Field Spectroscopy for Monitoring Nitrogen Fertilization and Estimating Cornstalk Nitrate Content in Maize

Abstract

Evaluating the response of maize crops to different nitrogen fertilization rates is essential to ensure their agronomic, environmental, and economic efficiency. In this study, the spectral information of maize plants subjected to five distinct nitrogen fertilization strategies was analyzed. The fertilization strategies were based on the practices commonly used in maize fields in the study area, with the aim of ensuring the research findings’ applicability. The spectral reflectance was measured using a spectroradiometer covering the 350–2500 nm range, and the results enabled the identification of optimal spectral regions for monitoring plants’ nitrogen status, particularly in the visible and infrared ranges. A Principal Component Analysis (PCA) of the reflectance data revealed the key wavelengths most sensitive to the nitrogen availability: 555 nm and 720 nm during the vegetative stage and 680 nm during the reproductive stage. This information will support the development of drone-mounted multispectral sensor systems for large-scale monitoring, as well as the design of low-cost sensors for early nitrogen deficiency detection. Furthermore, the study demonstrated the feasibility of estimating the cornstalk nitrate content based on direct reflectance measurements of maize stems. The prediction model showed satisfactory performance, with a coefficient of determination (R²) of 0.845 and a root mean square error of prediction (RMSECV) of 2035.3 ppm, indicating its strong potential for predicting the NO₃-N concentrations in maize stems.

1. Introduction

Maize (Zea mays L.) has long been a key component of agrifood systems, and its significance continues to grow. In recent decades, its global production has expanded rapidly due to increasing demand, technological innovations, improved yields, and larger cultivation areas. It is the most produced cereal and is expected to become the most widely grown and traded crop in the near future [1]. While primarily used for animal feed, it also serves as a key raw material in sectors like nutrition, healthcare, chemicals, and bioenergy. The substantial crop residue it generates supports nutrient recycling, increases the organic matter, and helps to protect the soil, while its roots contribute to improvements in the soil structure [2].

One of the major challenges in corn production is ensuring an adequate nutrient supply. Nitrogen (N) plays a crucial role in maize plants’ development, and a deficit of it can potentially cause substantial reductions in maize growth and yields [3]. Nitrogen management is particularly complex due to the numerous chemical and biological processes affecting nitrogen’s availability, as well as the plant uptake’s strong dependence on the soil and climatic conditions. Specifically, the fertilization of corn with nitrates (NO₃-N) can cause several problems, both environmental and health-related, if not managed properly. In fact, N from fertilizers has significant potential to end up in the environment, as most applied fertilization is not immediately absorbed by crops [4]. The main issues regarding nitrate use in corn fertilization are water contamination, nitrogen oxide emissions, its effects on soil salinity, its negative impacts on human health, and the loss of fertilizer efficiency, among others [5,6]. Therefore, improving the nitrogen use efficiency in agricultural fields can help to ensure both food and environmental security [7]. To mitigate the aforementioned problems, it is important to properly manage fertilization, considering the specific needs of the crop, soil type, and climatic conditions. This goal can be achieved through the widespread implementation of existing and emerging technologies in agricultural systems. For example, the use of GNSS technologies and yield monitors can help to ensure more efficient nitrogen application, such as through variable-rate fertilization [8,9]. In this context, field spectroscopy offers a valuable tool for assessing the nitrogen fertilization efficiency, as it is a rapid, non-invasive, and reliable technique for estimating the nitrate content in maize plants.

Traditional approaches to monitoring nitrogen levels involve analyzing plant tissues, a widely used method for assessing the nutrient content in leaves and optimizing nitrogen fertilizer applications [2]. However, these conventional methods are costly and face limitations regarding their large-scale use, as they require extensive sampling and involve a slow data acquisition process. As an alternative, hyperspectral remote sensing (RS) has emerged as an efficient and cost-effective tool to assess the nitrogen content in agriculture [10]. Spectroscopy provides an accurate, non-destructive estimation of the leaf nitrogen accumulation. This is essential for monitoring crop growth and optimizing nitrogen management, with direct implications for real-time fertilization and yield prediction [11]. The spectral reflectance of plants within the visible (VIS), near-infrared (NIR), and shortwave infrared (SWIR) ranges provides valuable insights into their nitrogen status [12]. NIR spectroscopy is widely used for quantifying compounds in biological materials because it can detect specific molecular vibrations. These include the vibrations of the N-H bonds found in proteins and other nitrogen-containing compounds in plants [13]. Remote sensing indicators related to nitrogen levels are often associated with the chlorophyll or nitrogen concentration, and hyperspectral data can be collected from individual leaves or the entire crop canopy. Research has shown a significant correlation between the leaf/canopy characteristics and grain yield in cereal crops [14]. Since photosynthesis is a key factor in converting light into biomass and directly influences both the biomass and grain production, remote sensing indicators derived from plant aerial structures offer essential information for crop monitoring and nutrient assessment [15].

