Research on Water and Fertilizer Diagnosis of Maize Using Visible–Near-Infrared Hyperspectral Technology

Abstract

This study focuses on maize to explore spectral estimation methods for agricultural traits in maize leaves under water-saving and fertilizer-reduction strategies. A factorial experiment was conducted with different nitrogen application rates (N0–N4) and irrigation levels (W1–W4). Hyperspectral data were collected at V12, R1, and R3 stages, alongside measurements of agricultural traits ((relative chlorophyll content) SPAD values, leaf water content (LWC), and leaf nitrogen content (LNC)). Results indicated that reducing nitrogen by 10% (N3) had no significant effect on physiological indicators, whereas reducing irrigation by 10% (W3) led to significant differences. First- and second-derivative transformations of spectral data enhanced the correlation with agricultural traits. Support vector regression (SVR) and random forest (RF) models were developed for estimation. RF outperformed SVR in predicting agricultural traits (SPAD, LWC, and LNC), with estimation accuracy R² values reaching 0.92, 0.94, and 0.95, respectively. The RF model demonstrated higher accuracy, providing technical support for growth monitoring and precise water and nutrient management in maize.

1. Introduction

China’s maize planting area spans 663 million mu, accounting for 37.1% of the total grain planting area; its yield reaches 290 million tons, representing 41.5% of the total grain production [1]. As China’s most extensively cultivated staple grain crop, maize holds significant importance for agricultural development and national food security [2]. Maize growth directly impacts yield and quality, making effective growth monitoring essential for field management and yield estimation [3].

Hyperspectral technology, based on spectral characteristics, plays a vital role in studying crop growth, nutrition, and predicting crop yields. Crop plants exhibit specific spectral reflectance characteristics, meaning their absorption, transmission, and reflection of light change with alterations in internal structure and physiological traits [4]. Hyperspectral technology acquires spectral characteristics of targets, analyzes spectral data, and through model inversion, reveals changes in agricultural traits during crop growth [5]. This enables non-destructive, precise, and real-time monitoring of crop development. Scholars have conducted in-depth research on hyperspectral technology theories and applications related to crops, establishing numerous corresponding models [6]. However, factors such as varying nutritional conditions, external climatic influences, and crop varieties can reduce the accuracy of growth estimation models. Therefore, developing high-precision crop growth estimation models for different developmental stages holds significant importance for guiding agricultural production.

Chlorophyll, moisture, and nitrogen content—key indicators of plant physiological and biochemical processes—can be non-destructively monitored through spectral reflectance characteristics [7,8]. SPAD values serve as an indicator for assessing chlorophyll content in plant leaves. These values are significantly influenced by nitrogen levels within the leaves, as nitrogen is a key constituent element of chlorophyll. Chlorophyll content directly impacts photosynthetic efficiency in plants, making SPAD values an indirect indicator of crop nitrogen nutrition status [9]. For example, leaf SPAD values increased with rising nitrogen supply levels, peaking at the heading stage before gradually declining. This indicates that nitrogen supply directly promotes chlorophyll synthesis and SPAD values. Different crops exhibit specific SPAD value trends across their growth stages. The SPAD readings in rice exhibit a unimodal pattern through the jointing, heading, and milk ripening stages. They rise initially, reach a maximum at heading, and subsequently decline [10]. This correlates with physiological demands during plant development. The heading stage marks a critical transition from vegetative to reproductive growth in rice, characterized by high nitrogen requirements and peak chlorophyll content. SPAD values are widely applied in agricultural practices for nitrogen management, crop growth assessment, and quality regulation. For instance, in maize research, linear regression models linking SPAD values at different growth stages and leaf positions to leaf nitrogen content reached highly significant levels, enabling precision fertilization guidance [11].

Leaf water content, as a core agronomic trait parameter, plays a crucial indicative role in crop physiology research and agricultural production. It serves not only as the most direct “barometer” of plant water status, sensitively reflecting the severity of water stress and guiding precision irrigation, but also tightly couples with key physiological processes such as photosynthesis and transpiration, collectively determining crop water use efficiency. Plant water content exhibits distinct absorption features in the near-infrared and shortwave infrared regions, with pronounced absorption troughs near 1450 nm and 1940 nm [12]. He et al., (2023) found that absorption depth at 1450 nm positively correlates with wheat leaf water content [8]. Leaf water status does not exist in isolation but is synergistically regulated alongside numerous traits such as specific leaf area, nitrogen content, and stomatal characteristics, collectively forming comprehensive strategy for plants to respond to environmental changes [13]. Ultimately, this synergistic interaction directly influences crop biomass accumulation, economic yield, and quality formation by affecting leaf growth, development, and function. Therefore, whether in field water management, drought-resistant variety breeding, or applications of high-throughput phenotyping and smart agriculture, monitoring and understanding leaf water content has become a critical basis for optimizing water resource allocation, enhancing crop stress tolerance, and achieving sustainable agricultural development.

