Enhanced Spring Wheat Soil Plant Analysis Development (SPAD) Estimation in Hetao Irrigation District: Integrating Leaf Area Index (LAI) Under Variable Irrigation Conditions

Abstract

Non-destructive monitoring of chlorophyll content through Soil Plant Analysis Development (SPAD) values is essential for precision agriculture in water-limited regions. However, current estimation methods using spectral information alone face significant limitations in sensitivity and transferability under variable irrigation conditions. While integrating canopy structural parameters with spectral data represents a promising solution, systematic investigation of this approach throughout the entire growth cycle of spring wheat under different irrigation regimes remains limited. This study evaluated three machine learning algorithms (Random Forest, Support Vector Regression, and Multi-Layer Perceptron) for SPAD estimation in spring wheat cultivated in the Hetao Irrigation District. Using a split-plot experimental design with two irrigation treatments (conventional: four irrigations; limited: two irrigations) and five nitrogen levels (0–300 kg·ha⁻¹), we analyzed ten vegetation indices derived from Unmanned Aerial Vehicle (UAV) multispectral imagery, with and without Leaf Area Index (LAI) integration, across six growth stages. Results demonstrated that incorporating LAI significantly improved SPAD estimation accuracy across all algorithms, with Random Forest exhibiting the most substantial enhancement (R² increasing from 0.698 to 0.842, +20.6%; RMSE decreasing from 5.025 to 3.640, −27.6%). Notably, LAI contributed more significantly to SPAD estimation under limited irrigation conditions (R² improvement: +17.6%) compared to conventional irrigation (+11.0%), indicating its particular value for chlorophyll monitoring in water-stressed environments. The Green Normalized Difference Vegetation Index (GNDVI) emerged as the most important predictor (importance score: 0.347), followed by LAI (0.213), confirming the complementary nature of spectral and structural information. These findings provide a robust framework for non-destructive SPAD estimation in spring wheat and highlight the importance of integrating canopy structural information with spectral data, particularly in water-limited agricultural systems.

1. Introduction

Monitoring crop physiological parameters is essential for precision agriculture, providing critical information for optimizing management practices and maximizing yield potential under varying environmental conditions. Among these parameters, SPAD (Soil Plant Analysis Development) values serve as a reliable proxy for leaf chlorophyll content, allowing for rapid assessment of plant nitrogen status and photosynthetic capacity [1,2]. Similarly, Leaf Area Index (LAI), defined as the total one-sided green leaf area per unit ground surface area, provides crucial information about canopy structure and light interception efficiency [3,4]. These complementary physiological indicators together offer comprehensive insights into crop health and productivity potential. Traditional methods for SPAD determination, while accurate, are limited by their point-based nature, labor intensiveness, and inability to capture spatial variability across large agricultural fields [5]. Remote sensing technologies, particularly those deployed on Unmanned Aerial Vehicles (UAVs), have emerged as promising alternatives for non-destructive, spatially explicit monitoring of crop physiological status [6,7]. UAV-based multispectral imaging enables the calculation of various vegetation indices that correlate with chlorophyll content, offering the potential for high-throughput SPAD estimation across agricultural landscapes [8,9].

Spring wheat cultivation faces significant challenges in arid and semi-arid regions, particularly where water resource constraints limit agricultural productivity. The Hetao Irrigation District represents one of China’s largest spring wheat production areas, contributing significantly to national food security while operating under severe water limitations typical of northern China’s agricultural regions [10,11,12,13]. Recent research has demonstrated that spring wheat cultivation in these environments requires sophisticated monitoring approaches to optimize both water use efficiency and crop productivity under variable irrigation conditions. The development of limited irrigation strategies represents an important advancement in agricultural water management, but it also introduces new challenges in understanding crop physiological responses [14,15]. Monitoring SPAD values under different irrigation regimes provides valuable insights into crop water status and nitrogen utilization efficiency, enabling more precise management decisions in water-limited environments [16]. Despite significant advances in UAV-based remote sensing for SPAD estimation, current approaches face several limitations. Most existing methodologies rely solely on spectral vegetation indices, which often experience saturation effects in dense canopies, particularly during later growth stages [17,18]. Additionally, variable illumination conditions, calibration instability, and the heterogeneous vertical distribution of chlorophyll within crop canopies compromise estimation accuracy [19,20]. Perhaps most challenging is the limited transferability of models across different locations, seasons, and crop varieties, restricting their practical utility for precision agriculture applications [21].

