beijing; 39°39′N, 116°15′E). Leaf samples were collected from three maize cultivars—Jiyuan 168 (n = 88, JY168), Zhengdan 958 (n = 66, ZD958), and Jingke 968 (n = 74, JK968)—during the heading stage. In total, 90 fully expanded and matured ear leaves were selected from JY168, 75 from JK968, and 70 from ZD958. All leaves were harvested before dawn to ensure they were close to full turgor. Data with a Ψ_leaf value below −4 MPa were excluded from the analysis. Immediately after sampling, the plant samples were placed in self-sealing bags and promptly transported to the laboratory for analysis. The temperature was controlled at 26 °C in the laboratory.

A stable and consistent controlled environment is crucial for ensuring the accuracy of the test results. Maize leaves were taken out from the bags and placed on a drying bench for subsequent spectral reflectance and Ψ_leaf measurements. The spectral reflectance of three leaves from each cultivar was measured immediately after being taken out and Ψ_leaf was measured immediately after the spectral reflectance measurement. Measurements were taken at 5 min intervals for the first 15 min, then at intervals of 10 min, 30 min, 1 h, and 2 h in sequence until the Ψ_leaf value dropped below −4 MPa. The timing of the intervals was based on the observed change rate of the Ψ_leaf values. Ψ_leaf was measured with a pressure chamber (PMS 600; PMS Instrument Co., St Albany, OR, USA; −4~0 MPa).

Leaf spectral reflectance data were collected at 380~2500 nm using a portable spectroradiometer (FieldSpec4 Wide-Res, ASD, Boulder, CO, USA) equipped with a leaf clamp (Equipped with a fixed light source). Reflectance spectra were derived through calibration with a white reference board. The measurements were repeated three times for each leaf and averaged as the final spectral reflectance of the leaf. The spectrometer was calibrated using a whiteboard before each reflectance measurement.

2.2. Selection of Hyperspectral VIs

VIs were prioritized based on their ability to reflect leaf water content, particularly those corresponding with water-sensitive bands such as 980 nm, 1200 nm, 1400 nm, and 1900 nm. These bands exhibit distinct absorption features during variations in leaf water potential, making them effective for reflecting dynamic changes in leaf water status. Additionally, certain VIs were selected to capture spectral variations in other regions during leaf dehydration, providing complementary information on leaf water changes, although they may not have directly corresponded with water absorption features. In this study, 17 VIs reported in the literature were selected and evaluated to verify their effectiveness in monitoring leaf moisture status during dehydration (

Table 1. VIs used in this study, their formulas (where R is the reflectance of the subsequent wavelength), and references.

VIs	Formula	Reference
WI	R970/R900	[23]
RSI(1402,2272)	R1402/R2272	[24]
NDSI(1402,2272)	(R1402 − R2272)/(R1402 + R2272)	[24]
WI/NDVI	(R900/R970)/[(R900 − R680)/(R900 + R680)]	[25]
VARI	(R555 − R645)/(R555 + R645 − R469)	[26]
SRWI	R860/R1240	[27]
NDWI	(R860 − R1240)/(R860 + R1240)	[28]
NDWI2130	(R858 − R2130)/(R858 + R2130)	[29]
NDWI1640	(R858 − R1640)/(R858 + R1640)	[29]
NDII	(R850 − R1650)/(R850 + R1650)	[30]
SR(1300,1450)	R1300/R1450	[31]
MSI	R1600/R820	[32]
R(810,460)	R810/R460	[33]
R(610,560)/ND(810,610)	(R610/R560)/[(R810 − R610)/(R810 + R610)]	[34]
mSR705	(R750 − R445)/(R705 − R445)	[35]
mND705	(R750 − R705)/(R750 + R705 − 2 × R445)	[35]
DDn(1530,525)	2 × R1530 − R1005 − R2055	[14]

).

2.3. Determination of the Best Indices

Two types of VIs, the normalized ratio (ND) and double-difference (DDn), were selected to estimate Ψ_leaf.