Leaf nitrogen content serves as a core agronomic parameter characterizing maize nitrogen metabolism and nutritional status, acting as a crucial bridge between nitrogen fertilizer application and photosynthetic products. As a central constituent element of chlorophyll, proteins, and enzymes, nitrogen content exhibits significant positive correlation with photosynthetic rate [14]. By monitoring leaf nitrogen status, real-time diagnosis of plant nitrogen nutrition deficits or surpluses enables precision fertilization tailored to actual needs. Leaf nitrogen content serves as an ideal target for high-throughput remote sensing monitoring, providing critical technical parameters for smart agriculture. Leaf nitrogen exhibits distinctive spectral response characteristics in the visible–near-infrared (VIS-NIR) wavelength range [15]. Such as, the reflectance spectrum of electromagnetic waves can be affected by the chlorophyl pigment especially in blue (450 nm) and red (670 nm) bands in wheat plants which is related to leaf nitrogen content [16]; it measured spectral reflectance from wheat canopy in the red (671 nm) and NIR (780 nm) bands [17]. By mounting hyperspectral sensors on drone or satellite platforms, large-scale, non-destructive acquisition of canopy spectral information can be achieved. Utilizing nitrogen-sensitive spectral indices such as NDRE, quantitative inversion models for leaf nitrogen content can be constructed [18]. This technology transforms traditional point-based destructive sampling into real-time diagnostic mapping, achieving a leap from “post-event analysis” to “process monitoring.” It provides unprecedented data support for managing nitrogen spatial variability across large-scale fields.

Therefore, this study focused on spring corn in Inner Mongolia, employing different fertilization rates and irrigation levels as treatments. Using the FieldSpec4 portable hyperspectral sensor produced by ASD Inc. (USA), spectral reflectance data from maize leaves at various growth stages were collected. Combined with field measurements and experimental analysis, this study investigated the correlation between agricultural traits (SPAD, LWC, and LNC) of maize leaves and their reflectance spectra during different growth stages. Hyperspectral technology was employed to derive vegetation indices correlated with maize leaf growth. Various vegetation indices served as independent variables, while dependent variables included SPAD, LWC, and LNC at different growth stages. Optimal fitting models (RF and SVR) for maize leaves growth at each stage were established and validated using verification data to enhance estimation accuracy. The research findings can provide timely corn growth information for the study area and offer relevant references for fertilization and irrigation practices in farmland.

2. Materials and Methods

2.1. Experimental Site Overview

The experiment was conducted at the China Chilechuan Modern Agricultural Expo Park (40°33′ N, 110°31′ E) in Tumed Right Banner, Baotou City, Inner Mongolia, China (Figure 1). The climate of Tumed Right Banner is characterized by a typical continental semi-arid monsoon climate. The soil type is sandy loam, and the previous crop was maize. Rainfall during the corn growing season (May–September 2023) in the experimental area totaled 267.7 mm. Soil basic fertility data are presented in

Table 1. Soil basic fertility of test site in 2023.

Organic Matter (g/kg)	Available N (mg/kg)	Available P (mg/kg)	Available K (mg/kg)	pH
14.17	67.83	23.18	157.35	7.18

.

2.2. Experimental Design

This experiment employed a split-plot design, with nitrogen application rate and irrigation volume as two interacting factors. The main plots were assigned to nitrogen application rates, while the subplots were assigned to irrigation volumes. N0 (no nitrogen application), N1 (pure nitrogen application at 210 kg/ha, 30% nitrogen reduction), N2 (240 kg/ha, 20% nitrogen reduction), N3 (270 kg/ha, 10% nitrogen reduction), N4 (300 kg/ha), and irrigation rate as the subplot. The irrigation rate had four gradients: W1 (2160 m³/a, 40% water saving), W2 (2880 m³/ha, 20% water saving), W3 (3240 m³/ha, saving 10%), and W4 (3600 m³/ha, conventional irrigation). Urea was used as the nitrogen fertilizer. Irrigation commenced when soil relative moisture content fell below 70%, with approximately 6–7 irrigations throughout the entire growth period after seedling emergence. The variety used was Xianyu 696. Plowing depth was 35 cm. Fertilizer contained P₂O₅ and K₂O at rates of 150 kg/ha and 75 kg/ha, respectively. Planting density was 82,500 plants/ha. Weed control and pest/disease management were strictly implemented throughout the growing season. SPAD values, LWC, and LNC were measured on maize leaves at key growth stages—the bell-mouth stage (V12), silking stage (R1), and milk stage (R3)—under different water–nitrogen combinations. The experiment employed shallow-buried drip irrigation with integrated water–fertilizer management without plastic mulch. Plants were grown in equal row spacing, with one lateral pipe per row positioned 20 cm from the seedling belt.

2.3. Measurement Indicators and Methods

2.3.1. Measurement Indicators

SPAD value: The relative chlorophyll content was measured at the bell-mouth stage (V12), silking stage (R1), and milk stage (R3). Using a handheld chlorophyll meter (SPAD-502, Minolta Osaka Company Ltd., Tokyo, Japan), select three leaf layers (upper, middle, and lower) from the ear leaves of the tested maize plants. For each leaf, take three measurements, avoiding the leaf veins, and calculate the average as the relative chlorophyll content (SPAD) value for that leaf.

Leaf Water Content (LWC): LWC was measured at the bell-mouth stage (V12), silking stage (R1), and milk stage (R3). Three plants meeting density standards were selected. Leaves were blanched at 105 °C for 30 min, dried at 80 °C to constant weight, and then weighed.

Leaf Nitrogen Content (LNC): LNC was measured at the bell-mouth stage (V12), silking stage (R1), and milk stage (R3). From each experimental plot, three plants with identical markers were selected, and leaf samples were collected. The samples were blanched in a 105 °C oven for 30 min, after which the temperature was reduced to 80 °C for drying until a constant weight was achieved. The dried leaves were ground in a pulverizer, and the resulting powder was thoroughly mixed and then digested using the H₂SO₄-H₂O₂ digestion method. Finally, the leaf nitrogen content was determined via the semi-micro Kjeldahl method.