The integration of LAI with spectral approaches represents a promising solution to overcome these limitations. Both SPAD and LAI reflect the plant’s allocation of resources—particularly nitrogen—to photosynthetic machinery, but their response to environmental stresses differs [22]. Water stress typically reduces LAI more rapidly than chlorophyll content, while nitrogen deficiency may affect SPAD values before noticeable changes in LAI occur [23]. This differential response makes the integration of both parameters particularly valuable for comprehensive crop monitoring, especially under variable irrigation conditions [24]. Recent studies have demonstrated that incorporating LAI as an auxiliary variable significantly improves SPAD estimation accuracy through multiple mechanisms [25,26]. LAI information helps account for canopy structural effects on spectral measurements and corrects for the non-uniform vertical distribution of chlorophyll in the canopy [27]. Furthermore, LAI data helps overcome the saturation problem in chlorophyll estimation, particularly in later growth stages when traditional vegetation indices lose sensitivity at high biomass levels [28]. Quantitative improvements from LAI integration are substantial, with models using spectral information alone typically achieving R² values of 0.65–0.75, while those integrating LAI features reach R² values of 0.78–0.92 [2,29]. Despite the potential benefits of integrating LAI data with spectral information for SPAD estimation, few studies have systematically investigated this approach under different irrigation regimes throughout the entire growth cycle of spring wheat. This research gap limits our understanding of how irrigation treatments affect the relationship between canopy structure, chlorophyll content, and spectral characteristics, as well as the potential improvements in SPAD estimation accuracy through LAI integration.

Therefore, the objectives of this study are to do the following: (1) evaluate the performance of different machine learning algorithms for SPAD estimation using spectral information alone; (2) quantify the improvement in estimation accuracy when incorporating LAI as an auxiliary variable; and (3) analyze how the contribution of LAI to SPAD estimation varies across different irrigation modes. The findings will provide valuable insights for developing more accurate and robust methods for non-destructive monitoring of crop physiological status in precision agriculture applications, particularly in water-limited environments. These insights can directly benefit farmers and agricultural managers in the Hetao Irrigation District by enabling more precise irrigation scheduling and nutrient management decisions throughout the growing season. While this study focuses on a specific cultivar and region, the methodological framework and findings have broader implications for global agriculture. The integration of LAI with spectral indices represents a universally applicable approach for improving crop monitoring accuracy, particularly relevant for the 40% of global agricultural land facing water stress. The demonstrated relationship between irrigation stress and the importance of structural parameters provides insights applicable to diverse arid and semi-arid agricultural systems worldwide.

2. Materials and Methods

2.1. Study Site and Experimental Design

This research was conducted at the Wuyuan Agricultural Technology Extension Center (107°35′ E, 40°30′50″ N) in Bayannur City, Inner Mongolia, China, during the 2020–2021 growing seasons (Figure 1). The experimental site is characterized by a temperate continental climate with long-term (1990–2020) average annual precipitation of 235 mm, mean annual temperature of 8.2 °C, a frost-free period ranging from 117 to 150 days, annual evaporation of 1992–2400 mm, and 3100–3450 h of annual sunshine. The soil at the experimental site was classified as loamy, with fertility characteristics presented in

Table 1. Soil fertility characteristics of the experimental site.

Year	Organic Matter (g/kg)	Alkaline-N (mg/kg)	Available-P (mg/kg)	Available-K (mg/kg)	pH
2020	14.31	59.45	21.32	137.71	7.62
2021	13.41	55.24	20.23	108.84	7.41

.

The study employed spring wheat cultivar “Yongliang 4” in a split-plot experimental design. The main plots comprised two irrigation regimes: conventional irrigation (4W)—four irrigations applied at tillering, heading, flowering, and grain filling stages—and limited irrigation (2W)—two irrigations applied at tillering and flowering stages. Each irrigation event delivered 900 m³·ha⁻¹ of water. The subplots consisted of five nitrogen fertilization levels strategically designed to create comprehensive physiological variability: N0 (0 kg·ha⁻¹), N1 (75 kg·ha⁻¹), N2 (150 kg·ha⁻¹), N3 (225 kg·ha⁻¹), and N4 (300 kg·ha⁻¹). This gradient spans from nutrient deficiency through luxury consumption, encompassing the full spectrum of SPAD responses typically encountered in commercial spring wheat production systems ranging from organic systems with minimal external inputs to intensive production systems with high fertilizer applications. These nitrogen application rates were designed to encompass the optimal range established through previous research in the Hetao Irrigation District, where optimal rates range from 185.5–209.5 kg·ha⁻¹ under water-saving conditions and 212.5–240.9 kg·ha⁻¹ under conventional irrigation management [30]. This treatment design ensures robust model training across the complete range of chlorophyll concentrations while maintaining statistical power for detecting irrigation-dependent effects and capturing non-linear SPAD responses to validate model performance across diverse agricultural management scenarios. The experiment included three replications of each treatment combination, resulting in 30 experimental plots (two irrigation regimes × five nitrogen levels × three replications), each measuring 42 m². Seeds were mechanically sown at a density of 375 kg·ha⁻¹. Phosphorus fertilizer was applied as a basal application at sowing, while no potassium fertilizer was applied throughout the growing season.

