Field experiments were conducted across Jiangsu, Zhejiang, and Jiangxi provinces in China. These experiments followed a randomized block design, incorporating 3 to 6 different nitrogen application rates, each with 3 to 4 replicates. A total of 15 rice varieties were tested, including 10 japonica, 4 indica, and 1 indica-japonica hybrid, covering a range of regions and cultivation systems. Basic details of these varieties are shown in Table 1 . Except for variations in nitrogen application, all other agronomic practices followed local conventional standards.

Samples were obtained every 10-15 days from the Tillering (TI) stage to the Heading (HD) stage, corresponding to key growth stages: TI, Stem Elongation (SE), Panicle Initiation (PI), and HD. Due to differences in growth durations among the rice varieties, some early-maturing varieties had entered the Grain Filling (GF) stage when most varieties were at HD. Consequently, a small number of samples were collected during the GF stage for these varieties. At each sampling time, 3 to 6 hills of uniformly growing plants were destructively sampled from each plot. Samples were separated into leaves, sheaths, and panicles, then oven-dried at 105°C for 30 minutes, followed by 75°C until a constant weight was achieved. Nitrogen concentration was determined using the Kjeldahl method.

The NNI was derived by calculating the ratio of Plant Nitrogen Concentration (PNC) to the Critical Nitrogen concentration (N_c). NNI values greater than 1 indicate nitrogen surplus, 1 indicates sufficiency, and less than 1 indicates deficiency.

where N_c is calculated using the nitrogen dilution curve established from our dataset:

where W represents the above-ground dry matter (t·ha^-¹). The parameters (3.44 and -0.43) were derived from our experimental data using the method described by Justes et al. (1994). This method involves plotting plant nitrogen concentration against biomass and fitting a power function to the minimum nitrogen concentration observed at each biomass level, representing the N_c required for optimal growth.

On the same day as the destructive sampling, 10 to 20 rice plants were selected from each plot to measure the 1-5 LFT using a Dualex Scientific+ instrument (Force-A Inc., France). Each leaf was measured at three positions: middle, upper one-third, and lower one-third, and the average value was calculated for that leaf. The Dualex Scientific+ is a portable optical sensor designed for rapid, non-destructive detection of various physiological parameters in plant leaves. By measuring the transmission and reflection of specific wavelengths of light, it provides readings for Chl, Flav, Anthocyanin content (Anth), and the NBI. NBI is the ratio of Chl to Flav, serves as an indirect measure of the plant’s nitrogen nutritional status. Although the Dualex Scientific+ instrument also measures Anth, preliminary analyses indicated that anthocyanin levels in rice leaves were consistently low and showed minimal variation across treatments and growth stages. Consequently, Anth was not included in the subsequent analyses.

During early growth stages, such as the TI stage, it was sometimes not possible to identify five fully expanded leaves in a shoot due to the limited number of leaves. In these cases, we measured at least the 1-4 LFT and calculated the average values. This average was then used as the value for the 5 LFT. This data preprocessing step ensured consistency across samples and prevented errors due to unequal data when applying machine learning models.

The study employed several modeling approaches including Simple Linear Regression (LR), Quadratic Curve Regression (QCR), Partial Least Squares Regression (PLS), and four machine learning models—Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGB), and Neural Network (NN)—to estimate rice Leaf Nitrogen Concentration (LNC), PNC, and NNI. The hyperparameters for each machine learning model were selected based on preliminary analyses that involved testing various parameter combinations using grid search and cross-validation. This approach helped balance model complexity and performance, ensuring robust generalization. By comparing the performance of these models, their applicability in estimating rice nitrogen nutrition indices was evaluated.

To evaluate the saturation effects of different spectral measurements on rice nitrogen indices, we conducted an incremental change analysis that included the following steps:

To gain deeper insights into the contributions of each input variable to the model’s predictions, we employed the SHAP (SHapley Additive exPlanations) method (Lundberg and Lee, 2017), which is based on game theory, to analyze feature importance for the RF and XGB models. SHAP values quantify the impact of each feature on the prediction outcomes by calculating its marginal contribution across all possible feature combinations. Compared to traditional feature importance metrics, the SHAP method offers consistency and additivity, enhancing model interpretability and providing guidance for optimizing measurement strategies in practical applications.