There is a large body of literature on estimating the nitrogen content of plants and monitoring crops using field spectroscopy. In this regard, Berger et al. [16] provide an excellent literature review on crop nitrogen monitoring in the context of spectroscopy. Oliveira et al. [2] reviewed the literature on nitrogen predictions in corn leaves and reported promising results, with the models’ R² values ranging from 0.6 to 0.9. Graeff & Claupein [17] estimated the total N concentration using reflectance data models that showed a strong correlation between the nitrogen levels and leaf reflectance. This correlation was particularly evident at wavelengths of 510–780 nm, 516–1300 nm, and 540–1300 nm, with the reflectance at these wavelengths consistently corresponding to the N status across different years and sampling dates. Gautam and Panigrahi [18] estimated the leaf nitrogen content in corn plants using aerial images, achieving strong predictive performance, with the models tested yielding an RMSE of 6.6% and a correlation coefficient of 78%. Xie et al. [19] found that the best spectral region for N content estimation in corn leaves was 350–1000 nm, developing prediction models that achieved an R² of 0.65 and an RMSE of 0.26. Lu et al. [20] compared the response of two different maize varieties to N stress across different growth stages. The spectral reflectance in the 548–556 nm and 706–721 nm ranges showed higher sensitivity to the leaf nitrogen content of the first variety at the V12 stage. In contrast, the second variety exhibited stronger correlations between the nitrogen content and reflectance in the 760–1142 nm range at the R1 stage. These findings support the potential for a variety of stage-specific nitrogen management strategies to improve the fertilization precision, supported by hyperspectral data. Recently, some studies showed that considering both spectral bands and vegetation indices (VIs) improved the prediction accuracy of models (R² between 0.74 and 0.87; RMSE ranged from 1.00 to 1.50 kg ha⁻¹) compared to using bands (R² from 0.69 to 0.85; RMSE between 1.00 and 1.77 kg ha⁻¹) or VIs alone (R² ranged from 0.55 to 0.68; RMSE between 1.00 and 1.78 kg ha⁻¹) [21]. Ramos-García et al. [22] reported that both the phenological stage and nitrogen application rate significantly influenced the chlorophyll and nitrogen contents, as well as the leaf spectral responses. Indices based on the red edge range exhibited strong correlations with these physiological traits. Additionally, the grain yield was highly correlated with the nitrogen and chlorophyll levels.

One aspect of estimating biochemical variables from spectral signatures that generates debate is whether or not reflectance data should be preprocessed before model development. Some authors, such as Martin et al. [23], argue against applying any transformations to the data to ensure that the models remain robust and applicable across multiple scenarios. Following this approach, successful predictive models of the leaf nitrogen content have been developed [2], and crops’ responses to different fertilizer doses have been evaluated [19]. However, other authors recommend using various preprocessing techniques to enhance the predictive accuracy. These include converting the reflectance to absorbance [24], removing spectral regions with a low signal-to-noise ratio [25], applying smoothing filters, and resampling using spline interpolation [26].

Nevertheless, these studies suggest that further research is needed on nutritional analysis. Given the complexity of the nitrogen dynamics in the soil and crops, robust and reliable techniques for assessing the leaf nitrogen content require deeper investigation to improve our understanding of the information provided by spectral responses. Most of the studies mentioned above were conducted on plants grown under highly controlled experimental plot conditions or used laboratory measurements. There is a lack of studies focusing on maize plants grown under conditions similar to those of commercial maize fields and using measurements taken directly in the field. Also relevant and requiring further investigation is the suitability of applying some form of preprocessing to reflectance data prior to model calibration.

Therefore, the aim of this study was to characterize the spectral signature of maize plants subjected to different nitrogen fertilization treatments with both varying application rates and timing. Special emphasis was placed on the spectral identification of nitrogen-deficient plants based on measurements taken directly in the field. Simultaneously, the fertilization efficiency was assessed through the estimation of the NO₃-N concentration in the maize stem. The nitrogen treatments applied were similar to those used in commercial maize fields in the study area.

2. Materials and Methods

2.1. Study Site and Experiment Design

The study was conducted during the 2024 growing season in a commercial corn plot located in Grisuela del Páramo (Léon, Spain; coordinates: 42°24′58″ N, 5°47′50″ W WGS84) (Figure 1). The soil had a sandy clay loam texture, with an electrical conductivity of 0.084 dS/m and an organic matter content of 1.94%. The previous crop was corn.

Corn requires nitrogen availability throughout its growth cycle, although approximately one-third of its total nitrogen uptake occurs after pollination. Traditional corn fertilization involves two main nitrogen applications: the first before seeding or during the early growth stages, and the second when there are 6–7 leaves and the leaf collars are visible. In this experimental plot, three different N rates were used: 0, 320, and 382 kg N/ha. The fertilizer used contained nitrification inhibitors such as 3,4-dimethylpyrazole phosphate—DMPP (ENTEC^®: EuroChem Agro GmbH, Mannheim, Germany)—and 3,4-dimethylpyrazole succinic—DMPSA (ENTEC^® EVO™: EuroChem Agro GmbH, Mannheim, Germany). The nitrogen was applied at three different time points (pre-seeding, V2, and V6). In each treatment, a specific amount of nitrogen was applied on different application dates (

Table 1. N rate (kg N/ha), growth stages, and dates of application.

Treatment	Before Seeding * (8 May 2024)	V2—Second Leaf * (12 June 2024)	V6—Sixth Leaf * (3 July 2024)	N Rate (kg N/ha)
T0	0	0	0	0
T1	91	0	229	320
T2	117	0	203	320
T3	140	0	180	320
T4	0	320	0	320
T5	220	0	162	382

* Growth stages following Ritchie et al. [27].