2.2. Data Collection

2.2.1. UAV Multispectral Data Acquisition

Multispectral imagery was acquired using a DJI Phantom 4 Multispectral UAV (DJI Technology Co., Ltd., Shenzhen, China). Data collection was performed at six critical growth stages: tillering, jointing, heading, flowering, grain filling, and maturity. The UAV was equipped with a multispectral camera system comprising one RGB sensor and five multispectral sensors capturing reflectance at 450 ± 16 nm (blue), 560 ± 16 nm (green), 650 ± 16 nm (red), 730 ± 16 nm (red edge), and 840 ± 26 nm (near-infrared).

All flights were conducted on clear, windless days between 09:00 and 11:00 local time to minimize shadow effects and hotspot phenomena. Prior to each flight, radiometric calibration was performed using three reflectance calibration panels (20%, 40%, and 60% reflectivity). Flight missions were programmed using DJI GS Pro software (ver. 2.0.10, DJI Technology Co., Ltd., Shenzhen, China) with the following parameters: flight altitude of 30 m, resulting in a ground sampling distance (GSD) of 1.59 cm·pixel⁻¹, 85% forward overlap, and 80% side overlap to ensure complete coverage and sufficient redundancy for image reconstruction. A D-RTK 2 high-precision GNSS mobile station (DJI Technology Co., Ltd., Shenzhen, China) was used to enhance positional accuracy. Following data acquisition, imagery underwent radiometric correction in DJI Terra software (ver. 3.0.2, DJI Technology Co., Ltd., Shenzhen, China) before being mosaicked into single-band reflectance orthophotos for subsequent analysis.

2.2.2. Field Biophysical Parameter Measurements

Leaf Area Index (LAI) was measured concurrently with UAV flights at each growth stage using an LI-2200C Plant Canopy Analyzer (LI-COR Inc., Lincoln, NE, USA). For each experimental plot, five measurements were taken following a diagonal sampling pattern to account for spatial variability. Measurements were conducted under diffuse light conditions (8:30–10:30) to minimize direct sunlight interference. The LI-2200C was configured with a 45° view cap to restrict the azimuthal view and reduce measurement errors from neighboring plots. Each measurement sequence consisted of one above-canopy reading followed by four below-canopy readings. LAI values were calculated automatically by the instrument’s software based on light interception principles, and the average of five sampling points was used to represent the LAI value for each plot.

Relative chlorophyll content was measured using a SPAD-502 Plus chlorophyll meter (Konica Minolta Inc., Tokyo, Japan) on the same days as UAV flights and LAI measurements (10:30–14:00). The “five-point sampling method” was implemented to select 25 uniformly growing wheat plants in each plot. For each plant, SPAD readings were taken at three positions (top, middle, and bottom) on the flag leaf (or the uppermost fully expanded leaf for early growth stages). The three measurements were averaged to represent the SPAD value for each plant, and the mean of all 25 plants was calculated as the representative SPAD value for the treatment plot.

2.3. Data Processing and Analysis

2.3.1. Spectral Data Processing

The multispectral orthophotos were processed using ENVI software (ver. 5.6, L3Harris Geospatial, Broomfield, CO, USA) to extract reflectance values for each experimental plot. Plot boundaries were digitized based on field markers visible in the imagery, and a 0.5 m buffer was applied to exclude edge effects. The zonal statistics function was used to extract mean reflectance values for each spectral band within each plot.

2.3.2. Calculate Vegetation Indices

Ten vegetation indices (VIs) were calculated from the extracted reflectance values to reduce the influence of soil background and enhance the sensitivity to plant biophysical parameters. The selected VIs have demonstrated utility for chlorophyll content estimation in previous research and are listed in

Table 2. Vegetation indices calculation method.