To comprehensively evaluate the predictive performance of each model, we used the coefficient of determination (R²), Root Mean Square Error (RMSE), and Relative Root Mean Square Error (RRMSE) as evaluation metrics. The formulas for these indicators are as follows:

• Coefficient of determination:

• Root mean square error:

• Relative root mean square error:

where Y_i is the actual value, Y ^ i is the predicted value, Y ¯ is the mean of the actual values, and n is the number of samples.

All statistical analyses and model construction in this study were performed using Python 3.12.4. Data processing and analysis utilized libraries such as Pandas, NumPy, and SciPy. Machine learning modeling was done using Scikit-learn, XGB, TensorFlow, and Keras, while visualization was carried out using Matplotlib. Statistical tests were conducted with Statsmodels and SciPy.stats, and feature importance analysis was performed using the SHAP library. To ensure reproducibility of results, all random processes were set with the same random seed during modeling and testing.

Average Dualex values (Chl, Flav, NBI) from the first to fifth LFT were analyzed using LR and QCR models to explore their relationship with rice nitrogen nutritional indicators—LNC, PNC, and the NNI ( Figure 1 ). The findings revealed notable differences in the effectiveness of these indicators for nitrogen estimation depending on the specific Dualex metrics used.

Chl displayed a nonlinear relationship with both LNC and PNC ( Figures 1A, D ). As Chl levels increased, both LNC and PNC rose, showing accelerated growth rates at Chl values around 30. However, a saturation effect was evident between Chl values of 35 to 40, where further increases in Chl did not lead to corresponding changes in LNC and PNC.

Conversely, Flav showed a stable and linear relationship with LNC and PNC, without notable saturation effects ( Figures 1B, E ). Flav levels increased steadily as nitrogen content decreased, demonstrating a strong negative correlation with LNC and PNC across the full range, with high regression fits (R² = 0.70 and 0.71). Even when accounting for differences across rice varieties and growth stages, Flav consistently provided an accurate assessment of nitrogen status.

NBI, defined as the ratio of Chl to Flav, exhibited a significant linear relationship with LNC and PNC ( Figures 1C, F ), achieving R² values of 0.74 and 0.70, respectively. This underscores NBI’s robustness and stability in estimating nitrogen nutritional status. As a composite indicator, NBI effectively mitigates the saturation effect observed when using Chl alone, making it a highly reliable tool for nitrogen diagnosis.

For NNI estimation, which represents a relative value (the ratio of PNC to the N_c), measurements across different growth stages showed consistent NNI values, primarily ranging between 0.4 and 1.4. However, Dualex readings varied across stages, leading to lower estimation accuracy for NNI ( Figures 1G–I ). Although NNI tended to increase with changes in Chl, Flav, and NBI, this trend exhibited significant variability across growth stages, resulting in inconsistent patterns.

The saturation effect analysis of Dualex values in estimating rice nitrogen nutritional indicators is illustrated in Figure 2 . To minimize the impact of model fitting differences when evaluating saturation effects across various Dualex indices, we applied an incremental analysis method using raw and standardized data. The standardized rate of change (ΔY/ΔX) was calculated based on these original data points. Each subplot displays the rate of change of nitrogen nutritional indicators relative to standardized Dualex values. In the analysis of Chl versus nitrogen indicators ( Figures 2A, D, G ), an initial low rate of change was observed in the lower Chl range, followed by a gradual increase, peaking at moderate Chl values. However, as Chl values continued to rise, the rate of change began to decline, particularly in the high Chl region. In the low Chl region, the low nitrogen indicator values were primarily responsible for the reduced rate, while in the high Chl region, a declining rate suggested that further increases in Chl no longer led to corresponding rises in nitrogen indicators, indicating a clear saturation effect. The density curves at the bottom further validated this: standardized Chl values peaked around 0.68 (for LNC and PNC) and 0.74 (for NNI), showing a marked decline in estimation accuracy at these levels.

In contrast, Flav exhibited a negative correlation with nitrogen indicators ( Figures 1B, E, H ). The rate of change remained relatively stable in the low Flav region ( Figures 2B, E, H ), indicating effectiveness in estimating nitrogen indicators. The density peaks for Flav values were observed at 0.22 (for LNC and PNC) and 0.39 (for NNI), with a gradual distribution, suggesting that even at high nitrogen indicator levels, Flav measurements were evenly spread without a significant saturation effect. However, for NNI, as Flav values decreased, a rapid decline in the rate of change was noted, with the density peak at 0.39 indicating a more pronounced saturation. Comparing this to Figures 1H , NNI exhibited almost horizontal distribution in the low Flav range, indicating diminished sensitivity to changes in Flav.