The ND is calculated with Equation (1).

$$\mathit{ND}={\displaystyle \frac{R_i-R_j}{R_i+{ R}_j}}$$

(1)

The DDn is calculated with Equation (2), as follows:

$$\mathit{DDn}={2R}_i-R_{i + △}-R_{i - △}$$

(2)

where R_i and R_j represent the reflectance at wavelengths i and j, respectively, and △ denotes the interval step length.

2.4. Segmented Regression and Ψ_TLP

Through observation, distinct slope changes were found in the linear relationships between the VIs and Ψ_leaf, suggesting a breakpoint. Therefore, an additional objective of this study was to explore this using discontinuous linear segmentation fitting and to examine whether the associated breakpoints could indicate Ψ_TLP. Discontinuous linear segmentation fitting was implemented using R statistical software, following the three steps. (1) The dataset was divided into two subintervals, with linear regression applied separately to each interval to enhance the accuracy of the model fit. (2) The segmentation points were traversed and the residual sum of squares was calculated at each point. (3) The point with the smallest value was selected as the optimal segmentation, yielding the best-fitting model. This method effectively optimizes the data-fitting process by integrating ridge regression with segment fitting, thereby improving model accuracy while mitigating overfitting. Data from all three cultivars were combined for evaluation. To prevent the occurrence of random breakpoints, a newly identified VI and two previously published VIs with the highest correlation coefficient to Ψ_leaf were used for the regression analysis.

2.5. Statistics

This article mapped the color scale of the measured hyperspectral reflectance according to the gradient of the Ψ_leaf variation using Origin 2024b.

The selection of modeling algorithms was based on their complementary strengths for hyperspectral data analysis, as follows: (1) partial least squares regression (PLSR) was employed due to its dual advantages in handling high-dimensional datasets with inherent multicollinearity and maintaining stable performance under limited sample size conditions—particularly valuable for plant phenotyping studies where destructive sampling often restricts sample availability; (2) random forest (RF) was adopted to capture potential non-linear relationships while providing inherent resistance to overfitting through ensemble learning; and (3) multiple linear regression (MLR) served as a baseline model to quantitatively evaluate the performance gains achieved by the more sophisticated algorithms.

The dataset of VIs was based on the following two conditions: each cultivar and all three cultivars. VIs were selected based on their correlation coefficients with Ψ_leaf based on the following three steps: (1) for the published VIs, these were selected based on their correlation coefficients with Ψ_leaf, with a threshold of 0.78; (2) for the new VIs, correlation analyses were conducted on all possible wavelength combinations with Ψ_leaf, and the VIs with the highest correlation coefficient to Ψ_leaf were selected; and (3) the selected VIs were used as input variables to construct the PLSR, RF, and MLR models. To reconcile the robustness of the varietal model, the model was built using samples from all three cultivars. All model constructions were conducted using R statistical software.

Two-thirds of the sample data collected from each variety were used for model development while one-third was used for validation. The accuracy of the model was mainly determined by the coefficient of determination (R²), the mean absolute error (MAE), the root mean square error (RMSE), and the normalized root mean square error (NRMSE). If R² tends to be 1, it indicates that the model is more stable in its predictive ability; the RMSE, MAE, and NRMSE responded to the predictive ability of the model, and the smaller their values, the stronger the model estimation ability; and a NRMSE less than 20% indicated that the model was suitable for estimating Ψ_leaf whereas greater than 30% indicated that the model was unreliable [36].

$$\mathit{MAE}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|O_i-P_i\right|$$

(3)

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

(4)

$$\mathit{NRMSE}={\displaystyle \frac{\mathit{RMSE}}{O_{\mathit{max}}-O_{\mathit{min}}}}\times 100\%$$

(5)

$$R^2={\left[{\displaystyle \frac{\sum _{i=1}^n(O_i-O_{\mathit{avg}}\left)\right(P_i -P_{\mathit{avg}})}{\sqrt{{\sum }_{i=1}^n{\left(\right(O_i -O_{\mathit{avg}})}^2}\sqrt{{\sum }_{i=1}^n{\left(\right(P_i -P_{\mathit{avg}})}^2}}}\right]}^2$$

(6)

Here, O_i and P_i are the ith measured and predicted values, respectively; O_max and O_min are the maximum and minimum values of the measured values, respectively; and n is the number of observations.