Vegetation Indices	Calculation Formula	References
Normalized Difference Vegetation Index	$NDVI=\left(R_{nir }-R_{red}\right)/\left(R_{nir}+R_{red}\right)$	[31]
Enhanced Vegetation Index	$EVI=2.5\times \left(R_{nir}-R_{red}\right)/\left(R_{nir}+6\times R_{red}-7.5\times R_{blue}+1\right)$	[32]
Green Normalized Difference Vegetation	$GNDVI =\left(R_{nir }-R_{green}\right)/\left(R_{nir }+R_{green}\right)$	[33]
Modified Soil Adjusted Vegetation Index1	$MSAVI1=\left(1+L\right)\left({\displaystyle \frac{R_{nir}-R_{red}}{R_{nir}+R_{red}+L}}\right)\left(L=0.1\right)$	[34]
Optimized Soil Adjusted Vegetation Index	$OSAVI=\left(1+0.16\right)\times {\displaystyle \frac{\left(R_{nir}-R_{red}\right)}{\left(R_{nir}+R_{red}+0.16\right)}}$	[35]
Leaf Chlorophyll Index	$LCI=(R_{nir}-R_{rededge})/\left(R_{nir }+R_{red}\right)$	[36]
Modified Chlorophyll Absorption Reflectance Index	$MCARI =\left(R_{rededge }-R_{red }-0.2\times \left(R_{rededge }-R_{green}\right)\right)\times \left({\displaystyle \frac{R_{rededge}}{R_{red}}}\right)$	[37]
Structure-Insensitive Pigment Index	$SIPI=\left(R_{nir }-R_{blue}\right)/\left(R_{nir}+R_{red}\right)$	[38]
Modified Triangular Vegetation Index 2	$MTVI2={\displaystyle \frac{1.5\times \left[1.2\times \left(R_{nir }-R_{green}\right)-2.5\times \left(R_{red }-R_{green}\right)\right]}{\sqrt{{\left(2\times R_{nir}+1\right)}^2-\left(6\times R_{nir}-5\times \sqrt{R_{red }}\right)-0.5}}}$	[39]
Normalized Difference Red Edge Index	$NDRE=\left(R_{nir }-R_{rededge}\right)/\left(R_{nir }+R_{rededge}\right)$	[40]

.

2.3.3. SPAD Estimation Models

Three machine learning algorithms were implemented to develop SPAD estimation models using the scikit-learn library in Python (ver. 3.8.10). For all algorithms, hyperparameter optimization was conducted using grid searches with fivefold cross-validation to identify optimal parameter configurations:

Random Forest Regression (RF): Grid search explored combinations of estimator counts (100, 200, 300), minimum samples per leaf (1, 2, 5), and maximum features ratios (0.5, 0.7, “sqrt”). The optimal configuration was 200 estimators, two minimum samples per leaf, and a maximum features ratio of 0.7, with Gini impurity as the splitting criterion and no restrictions on maximum tree depth.

Support Vector Regression (SVR): Grid search explored combinations of C values (1, 10, 100), gamma parameters (0.01, 0.05, 0.1), and epsilon values (0.1, 0.2, 0.3). The optimal configuration employed an RBF kernel with C = 10, gamma = 0.05, and epsilon = 0.2, trained using the Sequential Minimal Optimization algorithm with a convergence tolerance of 0.001.

Multi-Layer Perceptron (MLP): Grid search explored combinations of hidden layer configurations ((5,5), (10,10), (15,15)), alpha values (0.01, 0.05, 0.1), and learning rates (0.001, 0.01, 0.1). The optimal architecture consisted of two hidden layers with 10 neurons each, ReLU activation functions, and alpha = 0.05 for L2 regularization, trained with the Adam optimizer using a batch size of 32 and early stopping (patience = 20).

For each algorithm, two distinct models were developed: spectral-only models (using only the ten vegetation indices as input features) and combined models (integrating both vegetation indices and LAI as input features). This dual modeling approach enabled us to quantitatively assess the contribution of LAI to SPAD estimation accuracy across different irrigation treatments.

2.4. Model Evaluation

The dataset was randomly divided into training (70%) and testing (30%) sets with stratified sampling to maintain the distribution of irrigation and nitrogen treatments. Prior to model development, all input features were standardized using a standard scaler (z-score normalization) to transform variables to have zero mean and unit variance, ensuring equal weighting during training. Model performance was evaluated using two statistical metrics: coefficient of determination (R²) and root mean square error (RMSE). R² measures the proportion of variance explained by the model, with higher values indicating better model fit. RMSE quantifies the average prediction error magnitude, with lower values indicating higher model fitting accuracy.

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-{\overline{y}}_i\right)}^2}}$$

(1)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{n}}}$$

(2)

where ${\widehat{y}}_i$ is the predicted value, ${\mathrm{y}}_{\mathrm{i}}$ is the observed value, ${\overline{y}}_i$ is the mean of the observed value, and n is the number of samples.