As a composite indicator, NBI (Chl to Flav ratio) displayed a more consistent trend in estimating LNC, PNC, and NNI ( Figures 2C, F, I ). The rate of change initially increased, then decreased with rising NBI, resembling the pattern seen in Chl but with density peaks closer to boundary values (0.76, 0.77, and 0.74, respectively). This suggests that NBI effectively estimates nitrogen nutritional status even at high nitrogen levels, exhibiting significantly less saturation than Chl alone.

In summary, Chl demonstrated a clear saturation effect, particularly at high nitrogen levels, where its sensitivity diminished. Conversely, Flav and NBI retained considerable variability in their rates of change at high nitrogen levels, showing enhanced sensitivity to changes in nitrogen status and providing a more accurate reflection of variations in rice nitrogen nutrition.

Figures 3 – 5 illustrate the performance of five different models in estimating rice nitrogen nutritional indicators using Dualex readings. Each model utilized measurements from the 1-5 LFT, including Chl, Flav, and NBI values for individual leaves and their averages, totaling 18 input variables. The models showed distinct differences in their performance across training and validation sets, particularly when estimating different nitrogen indicators. Overall, estimation accuracy was highest for LNC, followed by PNC, with NNI being the most challenging to estimate accurately.

For LNC estimation ( Figure 3 ), all models demonstrated strong fitting accuracy, achieving R² values above 0.81 in training and 0.78–0.87 in validation datasets. SVR, RF, and XGB models demonstrated the highest validation accuracy, with R² values between 0.86 and 0.87, and RRMSE below 10%. Data from all sampling periods, except for minor overestimations at the tillering stage, were evenly distributed around the 1:1 line.

In estimating PNC ( Figure 4 ), model performance was similar to that observed for LNC, despite with slightly reduced accuracy. RF and XGB remained the top performers, with validation R² values of 0.82 and RMSE at 0.29. The SVR model followed closely with an R² of 0.81. Conversely, PLS regression model demonstrated slightly lower performance.

For NNI estimation ( Figure 5 ), all models showed a noticeable decline in fitting accuracy, which was consistent with regression analysis results from quadratic models. Despite the challenges, RF and XGB continued to lead, with validation R² values ranging from 0.52 to 0.54, marking a 71% improvement over QCR models. However, estimating NNI proved to be more difficult than LNC and PNC, with apparent biases in both low and high-value regions.

Radar charts were used to visualize the estimation accuracy of different models and input variable combinations on the validation dataset ( Figure 6 ). The R² values decreased progressively from LNC to PNC and then to NNI ( Figures 6A-C ). While LNC and PNC estimations showed near-equal accuracy, R² values for NNI dropped substantially. Across all input variable combinations, R² for NNI estimation declined by an average of 57% and 55% compared to LNC and PNC, respectively.

In comparing input variable combinations, the PLS model using multi-leaf NBI value was most effective in estimating LNC and PNC. For the other models, the all-variable combination (Combination 4) and NBI combination consistently provided high accuracy. Changes in RRMSE mirrored R² trends inversely ( Figures 6D-F ); models with higher R² exhibited lower RRMSE. RRMSE increased sequentially when estimating LNC, PNC, and NNI. Among the input variable combinations, models using only the Chl combination had the highest RRMSE for LNC and PNC estimations, indicating lower accuracy in these cases. In contrast, for NNI estimation, models using only the Flav combination had the highest RRMSE ( Figure 6F ). The NBI and all-variable combinations consistently achieved similar, superior accuracy across all nitrogen indicators.

The feature importance analysis results based on SHAP values for the RF and XGB models in nitrogen estimation are presented in Figure 7 . Both models used Combination 4 as input to assess the impact of each variable on their predictions.

For LNC prediction, Figures 7A, D show the SHAP analysis for the RF and XGB models, respectively. In the RF model ( Figure 7A ), the NBI value from the 2 LFT was the most influential variable, followed by the Flav value from the 1 LFT and the NBI value from the 1 LFT. The XGB model ( Figure 7D ) also highlighted these three variables as having the highest impact, with their SHAP values significantly surpassing those of other variables. This suggests that these variables were critical for accurate LNC prediction.