).

The corn (var. DEKALB 5362) was planted on 24 May at a density of 99,000 seeds/ha, in rows spaced 0.65 m apart, and emerged on 31 May 2024. The experiment had a randomized complete block design with four replications. The nitrogen rates, application dates, and corn growth stages for each treatment are shown in Table 1. The main plots had an area of 32 m² and included 6 crop rows of 8 m. All the data were collected from the two central rows to avoid interference between the treatments.

These fertilizer doses and application dates are commonly used by local farmers and follow the recommendations set out in the regional regulations (DECREE 5/2020 of June 25, which designates zones vulnerable to water contamination by nitrates from agricultural and livestock sources and outlines the Code of Good Agricultural Practices; document available at https://bocyl.jcyl.es/boletines/2020/06/30/pdf/BOCYL-D-30062020-1.pdf; accessed on 27 March 2025). This experimental design made it possible to evaluate the crop growth under each treatment, as well as the nitrogen use efficiency.

2.2. Data Collection

Reflectance data for the samples were collected using a FieldSpec 4 (ASD-FS4) portable spectroradiometer (Analytical Spectral Devices, Inc., Boulder, CO, USA), which captures spectral data at wavelengths between 350 and 2500 nm. This device has three detectors: a visible near-infrared (VNIR) silicon photodiode array (350–1000 nm) and two graded Indium Gallium Arsenide photodiode detectors for short-wavelength infrared (SWIR) measurement (1000–1800 nm and 1800–2500 nm).

A plant probe was used to measure the leaf reflectance. This handheld device consisted of a quartz–halogen bulb, a grip to attach the fiber optic cable input to the spectroradiometer [28], and a quartz window that was pressed against the leaf surface. This accessory, used for the contact measurement of the leaves, minimized the errors associated with stray light. The probe had a spot size of 10 mm and plugged directly into the field spectroradiometer. A leaf clip was also attached to the plant probe to facilitate spectral data collection without damaging the leaf, thereby ensuring the non-destructive contact measurement of living leaves.

Spectral reflectance data were collected during both the vegetative and reproductive stages. The first three measurements were carried out on leaves in the field during the V4 (fourth leaf (June 24)), V8 (eighth leaf (July 22)), and R1 (silking (August 28)) stages. The fourth round of measurements was conducted in the laboratory on samples collected on October 18 (R6—maturity stage). These samples were dried at 60 °C until reaching a constant weight before being measured on November 11. This last set of spectral data was complemented with additional measurements of the dried stem samples.

In the field, ten plants per plot were measured, with mature (most developed) and healthy leaves selected. The spectroradiometer was warmed up for fifty minutes and set up (optimization and calibration) using a white reference panel, following the recommendations of ASD Inc. [29]. It was recalibrated before measuring the first leaf of each plot. The spectral reflectance of the leaf’s upper face was measured three times, avoiding veins and leafspots, and the mean value was recorded. Consequently, a total of 240 fresh leaves were measured during each of the three field campaigns, yielding a total of 720 samples. To measure the samples in the laboratory, the same procedure was followed as in the field. However, only 3 stem samples were taken from each plot, resulting in the measurement of 72 dried stems and 240 dried leaves in total.

At the end of the crop cycle, the plots were harvested. Plants from the two central rows were collected to determine the number and weight of their ears, as well as the weight and moisture content of their grain. Additionally, the nitrate content in the stalks (“end-of-season cornstalk test”) was determined following the protocol defined by [30,31]. To verify whether there were statistically significant differences between the treatments, a non-parametric Mann–Whitney U test [32] was used.

2.3. Spectral Analysis

All the spectral signatures were plotted and reviewed to detect possible failures in data capture. The spectral range of 350–400 nm exhibited a higher noise-to-signal ratio and was excluded from the spectral analysis [26]. Before statistical analysis, the raw reflectances (R) were transformed into the absorbance (A = log(1/R)). In this way, the raw reflectance and absorbance results could be compared. The log(1/R) transformation is a valuable preprocessing step in field spectroscopy, as it improves the spectral clarity, enhances weak absorption features, and enables better material identification and quantification [33]. This technique is widely used in applications such as soil analysis, mineral mapping, and vegetation assessment [34].

In this study, the spectral signatures were analyzed using two approaches: (1) Principal Component Analysis (PCA) to identify the differences between the treatments and (2) Partial Least Squares Regression (PLSR) to predict the nitrate content (NO₃-N) in the cornstalks.

PCA is a modeling method that provides an interpretable overview of the main information contained in a spectral signature dataset. It projects the original data onto a smaller number of variables called principal components (PCs). Because each PC explains a certain amount of the total information contained in the original spectrum, plotting PCs can reveal important interrelationships between the reflectance values or samples. This, in turn, enables the interpretation of the sample groupings and the similarities and differences between the treatments. Additionally, PCA loading plots were analyzed to select the most informative wavelengths, since the highest positive and lowest negative peaks indicate wavelengths containing the most relevant information on the differences in the N content [19].