2.3.2. Hyperspectral Data Acquisition

Data acquisition was performed using the FieldSpec4, Pro FR portable spectroradiometer (ASD Inc., Analytical Spectral Devices Boulder, Boulder, CO, USA). The instrument’s basic parameters are as follows: wavelength range: 350–2500 nm; spectral resolution: 3 nm@700 nm, 6 nm@1400/2100 nm; sampling interval: 1.4 nm@350–1000 nm, 1.1 nm@1001–2500 nm.

At each plot, two sampling points were selected for hyperspectral data acquisition of maize leaves. At each point, hyperspectral data were collected from two maize plants. For each plant, leaves from the upper, middle, and lower sections of the ear-positioned leaves were selected, avoiding leaf veins. Five spectra were acquired per measurement. The total number of spectral data points per period was 5 × 2 × 3 × 56 (plots) = 1680. Measurement periods and agronomic parameters were synchronized and acquired concurrently.

When measuring corn leaves’ spectral information, ensure measurements occur on clear, windless, and cloudless days between 10:00 AM and 2:00 PM. Position the sensor probe vertically downward during observation. Calibrate promptly using standard a whiteboard throughout the measurement process. Calibrate against the standard reference board before and after observing each target group. After data collection, raw data undergoes various preprocessing steps using the ASD ViewSpecPro 6.2 software. The measured data from all sampling areas are averaged to obtain the final measurement value for each region.

2.3.3. Spectral Preprocessing

To eliminate background and noise interference in raw spectra and enhance the accuracy and adaptability of spectral information for parameter inversion models, raw spectra (R) of maize leaves at different growth stages require transformations such as logarithmic (logR), first-order derivative (FOD), and second-order derivative (SOD) processing.

2.3.4. Spectral Index Extraction

Spectral indices are combinations of spectral bands obtained through mathematical operations on two or more characteristic bands, representing specific information about the measured material. This study employs the 33 spectral indices listed in

Table 2. Normal spectral indices and calculation formula.

Vegetation Index	Calculation Formula	Citation
NDVI	(Rλ1 − R670)/(Rλ1 + R670)	Rouse [16]
GNDVI	(Rλ1 − R550)/(Rλ1+ R550)	Gitelson et al. [17]
DVI	Rλ1 − R670	Richardson and Wiegand [18]
RVI	Rλ1/R720	Jordan [19]
SAVI	1.5 × (Rλ1 − R670)/(Rλ1 + R670 + 0.5)	Huete [20]
HNDVI	(Rλ1 − R668)/(Rλ1 + R668)	Oppelt and Mauser [21]
PRI	(Rλ1 − R531)/(Rλ1 + R531)	Penuelas et al. [22]
SIPI	(Rλ1 − R450)/(Rλ1 +R450)	Penuelas et al. [22]
PSNDa	Rλ1 − R680)/(Rλ1 + R680)	Blackburn [23]
PSNDb	(Rλ1 − R635)/(Rλ1 + R635)	Blackburn [23]
PSSRa	Rλ1/R680	Blackburn [23]
PSSRb	Rλ1/Rλ2	Blackburn [23]
CIred_edge	Rλ1/Rλ2-1	Gitelson et al. [24]
SR	Rλ1/Rλ2	Jordan [19]
VOG1	Rλ1/Rλ2	Vogelmann et al. [25]
MRESR	(Rλ1 − R445)/(Rλ1 + R445)	Datt [26]
NPCI	(Rλ1 − Rλ2)/(Rλ1 + Rλ2)	Penuelas et al. [27]
GRVI	Rλ1/Rλ2	Gitelson et al. [28]
RNDVI	(Rλ1 − Rλ2)/sqrt(Rλ1 + Rλ2)	Wang et al. [29]
MSR	(Rλ1/Rλ2 − 1)/(sqrt(Rλ1/Rλ2) + 1)	Haboudane et al. [30]
NPQI	(Rλ1 − Rλ2)/(Rλ1 + Rλ2)	Barnes et al. [31]
IPVI	Rλ1/(Rλ1 + Rλ2)	Kooistra et al. [32]
RENDVI	(Rλ1 − Rλ2)/(Rλ1 + Rλ2)	Gitelson and Merzlyak [33]
NDNI	(log(1/Rλ1) − log(1/Rλ2))/(log(1/Rλ1) + log(1/Rλ2))	Serrano et al. [34]
MSI	Rλ1/Rλ2	Hunt et al. [35]
NDII	(Rλ1 − Rλ2)/(Rλ1 +Rλ2)	Serrano et al. [34]
NDWI	(Rλ1 − Rλ2li)/(Rλ1 + Rλ2)	Gao [36]
WBI	Rλ1/Rλ2	Penuelas et al. [37]
mSR	(Rλ1 − Rλ2)/(Rλ1 + Rλ2)	Sims and Gamom [38]
PPR	(Rλ1 − Rλ2)/(Rλ1 + Rλ2)	Kooistra et al. [32]
NDSI	(Rλ1 − Rλ2)/(Rλ1+ Rλ2)	Richardson et al. [39]
WI	Rλ1/Rλ2	Penuelas et al. [37]
GVWI	[(Rλ1 + 0.1) − (Rλ2 + 0.02)]/[(Rλ1 + 0.1) + (Rλ2 + 0.02)]	Ceccato et al. [40]

Note: In the formula, R_λ1 denotes the spectral reflectance at wavelength λ1, R_λ2 denotes the spectral reflectance at wavelength λ2, and so on.

