The two field experiments for this study were conducted in the Pukou District (32°04’15” N, 118°28’21” E) and Liuhe District (32°25’04” N, 118°59’18” E) of Nanjing City, Jiangsu Province, China ( Figure 1 ). Each experiment was organized into treatment plots with multiple replications to study the effects of varying rates and types of nitrogen input ( Table 1 ). Specifically, the Pukou experiment consisted of five treatments that examined different rates of controlled-release fertilizer input, each with four replications (20 plots in total), while the Liuhe experiment involved four treatments, each with four replications except for the control (CK), which had three replications, resulting in a total of 15 plots (each measuring 200 m²).

The experimental sites in Pukou and Liuhe featured paddy soils (anthrosols) with distinct fertility profiles. Pukou soils exhibited organic matter content of 22.3 g kg^-1, total nitrogen of 1.3 g kg^-1, Olsen phosphorus of 15.4 mg kg^-1, and available potassium of 146.4 mg kg^-1. Liuhe soils contained higher nutrient levels: 26.6 g kg^-1 organic matter, 1.6 g kg^-1 total nitrogen, 15.2 mg kg^-1 Olsen phosphorus, and 166.2 mg kg^-1 available potassium. Regarding agronomic practices, two japonica rice cultivars (Wuyunjing 23 and Nanjing 5055) were transplanted in Pukou on June 12 during the 2018 and 2019 growing seasons, with harvests conducted on October 22, 2018 and November 12, 2019, respectively. In Liuhe, the single cultivar Nanjing 5055 was transplanted on June 25, 2019 and harvested on November 13, 2019, maintaining consistent phenological scheduling across sites.

In this study, the SMLR technique was employed to establish a correlation between color indices (predictor variables) and LNC (response variable). The SMLR method integrates both forward selection and backward elimination processes to optimize the model. Initially, the forward selection process identifies and includes the predictor variable with the highest statistical significance, provided it meets the 5% significance criterion. At each subsequent step, the process continues to add the most significant remaining variables. Once a new variable is added, the backward elimination process evaluates all included variables, applying a 10% significance criterion to decide whether to remove any variable that no longer meets this threshold. This iterative process of adding and removing variables ensures that only the most statistically significant predictors are retained in the model, enhancing its predictive accuracy and interpretability (Jin et al., 2012).

To further validate the robustness of the SMLR models and mitigate the risk of overfitting, the leave-one-out cross-validation (LOOCV) method was employed. In LOOCV, the model is trained on all data points except one, which is used as the validation set. This process is repeated such that each data point is used once as the validation set. The results are then averaged to provide an estimate of the model’s performance, ensuring that the SMLR model maintains high generalization capability on unseen data. Additionally, to assess the model’s transferability across different sites, cross-site validation was performed. Models were trained on data from the Pukou site and tested on the independent Liuhe site.

The correlations between measured LNC and color indices at different growth stages were analyzed using SAS software (version 9.2, SAS Institute Inc., Cary, NC, USA). Model fitness was evaluated by comparing estimated and measured LNC values in a 1:1 plot. The performance of SMLR models for LNC estimation was assessed using the coefficient of determination (R²), root mean square error (RMSE), and normalized root mean square error (NRMSE). These statistical indicators were defined in Equations 2–4, respectively, as follows:

where n represents the number of observations, O ¯ denotes the average value of measured LNC, P i and O i are the estimated and observed values of LNC, respectively. Typically, the simulation is classified as excellent when NRMSE is less than 10%, good if NRMSE is between 10% and 20%, fair if NRMSE is between 20% and 30%, and poor if NRMSE exceeds 30%.

The measured LNC across growth stages, experimental sites, years, and nitrogen treatments exhibited substantial variability, ranging from 0.93% to 3.95% ( Table 4 ). Consistent with the nitrogen dilution effect, mean LNC declined progressively from booting to filling stages at both sites. At Pukou, mean LNC decreased by 14.3% from 2.31% at the booting stage to 2.01% at the flowering stage, followed by a further 15.4% reduction to 1.71% at the filling stage. Liuhe displayed higher baseline LNC values than Pukou at equivalent phenological phases, with a 15.0% decline from 3.00% at booting to 2.55% at flowering and a 22.4% decrease to 1.98% at filling. This site-specific disparity likely reflects differences in soil fertility, as Liuhe soils contained 19.3% higher organic matter and 23.1% more total nitrogen compared to Pukou.

