Introduction

Nitrogen (N) is a vital macronutrient crucial for plant growth, development, and overall productivity. It is an essential constituent of chlorophyll, amino acids, proteins, and nucleic acids, playing a pivotal role in photosynthesis and metabolic processes in crops1,2,3. Among staple food crops, rice (Oryza sativa L.) is particularly dependent on adequate nitrogen availability, as its yield and grain quality are closely linked to the nitrogen content of its leaves2,3,4. Assessing Leaf Nitrogen Content (LNC) provides valuable insight into the nutritional status of rice plants, allowing for precise nitrogen management to optimize fertilization strategies, improve yield, and enhance sustainability in agricultural systems5,6,7,8,9. Given the global emphasis on food security and environmental conservation, efficient nitrogen management in rice production is critical for balancing high productivity with reduced ecological impact2,9.

The accurate assessment of LNC is a key component of precision fertilization strategies, which aim to enhance nitrogen use efficiency (NUE) while minimizing environmental pollution. Excessive nitrogen fertilization has been linked to severe ecological consequences, including nitrate leaching, soil degradation, and greenhouse gas emissions2,4,5. Conversely, nitrogen deficiency can lead to yield losses and reduced grain quality, affecting food security and economic returns for farmers. Implementing site-specific nitrogen management based on real-time LNC estimation can help optimize fertilizer application, reducing waste and mitigating environmental risks2,5,9.

Traditionally, LNC has been estimated through destructive chemical laboratory methods such as Kjeldahl digestion, Dumas combustion, and spectrophotometric techniques considered the gold standard due to their precision and well-established protocols10. However, while these approaches provide accurate quantification, they require extensive sample preparation, hazardous reagents, and specialized equipment, making them time-consuming, costly, and impractical for large-scale field applications11,12. The need for a rapid, non-destructive, and cost-effective alternative has driven the development of spectroscopic techniques for real-time nutrient assessment in crops. Among these, the green technology, Near-Infrared Spectroscopy (NIRS) has become a powerful analytical technique, enabling the detection of complex biochemical information by analyzing light interactions with plant tissues11,12.

NIRS operates in the 750–2500 nm range of the electromagnetic spectrum, detecting molecular overtones and combination vibrations associated with chemical bonds, particularly those involving O–H, C–H, and N–H functional groups11,13,14. The technique allows for the indirect estimation of biochemical constituents such as proteins, carbohydrates, moisture, and pigments in plant tissues14. When combined with chemometric modeling, NIRS provides a robust framework for predicting LNC, enabling rapid, high-throughput assessment without the need for destructive sampling or chemical processing12,15. In the context of NIRS, chemometrics involves the use of multivariate statistical techniques to analyze spectral data and uncover relationships between spectral features and chemical properties. By reducing data dimensionality and handling collinearity, chemometric approaches such as Principal Component Analysis (PCA), Partial Least Squares Regression (PLSR), and Linear Discriminant Analysis (LDA) enhance the interpretability and predictive power of NIR-based models for estimating traits like leaf nitrogen content11,16,17.

The application of NIRS for plant analysis has been extensively studied, demonstrating strong predictive capabilities for various biochemical parameters. Several studies have reported the successful use of NIRS to estimate nitrogen content in cereal crops, including wheat, maize, and rice, showcasing its potential for non-destructive nutrient assessment11,12,15,18,19. Despite the growing use of NIRS in crops, most previous studies have focused on single rice genotypes or broad nitrogen gradients, often overlooking key sources of variability such as genotype-specific spectral responses and temporal shifts in leaf composition15. Additionally, limited attention has been given to the comparative performance of distinct spectral regions or the use of classification-based chemometric models for nitrogen level discrimination15,18,20. To address these gaps, this study evaluates the effectiveness of PCA-LDA models for classifying nitrogen treatments (0%, 25%, 50%, 75%, and 100%) in five diverse rice genotypes (Nerica, Rufipogon, IR64, Ciherang, and Curinga), using fresh-leaf spectra collected across three growth stages and multiple wavelength regions. By systematically comparing spectral performance across genotypes, nitrogen levels, and timepoints, this work advances the precision and generalizability of NIRS-based nitrogen estimation. The study also identifies the most informative spectral subregion (1700–2200 nm) for classification accuracy, offering practical insights for future model development.

Implementing NIRS-based LNC estimation in rice production systems has the potential to improve nitrogen management, optimize fertilizer use, and reduce environmental impacts, aligning with the broader objectives of sustainable agriculture and resource conservation2,3,4,6,9,21. The results of this research provide a foundation for further exploration of NIRS applications in crop monitoring, paving the way for scalable and data-driven solutions in modern agronomy17,22,23. Future research should focus on expanding the applicability of NIRS to other crops and environmental conditions, enhancing predictive models, and incorporating advanced technologies, including artificial intelligence (AI) and machine learning (ML), to improve spectral analysis11,12,24. The combination of spectroscopy, chemometrics, and agronomic expertise represents a transformative approach to nutrient management, fostering innovation in sustainable agriculture2,22.