.

2.3.5. Model Construction and Validation

RF and SVR models were employed to construct estimation models for SPAD, LNC, and LWC in maize leaves. Decisions were made based on the expression of agronomic phenotypic traits during critical growth stages of maize, and machine learning models were trained accordingly. The scikit-learn library [41,42] was used to establish models for the estimation of SPAD, LNC, and LWC using common machine learning methods: RF and SVR regression models. In this study, mtry was set to 1/3 of the number of independent variables, and ntree was set to 500, for the RF model, and timized within [10−2, 10−1, 1, 10, 100] and [10−4, 10−3, 10−2, 10−1, 1, 10], respectively, for the SVR model. Ten-fold cross-verification and grid searching were used to identify the optimal parameters during model development. Maize leaves’ spectral data were collected and then randomly divided into a calibration dataset (2/3) and a validation dataset (1/3). Model accuracy was then evaluated using the coefficient of determination (R²) and root mean square error (RMSE).

Coefficient of Determination (R²): A statistical measure indicating the closeness of the relationship between dependent and independent variables. The name of this statistical indicator varies based on the characteristics of the correlation phenomenon. The calculation formula is as follows:

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

(1)

Root Mean Square Error (RMSE): The RMSE is the square root of the ratio of the sum of squares of the differences between predicted and actual values to the number of observations. It measures the deviation between predicted and actual values and is sensitive to outliers in the data. In the formula, $y_i$ represents the actual value, ${\widehat{y}}_i$ represents the predicted value, $\mathrm{n}$ represents the number of samples, and $\overline{y}$ represents the mean of the sample values. The calculation formula is as follows:

$$RMSE=\sqrt{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2/n}$$

(2)

2.3.6. Data Analysis

Data processing was performed using Microsoft Excel 2021, and split-plot analysis of variance was conducted using SAS 9.4. Multiple factor analysis of variance was performed using the LSD (least-significant difference) method. Origin 2021 was used for graphing. Spectral data underwent preprocessing using the ASD data post-processing software ViewSpecPro and ENVI 5.3.

3. Results

3.1. Spectral Characteristics Analysis of Maize Leaves

3.1.1. Spectral Characteristics of Maize Leaves at the Different Spectral Transformations

The application of a suite of transformations—logarithmic, first-order derivatives, and second-order derivatives—to the raw spectral data of maize leaves elucidates subtle features in reflectance profiles, thereby enhancing spectral detail resolution throughout key phenological phases. This effect is particularly pronounced in the first- and second-order derivative spectra, which amplify spectral details more significantly (Figure 2).

3.1.2. Spectral Characteristics of Maize Leaves Under Water–Nitrogen Coupling Conditions

The spectral reflectance curves of maize leaves under different nitrogen application rates and irrigation conditions reveal a gradual increase in spectral reflectance with increasing nitrogen application, exhibiting the pattern N0 < N1 < N2 < N3 < N4. As irrigation rates decreased, maize leaf spectral reflectance showed a gradually increasing trend, exhibiting the pattern W1 > W2 > W3 > W4 (Figure 3).

3.1.3. Spectral Characteristics of SPAD Value in Different Spectral Transformations

Spectral transformations such as logarithmic (logR), first-order derivative (FODR), and second-order derivative (SODR) were applied to the raw spectral data (R). First-order derivative and second-order derivative spectral transformations effectively eliminate linear noise during processing while highlighting variations in spectral reflectance. When SPAD values of maize leaves change, spectral shifts occur across the entire wavelength range. However, since chlorophyll primarily absorbs and utilizes light energy in the visible spectrum, spectral data in this region exhibits heightened sensitivity to SPAD value fluctuations. Figure 4 demonstrates that within the 350–700 nm range, spectral patterns exhibit consistent, regular variations corresponding to SPAD changes. Although significant spectral shifts occur across 700–2500 nm with SPAD variations, these lack pronounced regularity or consistency.

Within the visible spectrum, corn leaves exhibit high chlorophyll content, resulting in spectral reflectance curves that increase as SPAD values decrease. Due to chlorophyll’s strong reflection in the near-infrared range, spectral reflectance gradually rises in the near-infrared band. Enhanced red edge bands are observable in the reflectance curves derived from first-derivative and second-derivative spectral transformations.

3.1.4. Spectral Characteristics of Maize Leaves at Different Moisture Contents

Figure 5 shows the spectral curve characteristics of maize leaves at different moisture contents. As shown in Figure 5, the spectral reflectance of maize leaves exhibits a decreasing trend with increasing leaf water content. This decrease is most pronounced in the 350–1350 nm wavelength range, while no significant changes occur between 1350 and 2500 nm. Two water absorption troughs are observed at 1450 nm and 1960 nm. The reflectance at these absorption troughs gradually increases as leaf water content decreases.

3.1.5. Spectral Characteristics of Maize Leaves with Varying Nitrogen Content

Figure 6 shows the spectral characteristics of maize leaves with different nitrogen contents after various spectral transformations. As shown in Figure 5, the spectral reflectance of maize leaves increases with rising nitrogen content across the 760–1150 nm range. In the first-order derivative spectral transformation of the red edge region, the red edge amplitude increases with higher nitrogen content, and a “red shift” phenomenon occurs at the red edge position. In the second-derivative spectral transformation, nitrogen content variations exert a significant influence on the visible light band.