2.5. Statistical Analysis and Visualization

Pearson correlation analysis was conducted to quantify the relationships between spectral vegetation indices, LAI, and SPAD values at different growth stages. The correlation matrices were visualized using heatmaps to identify the most relevant indices for SPAD estimation. All statistical analyses and visualizations were implemented in Python (ver. 3.8.10) using the NumPy (ver. 1.21.0), pandas (ver. 1.3.3), scikit-learn (ver. 1.0.2), Matplotlib (ver. 3.4.2), and seaborn (ver. 0.11.1) libraries.

3. Results

3.1. Temporal Dynamics of LAI and SPAD in Spring Wheat Under Two Irrigation Modes

The temporal patterns of LAI and SPAD values throughout the spring wheat growth cycle demonstrated clear responses to irrigation treatment (Figure 2). SPAD values displayed an initial increase from tillering to heading stages in both irrigation treatments, with peak values observed during heading. The 4W treatment maintained higher SPAD values than the 2W treatment throughout the growth cycle, with the most pronounced differences during the heading and flowering stages. Both treatments exhibited a decline in SPAD values after flowering, corresponding to natural senescence and nitrogen remobilization from leaves to developing grains, with a slightly steeper decline observed in the 2W treatment, indicating accelerated senescence under water limitation. LAI values followed a classic dome-shaped pattern in both irrigation treatments, with values increasing from tillering through jointing and heading and reaching maximum values at flowering, followed by a decline toward maturity. The 4W treatment consistently produced higher LAI values compared to 2W, with differences most pronounced during heading and flowering stages when canopy development was most active. This response directly reflects the positive effect of increased water availability on vegetative growth and leaf expansion.

3.2. Spectral Reflectance Characteristics Under Different Irrigation Modes

Spring wheat canopy reflectance exhibited typical vegetation spectral patterns across growth stages while revealing important differences between irrigation treatments (Figure 3). Both treatments showed low reflectance in visible wavelengths, a sharp increase in the red edge region, and high reflectance in NIR. The blue band maintained consistently low reflectance (0.021–0.035) with minimal treatment differences. The green band displayed the characteristic “green peak.” Red band reflectance remained low (0.036–0.060) due to chlorophyll absorption. The most significant treatment differences appeared in red edge and NIR regions, where 4W consistently exhibited higher reflectance, particularly during heading and flowering stages. Temporally, NIR reflectance increased from tillering to heading/flowering before declining during grain filling and maturity, corresponding to canopy development and senescence. During maturity, reflectance decreased across all bands, with treatment differences diminishing as canopy characteristics converged during senescence.

3.3. Correlation Analysis Between SPAD, LAI and Vegetation Indices

Correlation analysis revealed significant relationships between SPAD, LAI, and vegetation indices derived from multispectral imagery (

Table 3. Pearson correlation coefficients between SPAD, LAI, and vegetation indices derived from multispectral imagery of spring wheat.

Index	LAI	NDVI	OSAVI	MCARI	SIPI	NDRE	LCI	EVI	GNDVI	MSAVI1	MTVI2
SPAD	0.74	0.75	0.75	0.73	−0.71	0.80	0.81	0.75	0.83	0.74	0.73
LAI	1.00	0.68	0.72	0.69	−0.59	0.78	0.79	0.72	0.79	0.72	0.71

, Figure 4). Among all vegetation indices, GNDVI showed the strongest correlation with SPAD values (r = 0.83), followed by LCI (r = 0.81) and NDRE (r = 0.80). This finding suggests that indices incorporating green and red-edge bands are particularly sensitive to chlorophyll content variations in spring wheat. SIPI exhibited a negative correlation with SPAD (r = −0.71), likely due to its focus on carotenoid/chlorophyll ratios rather than absolute chlorophyll content. LAI demonstrated a strong positive correlation with SPAD (r = 0.74), indicating substantial covariation between canopy structural characteristics and leaf chlorophyll content. LAI also correlated strongly with several vegetation indices, particularly GNDVI (r = 0.79), LCI (r = 0.79), and NDRE (r = 0.78), suggesting that these indices respond to both structural and biochemical canopy properties.

The LAI-SPAD relationship varied across growth stages and irrigation treatments (Figure 4). Both irrigation treatments showed consistent positive correlations at all growth stages, with correlation coefficients ranging from r = 0.83 to r = 0.92 for limited irrigation (2W) and r = 0.86 to r = 0.95 for conventional irrigation (4W). The strongest correlations occurred during the heading stage for 2W (r = 0.92) and 4W (r = 0.95).