PLSR was used to predict the NO₃-N contents in corn stems. This regression model was effective for predicting collinear variables, such as spectral signatures, by identifying the latent variables (or factors) in the reflectance that best predicted the NO₃-N content. The performance of the prediction models was evaluated using the coefficient of determination (R²) and the RMSECV, obtained through leave-one-out cross-validation (CV).

3. Results

3.1. Spectral Signatures

Figure 2 shows the representative spectral signatures of the reflectance (R) (Figure 2a–d) and absorbance (calculated as log(1/R)) (Figure 2e–h), obtained for each treatment and replicate across the four measurement dates. Each line represents the average values for each replicate, and each treatment is shown in a different color (see the legend in Figure 2). A comparative analysis of the spectral curves across the four dates reveals that the magnitude and consistency of the differences between the treatments progressively increased from June to October. The greatest differences between the treatments were observed in the visible spectrum (VIS: 400–700 nm), particularly around the green peak (550 nm) and the blue (450 nm) and red (670 nm) absorption bands. Other minor differences were observed across wavelength ranges, such as the red edge (RE: 700–750 nm) and near-infrared regions (NIR: 750–1300 nm), and the water absorption bands (around 1450 nm and 1950 nm) were identified in the short-wave infrared region (SWIR: 1300–2400 nm). These ranges could be indicative of variations in the chlorophyll content (VIS), photosynthetic efficiency (RE), internal leaf structure or biomass (NIR), and water content (SWIR).

As expected, during the early stages of growth, the differences between the spectral signatures of the control (T0) and treated plants (T1 to T5) were minimal (Figure 2a,e). The maize plants were in their initial phase of development, primarily focused on root establishment and early leaf expansion. The applied treatments may not have had sufficient time to induce significant physiological or structural changes that would manifest strongly in their spectral signatures. Consequently, the spectral curves for all the treatments were expected to appear relatively similar, exhibiting the typical patterns of young, healthy vegetation with high chlorophyll absorption. However, as the crop developed, the differences between the treatments became evident, especially in the 400–750 nm range. On 22 July (Figure 2b,f) and 28 August (Figure 2b,f), the plants were actively growing, and the nitrogen treatments (T1 to T5) had more time to influence the physiological processes affecting their spectral signatures. The treatment groups showed slight deviations from the controls (T0), particularly in the VIS region, indicating pigment changes or subtle variations in the RE slope. The last spectral measurements (on 10 October) showed clear signs of senescence across all the treatments: increased reflectance in the red region due to chlorophyll degradation and a significant decrease in the NIR reflectance due to a loss of cell structure, dehydration, and reduced vigor (Figure 2d,h). Clear differences between the control (T0) and treatment groups (T1 to T5) remained, mainly in the VIS and RE ranges (400–750 nm).

Figure 3a–d show scatter plots of the PC1 and PC2 scores from a PCA performed on the raw reflectance data. PC1 and PC2 explained more than 82% of the variation in the reflectances. The closer the samples were to each other in the score plot, the more similar their PC1 and PC2 scores were. The effects of nitrogen fertilization are evident in the scatter plots. On the first measurement date (Figure 3a,e), before applying the topdressing fertilization, no distinct groups were observed for any treatment or the control. On the second measurement date (Figure 3b,f), the control (T0) began to differentiate, and these differences became more pronounced on the subsequent measurement dates (Figure 3c,g). The PCA of the dried leaf reflectances (Figure 3d,h) clearly distinguished T0 from the other treatments (T1 to T5). Converting the reflectance into the absorbance (using log(1/R)) allowed for earlier cluster identification in the scatter plots (Figure 3e–h). In fact, the control group was already clearly separate from the rest by the second spectral sampling date (Figure 3f). These differences were maintained in the scores calculated using the absorbance data for the two following dates (Figure 3g,h).

Figure 4 shows the loadings of the PCs, indicating which wavelengths contributed the most to the variance explained by each PCA component. The importance of the reflectance at different wavelengths varied across the dates (Figure 4a–h). In general, significant wavelengths were distributed throughout the entire spectral range, especially in the analysis of the raw reflectance (Figure 4a–d). However, the 400–750 nm range contained a particularly notable source of information, which was the most evident in the data from the dried leaves (Figure 4d,h).

The loadings for PC1 based on the raw reflectance were consistently positive, while those for PC2 exhibited both positive and negative values. On the first measurement date, the important wavelengths for PC1 were distributed across the entire spectrum, specifically at around 760–1100, 1670, and 2205 nm. The most prominent peaks in PC2 were centered at 850, 1075, 1265, 1395, 1540, 1860, and 2225 nm (Figure 4a). However, from the second measurement date onward, the significant loadings became more concentrated, forming distinct peaks within narrower wavelength ranges. On dates 2 and 3 (Figure 4b,c), the most important wavelengths for PC1 were located at 555 and 720 nm. On the same dates, the strongest positively signed loadings for PC2 occurred at 555 and 715 nm, while the most prominent negatively signed loadings were observed at 910, 1070, and 1265 nm. These changes suggest that the wavelengths mentioned earlier were the most influenced by nitrogen fertilization. When analyzing the results from the dried leaves, the significant loadings were concentrated in fewer wavelengths: 685 nm for PC1 and 595, 635, and 710 nm for PC2 (Figure 4d). This indicates that the 555–720 nm range is the most related to the nitrogen availability, while reflectance values between 910 and 1900 nm are primarily associated with the plant water content [35].