Materials and methods

Trial design and crop establishment

The experiment was conducted at the Centro Internacional de Agricultura Tropical (CIAT) research station in Palmira, Valle del Cauca, Colombia, at an altitude of 970.67 m.a.s.l (3°30′29.21″N, 76°20′53.98″W). The experiment was designed using a complete block randomization approach (RCBD) incorporating five rice genotypes, four replicates, and five nitrogen treatments under the N omission platform25. Each plot measured 1.5 m in width × 2.75 m in length, arranged in a grid of 12 rows and eight columns, with a row-to-row spacing (R–R) of 22.15 cm. Genotypes were planted in single rows, with eight plants per row, and 12 rows per plot (Fig. 1). The trial was sown in May 2024. The selected genotypes, Nerica, Rufipogon, IR64, Ciherang, and Curinga, were chosen based on their contrasting morphological characteristics and tillering capacity21,24,26,27,28. Seedlings were initially grown in trays and subsequently transplanted into flooded soil under continuous irrigation to mimic typical lowland rice production systems.

Fig. 1
figure 1

Experimental design and field layout for nitrogen treatment classification. a Schematic representation of the experimental field showing the distribution of five rice genotypes (G1: Nerica, G2: Rufipogon, G3: IR64, G4: Ciherang, G5: Curinga) under five nitrogen levels (0%, 25%, 50%, 75%, 100%), with four replicates (R1–R4) per treatment. Each cell indicates the genotype (GT), plot number (Plot), and replicate (Rep). b Photographs of the experimental plots at the CIAT-Palmira research station during field data collection. The layout facilitated consistent sampling for NIRS and ensured genotype-level resolution for spectral modeling.

Full size image

Nitrogen treatments were established based on a standard commercial fertilization rate of 180 kg N ha⁻¹, considered 100% nitrogen application. While this value is slightly higher than typical recommendations for rice (120–160 kg N ha⁻¹), it reflects local agronomic conditions at the CIAT field site, where high soil alkalinity and calcium-induced antagonism can reduce nitrogen uptake. Four additional treatments were implemented by reducing nitrogen application by 25% increments, resulting in final levels of 100%, 75%, 50%, 25%, and 0% N. All nitrogen treatments were applied within a single field block to minimize variability in soil texture, nutrient content, and mechanical management practices, ensuring uniform growing conditions across the entire experiment. To establish the 0% nitrogen treatment, the field was previously subjected to three successive maize cropping cycles without nitrogen fertilization, ensuring severe nitrogen depletion in the soil. Additionally, the 0% N plots were positioned at the highest elevation within the field, taking advantage of the natural topography to prevent lateral nitrogen movement from higher-dose treatments. This design minimized the risk of leaching or contamination via surface water, especially under puddled field conditions.

A pre-defined nutrient management protocol was followed, applying NPK and micronutrients at specific growth stages: at transplanting, 8 days after transplanting (DAT), 20 DAT, and 40 DAT. To ensure adequate nutrient availability, the fertilization scheme incorporated two nitrogen sources, one phosphorus source, pH-corrective sulfates, a complete micronutrient blend, and a potassium source. For the 0% nitrogen treatment, no nitrogen fertilizers were applied throughout the experiment. Instead, mono-potassium phosphate (MKP, 53% P₂O₅, 34% KCl) replaced DAP at transplanting to maintain phosphorus and potassium availability in the absence of nitrogen inputs.

At transplanting, soil pH correction was carried out using iron sulfate (Sulfex Fe) and zinc sulfate (Sulfex Zn), which contain 17% sulfur and 28% of the respective micronutrient (Fe or Zn). Additionally, a micronutrient mixture (Agrimins), potassium chloride (KCl, 60% K), and diammonium phosphate (DAP, 18% ammoniacal nitrogen and 48% phosphorus as P₂O₅) were applied to facilitate the slow release of nitrogen and phosphorus. At 8 DAT, nitrogen was applied exclusively as urea (46% ammoniacal nitrogen). This fertilization regime was repeated at 20 DAT, with an additional KCl application, and the same treatment was administered at 40 DAT.

The selected genotypes exhibited variation in flowering time and maturity: Nerica flowered at 56.2 DAT and matured in 82–87 days, Rufipogon at 61 DAT (maturity: 75 days), Curinga at 59.5 DAT (maturity: 85–90 days), IR64 at 67.15 DAT (maturity: 80 days), and Ciherang at 70.35 DAT, reaching maturity between 90 and 95 days. This variation reflects their differing developmental rates and nitrogen use patterns.