3.4. SPAD Estimation Model Performance Analysis

All three machine learning algorithms showed significant performance improvements when LAI was integrated with spectral features for SPAD estimation (Figure 5). Random Forest Regression (RF) demonstrated the most substantial gains on the independent test dataset, with test R² increasing from 0.698 to 0.842 (+20.6%) and test RMSE decreasing from 5.025 to 3.640 (−27.6%). Support Vector Regression (SVR) showed more modest test set improvements (test R² from 0.692 to 0.756, +9.2%; test RMSE from 5.078 to 4.514, −11.1%), while Multi-Layer Perceptron (MLP) achieved intermediate enhancements on the test data (test R² from 0.728 to 0.785, +7.8%; test RMSE from 4.770 to 4.237, −11.2%). The combined RF model delivered the best overall performance on unseen test samples, suggesting its superior ability to capture complex non-linear relationships between canopy structure, spectral properties, and chlorophyll content. All models demonstrated good generalization capability when evaluated on the independent test set, confirming the value of integrating LAI with spectral data for more accurate SPAD estimation in spring wheat under field conditions.

3.5. Feature Importance Analysis of RF Model Combined with LAI

Feature importance analysis of the optimal Random Forest model revealed the relative contribution of each input variable to SPAD estimation accuracy (Figure 6). GNDVI emerged as the most influential feature with an importance score of 0.347, accounting for approximately 35% of the model’s predictive power. This finding aligns with the strong correlation between GNDVI and SPAD observed in the correlation analysis. LAI ranked as the second most important feature with a score of 0.213, highlighting its substantial contribution to SPAD estimation despite not being a spectral index. This significant ranking validates the integration of structural parameters with spectral information for improved model performance. Combined, GNDVI and LAI accounted for more than 55% of the total feature importance, demonstrating their dominant role in SPAD estimation for spring wheat.

Among the remaining vegetation indices, SIPI, LCI, and NDRE showed moderate importance, while NDVI, MCARI, MSAVI1, and MTVI2 contributed less significantly. The importance of LAI suggests that canopy architecture provides complementary information about chlorophyll distribution that cannot be captured by spectral indices alone.

3.6. LAI Contribution to SPAD Estimation Under Contrasting Irrigation Modes

The contribution of LAI to SPAD estimation accuracy varied notably between irrigation treatments when separate models were developed for each irrigation regime (

Table 4. Test set performance comparison of Random Forest models for SPAD estimation using spectral features only versus combined with LAI under two irrigation modes.

Irrigation Mode	R2			RMSE
	Spectral	Combined LAI	Increase	Spectral	Combined LAI	Increase
2W	0.694	0.817	17.6%	4.967	3.847	−22.6%
4W	0.733	0.813	11.0%	4.506	3.768	−16.4%

). Under limited irrigation (2W), incorporating LAI with spectral features yielded a substantial improvement in model performance, with R² increasing from 0.694 to 0.817 (+17.6%) and RMSE decreasing from 4.967 to 3.847 (−22.6%). Under conventional irrigation (4W), LAI integration also improved model accuracy, R² increased from 0.733 to 0.813 (+11.0%) and RMSE decreased from 4.506 to 3.768 (−16.4%). The spectral-only model performed slightly better under conventional irrigation compared to limited irrigation, suggesting that spectral indices alone may be more reliable predictors of chlorophyll content under optimal water conditions. The more pronounced improvement under water-limited conditions indicates that LAI provides particularly valuable complementary information when plants experience moderate water stress.

3.7. Nitrogen Treatment Effects and Experimental Design Validation

Two-way ANOVA revealed significant main effects for both irrigation and nitrogen treatments on SPAD values and LAI, with notably different response patterns between these physiological parameters (

Table 5. Two-way ANOVA results.

Sources of Variation	SPAD		LAI
	F Value	p Value	F Value	p Value
Irrigation (I)	39.91	<0.001	31.74	<0.001
Nitrogen (N)	80.84	<0.001	17.54	<0.001
I × N	3.82	0.007	4.15	0.004

). For SPAD values, nitrogen treatments demonstrated the strongest effect (F = 80.84, p < 0.001), followed by irrigation treatments (F = 39.91, p < 0.001). In contrast, for LAI, irrigation effects dominated (F = 31.74, p < 0.001), substantially exceeding nitrogen effects (F = 17.54, p < 0.001). Both parameters showed significant irrigation × nitrogen interactions (SPAD: F = 3.82, p = 0.007; LAI: F = 4.15, p = 0.004). This differential response pattern validates our research focus on irrigation effects and explains why LAI integration proved particularly valuable for SPAD estimation under different irrigation regimes. The predominant irrigation influence on LAI reinforces the importance of incorporating canopy structural parameters when estimating chlorophyll content across varying water availability conditions. Nitrogen main effects (

Table 6. Nitrogen main effects (averaged across irrigation treatments and stages).