3. Results

3.1. Response of Leaf Spectral Reflectance to Different Ψ_leaf Values

The results showed that the leaf spectral reflectance of all three maize cultivars was 0 to 0.55 (Figure 1). The spectral reflectance gradually increased with decreasing Ψ_leaf values in the visible (350~700 nm) and short-wave near-infrared (1450~2500 nm) ranges. However, the hyperspectral reflectance of leaves in the near-infrared wavelength range (700~1450 nm) did not show a regularity that significantly correlated with changes in Ψ_leaf. As shown in Figure 1, the depth of these absorption bands decreases with leaf dehydration, indicating a strong relationship with Ψ_leaf. We recognize that further research on the specific absorption band depths, combined with vegetation indices (VIs) such as WI, NDWI, and MSI, could better predict Ψ_leaf. For example, the depth features at 980 nm, 1200 nm, 1400 nm, and 1900 nm were indirectly reflected in the selected VIs, providing a promising direction for more accurate Ψ_leaf prediction in the future.

3.2. Screening of Spectral Vegetation Index

The sensitive VIs screened for JY168 and JK968 included SR(1300,1450), NDWI, SRWI, NDII, DDn(1530,525), MSI, NDWI1640, and NDWI2130, and those for ZD958 cultivars were SR(1300,1450), NDII, DDn(1530,525), MSI, NDWI1640, and NDWI2130 (Figure 2).

The same VIs selected for each cultivar were SR(1300,1450), NDII, DDn(1530,525), MSI, NDWI1640, and NDWI2130. Thus, the model constructed for all three cultivars using PLSR and MLR methods was built using these VIs.

In addition to the published VIs, normalization and double-differencing operations were performed on any band and step size of the raw spectral reflectance for each cultivar and all three cultivars. Subsequently, the calculated VIs were correlated with the corresponding Ψ_leaf, and contour plots of the correlation coefficients of ND, DDn, and Ψ_leaf were plotted (Figure 3). In general, specific wavelengths that significantly correlated with Ψ_leaf could be determined for each type of index and each cultivar, with most of the Pearson correlation coefficient values greater than 0.90 for each cultivar dataset and greater than 0.80 for all three cultivars.

3.3. Diagnosis of Ψ_leaf

3.3.1. PLSR, RF, and MLR Analyses

The R² of the validation set for each maize cultivar exceeded 0.77 and the NRMSE was lower than 11.98% (

Table 2. Results of PLSR and RF modeling.

		Calibration Dataset			Verification Dataset
Crop	Cultivar	R2	MAE	RMSE	R2	MAE	RMSE	NRMSE
Maize	JY168	0.8921	0.2826	0.3546	0.9574	0.1844	0.2544	6.99%
	JK968	0.7897	0.2801	0.3695	0.8802	0.2163	0.2758	7.81%
	ZD958	0.7107	0.3403	0.4408	0.7739	0.2629	0.3150	11.98%
All cultivars	PLSR	0.8446	0.2901	0.3720	0.8480	0.2846	0.3694	9.45%
	RF	0.9113	0.2094	0.2811	0.8615	0.2677	0.3526	9.02%

and

Table 3. Results of the MLR model.

Crop	Cultivar	Parametric	Empirical Formula Model	Verification Dataset
				R2	MAE	RMSE	NRMSE
Maize	JY168	x: DDn(1530,525)	y = −31.385x − 3.241	0.9453	0.2409	0.2846	7.82%
	JK968	x: NDWI1640	y = 10.694x − 6.130	0.7977	0.2867	0.3658	11.12%
	ZD958	x: SR(1300,1450)	y = 1.574x − 5.232	0.8229	0.3150	0.4192	11.68%
All cultivars		x1: SR(1300,1450) x2: NDII x3: MSI x4: NDWI2130	y = 2.728x1 − 193.92x2 − 144.304x3 − 16.403x4 + 134.326	0.8575	0.2831	0.3713	9.50%