The overall effect of the log(1/R) transformation was a notable reduction in the number of peaks in the loadings (Figure 4e–h) compared to the PCA results based on the raw reflectance (Figure 4a–d). This reduction was particularly evident in the measurements taken after the application of topdressing fertilization (Figure 4f–g). Additionally, the transformation also changed the sign of the PC2 values. Therefore, the log(1/R) transformation contributed to a reduction in the number of predictive variables and facilitated the development of simpler, more interpretable models.

3.2. Effects of Nitrogen Treatments on Corn Yield Variables

Table 2. Descriptive statistics of the variables measured at corn harvest.

Statistics	Cornstalk NO3-N Content (ppm)	Grain Yield (kg/ha)	Grain Moisture (%)	Corncobs (#N)	Weight of Corncob (g)
Mean	6446.08	16,754.90	20.85	70.08	30.51
Median	5860.50	18,139.42	20.90	72.00	31.16
SD	5194.98	4346.14	0.67	8.99	2.29
Minimum	67.00	6855.35	19.70	51	25.48
Maximum	18,702.00	22,526.21	22.10	83	33.54
Range	18,635.00	15,670.86	2.40	32	8.06
IQR	7125.25	2693.92	1.15	10.25	1.24

presents the descriptive statistics of five variables measured at corn harvest. The cornstalk NO₃-N content exhibited high variability, with a wide range (67.00 to 18,702.00 ppm) and a large standard deviation (5194.98 ppm), suggesting substantial differences in the nitrate accumulation between the treatments. The grain yield also showed considerable variation (range: 15,670.86 kg/ha; SD: 4346.14 kg/ha), with a mean yield of 16,754.90 kg/ha. In contrast, the grain moisture was almost uniform across the samples (mean: 20.85%; SD: 0.67%), showing minimal variation. The number of corncobs per plot (mean: 70.08) showed moderate variability (SD: 8.99), while the weight of the individual corncobs was the most consistent variable measured (mean: 30.51 g; SD: 2.29 g).

The effects of the treatments compared to the control were evident in both the nitrate accumulation in the stalk and the grain yield (Figure 5). The boxplots of the cornstalk NO₃-N levels (Figure 5: red boxes, left y-axis) show clear variation across the treatments. The control (T0) had the lowest NO₃-N concentration, indicating minimal nitrate accumulation in the stalks. The treatment groups showed moderate levels, with a gradual increase from T1 to T3 and a slight drop at T4. Treatment 5 exhibited the highest variability and median NO₃-N concentration, suggesting possible over-fertilization or inefficient nitrogen uptake. The grain yield results (Figure 5: orange boxes, right y-axis) were more consistent across the treatments, except for in the control group (T0), which showed significantly lower yields. All the treatment groups (T1 to T5) exhibited high and stable grain yields, with T3 and T4 showing slightly higher medians. The low yield observed in the control group corresponded to its low NO₃-N levels, indicating a clear nitrogen deficiency.

The results of the Mann–Whitney U test indicated that statistically significant differences (p < 0.05) were primarily observed when comparing the control (T0) with the other treatments (

Table 3. Comparison of variables between treatments. p values from Mann–Whitney U test (if p < 0.05, differences between treatments were statistically significant, and they are highlighted in bold).

Case	Cornstalk NO3-N Content (ppm)	Grain Yield (kg/ha)	Grain Moisture (%)	Corncobs (#N)	Weight of Corncob (g)
T0 vs. T1	0.021	0.021	0.248	0.042	0.021
T0 vs. T2	0.021	0.021	0.309	0.019	0.021
T0 vs. T3	0.021	0.021	0.884	0.020	0.021
T0 vs. T4	0.083	0.021	0.468	0.020	0.021
T0 vs. T5	0.043	0.021	0.384	0.020	0.021
T1 vs. T2	0.564	0.468	0.773	0.655	0.885
T1 vs. T3	0.564	0.083	0.146	0.059	0.773
T1 vs. T4	0.773	0.021	0.564	0.309	0.191
T1 vs. T5	0.564	0.386	0.384	0.149	0.663
T2 vs. T3	0.564	0.248	0.465	0.076	0.564
T2 vs. T4	0.564	0.149	1.000	0.375	0.110
T2 vs. T5	0.564	1.000	0.561	0.146	0.559
T3 vs. T4	0.564	0.386	0.655	0.663	0.149
T3 vs. T5	0.564	0.386	0.559	0.468	0.773
T4 vs. T5	0.564	0.083	1.000	0.564	0.059

). Specifically, T0 showed significant differences from T1, T2, T3, T4, and T5 in terms of the cornstalk NO₃-N content, grain yield, number of corncobs, and corncob weight. No significant differences were found in the grain moisture between T0 and any other treatment. In contrast, the comparisons between treatments T1 to T5 generally showed no statistically significant differences across most variables, suggesting that these fertilization treatments had similar effects on the measured parameters. The only exception was the grain yield, where T1 differed significantly from T4 (p = 0.021). Overall, the data suggest that nitrogen fertilization had a marked impact when compared to the control (T0), but the differences between the fertilized groups (T1–T5) were minimal and not statistically significant.