Treatment	SPAD	LAI
CK	28.2 ± 0.64	1.5 ± 0.08
N1	38.5 ± 0.69	2.5 ± 0.14
N2	40.3 ± 0.68	2.7 ± 0.14
N3	43.2 ± 0.72	3.0 ± 0.16
N4	44.5 ± 0.77	3.2 ± 0.16

) demonstrated clear dose-response relationships for both parameters. SPAD values increased progressively from 28.2 ± 0.64 at the control level (0 kg·ha⁻¹) to 44.5 ± 0.77 at the highest nitrogen application rate (300 kg·ha⁻¹), representing a 58% increase. Similarly, LAI increased from 1.5 ± 0.08 to 3.2 ± 0.16, showing a 113% enhancement. The experimental design effectively captured physiological variability while maintaining focus on irrigation-related effects, ensuring that our SPAD estimation models remain applicable across diverse nutrient management scenarios in water-limited agricultural systems.

4. Discussion

The Random Forest (RF) algorithm demonstrated superior performance compared to other machine learning approaches when combining spectral indices with LAI for SPAD estimation in spring wheat. Shah et al. [41] found that RF achieved R² of 0.89 and RMSE of 5.49 μg·cm⁻² when estimating chlorophyll content in wheat using hyperspectral data. Belgiu and Drăguţ [42] highlighted RF’s multiple advantages in remote sensing applications, including robustness to noise and outliers, capacity to handle high-dimensional data without feature selection, resistance to overfitting, and built-in feature importance measures. Wang et al. [24] demonstrated that RF maintained stable performance (R² around 0.8 for both training and test sets), while SVR showed signs of overfitting (R² decreasing from 0.8651 to 0.5769). The integration of LAI as an auxiliary variable substantially enhanced SPAD estimation performance, attributable to RF’s ensemble approach, which combines multiple decision trees to better capture complex relationships between spectral indices, LAI, and SPAD values [28,43].

LAI contributed more significantly to SPAD estimation under limited irrigation conditions than under conventional irrigation due to several physiological mechanisms. Under water stress, leaf morphological and anatomical changes affect the relationship between LAI and chlorophyll content. Jahan et al. [44] noted that water stress induces significant changes in wheat leaves, including reduced lamina thickness, decreased mesophyll tissue thickness, and altered vascular bundle dimensions. Han et al. [29] demonstrated stronger relationships between LAI and SPAD values under limited irrigation (R² = 0.81) compared to conventional irrigation (R² = 0.53). The stronger contribution of LAI under water-limited conditions likely reflects the complex physiological adjustments that occur when plants manage both water and nutrient stress simultaneously. Under these conditions, the relationship between canopy structure and chlorophyll distribution becomes more pronounced, as plants optimize resource allocation between vegetative growth and photosynthetic capacity. Yang et al. [45] found that water stress causes wheat leaves to become smaller but relatively thicker, with higher cell density and altered chloroplast distribution, increasing chlorophyll concentration per unit leaf area. Furthermore, water stress triggers a “concentration effect” of chlorophyll in smaller leaves, as plants prioritize maintaining vital functions in reduced leaf area [46,47].

These physiological adjustments explain why incorporating LAI into SPAD estimation models becomes particularly valuable under water-limited conditions. The underlying mechanisms driving these irrigation-dependent relationships can be better understood by examining the temporal dynamics of both physiological parameters and their spectral signatures throughout the growing season. The temporal patterns observed in Figure 2 and Figure 3 provide valuable insights into wheat physiological responses to irrigation management. The characteristic peak of SPAD values during the heading stage followed by subsequent decline reflects sophisticated physiological programming involving coordinated nitrogen remobilization rather than simple senescence [48]. During this critical period, wheat achieves maximum photosynthetic capacity to support grain filling, with pre-anthesis nitrogen accumulation peaking around heading and up to 80% of grain nitrogen subsequently derived from remobilization [49]. The dome-shaped LAI pattern observed in both irrigation treatments corresponds to well-documented canopy development dynamics, where water stress fundamentally alters expansion rates through linear responses when relative soil water availability drops below 0.7 [50,51]. The spectral reflectance characteristics demonstrate irrigation-dependent responses consistent with recent findings that water stress induces a 5–15 nm blue shift in red edge position and reduces NIR reflectance from 45–55% under well-irrigated conditions to 25–35% under severe water limitation [45,52]. The more pronounced differences between irrigation treatments during heading and flowering stages reflect the period when canopy structural effects on spectral measurements are most pronounced, as NIR reflectance responds to both enhanced leaf internal scattering from healthy mesophyll cells and increased canopy light interception under adequate water supply [29]. These temporal dynamics explain why LAI integration provides greater improvement in SPAD estimation under limited irrigation conditions, as water stress creates stronger coupling between canopy structural parameters and chlorophyll distribution patterns [53].