3.2. Variation Patterns in Phenotypic Parameters of Maize Leaf Agronomic Traits

Nitrogen application and irrigation volume each exerted a highly significant effect on SPAD values at all growth stages (P < 0.01); however, their interaction did not reach statistical significance (P > 0.05). At the bell-mouth stage, no significant differences were detected among treatments. The SPAD value under N4W4 was 64.45, which was statistically comparable to that under N1W4, N2W3, and N2W4. By the silking stage, the value in the N4W4 treatment reached 71.1, showing no significant differences from N3W4, N2W3, and N4W3. At the milking stage, the highest chlorophyll content (66.1) occurred under N3W4 and did not differ significantly from that under N2W4, N4W3, or N4W4. Overall, SPAD value exhibited a unimodal trend across the three growth stages, increasing initially and then declining. Under the no-nitrogen condition (N0), low irrigation levels (W1 and W2) resulted in a gradual decline, whereas moderate and high irrigation (W3 and W4) led to an initial increase followed by a decrease. In contrast, under nitrogen application (N1–N4), all treatments consistently showed an increase–decrease pattern.

Nitrogen application significantly influenced leaf water content at the bell-mouth, silking, and milking stages (P < 0.01). Irrigation volume also had a highly significant effect on water content at the silking and milking stages, but not at the bell-mouth stage. Their interaction significantly affected water content at the bell-mouth and milking stages, but not at the silking stage. Leaf water content increased with advancing growth stages, peaking at the silking stage before declining during the milking stage.

Nitrogen, irrigation, and their interaction all significantly affected leaf nitrogen content at each growth stage (P < 0.01). At the bell-mouth stage, the nitrogen content under N1W4 was 1.477%, which did not differ significantly from that under N2W4, N3W4, or N4W4. At the silking stage, the value measured in N4W3 was 1.44% and was statistically similar to that in N4W4. At the milking stage, the nitrogen content in N2W3 was 1.19%, showing no significant differences compared to N2W2, N2W4, or N3W4. Leaf nitrogen content consistently followed a pattern of initial increase followed by a decrease across growth stages, with the maximum value occurring at the silking stage (

Table 3. Analysis of variance of SPAD, LWC, and LNC in maize leaves.

Agronomic Trait Parameters	Source of Variation	V12	R1	R3
SPAD
	Nitrogen application rate(N)	9.48 **	17.96 **	38.27 **
	Main zone error	1.99	5.47	1.71
	Irrigation amount(W)	18.32 **	17.59 **	8.86 **
	Secondary zone error	3.53	7.92	5.38
	Nitrogen application rate × Irrigation amount (N × W)	1.01 ns	0.49 ns	0.2 ns
LWC
	Nitrogen application rate(N)	47.12 **	5.8 **	18.11 **
	Main zone error	0.00	0.00	0.00
	Irrigation amount(W)	2.62 ns	24.97 **	181.59 **
	Nitrogen application rate × Irrigation amount (N × W)	6.63 **	1.04 ns	9.66 **
	Secondary zone error	0.00	0.00	0.00
LNC
	Nitrogen application rate(N)	282.17 **	45.92 **	50.20 **
	Main zone error	0.00	0.00	0.00
	Irrigation amount(W)	79.75 **	153.57 **	177.55 **
	Nitrogen application rate × Irrigation amount (N × W)	15.83 **	17.12 **	7.54 **
	Secondary zone error	0.00	0.00	0.00

Note: ** represents 0.01 significant level; ns stands for not significant V12: the bell-mouth stage; R1: the silking stage; R3: the milk stage.

).

3.3. Screening Optimal Spectral Indices

3.3.1. Correlation Analysis Between Spectral Indices and SPAD Values of Maize Leaves

Correlation analysis was conducted between various spectral indices from different spectral processing methods and SPAD values of maize leaves at different growth stages. The correlation coefficients (r) obtained from analyzing the spectral indices of the raw spectrum (R) with SPAD values at the V12, R1, R3, and All growth stages ranged from −0.10 to 0.12, −0.35 to 0.28, −0.51 to 0.48, and −0.18 to 0.22, respectively. Correlation analysis between logarithmic spectra (logR) and SPAD values at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.12 to 0.10, −0.25 to 0.25, −0.38 to 0.46, and −0.07 to 0.19, respectively; correlation analysis between the first-order derivative spectrum (FODR) and SPAD values at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.17 to 0.28, −0.37 to 0.38, −0.26 to 0.39, and −0.17 to 0.32, respectively; correlation analysis between second-order derivative spectra (SODR) and SPAD values at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.36 to 0.43, −0.28 to 0.37, −0.62 to 0.41, and −0.41 to 0.36, respectively (Figure 7). The spectral indices showing significant correlations with SPAD values for each spectral treatment are listed in

Table 4. Spectral indices sensitive to SPAD values extracted from maize leaves at different growth stages using different spectral treatments.

	R	logR	FODR	SODR
V12	NDSI	PRI, NPQI, mSR	GNDVI, PSNDb, GRVI, NPQI	PSNDa, PSSRa, NDII, NDWI, GVWI
R1	SAVI, PRI, PSNDa, SR, NDSI	SAVI, PRI, SR, VOG1, RNDVI, NPQI, RENDVI, mSR	MSI, NDII, WBI, WI	MRESR, NDII, WBI, NDSI, WI
R3	GNDVI, RVI, PSSRb, VOG1, MRESR, GRVI, NPQI, RENDVI, NDNI, MSI, NDWI, mSR	NPQI, NDNI, PPR	NDVI, DVI, SAVI, VARIgreen, RNDVI, IPVI, NDSI	DVI, SAVI, VARIgreen, SR, NPCI
All	NPCI, NDSI	SR, MRESR, NPCI, mSR	DVI, SAVI, VARIgreen, RNDVI	DVI, SAVI, VARIgreen, SR, NPCI

Note: V12: the bell-mouth stage; R1: the silking stage; R3: the milk stage; All: full growth period; R: raw spectral data; logR: logarithmic spectral data; FODR: first-order derivative spectral data; SODR: second-order derivative spectral data.