Acquisition of NIR spectral signatures

Data collection was conducted at the maximum tillering and early flowering stages, as these represent periods of peak leaf area and chlorophyll content, providing optimal spectral sensitivity to nitrogen status. The NIR spectra (n = 2,700) were collected from fresh rice leaves on Day 27, Day 41, and Day 53 after transplanting, corresponding to three measurement time points (D1, D2, and D3). In the growth cycle of a rice plant, Day 27, Day 41, and Day 53 after transplanting (DAT) correspond to key vegetative and early reproductive stages, which are critical for nitrogen uptake and assimilation7,29,30,31. By Day 27, the rice plant is in the early to mid-vegetative stage, actively growing, developing tillers, and expanding its leaf area, with a high nitrogen demand to support chlorophyll synthesis, protein formation, and structural development30,31. By Day 41, the plant transitions from the late vegetative to the early reproductive stage, nearing the completion of tillering and initiating panicle differentiation, making nitrogen availability essential for ensuring proper grain formation potential30,31. By Day 53, the plant enters the early reproductive stage, where panicle initiation and booting occur, and nitrogen plays a crucial role in panicle formation, grain number, and overall yield potential30,31. At this stage, nitrogen uptake from the soil is near its peak; following fertilization, external nitrogen absorption declines as the plant begins to remobilize stored nitrogen from vegetative tissues to developing grains30,31. These stages were selected for NIR spectral analysis because they capture critical points in nitrogen uptake, redistribution, and utilization, impacting growth, yield potential, and NUE.

Spectral data acquisition was performed using a portable ASD LabSpec®4 Standard-Res spectrophotometer (Malvern Panalytical, ASD Analytical Spectral Devices Inc., Boulder, CO, USA) operated with Indico® Pro Spectral Acquisition software. The spectral resolution was 1.4 nm (350–1000 nm) and 2 nm (1000–2500 nm). Each spectrum was obtained by averaging 50 consecutive scans, with an integration time of 34 ms per scan to enhance signal quality and minimize noise. Three representative plants (P1, P2, P3) per plot were selected on each sampling date for spectral analysis based on uniformity in height and leaf development stage. The plants were collected from the field during the morning hours and stored in humid chambers at 5 °C until spectral collection. Before the collection, plants were left at room temperature for 30 min, and excess water was removed from the leaves using Kimwipes®tissue. For each plant, three independent spectral signatures were recorded from three different spots of the adaxial leaf surface in reflectance mode over the 350–2500 nm wavelength range using a contact probe to ensure consistent spectral collection. A total of 900 spectral signatures per measurement day were acquired, resulting in 2700 spectral signatures analyzed in this study (Table 1). Spectral data acquisition was conducted under controlled conditions to minimize environmental interferences.

Table 1 Distribution of NIR spectra collected for fresh leaves from five rice genotypes across different nitrogen treatment levels. Each nitrogen treatment (0%, 25%, 50%, 75%, and 100%) includes an equal number of spectra per genotype (D1, D2, and D3, representing different days of measurement).
Full size table

Chemometrics

The raw reflectance spectra (350–2500 nm) were extracted and examined for potential outliers before undergoing mathematical transformations. Absorbance values were computed using R Studio (R version 4.4.1, “Race for Your Life”, The R Foundation for Statistical Computing), employing the prospectr and stats libraries. To minimize spectral variability and enhance feature extraction, spectral pre-processing techniques were applied, including Baseline Correction, Standard Normal Variate (SNV), De-trending, and first derivative (polynomial order: 1, Savitzky-Golay smoothing points: 25). The impact of these pre-treatments on model performance was evaluated through cross-validation and independent testing, enabling the selection of the most suitable preprocessing methods and algorithms for nitrogen estimation in fresh rice leaves22.

A balanced distribution of NIR spectra was ensured for each rice genotype across nitrogen treatments, following a structured dataset allocation. Spectral data were organized as outlined in Table 2, maintaining equal representation across nitrogen levels. Three tables (one for each model) were constructed to carry out the chemometrics for each genotype. In Model 1, spectra from Plant 1 and Plant 2 were used for training and cross-validation, while Plant 3 was reserved for testing. In Model 2, Plant 1 and Plant 3 were designated for training and cross-validation, with Plant 2 serving as the test set. In Model 3, Plant 2 and Plant 3 were used for training and cross-validation, leaving Plant 1 for testing. For each model, the dataset was stratified into training (TR, n = 90), cross-validation (CV, n = 30), and test (TE, n = 60) subsets, ensuring a representative distribution of spectral variability across nitrogen treatments, where the training and the cross-validation set presented a 75%/25% random distribution. Models were trained and tested using spectral data from four different wavelength regions (800–2400 nm, 350–1000 nm, 1300–1600 nm, and 1700–2200 nm), ensuring a comprehensive assessment of classification accuracy across spectral variations.

Principal Component Analysis (PCA) utilizing Singular Value Decomposition (SVD) algorithm and full random cross-validation, was performed for feature extraction and outlier detection. The extracted principal components explaining 99.5% of the variance were subsequently used as input variables in Linear Discriminant Analysis (PCA-LDA) to classify spectral variations associated with different nitrogen treatments. The integration of PCA and LDA facilitated spectral differentiation and pattern recognition, ensuring a robust classification framework for nitrogen estimation in fresh rice leaves16,17,32. Two-dimensional figures were created using Microsoft Excel (Microsoft®365™ 2024, Microsoft Corporation®).