3.3. Estimation of Cornstalk Nitrate Content Using Spectroscopy

Figure 6 presents scatter plots comparing the measured NO₃-N content and that predicted by PLSR models, using the raw reflectance and absorbance as predictor variables. Both models successfully predicted the nitrate content in maize stalks. The model using raw reflectance values (Figure 6, blue squares) considered 10 factors and yielded a coefficient of determination (R²) of 0.752 and an RMSECV of 2610.1 ppm. While this model demonstrated a moderate correlation between the predicted and observed NO₃-N values, the predictions deviated substantially from the ideal 1:1 line, especially at higher concentrations. This level of predictive accuracy suggests that the raw reflectance data captured some, but not all, of the spectral variation associated with the NO₃-N content in cornstalks, possibly due to noise or overlapping spectral features that limited the sensitivity. In contrast, the model based on the absorbance data (Figure 6, green triangles) showed a stronger correlation, with an R² of 0.845 and a lower RMSECV of 2035.3 ppm, but was more complex (15 factors). This improved performance indicates that transforming reflectance data into absorbance better linearizes the relationship between the spectral response and NO₃-N concentration. The use of logarithmic transformation likely enhanced the detection of subtle spectral differences related to the nitrate content, making the model more robust and accurate, particularly across a wider concentration range.

The prediction models based on the reflectance were influenced by data preprocessing and the number of factors included in the model. Figure 7 shows the performance (in terms of the R² and RMSECV) of the PLSR models when the number of factors (F on the x-axis) considered varied. The RMSECV values for both models exhibited distinct behaviors as the number of PLSR factors increased. For the model based on the raw reflectance (Figure 7, solid blue line), the RMSECV started relatively high and decreased moderately until at eleven factors (F-11), where it reached its minimum before increasing again with the addition of more factors. In contrast, the absorbance-based model using the log(1/R) (solid green line) showed a sharper decrease in the RMSECV, reaching its lowest value at F-5, indicating better predictive performance with fewer latent variables. Beyond F-5, the RMSECV of the absorbance-based model increased more sharply than in the raw reflectance model, indicating that the inclusion of additional factors led to model overfitting and a decline in the prediction accuracy.

The R² values (dashed lines in Figure 7), which indicate the proportion of the variance explained by the model, showed complementary trends. For the raw reflectance model (dashed blue line), the R² increased steadily up to F-11, peaking at around 0.75, before declining slightly. In contrast, the absorbance model (dashed green line) demonstrated a more rapid increase in the R², surpassing 0.85 by F-14. These results highlight that the log(1/R) transformation captured more of the variance in the NO₃-N data with fewer factors, resulting in a more efficient and powerful model.

4. Discussion

4.1. Effect of Nitrogen Fertilization on the Spectral Signature

The maize leaves’ spectral signature changed throughout the vegetative growth cycle, and these changes differed markedly between the fertilized and control plants (Figure 2); therefore, some of these changes are related to the nitrogen content in the plants. The nitrogen content in maize plants increases progressively from the beginning of the vegetative cycle until shortly after the start of the flowering stage (R1), reaching a peak approximately three weeks after R1, and then decreases as the plant approaches physiological maturity due to the remobilization of nitrogen to the developing grain [36]. However, the foliar nitrogen content (%) decreases over time, from high levels at the V3/V4 stages to moderate levels at V6 [2,37]. This reduction is attributed to the dilution of nitrogen resulting from the accelerated growth of biomass, which is typical in rapidly developing crops.

The differences observed between spectral reflectance curves, particularly at around 555 nm within the VIS spectral range, are closely associated with variations in the pigment concentration and biochemical leaf properties [38,39]. In contrast, the NIR and SWIR regions predominantly capture structural attributes such as the leaf thickness, dry matter accumulation, water status, and the content of other biochemicals [35]. Consequently, the latter regions exhibit limited responsiveness to nitrogen variation.

In the visible range (400–700 nm), the spectral signature of the control (T0) showed clear deviations from that of the fertilized plants, consistent with the reduced chlorophyll content typically associated with nitrogen deficiency. Moreover, the control exhibited lower absorption in the blue (475 nm) and red (670 nm) regions, which is indicative of a reduced chlorophyll content, as nitrogen deficiency directly impairs chlorophyll biosynthesis.

Due to its high sensitivity to nitrogen-driven pigment dynamics, light in the visible spectrum could be used to estimate the maize leaf nitrogen concentration employing reflectance-based methods, consistent with the findings previously published for maize [2,19]. The first measurements were made of corn plants in the V4 stage (June 24), before the application of topdressing. For this reason, the spectral signatures were similar, and no differences were seen between the treatments (Figure 2a,e). The other two field measurements were carried out after topdressing and showed that the reflectance values are sensitive to differences in nitrogen fertilization (Figure 2b,c,f,g). The highest reflectance values were obtained in the control plots (T0) without nitrogen fertilization, showing that a nitrogen deficit makes it difficult to create chlorophyll in plants. The main sensitive range was a broad wavelength region from 530 to 750 nm, with a peak centered at 555 nm. The differences between the treatments were also very evident in this range in the laboratory measurements of the dried leaves (Figure 2d,h), where the peaks were centered on 590 and 635 nm. The broad nitrogen-sensitive region centered near 555 nm, which overlapped with the sensitivity observed at around 650 nm, suggests that a high spectral resolution is less critical in this range. Conversely, the sharp reflectance peak near 710 nm necessitates greater precision in wavelength selection due to its narrow spectral response. The 650 nm wavelength coincides with the value at which chlorophyll strongly absorbs light. Chlorophyll also absorbs light strongly near 450 nm, but this area of the spectrum does not appear to be a good N status indicator. These findings align with the observations of Blackmer et al. [40], who identified spectral regions near 550, 650, and 710 nm to be the most effective for detecting nitrogen deficiency.