). The best-fit models varied for the three maize cultivars. Overall, the PLSR model had higher prediction accuracy for the JY168 cultivar (R² = 0.96, RMSE = 0.25, and NRMSE = 6.99%) and JK968 cultivar (R² = 0.88, RMSE = 0.28, and NRMSE = 7.81%), whereas, for ZD958, the MLR model had better prediction performance and the ZD958 MLR model output parameters were SR(1300,1450) (R² = 0.82, RMSE = 0.42, and NRMSE = 11.68%) (Figure 4). The samples with all three cultivars were modeled with the PLSR, RF, and MLR methods, respectively, and the results showed that the model constructed with RF was better (R² = 0.86, RMSE = 0.35, and NRMSE = 9.02%), exceeding that of the individual model for ZD958 (Table 2 and Table 3).

3.3.2. Modeling of Newly Identified VIs

Based on the contour plots shown in Figure 3, the most relevant band combinations were screened (

Table 4. Newly built VI regression model results.

Crop	Cultivar	Parametric	Empirical Formula Model	${\mathit{R}}^2$	MAE	RMSE	NRMSE
Maize	JY168	x: DDn(2028,40)	y = 166.98x − 3.183	0.9447	0.2118	0.2694	7.09%
		x:ND(1771,1773)	y = 189672x2 − 2744.6x + 0.6859	0.9229	0.2567	0.3182	8.37%
	JK968	x: DDn(1880,10)	y = −168.6x − 2.1552	0.8267	0.2705	0.3299	10.03%
		x:ND(1130,1134)	y = 351637x2 + 2998.3x − 6.0636	0.8424	0.2728	0.3199	9.72%
	ZD958	x: DDn(2372,26)	y = −797.43x − 0.7728	0.8692	0.2501	0.3096	8.62%
		x: ND(1131,1134)	y = 237609x2 + 3036.3x − 5.4458	0.8745	0.2380	0.2946	8.21%
All cultivars		x: DDn(2030,45)	y = −132.84x − 1.0743	0.8962	0.2331	0.3057	7.94%
		x: ND(1000,1888)	y = −9.5246x2 − 19.971x − 9.6646	0.8791	0.2383	0.3284	8.53%

) and inverse models for the Ψ_leaf values were constructed with and without distinguishing between different cultivars. The results showed that although the distribution of correlation coefficient matrices for predicting Ψ_leaf was relatively similar for other cultivars, there were differences in the optimal VIs screened, with more minor differences in the model performance effects for DDn and ND. Without distinguishing between cultivars, the DDn-based model showed the best prediction compared with ND (Figure 5). A further analysis showed that the model constructed based on the DDn(2030,45) vegetation index (R² = 0.90, RMSE = 0.3057, and NRMSE = 7.94%) had smaller NRMSE values than the constructed PLSR and multiple stepwise regression models. Overall, the Ψ_leaf inversion model built based on the self-constructed vegetation index showed better predictive performance.

3.4. Discontinuous Segmented Fitting

Clear changes in slopes were observed in the scatter plots of VIs and Ψ_leaf, with a breakpoint being identified. To interpret the representativeness of the breakpoints, two previously published VIs were also subjected to discontinuous segmented fitting with Ψ_leaf. The results showed that the Ψ_leaf values corresponding with the breakpoints of NDWI2130, DDn (2030,45), and MSI were −1.21, −1.20, and −1.21 MPa (Figure 6), respectively.

4. Discussion

4.1. Spectral Diagnosis of Ψ_leaf

Several studies on leaf dehydration have demonstrated similar leaf reflectance responses to our results, showing an increase in reflectance across the 350~2500 nm range as leaf water content decreases, particularly in the 350~700 nm and 1300~2500 nm ranges [22,31,37]. Carter [38] referred to this phenomenon as the ‘primary effect’, indicating that leaf water status is the main driver of changes in leaf reflectance, which enables the assessment of Ψ_leaf. Santos et al. [39] successfully used a spectral range (1100~2300 nm) to predict Ψ_leaf in several grapevine cultivars. Light absorption in the 350~750 nm range is influenced by pigments [40], with chlorophyll breakdown and stomata closure [10] occurring during dehydration, which may influence leaf reflectance. Rehydrating experiments have confirmed this, identifying it as a ‘secondary effect’, which is evidenced by a blue shift [38], distinguishing the exact reason why leaf reflectance within 350~750 nm varies as leaf water status identification remains challenging. In addition, our results indicate that leaf dehydration experiments can provide a wider range of Ψ_leaf values (−4~−0.05 MPa; Figure 1) than other field-measured datasets, with Ψ_leaf ranging from −2.5 to −0.9 MPa [41].