PCA-LDA performance metrics

The classification performance of the PCA-LDA models for nitrogen treatment discrimination across rice genotypes was evaluated using R Studio, employing the MASS library on NIR spectral data collected on different measurement days. Model performance was assessed using accuracy (Ac), sensitivity (Se), and specificity (Sp) as key evaluation metrics16,17,32. Accuracy measured as the ratio of correctly classified spectra to the total spectra analyzed16,17,32. Sensitivity assessed the ability to correctly classify positive cases of each nitrogen treatment level, computed as TP/(TP + FN), where TP (true positives) represented correctly classified spectra, and FN (false negatives) corresponded to misclassified positive cases16,17,32. Specificity evaluated the ability of the model to correctly identify negative cases, calculated as TN/(TN + FP), where TN (true negatives) denoted correctly rejected negative cases, and FP (false positives) represented misclassified negative instances16,17,32.

Performance evaluation was conducted separately for the training, cross-validation, and test subsets to assess the generalization ability of each predictive model. Metrics were computed both at the individual nitrogen treatment level and the overall model level, ensuring a robust classification assessment. The presented values indicate the average and standard deviation, calculated from the three models developed for each measurement day (D1, D2, and D3). This evaluation framework provided a comprehensive assessment of the PCA-LDA capability to classify nitrogen treatments based on NIR spectral data, ensuring reliable classification and supporting its application for non-destructive nitrogen estimation in fresh rice leaves.

Table 2 Balanced distribution of NIR spectra used for each rice genotype across nitrogen treatments. The table outlines the allocation of spectra for each PCA-LDA model (Model 1, model 2, and model 3), ensuring equal representation across training (TR, n = 90), cross-validation (CV, n = 30), and test (TE, n = 60) sets. Spectra were evenly distributed among three individual plants (P1, P2, and P3) within each dataset split, maintaining a balanced representation across nitrogen treatments.
Full size table

Results

Fresh rice leaves NIR spectral profile

The following results correspond to the full NIRS range (800–2400 nm), encompassing spectral data from all rice genotypes analyzed in this study (Fig. 2). Similar spectral trends were observed across individual genotypes and different spectral regions, as detailed in the supplementary figures S1–S5 online. The observed mean raw NIR spectra of fresh rice leaves under different nitrogen treatments (Fig. 2a) displayed distinct absorption features attributed to fundamental molecular vibrations of key biochemical constituents8,33,34. Prominent absorbance peaks were identified at 970, 1200, 1450, 1750, 1930, 2100, and 2350 nm, corresponding to fundamental vibrational modes associated with O-H, C-H, and N-H bonds in plant tissues11,14.

The absorption peaks at 970, 1200, and 1450 nm can be primarily attributed to O-H stretching vibrations, reflecting the presence of water and hydroxyl-containing compounds, such as cellulose, hemicellulose, and other structural polysaccharides35. Peaks observed at 1750 and 1930 nm correspond to C-H stretching and combination vibrations, often associated with lipids, proteins, and organic compounds integral to plant metabolism, leaf biochemistry, and cell membrane structure11. Notably, the absorption features in the 2100–2350 nm range correspond to a combination of O-H, C-H, and N-H functional groups, which are indicative of proteins, amino acids (RCH(NH₂)COOH), structural carbohydrates (Cx(H₂O)y), and nitrogen-containing compounds11,13,39. This region is highly relevant for detecting proteins and amino acids, due to the presence of amide bond overtones and N–H bending vibrations that appear prominently in this spectral range. Specific groups such as RCH(NH₂)COOH contribute to absorption in this region and are strongly associated with nitrogen metabolism in plant tissues36.

In addition to proteins and amino acids, spectral features within this range also reflect the presence of chlorophyll-related molecules, which contain nitrogen as a central element and are closely linked to photosynthetic capacity8,33,34. The presence and variation of absorption in this region are consistent with differences in nitrogen treatment levels, with higher nitrogen availability generally enhancing signal intensity around these bands (Fig. 2a and b). This spectral behavior supports the association between NIR response and nitrogen-driven biochemical changes in leaf tissues8,33,34.To enhance spectral differentiation and minimize baseline shifts, light scattering, and variability in path length, baseline correction, SNV, and first derivative pre-treatments were applied (Fig. 2b). Baseline correction removed additive offsets across the spectrum, SNV normalized the spectra to reduce scattering effects due to particle size and sample morphology, and the first derivative transformation improved peak resolution and helped isolate subtle absorption features, particularly in the nitrogen-sensitive 2100–2350 nm region11,17. These transformations increased the visibility of spectral features associated with nitrogen-containing compounds such as proteins and amino acids. While these enhancements improved overall spectral clarity, clear visual separation between nitrogen treatment levels remained limited, especially in overlapping regions. Minor shifts in absorbance intensity were observed in higher nitrogen treatments, particularly near the 1450 nm and 2100–2350 nm ranges, but the differences were insufficient for consistent visual discrimination. Therefore, multivariate chemometric approaches were necessary to extract relevant patterns and achieve robust classification.