In the case of PNC estimation, Figures 7B, E illustrate the SHAP analysis for the RF and XGB models. The RF model ( Figure 7B ) indicated that the Flav value from the 1 LFT and the NBI value from the 2 LFT were the most significant variables. Similarly, in the XGB model ( Figure 7E ), the NBI of the 2 LFT and Flav of the 1 LFT were also identified as the most influential factors. This consistency across models underscores the importance of NBI and Flav as key indicators for both LNC and PNC estimation.

For NNI prediction, Figures 7C , F display the SHAP analysis results for the RF and XGB models. The RF model ( Figure 7C ) identified the Chl value from the 4 LFT as the most important variable, followed by the NBI value from the 4 LFT. The XGB model ( Figure 7F ) similarly showed these two variables as dominant; however, there were differences in their SHAP value distributions. In the RF model, the Chl value from the 4 LFT had a SHAP value notably higher than the other variables, whereas, in the XGB model, SHAP values of variables decreased in a stepwise manner, indicating that no single variable was overwhelmingly dominant.

Overall, the SHAP analysis results for LNC and PNC predictions revealed certain dominant variables that enabled the models to achieve high-precision estimates. However, for NNI prediction, the lack of clearly dominant variables and weaker associations between input variables and the target presented challenges in achieving satisfactory model accuracy.

Incremental analysis highlighted distinct differences in the saturation effects of different Dualex indices when estimating rice nitrogen indicators. The results showed that Chl experiences significant saturation at high measurement values, consistent with findings from studies using SPAD-502 (Yuan et al., 2016; Zhang et al., 2019a; Jiang et al., 2021) and chlorophyll-based multispectral/hyperspectral indices (Fu et al., 2021; Yao et al., 2023; Gao et al., 2024). Previous research has noted that SPAD values become less responsive to nitrogen at higher concentrations, complicating the differentiation of nitrogen levels (Yue et al., 2020). This saturation effect is documented across various crops, including rice (Wang et al., 2023), wheat (Jiang et al., 2021), and maize (Dong et al., 2021). Our findings reaffirm that Chl and SPAD measurements exhibit saturation under high nitrogen conditions.

When leaf chlorophyll content is low, an increase can enhance the leaf’s net photosynthetic rate (Tantray et al., 2020). However, as nitrogen availability rises and chlorophyll reaches a threshold level, photosynthetic rates plateau due to limits in light absorption area (Hou et al., 2019). At this point, excess nitrogen is allocated to non-photosynthetic reserves, altering the ratio of chlorophyll to leaf nitrogen (Hou et al., 2019; Mu and Chen, 2021). Consequently, while Chl is a sensitive indicator at low nitrogen levels, its effectiveness decreases at higher levels, introducing errors into predictive models. This trend was observed in our analysis, particularly in LNC predictions, where models relying solely on Chl values achieved lower R² scores, failing to capture nitrogen variations accurately. Therefore, Chl’s utility as a single diagnostic variable is limited, especially under high nitrogen conditions.

The saturation effect in this study was mainly observed when rice plants exhibited high nitrogen content. Since nitrogen concentration in rice leaves or plants gradually decreases during growth—a pattern known as the nitrogen dilution effect (Yao et al., 2021; Li et al., 2022a)—with the impact most noticeable around the tillering and jointing stages. This differs from canopy-level saturation observed after canopy closure, where high leaf area can cause remote sensing indices (e.g., NDVI) to saturate due to optical signal limitations, reducing their sensitivity to nitrogen changes (Tenreiro et al., 2021). Conversely, Chl saturation is primarily driven by physiological limits on chlorophyll content within leaves, necessitating a distinction between these two types of saturation in research.

In contrast, Flav and NBI showed no significant saturation effects, maintaining their effectiveness in estimating nitrogen indicators, even at high nitrogen levels. Flav is closely associated with plant stress responses, increasing under conditions of nitrogen stress (Li et al., 2021). When nitrogen is scarce, plants produce more flavonoids as a protective response (Gitelson et al., 2017). NBI combines the divergent responses of these indicators. In nitrogen-rich conditions, Chl levels are high while Flav remains low, leading to a high NBI. Under nitrogen deficiency, Chl decreases and Flav increases, resulting in a low NBI. By integrating complementary data from both indicators, NBI provides a robust reflection of nitrogen status, offering distinct advantages over Chl alone. This aligns with studies showing that NBI can overcome saturation and retain sensitivity under high nitrogen levels (Dong et al., 2022; Song et al., 2022). Our dataset, encompassing 15 rice varieties across different growth stages, confirmed that NBI and Flav maintained high predictive accuracy across various environmental conditions, suggesting these indicators can reliably reflect nitrogen status across diverse environmental conditions.