The challenges that need to be addressed in the use of spectral reflectance data for parameter diagnosis include spectral variability as well as the influence of various other factors. These limitations can be overcome by VIs, which enhance measurement accuracy, normalize data, and mitigate external influences, thereby reducing data noise by emphasizing band similarities and minimizing random variations [11]. Some types of indices used to diagnose leaf water status have been published, such as NDVI, WI, DDn, and dSR [23,28,33]. Our results show that similar VIs were selected for each cultivar and all three cultivar samples, particularly for JK968 and JY168. The VIs screened for JY168 and JK968 included SR(1300,1450), NDWI, SRWI, NDII, DDn(1530,525), MSI, NDWI1640, and NDWI2130. The selected VIs mostly included bands within water-sensitive regions, such as 1400 nm, 1600 nm, and 1900 nm. We constructed Ψ_leaf prediction models using PLSR and MLR methods for each cultivar. The results show that both JK968 and JY168 had more accurate estimations using PLSR. The observed improvement in accuracy can be attributed to the significant correlation between the input VIs and Ψ_leaf for these two cultivars, which enabled the PLSR model to effectively extract latent components that captured the underlying patterns in the data. Additionally, the reduced multicollinearity in these indices enhances model performance [42].

In this study, the results indicated that the spectral features sensitive to Ψ_leaf were similar across different cultivars, and the model incorporating samples from all three cultivars showed similar performance to the model based on a single cultivar. By combining samples from all three cultivars, a more robust sample set was created compared with using individual cultivars. An increase in sample size enhances the robustness and generalization capability of a model [43], leading to a more reliable and accurate estimation model applicable across cultivars. We constructed predictive Ψ_leaf models based on the screened VIs using PLSR, RF, and MLR. The results show that the NRMSE of the RF model in Ψ_leaf prediction was lower than that of the PLSR and MLR models, indicating that the RF model has higher accuracy in estimating Ψ_leaf. Most of the VIs screened in this study have strong correlations; RF is robust to covariance, is not susceptible to multiple covariances between features, and can effectively reduce the overfitting phenomenon, improving the generalization ability of the model [20]. In addition, the band and step combinations that exhibited the strongest correlation with Ψ_leaf were screened using two types of VIs, ND and DDn. Wang and Li [14] used six types of indices, including R (reflectance at a given wavelength), D (reflectance difference), SR (simple ratio), NDVI, mNDVI (modified normalized difference), mSR (modified simple ratio), and DDn, to diagnose chlorophyll content, leaf water thickness, and leaf mass per area in typical temperate deciduous forests. It was found in their study that the DD-type index model was the most robust. Similarly, in this study, the NRMSE of the simple linear regression model for Ψ_leaf based on DDn(2030,45) was 7.94%, which was lower than the 9.02% of the RF model. This suggests that the DDn index is particularly effective at reducing noise by leveraging adjacent-wavelength reflectances, thereby enhancing the reliability of spectral data that are often compromised by instrumental and environmental factors [44].

Although DDn(2030,45) has shown high predictive accuracy on specific datasets, this advantage may be limited to that particular sample set and its adaptability and stability on new samples need further validation. This suggests that the effectiveness of DDn may be constrained by the characteristics of the dataset, especially in different environments or cultivars, where its performance may not meet expectations. In contrast, the RF model, by integrating multiple decision trees and utilizing established vegetation indices as inputs, demonstrated strong generalizability in this study. This indicates that the RF model can maintain relatively stable and reliable prediction performance when facing changes in environmental conditions, cultivar differences, and unseen new data. Therefore, in practical applications, especially in diverse and dynamic scenarios, RF may prove to be more robust and adaptable.