Fig. 2
figure 2

Evaluation of NIR spectra acquired for fresh leaves from five rice genotypes under different nitrogen treatments (n = 2700), including spectra collected on three different days (D1, D2, and D3). (a) Raw averaged NIR spectra (350–2500 nm) with baseline correction across nitrogen treatments, showing the characteristic spectral patterns of fresh rice leaves. Nitrogen-associated absorption regions are highlighted in blue, indicating key bands linked to proteins, amino acids, and chlorophyll-related compounds involved in nitrogen metabolism. (b) Transformed averaged spectra (350–2500 nm) showed no significant differences between nitrogen treatments. (c) PCA scores plot displaying the distribution of spectra based on the first two principal components (PC-1 and PC-2), explaining the majority of the variance in the dataset (97.7%). Spectra from D1 are clustered on the left side, while spectra from D2 and D3 are grouped on the right. (d) PCA loadings illustrating the dominant wavelengths contributing to the observed variance in the PCA scores plot.

Full size image

The exploratory spectral analysis conducted using PCA revealed spectral similarities among nitrogen treatments in the transformed NIR spectra across the 800–2400 nm range. The two-dimensional PCA scores plot (Fig. 2c) illustrates the distribution of spectral variance, where Principal Component 1 (PC-1) accounted for 95.1% of the total variation, while Principal Component 2 (PC-2) explained 2.6%, together capturing 97.7% of the variability of the full dataset. This high level of explained variance indicates that the dominant spectral differences among nitrogen treatments are well represented within these two components.

The scores plot (Fig. 1c) reveals distinct separation patterns, with spectra from Day 1 (D1) primarily clustering on the left, while those from Day 2 (D2) and Day 3 (D3) are grouped on the right. This distribution suggests that temporal changes in leaf composition influence spectral characteristics. These shifts in clustering may reflect biochemical changes occurring over time, such as nitrogen redistribution, chlorophyll metabolism, or structural modifications in leaf tissues30,31. The Hotelling’s T² influence plot (not shown) confirmed that no significant outliers were present, ensuring that the observed spectral distribution is representative of natural variability rather than artifacts or measurement inconsistencies.

The PCA loadings plot (Fig. 2d) provided insights into the key spectral regions driving the differentiation observed in the scores plot. PC-1, which explained the majority of the spectral variance, was predominantly influenced by strong dominant peaks at 970, 1200, 1350, 1450, 1750, 1900, 1930, 2100, and 2300 nm, corresponding to O-H, C-H, and N-H vibrational modes11. Positive absorption peaks at 1200, 1450, 1930, 2100, and 2300 nm indicated spectral contributions from water content (O-H stretch at 1200 and 1930 nm), protein structures (N-H at 2100–2300 nm), and carbohydrate-linked C-H vibrations (1450 nm), all of which play key roles in nitrogen metabolism11,20. Conversely, negative absorption peaks at 970, 1350, 1750, and 1900 nm suggested the influence of structural carbohydrates and nitrogen-associated compounds, further differentiating nitrogen treatments in rice leaves8,20,33. PC-2, which accounted for a smaller portion of the variance, exhibited dominant peaks at 1000, 1100, 1350, 1400, 1750, 1850, 2050, and 2200 nm, corresponding to O-H, C-H, and N-H combination bands11. Positive peaks at 1100, 1400, 1850, and 2200 nm highlighted moisture-related and protein conformation changes, while negative peaks at 1000, 1350, 1750, and 2050 nm suggested biochemical variations influenced by nitrogen availability8,11,20,33.

Leaf nitrogen content discrimination

The PCA-LDA models developed for nitrogen treatment classification were evaluated across multiple spectral ranges, including the full NIRS range (800–2400 nm) and selected subregions: 350–1000 nm, 1300–1600 nm, and 1700–2200 nm. The reported performance metrics represent the mean and standard deviation, calculated from the three independent models generated for each measurement day, ensuring a comprehensive assessment of classification reliability across different spectral regions (Fig. 3). Detailed PCA-LDA tables for each evaluated genotype are provided in the supplementary tables S1–S6 online.

Analysis of the Nerica genotype established a clear benchmark for classification performance across nitrogen treatments. Using the full NIRS spectral range (800–2400 nm), the PCA–LDA model achieved training accuracies exceeding 95% and test accuracies consistently above 90%. The model effectively discriminated between nitrogen-deficient (0%) and nitrogen-saturated (100%) plants, with sensitivity and specificity approaching 100%. Intermediate nitrogen levels (25% and 50%) exhibited higher misclassification rates, consistent with spectral overlap due to biochemical similarity under moderate nitrogen supply8,20,33. Performance declined when the analysis was restricted to the visible-NIR range (350–1000 nm). Training accuracy decreased to 85–92%, with lower cross-validation and test set results, particularly for intermediate nitrogen levels. Sensitivity was lowest for the 50% nitrogen treatment, indicating that pigment-dominated features in this region are insufficient to fully capture nitrogen-associated biochemical variation. Classification improved in the shortwave-NIR region (1300–1600 nm), where training accuracy exceeded 90% and test accuracy remained above 88%. Sensitivity increased at intermediate nitrogen levels, suggesting that this region provides additional biochemical information related to nitrogen-containing compounds. The 1700–2200 nm range produced the strongest results, with training accuracy above 94% and test accuracy exceeding 91%. Low rates of misclassification at intermediate levels reflect gradual physiological transitions rather than discrete boundaries.