In the present work, absorption features within the 800–2500 nm spectral range were detected near 1450, 1950, and 2200 nm. However, no consistent associations with nitrogen deficiency were identified. These findings are in line with those of Xie et al. [19], who reported that only a few wavelengths—specifically, 1042, 1124, and 1125 nm—showed responsiveness to nitrogen fertilization in field-grown maize, while no significant spectral response to nitrogen was observed within the 2000–2400 nm region. Similarly, Olivera et al. [2] demonstrated that the near-infrared (NIR) and shortwave infrared (SWIR) regions exhibit lower sensitivity to nitrogen levels when compared to the visible and red edge spectral ranges. This reduced sensitivity is primarily attributed to the reflectance in the 800–2500 nm region being predominantly governed by general biochemical components, rather than the nitrogen content specifically [39].

The effect of the log(1/R) transformation was the enhancement of weak absorption features. However, this transformation requires the preprocessing of raw data, and its effectiveness may vary depending on the dataset’s characteristics. In fact, Martin et al. [23] suggest not using any transformation if the goal is to generate models that are operational on a global scale.

The most notable differences were observed on 28 August (Figure 2c,g) and on the last date the leaves were measured after being oven-dried (Figure 2d,h). The same general trend was observed in the log(1/R) signatures (Figure 2e–h), although the differences in the 400–670 nm range were enhanced, allowing for the better visualization of the differences between the control and treatment groups. Determining the optimal timing for detecting nitrogen deficiency in maize plants is a complex task, as it is influenced by both inherent crop-specific physiological factors and the measurement technique employed. For instance, Dong et al. [41] estimated the maize nitrogen concentration using a leaf fluorescence sensor across four growth stages. They concluded, as also observed in our study, that the model performance is strongly influenced by the phenological stage and recommended the development of models that incorporate data from all growth stages. Recent research [42] explored the effectiveness of multispectral imagery for monitoring the nitrogen content in corn crops. The findings highlight the importance of specific spectral indices at key phenological stages for reliably detecting nitrogen stress, but they do not identify the best phenological state. Therefore, no single vegetative stage can be identified as optimal for estimating nitrogen deficiencies. While specific stages may be optimal for making individual observations, combining multitemporal data from several phenological stages—including both the vegetative and reproductive stages—can remarkably enhance the accuracy of yield prediction and, by extension, nitrogen status assessment [41]. Furthermore, vegetation indices utilizing the red edge and near-infrared regions generally show strong correlations with the nitrogen content in maize due to their sensitivity to the chlorophyll content and canopy structure [2]. Considering the phenological stages analyzed (V4, V8, R1, and R6) and the preprocessing of the predictive variables (i.e., the raw reflectance (R) and its log(1/R) transformation), the most effective method of nitrogen deficit detection was based on reflectance values centered at 560 and 705 nm. The log(1/R) transformation enhanced the absolute magnitude of the reflectance values and reduced the number of local maxima, thus facilitating the detection of nitrogen stress. However, this transformation requires additional computational processing time and alters the original spectral peak positions.

To maximize the input use efficiency in maize cultivation, the early detection of stress-affected areas is of paramount importance. This is particularly critical in relation to nitrogen fertilization, given its decisive impact on crop productivity and environmental implications. Within this framework, the spectral information analysis methodology proposed in this study demonstrated significant utility. Variations in both the magnitude and wavelength position of the reflectance peaks were consistently observed throughout the crop’s phenological development, with discernible differences between the maize plants subjected to different nitrogen fertilization regimes and the unfertilized controls. Consequently, the reflectance values—especially within the 555–720 nm spectral range—exhibit strong potential for use as diagnostic indicators of nitrogen deficiency in commercial maize fields, as well as for the construction of spectral indices derived from these measurements. In fact, some authors have used the reflectance at these wavelengths to develop indices indicative of the nitrogen fertilization levels in maize, achieving coefficients of determination (R²) greater than 0.85 [43].

The spectral analysis (Figure 2, Figure 3 and Figure 4) revealed that maize plants grown without nitrogen fertilization can be effectively distinguished from fertilized plants, especially within the VIS and red edge regions. The control treatment consistently showed a higher reflectance in the blue, green, and red bands and a less pronounced red edge shift, both indicative of a lower chlorophyll content and diminished plant vigor. These findings support the utility of hyperspectral sensing for non-destructive nitrogen stress detection in precision agriculture [43].

4.2. Performance of Models Regarding Estimation of Cornstalk Nitrate Content Based on Field Spectroscopy

In general, all the studied treatments had a similar effect on the corn plants (Figure 5). In fact, there were only statistically significant differences between the control and the other treatments in the cornstalk NO₃-N content, corncob variables (number and individual weight), and grain yield (Table 3). The grain yields were similar across the five nitrogen fertilization treatments, although the early topdressing treatment (T4) appeared to be the most efficient option (Figure 5).