. In summary, applying second-order derivative transformation to raw spectra significantly enhances the correlation between spectral indices and SPAD values.

3.3.2. Correlation Analysis Between LWC and Spectral Index

Figure 8 shows that the correlation analysis between various spectral indices from different spectral processing methods and LWC at different growth stages yielded correlation coefficients (r) ranging from −0.45 to 0.31 for the raw spectral index (R) with LWC at the V12 stage, from −0.45 to 0.42 for R1 and R3 indices, and from −0.40 to 0.43 for R3 and all growth stages, −0.40 to 0.43, and −0.34 to 0.43, respectively. Correlation analysis between logR spectra and LWC at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.32 to 0.38, −0.40 to 0.41, −0.34 to 0.30, and −0.45 to 0.23, respectively; correlation analysis between the first-order derivative spectrum (FODR) and LWC at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.42 to 0.57, −0.37 to 0.36, −0.43 to 0.26, and −0.29 to 0.25, respectively. The correlation analysis between the second-order derivative optical spectrum (SODR) and corn leaf moisture content at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.23 to 0.45, −0.36 to 0.30, −0.46 to 0.55, and −0.38 to 0.40, respectively.

Spectral indices showing significant correlations with corn leaf moisture content for each spectral treatment are listed in

Table 5. Different spectral treatments used to extract spectral indices sensitive to water content of maize leaves at different growth stages.

	R	logR	FODR	SODR
V12	NDVI, RVI, HNDVI, PRI, SIPI, PSNDa, PSNDb, PSSRa, SR, MSR, IPVI, PPR	DVI, PRI, VARIgreen, MRESR, NDNI, mSR, PPR	GNDVI, PSSRb, GRVI, NDSI	HNDVI, VOG1, NDSI, GVWI
R1	NDVI, RVI, HNDVI, PRI, PSNDa, PSSRb, SR MSR, IPVI	DVI, PRI, VARIgreen	MSI, NDII, NDWI, PPR	NDSI, GVWI, RENDVI, PSNDa,
R3	MSI, NDII, NDWI, WBI, WI, GVWI	NDNI, MSI, NDII, NDWI, WBI, WI, GVWI	GNDVI, GRVI, NDII, WBI, WI, GVWI	NDSI, RENDVI, PRI
All	NPCI, RENDVI, WBI, PPR, NDSI, WI, GVWI	MRESR, NPCI, Msr, PPR	DVI, SAVI, PSNDa, VARgreen, NPCI, NDSI	NDSI, SR, VARIgreen, HNDVI, SAVI, DVI

Note: V12: the bell-mouth stage; R1: the silking stage; R3: the milk stage; All: full growth period; R: raw spectral data; logR: logarithmic spectral data; FODR: first-order derivative spectral data; SODR: second-order derivative spectral data.

. In summary, applying first-order derivative spectral transformation to raw spectra significantly enhances the correlation between spectral indices and corn leaf moisture content.

3.3.3. Correlation Analysis of LNC and Spectral Indices

Correlation analysis was conducted between multiple spectral indices processed through different spectral methods and LNC at various growth stages. The correlation coefficients (r) obtained from analyzing the spectral indices of the raw spectrum (R) with LNC at the V12, R1, R3, and All growth stages ranged from −0.22 to 0.33, −0.37 to 0.30, 0.46 to 0.34, and −0.49 to 0.45, respectively. Correlation analysis between logR spectra and LNC at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.33 to 0.22, −0.26 to 0.34, −0.35 to 0.39, and −0.42 to 0.52, respectively. Correlation analysis between first-order derivative spectra (FODR) and LNC at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.31 to 0.26, −0.19 to 0.24, −0.45 to 0.35, and −0.42 to 0.48, respectively. Correlation analysis between second-order derivative spectra (SODR) and LNC at V12, R1, R3, and All growth stages yielded correlation coefficients (r) ranging from −0.35 to 0.44, −0.37 to −0.26, −0.47 to −0.49, and −0.27 to −0.23, respectively (Figure 9). The spectral indices showing significant correlations with corn leaf nitrogen content for each spectral transformation are listed in

Table 6. Different spectral treatments were used to extract the sensitive spectral indices of nitrogen content in maize leaves at different growth stages.

	R	logR	FODR	SODR
V12	NDSI, NPQI, NPCI	PRI, NPCI, NPQI	RVI, PSNDa, PSSRa, NDSI,	GVWI, PPR, HNDVI
R1	NDNI, IPVI, MSR, SR, PSSRb, PSSRa, PSNDb, PSNDa, SIPI, HNDVI, SAVI, RVI, NDVI	DVI, SAVI, VARIgreen, RNDVI	GRVI, NDII, WI, GVWI	GVWI, WI, NDSI, PPR, WBI, HNDVI
R3	PPR, mSR, RENDVI, NPQI, GRVI, MRESR, VOG1, PSSRb, PRI, GNDVI	MRESR, NPQI, mSR, PPR	DVI, SAVI, VARIgreen, VOG1, MRESR, RNDVI, RENDVI, mSR	GVWI, RENDVI, NPCI, SR
All	WI, WBI, NDWI, NDII, MSI, NPCI	MRESR, NPCI, NDNI, mSR, PPR, GVWI	DVI, SAVI, PSNDa, PSSRa, PSSRb, VARIgreen, NPCI	GVWI, PPR, NPCI, SR, SIPI

Note: V12: the bell-mouth stage; R1: the silking stage; R3: the milk stage; All: full growth period; R: raw spectral data; logR: logarithmic spectral data; FODR: first-order derivative spectral data; SODR: second-order derivative spectral data.