The Rufipogon genotype displayed a classification pattern broadly similar to Nerica, though with slightly lower overall accuracy at intermediate nitrogen levels. Across the full spectral range, training accuracy exceeded 96% and test accuracy remained above 92%. As with Nerica, restricting the analysis to the visible-NIR region resulted in reduced performance, with training accuracy between 83% and 90% and lower sensitivity for 25% and 50% nitrogen treatments. This reduction is consistent with the limited biochemical information captured in pigment-dominated wavelengths33. Shortwave-NIR improved classification, with training accuracies surpassing 91% and test accuracies above 89%. Intermediate nitrogen levels were better resolved in this range, highlighting its sensitivity to molecular vibrations in nitrogen-associated compounds8,33. The 1700–2200 nm region again provided the highest performance, with training accuracy above 95% and test accuracy consistently exceeding 92%. Misclassification rates at intermediate nitrogen levels remained low, supporting the interpretation of nitrogen effects as a physiological gradient rather than discrete classes.

Among the evaluated genotypes, IR64 displayed the highest classification performance. Using the full spectral range, the PCA–LDA model achieved training accuracy above 97% and test accuracy over 93%, with clearly defined class boundaries at 0% and 100% nitrogen. When the spectral range was limited to 350–1000 nm, accuracy decreased to 84–90%, particularly for the 50% nitrogen level. The shortwave-NIR region improved classification, with training accuracy exceeding 92% and test accuracy above 89%, indicating effective detection of protein-related biochemical variation. The 1700–2200 nm range yielded the strongest results across all genotypes, with training accuracy above 96% and test accuracy exceeding 94%. Intermediate nitrogen levels exhibited low misclassification rates, confirming the robustness of IR64 spectral response to nitrogen availability.

Ciherang showed a performance profile comparable to Rufipogon but with slightly lower accuracy, particularly in the visible-NIR region. Using the full NIRS range, training accuracy exceeded 96% and test accuracy remained above 92%. In the visible-NIR region, performance dropped to 83–90%, with reduced sensitivity for the 50% nitrogen treatment. The shortwave-NIR region resulted in improved classification (training accuracy above 91%, test accuracy above 88%), especially at intermediate nitrogen levels. As with other genotypes, the 1700–2200 nm region produced the best results, with training accuracy above 95% and test accuracy above 93%. The model effectively separated extreme nitrogen levels, indicating clear spectral distinctions between nitrogen-deficient and nitrogen-rich plants.

The Curinga genotype followed a trend similar to Ciherang, with strong classification performance at extreme nitrogen levels but reduced discrimination at intermediate treatments. Full-range NIRS analysis produced training accuracies exceeding 97% and test accuracies above 93%. Visible-NIR classification performance was lower (82–90%), particularly for the 50% treatment. The shortwave-NIR region improved accuracy (training > 92%, test > 89%) and increased sensitivity for intermediate nitrogen levels. The 1700–2200 nm region again delivered the highest accuracy (training > 96%, test > 94%), confirming its effectiveness for distinguishing nitrogen status in this genotype.

Integrating the results across all five genotypes reveals clear differences in classification performance. IR64 consistently achieved the highest classification accuracy, particularly when using the full NIRS range (800–2400 nm) and the 1700–2200 nm spectral region. Nerica and Rufipogon also achieved high classification accuracy, with test accuracies marginally lower than IR64 (approximately 90–92%). Ciherang and Curinga maintained overall high classification accuracy at extreme nitrogen levels but showed greater misclassification for intermediate nitrogen treatments. The 1700–2200 nm range provided the most consistent and robust classification performance across all genotypes, reflecting its sensitivity to nitrogen-associated biochemical signals and water absorption features. By contrast, the visible-NIR region yielded the lowest classification accuracy, as its pigment-dominated spectral information is less directly linked to nitrogen-related biochemical variation. The shortwave-NIR range provided intermediate performance, with IR64 and Rufipogon demonstrating higher discriminative capacity than Ciherang and Curinga. Across spectral regions, IR64 consistently demonstrated superior separability of nitrogen treatments, with test accuracies exceeding 94% in the 1700–2200 nm range. Nerica and Rufipogon followed closely, while Ciherang and Curinga exhibited slightly lower test accuracies, particularly at intermediate nitrogen levels. These patterns indicate that genotype-specific physiological characteristics influence spectral discrimination of nitrogen status, with IR64 showing the most distinct spectral response to nitrogen availability.

The LDA score plots (Fig. 4) illustrate the classification structure of nitrogen treatments across the five genotypes using the full NIRS range (800–2400 nm). Similar clustering behavior was observed in the visible-NIR, shortwave-NIR, and 1700–2200 nm subregions. The three columns correspond to different measurement days (D1, D2, D3), capturing temporal variability in spectral responses. Across genotypes, nitrogen treatments at 0% and 100% formed well-defined, distinct clusters, reflecting clear spectral differences between nitrogen-deficient and nitrogen-saturated conditions. Intermediate treatments (25%, 50%, and 75%) showed varying degrees of cluster overlap, consistent with gradual physiological transitions rather than discrete nitrogen states. Genotypic differences in cluster structure indicate variability in nitrogen uptake efficiency and associated biochemical responses. The consistent classification trends across spectral regions confirm the robustness of the PCA–LDA model in distinguishing nitrogen treatments.