4.2. Breakpoints and Ψ_TLP

A meta-analysis of 317 species from 72 studies demonstrated a strong correlation between Ψ_TLP and water availability across biomes, further supporting its potential for predicting drought responses [4]. Traditionally, Ψ_TLP is estimated from a pressure–volume curve (PV curve), which describes the relationship between Ψ_leaf and the relative water content (RWC) of a dry leaf by plotting the inverse of Ψ_leaf versus the relative water deficit (100−RWC) [4]. A graph containing both a non-linear and a linear partition can be obtained, where the point of expansion pressure is the intersection point where the curve transitions from a non-linear to a linear part [11]. However, this approach is often resource-demanding, rendering it impractical for studies with large sample sizes, such as those focused on irrigation system development in irrigation districts or global forest-fire assessments [45]. At the point of turgor loss, a reduction in water in leaves leads to changes in their tissue structure, cell walls, and stomata, especially the tissue density and surface structure of leaves [6]. The space between cell membranes and cell walls may be increased or deformed, which affects the reflection and scattering of light, leading to changes in spectral characteristics [46,47]. In the present study, the plotted scatterplot of VIs versus Ψ_leaf often showed instability in the same range. This study used the discontinuous linear segmented fitting method to obtain more accurate values of Ψ_leaf corresponding with the breakpoints. This method accommodates varying trends in the data and mitigates overfitting by segmenting it into intervals for local linear fitting. This approach reduces model complexity and enhances fit accuracy, particularly for data with breakpoints or non-linear characteristics. In this study, the breakpoint, identified from Ψ_leaf using DDn(2030,45), was located at −1.20 MPa. To validate the stability and generalizability of this breakpoint, two previously published VIs (NDWI2130 and MSI) were employed to identify segmentation points, and consistent results were obtained. The segmentation points obtained for the three VIs using this method aligned with the known Ψ_TLP range for graminoids, which is between −1.3~−1.1 MPa [4], suggesting a correlation between Ψ_leaf at this segmentation point and Ψ_TLP. This result strongly supports the further development of crop irrigation regimes and provides a more collaborative solution for estimating Ψ_TLP in large-scale studies.

This study primarily measured Ψ_leaf and hyperspectral reflectance in flag leaves at the heading stage without considering the influence of other growth stages on the measurement results. However, leaf age significantly affects water content and variations exist across different plant types. Therefore, future studies should further investigate the impact of leaf age on remote sensing estimations. Additionally, the reflectance spectroscopy measurements in the laboratory were conducted under controlled conditions and were limited to the leaf scale. In contrast, in practical applications, hyperspectral and multispectral data collected from UAVs and satellites at the canopy scale are significantly influenced by atmospheric conditions, the canopy structure, and other environmental factors. Consequently, these factors should be carefully considered when applying such methods in real-world scenarios. Although this study provides preliminary data support for hyperspectral remote-sensing-based leaf water content estimation, further work is required to address the scale differences and the influence of factors such as leaf age before broader applications can be realized.

5. Conclusions

This leaf dehydration experiment provided broad representativeness by covering a wide gradient of data and multiple cultivars (including three maize cultivars), thereby enhancing the generalizability of the research findings. Compared with PLSR and MLR, the RF model, constructed using samples from three maize cultivars, exhibited the highest performance when estimating Ψ_leaf. The simple linear regression model constructed based on DDn(2030,45) demonstrated better stability than the RF model. Based on the samples of three maize cultivars, the breakpoints diagnosed by discontinuous linear segmentation fitting using DDn(2030,45), NDWI2130, and MSI could serve as an estimation of Ψ_TLP. A feasible hyperspectral approach for the rapid estimation of maize Ψ_leaf and Ψ_TLP has been provided.

This study offers a reliable hyperspectral approach for the efficient monitoring of maize leaf water potential, providing practical guidance for precision irrigation and drought early warning while also serving as a methodological reference for rapid physiological parameter detection in other crops.