. In summary, applying first-derivative and second-derivative spectral transformations to the raw spectral data (R) significantly enhances the correlation between spectral indices and maize leaf nitrogen content.

3.4. Development of Hyperspectral Estimation Model

Using the optimized spectral indices as multivariate variables, RF and SVR machine learning models were employed to estimate agronomic phenotypic traits of maize leaves (SPAD, LWC, and LNC). During model construction, the optimized spectral indices were first standardized to eliminate the impact of dimensional and magnitude differences among them on the models. For the RF model, key parameters such as the number of trees and maximum depth were adjusted to identify the optimal model architecture, thereby enhancing prediction accuracy and generalization capability. For the SVR model, optimization focused on kernel type, penalty coefficient, and insensitivity loss parameters to ensure the model accurately captured the complex nonlinear relationships between spectral indices and SPAD, LWC, and LNC. After multiple training and validation iterations, the final RF and SVR models were obtained for estimating SPAD, LWC, and LNC during the critical growth stages of maize leaves.

In the training datasets, the optimal ${\mathrm{R}}_{\mathrm{c}}^2$ and RMSE_c values for SPAD in the RF model were achieved during the V12 stage with SODR spectral processing, at 0.92 and 1.03, respectively, while the SVR model’s optimal values were obtained during All stage with FODR processing, at 0.73 and 0.08, respectively. For LWC, the optimal ${\mathrm{R}}_{\mathrm{c}}^2$ and RMSE_c values for the RF model were achieved during the V12 stage with FODR spectral processing, at 0.94 and 0.77, respectively, while the SVR model’s optimal values were obtained during the R1 stage with raw spectral (R) processing, at 0.69 and 0.09, respectively. For LNC, the ${\mathrm{R}}_{\mathrm{c}}^2$ and RMSE_c values under the optimal RF model with FODR spectral processing at the V12 stage were 0.95 and 1.02, respectively, while the optimal SVR model with FODR processing of raw spectral (R) across All stages yielded values of 0.73 and 0.08 (Figure 10).

In the validation datasets, the optimal values for SPAD’s ${\mathrm{R}}_{\mathrm{v}}^2$ and RMSE_v were 0.90 and 0.15, respectively, for the RF model with FODR spectral processing during the R3 period. For LWC, the optimal values of ${\mathrm{R}}_{\mathrm{v}}^2$ and RMSE_v were 0.77 and 1.77, respectively, for the RF model with logR spectral processing during the V12 period. For LNC, the optimal values of ${\mathrm{R}}_{\mathrm{v}}^2$ and RMSE_v were 0.88 and 1.95, respectively, for the RF model with SODR spectral processing during the All period (Figure 11).

In summary, for the agronomic phenotypic traits of maize leaves (SPAD, LWC, and LNC), the RF model proved to be the optimal estimation model in both the training and validation datasets during the critical growth stages.

4. Discussion

4.1. Effects of Water–Nitrogen Coupling Reduction on Agronomic Trait Parameters of Maize Leaves

This study systematically investigated the effects of nitrogen fertilizer application rates and irrigation water volumes on key physiological indicators of maize growth. Results indicate that water–nitrogen management exerts significant and systematic regulatory effects on leaf chlorophyll relative content (SPAD values), water content, and nitrogen content.

First, regarding leaf photosynthetic potential, both nitrogen (N) fertilization and irrigation volume independently exerted significant effects (P < 0.01) on SPAD values across all growth stages. This is consistent with established knowledge that nitrogen is fundamental for chlorophyll synthesis, while water stress impairs chloroplast integrity and nitrogen assimilation [43]. The absence of a significant interaction suggests that their effects on SPAD values were additive rather than synergistic under the experimental conditions. A characteristic unimodal trend in SPAD values was observed throughout the developmental cycle, increasing from the bell-bottom stage, peaking at silking, and subsequently declining during the milking stage, reflecting the natural progression of chlorophyll degradation during senescence [44]. Notably, under zero-N conditions, low irrigation levels led to a continual decline in SPAD, whereas medium to high irrigation supported an initial increase followed by a decrease. This indicates that sufficient water availability can partially alleviate the early-stage stress induced by severe nitrogen deprivation.

Second, concerning leaf water status, N application significantly enhanced leaf water content from the heading to the grain-filling stages, likely by fostering root development and canopy architecture, thereby improving plant water uptake and retention capacity [45]. The effect of irrigation was most pronounced during the mid-to-late stages (silking and grain-filling), which coincide with peak crop water demand. A significant interaction between water and N was detected at the heading and grain-filling stages, underscoring the critical role of water–nitrogen synergy in maintaining leaf hydration during specific phonological windows. The temporal pattern of leaf water content, peaking at tasseling and declining thereafter, aligns with the shift from vegetative to reproductive growth and the ensuing senescence process.