Fig. 3
figure 3

Performance metrics of PCA-LDA models for rice genotype nitrogen treatment classification based on NIR spectral data collected across different days. The models were trained and evaluated using spectral data from different wavelength regions. Different dataset partitions were evaluated: training (TR, n = 90), cross-validation (CV, n = 30), and test (TE, n = 60). Performance was assessed using accuracy (blue), sensitivity (orange), and specificity (green) for each dataset split. Numerical values corresponding to the mean and standard deviation of each metric are provided in Supplementary Table S1 online. Genotypes = G1: Nerica, G2: Rufipogon, G3: IR64, G4: Ciherang, G5: Curinga; D = Day; TR = Training set, CV = Cross-validation set, TE = Test set.

Full size image
Fig. 4
figure 4

Linear Discriminant Analysis (LDA) score plots using the 800–2400 nm region for fresh leaves from five rice genotypes under different nitrogen treatments. These figures represent the training set of Model 1 only, serving as an example to illustrate the observed trends. Each subplot corresponds to a specific genotype: (a–c) Nerica, (d–f) Rufipogon, (g–i) IR64, (j–l) Ciherang, and (m–o) Curinga. The three columns display spectral data collected on different days: D1 (left), D2 (middle), and D3 (right). Different nitrogen treatments (0%, 25%, 50%, 75%, and 100%) are color-coded as indicated in the legend. The separation of clusters highlights the effects of nitrogen treatments and temporal variability in spectral responses across measurement days.

Full size image

Discussion

Understanding how nitrogen availability alters plant biochemical composition across diverse rice genotypes is central to improving nutrient management strategies and optimizing fertilizer use. This study demonstrates, for the first time, that NIRS combined with chemometric modeling can effectively classify nitrogen treatments across five rice genotypes: Nerica, Rufipogon, IR64, Ciherang, and Curinga. The classification trends observed across these genotypes provide valuable insight into how nitrogen-driven biochemical changes are expressed spectrally, especially in the near-infrared region. Notably, classification performance was highest in the 1700–2200 nm range, which aligns with known absorption features associated with nitrogen-linked functional groups15. These findings confirm that NIRS can reliably capture nitrogen-driven biochemical variation in rice and support its use as a diagnostic tool for genotype-informed nutrient management.

The ability to classify nitrogen treatments based on spectral signatures reflects underlying biochemical shifts, particularly in protein, chlorophyll, and structural carbohydrate composition15,18,34,36,37,38,39. IR64 exhibited the highest classification accuracy across nitrogen treatments, while Rufipogon showed lower accuracy at intermediate nitrogen levels (25% and 50%), indicating genotype-specific differences in physiological responses to nitrogen availability. These variations are consistent with differences in nitrogen assimilation and allocation strategies among genotypes. The high predictive accuracy obtained across nitrogen levels (typically > 90%) underscores the potential of NIRS to distinguish nitrogen responses at the genotype level8,15,34,40,41. Beyond classification, these findings provide insight into biochemical pathways underlying nitrogen uptake and utilization, offering a basis for more precise, genotype-specific fertilization strategies aimed at improving nitrogen use efficiency (NUE)1,2,4,42.

The observed spectral differentiation is closely linked to biochemical features inherent to nitrogen metabolism. Proteins, amino acids, and chlorophyll-related pigments are primary contributors to spectral variation in the near-infrared region, particularly within 1700–2200 nm, where N–H, C–H, and O–H bond vibrations dominate8,11,12,18,34. These absorption features reflect changes in leaf nitrogen content, structural carbohydrates, and moisture levels, which together shape the reflectance profile of rice leaves under varying nitrogen availability. Genotypic differences in these spectral signatures are likely associated with differences in nitrogen assimilation rates, allocation patterns, and leaf biochemical composition15,42. For example, IR64’s stronger class separability at both low and high nitrogen levels suggests a more pronounced physiological response and possibly a more efficient nitrogen assimilation system, whereas the overlap observed in Rufipogon and Curinga at intermediate levels reflects more buffered or gradual biochemical adjustments.

Achieving consistent discrimination using NIRS remains technically challenging because of intrinsic variability in leaf structure, water content, and the overlapping nature of spectral absorption bands. In this study, baseline correction, standard normal variate (SNV), and Savitzky–Golay first-derivative preprocessing were applied to minimize noise, reduce baseline shifts, and enhance resolution. These steps effectively improved class separation, consistent with previous findings demonstrating the critical role of preprocessing in NIRS applications11,17. Despite these corrections, some genotypes still exhibited lower accuracy at intermediate nitrogen levels, suggesting that physiological responses at these stages may be less spectrally distinct. This emphasizes the importance of careful preprocessing and sample handling to minimize noise and enhance the robustness of chemometric models. Differentiating genotypes based on their spectral response therefore provides meaningful information on nitrogen absorption, allocation, and metabolism under variable fertilization strategies2,9,15,43.