We compared the performance of five machine learning models—PLS, SVR, RF, XGB, and NN—in predicting nitrogen indicators. Overall, XGB and RF models showed the highest accuracy for LNC, PNC, and NNI predictions, whereas PLS regression and SVR models had noticeably lower performance, lagging behind the tree-based approaches. This trend is consistent with findings from other studies comparing similar models (Fan et al., 2022; Kganyago et al., 2024). While PLS regression is effective at explaining linear relationships, it struggles with complex nonlinear patterns (Yang et al., 2023). SVR, though capable of managing nonlinear data, is highly sensitive to parameter selection and less efficient with high-dimensional inputs (Ben Ishak, 2016). Conversely, tree-based ensemble models like RF and XGB excel at handling nonlinear, high-dimensional datasets (Nzekwe et al., 2024), with XGB particularly excelling due to its gradient boosting optimization, which iteratively improves model performance.

The NN model, however, was less stable than the XGB and RF models in this study. This may be due to the relatively smaller dataset size, as neural networks generally require larger datasets to train effectively and achieve good generalization. While the NN performed adequately in estimating LNC and PNC, its predictions for NNI were less accurate than those of the other models. The RF and XGB models demonstrated strong performance, although the training R² values were slightly higher than the validation R² values, especially for NNI estimation, the differences were minimal. This highlights the importance of model selection and appropriate complexity in achieving reliable predictions (Moradi et al., 2020; Yates et al., 2023).

The lower estimation accuracy for NNI compared to LNC and PNC can be attributed to several factors. NNI is a composite index involving both PNC and N_c, where N_c is dependent on plant biomass and derived from the nitrogen dilution curve. This adds layers of complexity and potential sources of error. The dynamic nature of biomass accumulation and nitrogen dilution across growth stages introduces variability that is difficult to capture with current models. Additionally, the relatively narrow range of NNI values reduces the models’ ability to distinguish subtle differences. Measurement errors in biomass and nitrogen content further compound the challenge. Future research could explore incorporating growth stage information, more precise biomass measurements, and additional physiological parameters to enhance NNI prediction accuracy.

SHAP analysis was instrumental in interpreting the models by quantifying the contribution of each input variable to the predictions (Lundberg et al., 2017). This is essential for enhancing model transparency, which is particularly valuable in agriculture, where interpretable models facilitate practical application (Ryo, 2022). The SHAP results highlighted that spectral data from different leaf positions do not contribute equally to nitrogen prediction. For LNC and PNC, the most significant contributions came from measurements taken at the 1 LFT and 2 LFT, corroborating findings from related studies where spectral reflectance from top leaves was found to be most sensitive for nitrogen assessment (Dong et al., 2021; Lu et al., 2022; Wang et al., 2023). For NNI prediction, SHAP analysis pointed to the importance of Chl measurements from the 4 LFT, suggesting that chlorophyll content in middle to lower leaves is particularly relevant. This may stem from the generally lower estimation accuracy of NNI, which tends to make the significance of variables more uniform. Additionally, nitrogen translocation in middle and lower leaves closely aligns with the plant’s overall nitrogen nutritional status (Sun et al., 2018; Li et al., 2022b).

To improve NNI prediction accuracy, future strategies could include incorporating additional variables closely related to NNI, such as plant leaf area index (Tuccio et al., 2022), growth stages, and other physiological and environmental factors (Dong et al., 2022, 2024). Moreover, combining these models with crop growth simulations or radiative transfer models, as well as using multi-source data, could enhance the model’s ability to account for complex interactions (Raya-Sereno et al., 2024; Wang et al., 2024). Expanding the dataset’s size and diversity will also be crucial for improving generalization and predictive accuracy.

The differences in contributions from various leaf positions suggest the potential for optimizing measurement strategies in practical applications. By focusing on key leaf positions that provide the most significant input to the models, measurement efficiency can be improved. Reducing the number of measurement variables not only lowers costs but also enhances the computational efficiency of the models, making them more viable for real-world use.