Finally, for leaf nitrogen status, N fertilizer, irrigation, and their interaction all significantly influenced leaf N content at all stages (p < 0.01), strongly affirming the central role of water–nitrogen co-regulation in optimizing N uptake and utilization [46]. Water mediates N use efficiency by affecting its soil mobility, root absorption, and in planta translocation. Leaf N content also exhibited a unimodal dynamic, reaching its maximum at the tasseling stage—a period critical for ensuring pollen viability, grain set, and ultimate yield formation [47].

In summary, optimized water–nitrogen management (e.g., treatments N3W4 and N4W4) effectively synergized to enhance chlorophyll content, water status, and nitrogen nutrition in maize leaves, maintaining them at high levels during the critical reproductive phase (silking to tasseling). These findings provide a theoretical foundation for high-yielding, efficient maize cultivation through precise irrigation and fertilization strategies.

4.2. The Impact of Spectral Processing on Machine Learning Model Estimation of Corn Leaf Agronomic Parameters

Hyperspectral technology provides a key method for non-destructive, efficient monitoring of crop physiological parameters. This study systematically investigated the application and effectiveness of different spectral processing methods and machine learning models in estimating corn leaf chlorophyll content (SPAD), leaf water content (LWC), and leaf nitrogen content (LNC).

The visible and near-infrared (VIS–NIR) spectral regions are highly sensitive to variations in leaf chlorophyll content, and estimation models based on spectral indices in these regions are well established. To further improve model performance, spectral transformation techniques are often applied to preprocess raw spectral data. Studies have shown that derivative transformations can enhance inter-band discriminability and suppress instrumental noise, thereby significantly strengthening the correlation between spectral reflectance features and chlorophyll content [4,48]. In terms of model selection, the random forest (RF) algorithm demonstrated excellent performance in chlorophyll estimation, achieving a maximum R² of 0.96, substantially surpassing that of the support vector regression (SVR) model (R² = 0.82). As an ensemble method, RF mitigates data redundancy through multi-variable collaboration, thereby improving model stability and generalization ability [49].

For monitoring leaf water content (LWC), the near-infrared and shortwave infrared regions contain prominent water absorption bands, making them critical for detecting plant moisture status. This study confirms that raw spectral indices are already significantly correlated with LWC, while first- and second-order derivative transformations further minimize background effects and accentuate moisture-related spectral features, leading to improved correlation and estimation accuracy. In comparative analyses, the RF model consistently exceeded the SVR model in both calibration and validation accuracy. This advantage is largely attributable to RF’s ability to integrate multiple decision trees and optimize feature usage, resulting in more robust and accurate inversion models [50].

Leaf nitrogen content (LNC) also displays distinct spectral responses within the VIS–NIR range. Spectral transformations help reveal subtle spectral information otherwise obscured in raw data. By emphasizing spectral variability and reducing noise, these preprocessing methods enhance the relationship between spectral indices and LNC, leading to more reliable inversion models [51]. Feature selection remains a crucial step in developing machine learning models, as it alleviates data redundancy and directly elevates model performance [52,53,54,55,56,57]. Under multi-variable input scenarios, RF consistently achieved higher calibration and validation accuracy than SVR, underscoring the superiority of the ensemble RF approach in handling high-dimensional spectral data and enabling precise retrieval of crop nitrogen status.

In summary, preprocessing through spectral transformations—particularly derivative transformations—combined with advanced machine learning algorithms such as RF, can significantly enhance the accuracy and robustness of hyperspectral estimation models for key physiological parameters in maize leaves, including chlorophyll, moisture, and nitrogen content. This research approach provides reliable technical support for precise crop growth monitoring and nitrogen–water management.

5. Conclusions

Hyperspectral remote sensing serves not only as the foundation for precision agriculture but also as a key to overcoming bottlenecks in modern agricultural development. The experiments yielded the following conclusions:

As nitrogen application rates increase, maize leaf spectral reflectance exhibits a gradually increasing trend, showing the pattern N0 < N1 < N2 < N3 < N4. Conversely, as irrigation water volumes decrease, maize leaf spectral reflectance also shows a gradually increasing trend, exhibiting the pattern W1 > W2 > W3 > W4. Agricultural traits (SPAD, LWC, and LNC) of maize leaves changed regularly across growth stages. SPAD and LNC increased then decreased, peaking at tasseling, while LWC declined continuously. Under nitrogen treatments, all parameters showed N4 > N3 > N2 > N1 > N0. Under irrigation treatments, chlorophyll and nitrogen followed W1 > W2 > W3 > W4, opposite to water content. Reflectance between 350 and 1350 nm negatively correlated with water content, showing absorption features at 1450 nm and 1960 nm. First- and second-derivative spectral transformations enhanced correlations with water and nitrogen content while reducing noise. The random forest (RF) model using second-derivative (SODR) spectra outperformed support vector regression (SVR) in estimating chlorophyll.

This study provides theoretical foundations and technical support for hyperspectral inversion of agronomic parameters in maize leaves within the experimental area through simple spectral transformations. Due to constraints in manpower, resources, and specific weather/time conditions during the experiment, the application of non-imaging spectral data was limited to certain growth stages of maize leaves, focusing on the relationship between non-imaging spectral data and maize physiological parameters. Future work requires gradual refinement and in-depth exploration of novel modeling approaches. Concurrently, further investigations into the relationship between non-imaging spectral data and maize physiological parameters, alongside comprehensive studies using imaging spectral data across all growth stages, will provide robust theoretical and technical foundations for large-scale real-time maize monitoring.