Previous research has provided strong evidence supporting the applicability of NIRS for estimating leaf nitrogen concentration in rice, but these investigations often focused on a single genotype or broad nitrogen gradients8,15,20,33,34. As a result, they largely overlooked the significance of genotypic variation in spectral response. This study expands upon existing work by explicitly incorporating multiple genotypes and quantifying their distinct spectral signatures in response to nitrogen application. Earlier work has typically treated rice as a homogeneous species in spectral modeling, which has led to inconsistencies in nitrogen estimation accuracy when models are transferred between cultivars2,8. In reality, differences in leaf structure, chlorophyll content, and biochemical composition among genotypes introduce variability into reflectance measurements and contribute to errors in universal nitrogen prediction models12,34. The present findings demonstrate that nitrogen-related spectral responses are genotype-specific and cannot be accurately captured by generalized models. The 1700–2200 nm region consistently provided the highest classification accuracy across genotypes, reflecting its sensitivity to biochemical changes associated with nitrogen metabolism11,44,45. In contrast, the 350–1000 nm region exhibited lower accuracy, likely due to its reliance on chlorophyll-related features, which are more susceptible to environmental variation33.

The classification patterns observed across genotypes provide additional physiological context. IR64 and Ciherang showed clear separability of nitrogen levels, indicating coordinated biochemical responses to nitrogen availability. This was particularly evident in the 1700–2200 nm region, where amide and N–H absorption features are most prominent8,12,39,41. By contrast, Rufipogon and Curinga showed greater overlap in intermediate nitrogen levels, suggesting more gradual or buffered responses to nitrogen limitation or enrichment. Such differences highlight the physiological plasticity inherent to different rice genotypes and the value of genotype-specific models for improving predictive accuracy8,12,39,41. These results also suggest that nitrogen uptake efficiency and allocation patterns may be driving factors behind the observed spectral variation, an interpretation consistent with known genotype-specific differences in nitrogen physiology in rice. The implications of these findings extend beyond classification performance alone. Genotype-specific spectral responses can inform more precise and efficient nitrogen management strategies, moving away from uniform fertilization approaches toward targeted nutrient supply. Identifying the most informative spectral regions also supports sensor optimization, enabling the design of cost-effective, targeted sensing solutions for nitrogen monitoring16,32. The 1700–2200 nm region, in particular, provides a strong biophysical basis for selecting wavelengths sensitive to nitrogen-associated functional groups, which can improve both the efficiency and robustness of future sensing applications.

Although the classification models in this study performed well, the inherent variability in spectral acquisition must be acknowledged. Factors such as leaf angle, surface moisture, turgor, and diurnal variability can influence reflectance measurements12,34,46. Preprocessing steps such as SNV and first-derivative filtering were essential to minimize these effects and ensure consistent classification11,13. Some residual misclassification at intermediate nitrogen levels likely reflects true physiological gradients rather than model limitations. This characteristic highlight both the strength and the limitations of classification-based chemometric approaches in biological systems14,16. While the full NIR range used here allows precise nitrogen classification at the leaf scale, extending these approaches to canopy or satellite platforms will require addressing additional complexities such as atmospheric absorption, bidirectional reflectance effects, and sensor resolution12,29,36,40.

The broader significance of these findings is reflected in their potential to support sustainable agricultural intensification. This research contributes to the Green Campus Initiative led by the Alliance Bioversity International and CIAT, which promotes sustainable, data-centric agricultural solutions at the Palmira research station in Colombia; as well as with global Sustainable Development Goals (SDGs 2, 12, 13, and 15), reinforcing the role of technology-driven approaches in sustainable agriculture. The integration of spectroscopy and chemometrics offers a promising avenue for developing non-destructive, real-time diagnostic tools that support more efficient input use. By linking genotype-specific nitrogen responses to spectral signatures, this approach contributes to the design of targeted nutrient management strategies that minimize fertilizer waste and environmental impact while maintaining productivity.

The future potential of this work lies in its scalability. As embedded spectroscopy and machine learning technologies advance, models developed at the leaf scale can be transferred to handheld or remote sensing platforms. The integration of deep learning architectures with chemometric feature extraction could allow automated, real-time interpretation of nitrogen status under field conditions, reducing the need for extensive preprocessing. Deploying such tools at the farm level could transform nitrogen management from centralized laboratory analysis to site-specific decision-making. This decentralized model has the potential to enhance nitrogen use efficiency, reduce greenhouse gas emissions, and support climate-resilient cropping systems.

Conclusion

This study establishes the value of combining NIRS with chemometric classification for genotype-specific nitrogen assessment in rice (typically > 90%). The approach effectively captures physiological variability, identifies informative spectral regions, and supports the development of precision fertilization strategies. While technical and environmental challenges remain, particularly for large-scale deployment, the framework presented here offers a scientifically robust and operationally relevant pathway toward more sustainable nutrient management. Continued interdisciplinary collaboration among plant physiologists, agronomists, and data scientists will be essential to further refine these models and translate them into scalable, field-ready solutions. As sensing technologies evolve, AI-driven spectroscopy has the potential to become a key component of nitrogen management strategies, advancing both agricultural productivity and environmental sustainability.