Green Normalized Difference Vegetation Index (GNDVI) emerged as the most important vegetation index for SPAD estimation, followed by LAI itself. Huang et al. [54] found that GNDVI had the highest correlation with canopy SPAD values in fruit trees, while Wang et al. [24] demonstrated that RF regression models based on GNDVI performed best for wheat SPAD estimation, particularly at the jointing stage. GNDVI’s effectiveness can be attributed to its sensitivity to chlorophyll content across a wider range of values. Hunt et al. [55] noted that GNDVI was originally developed for chlorophyll concentration measurement and proved highly relevant for multiple crops. Its superior performance can also be explained by using green reflectance instead of red—as chlorophyll content increases, red absorption saturates more quickly than green absorption [2], making GNDVI more sensitive to variations in chlorophyll content, especially at medium to high concentrations typical in wheat canopies.

Unlike SPAD estimation studies conducted in regions with diverse wheat varieties, this research focuses on Yongliang 4, a single variety that has dominated large-scale cultivation in the Hetao Irrigation District for 30–40 years [56]. The experimental design successfully incorporated nitrogen-induced physiological variability while maintaining clear focus on irrigation effects. The five nitrogen treatments served their intended purpose of providing robust physiological variability for model training, ensuring that the observed benefits of LAI integration for SPAD estimation are transferable across varying nutrient conditions commonly encountered in precision agriculture applications. This distinctive research context allows us to eliminate genetic variability while investigating environmental and management effects on crop parameters. Single-variety studies benefit from reduced genetic variability, allowing for the isolation of environmental and management effects on crop parameters [57]. However, this approach has limitations regarding model transferability, as spectral features perform differently across wheat varieties. Models calibrated on a single variety often show poor transferability due to genotype-specific differences in canopy structure and pigmentation [58]. Despite these limitations, our approach is appropriate for the Hetao Irrigation District with its unique environmental conditions. The practical implications of our research are significant for precision agriculture. The superior performance of Random Forest with combined spectral indices and LAI provides a robust framework for non-destructive SPAD estimation in wheat. The finding that LAI contributes more significantly to SPAD estimation under limited irrigation has particular relevance for irrigation management in water-scarce regions, enabling more accurate yield prediction and irrigation scheduling.

5. Conclusions

This study demonstrated that incorporating Leaf Area Index with spectral indices significantly improves SPAD estimation accuracy in spring wheat under different irrigation regimes. Random Forest consistently outperformed other machine learning algorithms when combining these data types, achieving a 20.6% increase in R² values compared to spectral-only models. This superior performance reflects Random Forest’s capacity to capture non-linear relationships between vegetation indices, LAI, and chlorophyll content, with GNDVI emerging as the most important predictor (importance score: 0.347) followed by LAI (0.213).

The differential contribution of LAI under varying irrigation conditions represents a critical finding with immediate practical implications. LAI contributed significantly more under limited irrigation conditions (R² improvement: +17.6%) compared to conventional irrigation (+11.0%), demonstrating that water stress alters the relationship between canopy structure and chlorophyll content through resource allocation shifts and concentration effects. This indicates that structural parameters become increasingly important for accurate chlorophyll monitoring as water availability decreases, providing essential guidance for developing robust monitoring systems in water-stressed agricultural environments where 40% of global agricultural land operates.

These findings offer clear pathways for implementation in precision agriculture systems. The demonstrated benefits of UAV-based multispectral monitoring combined with LAI estimation provide a practical framework for non-destructive crop monitoring that can inform irrigation scheduling and nutrient management decisions throughout the growing season. The temporal dynamics observed across growth stages, with strongest LAI-SPAD correlations during heading and flowering stages, establish optimal windows for field monitoring campaigns and model calibration.

Future research should prioritize testing across multiple wheat varieties and growing seasons, incorporating additional data sources such as thermal or hyperspectral imagery, and exploring deep learning approaches for enhanced feature extraction. Long-term validation studies across diverse environmental conditions will strengthen the transferability and operational utility of these monitoring approaches for precision agriculture in water-limited systems.