Regarding cornstalk nitrate content estimation, the PLSR model based on the absorbance outperformed the raw reflectance model in terms of both the R² and RMSECV (R² of 0.845; RMSECV of 2035.3 ppm), indicating that it is more suitable for predicting NO₃-N concentrations in cornstalks (Figure 6). However, both models exhibited some negative predicted values, which are physiologically implausible. These negative values may have resulted from model overfitting, limitations in spectral preprocessing, or extrapolation beyond the calibration range, particularly for low observed concentrations near zero. Such results highlight the importance of applying appropriate data transformations, constraints, or corrections when developing regression models for spectral data analysis to avoid non-physical predictions. These results corroborate the observations of Rawal et al. [34], who developed models with an R² > 0.80 when estimating potato leaves’ nitrogen content using the absorbance and PLSR.

Increasing the number of factors increased the coefficient of determination and decreased the RMSECV (Figure 7). Therefore, the increased model complexity initially led to improved estimates. However, while the five-factor model performed worse than the four-factor model, from six factors onward, the performance improved again as the number of factors increased. It is possible that these models were affected by overfitting and that the precise results were based on spurious correlations found in the spectra and not on the effects of the nitrate content of each sample. This general trend affected both the reflectance and absorbance models, although the relationship between the model complexity and efficiency was greater for absorbance-based models. Therefore, the estimation of the stem NO₃-N content is heavily “data-driven”: while increasing the number of factors in PLSR can improve the model fit, it is important to monitor performance metrics like the RMSECV to prevent overfitting and ensure the model remains robust and predictive [44].

One noteworthy aspect is that all the groups with nitrogen fertilization exhibited a very high cornstalk NO₃-N value, indicating excessive fertilization [45]. In fact, at stalk NO₃-N concentrations of 2000 ppm, the yields did not respond to further increases in N fertilizer, and the efficiency of the total fertilizer application declined [46] according to Brouder et al., 2000. Some studies highlight the need for more data on how best management practices for nitrogen fertilization—such as split nitrogen (N) application, the fertilizer type, irrigation, and their interactions—affect the gaseous losses and nitrogen use efficiency across different agricultural systems [47]. In this context, the present study has demonstrated that field spectroscopy can provide essential information for developing tools to assess fertilization’s environmental costs and management strategies that optimize the crop yield while minimizing the environmental impact of excessive nitrogen fertilization [48,49].

Overall, the absorbance-based (log(1/R)) model outperformed the raw reflectance model in terms of both the RMSECV and R², particularly at F-14, where the balance between the predictive error and explained variance was optimal (Figure 7). However, it is important to note that at F-5, both the models showed a drop in their R² and a spike in their RMSECV, indicative of instability and possibly poor factor selection at this stage. This behavior may reflect the introduction of collinearity or noise by including an inappropriate latent variable, emphasizing the importance of careful factor selection to avoid degrading the model's performance. Thus, the log(1/R) model with fourteen factors represented the most reliable and accurate model configuration for predicting the cornstalk NO₃-N content. Given the growing global emphasis on sustainable agricultural practices, the present findings provide a valuable foundation for future research on developing in situ nitrate content estimation methods based on the spectral measurement of live plant tissues.

The capability to not only detect but also accurately quantify nitrogen deficiency would represent a significant advancement in the management of maize nutrition. However, such quantification remains a non-trivial task due to its strong dependence on the specific reflectance dataset and the environmental and measurement conditions under which the data are acquired, as evidenced by findings from previous research and corroborated in the present study. At this stage, the remote diagnosis of nitrogen stress—particularly the development and deployment of rapid, in-field sensing technologies readily available to end users—continues to pose a substantial challenge and constitutes a critical area for further investigation.

5. Conclusions

Monitoring nitrogen fertilization using field spectroscopy presents several advantages over other techniques, including non-destructive measurement, rapid and high-throughput analysis, cost-effectiveness, environmental sustainability, a high spatial and temporal resolution, the potential for integration with remote sensing, and the capacity for multi-parameter analysis.

The multitemporal analysis of reflectance data enabled us to identify the wavelength regions containing the most informative signals at different phenological stages, as well as their relationship with the nitrogen availability. In general, during the early stages of crop development, the most informative reflectance peaks were centered at around 555 and 720 nm. In contrast, during the grain-filling phase, the highest variability in the reflectance was observed at around 680 nm. These findings suggest that changes in the reflectance are strongly related to the chlorophyll content, which is a key indicator for identifying potential nitrogen deficiencies.

Importantly, this study demonstrated the feasibility of accurately estimating the nitrate concentrations in maize stems using reflectance data obtained under controlled laboratory conditions from dried stem samples. Partial Least Squares Regression (PLSR) models developed for this purpose yielded high correlation coefficients (R² > 0.75), although the complexity of the models suggests potential overfitting. While the results are promising, they are currently limited to a single growing season. To enhance the robustness and applicability of this approach, further validation is required across multiple seasons and a range of agronomic contexts.

These advantages highlight the potential of field spectroscopy as a powerful tool for precision agriculture, enabling optimized nitrogen management while minimizing both the costs and environmental impact. However, it is important to note that the model performance is highly dependent on the specific dataset used, making validation under different conditions essential.