To assess the efficacy of 13 color indices derived from the leaf scale, correlations between LNC and color indices across various growth stages were established. At the booting stage, nearly all color indices demonstrated significant correlations (p-value< 0.05) with LNC, with the exception of G/R ( Figure 4 ). The NBI index exhibited the highest positive correlation with LNC, achieving a coefficient of 0.86^**, while other color indices showed negative correlations ( Supplementary Table 1 ). At the flowering stage, color indices displayed slightly lower correlations with LNC than those observed at the booting stage. NBI and G/B exhibited the highest positive and negative correlations with LNC, achieving coefficients of 0.86^** and -0.86^**, respectively. Furthermore, G/R and GMR demonstrated no significant correlations with LNC. At the filling stage, NRI revealed the highest correlation with LNC, achieving a coefficient of -0.75^**. Additionally, all color indices demonstrated significant correlations with LNC in rice at the filling stage.

Subsequently, these color indices were utilized to construct SMLR models for LNC estimation, as described in Section 2.4. The selected SMLR models at the leaf scale demonstrated remarkable performance across both single-stage and combined-stages datasets. Statistical indicators such as R², RMSE, and NRMSE ranged from 0.84 to 0.87, 0.16% to 0.25%, and 6.85% to 10.79%, respectively, across various growth stages ( Figure 5 ).

To evaluate model generalizability across spatial heterogeneities, cross-site validation was performed using models trained on Pukou site data and tested on the independent Liuhe site. Performance metrics are stratified by spatial scale (leaf, canopy, plot) and flight altitude (for plot-scale models: 50 m, 100 m, 150 m), as summarized in Supplementary Table 5 .

For the leaf scale, Pukou-trained models showed strong cross-site performance, with training R² of 0.72, RMSE of 0.32%, and NRMSE of 12.3%. Test results were R² of 0.59, RMSE of 0.37%, and NRMSE of 16.8%, demonstrating robustness against inter-site spectral variability. At the canopy scale, transferability declined: training yielded R² of 0.63, RMSE of 0.34%, and NRMSE of 15.1%, while test values dropped to R² of 0.45, RMSE of 0.48%, and NRMSE of 19.8%. Plot-scale performance proved altitude-sensitive: at 50 m, test R² was 0.40 with RMSE 0.47% and NRMSE 21.4%. Slight improvements were observed at 100 m (R² = 0.39, RMSE = 0.46%, NRMSE = 21.1%), followed by a marginal decline at 150 m (R² = 0.38, RMSE = 0.46%, NRMSE = 20.9%), likely due to reduced spatial resolution limiting fine-scale nitrogen detection.

This study highlights the critical role of spatial scale in determining the efficacy of RGB imaging for rice LNC assessment. At the leaf scale, the SMLR models achieved superior accuracy (R² = 0.84–0.87) compared to canopy- and plot-scale approaches ( Figure 9 ). This advantage stemmed from stringent laboratory conditions: leaves were flattened against the scanner surface, ensuring uniform orientation and lighting, while the white background facilitated near-perfect segmentation via Otsu’s method. By eliminating field variability (e.g., shadows, soil interference), leaf-scale imaging isolated spectral responses directly linked to nitrogen content. These findings align with Saberioon et al. (2014), who demonstrated that controlled leaf-scale imaging enables rapid nitrogen diagnostics, though their focus on single leaves contrasts with our multi-stage integration. The cross-site validation (Section 3.5) further supported the leaf-scale model’s generalizability, showing relatively stable performance when applied to a different site. This indicates that the leaf-scale model can capture fundamental spectral-nitrogen relationships that persist across environments, making it promising for broader applications.

At the​canopy scale, the high-resolution DSLR camera (26 megapixels) captured intricate canopy details but introduced noise from heterogeneous backgrounds (soil, water residues). Model performance exhibited phenological dependency: flowering and filling stages achieved higher accuracy (R² improvement > 15% vs. booting stage), likely due to stabilized canopy architecture. During booting, dense foliage and vertical leaf angles increased intra-canopy shadowing, reducing spectral discriminability. This aligns with Bendig et al. (2015), who reported similar structural complexity challenges in barley biomass estimation. Additionally, visible reflectance saturation at booting stage—attributable to high leaf area index (LAI > 4.5) and chlorophyll density—limited model sensitivity, as observed in rice nitrogen accumulation studies (Zheng et al., 2018). Such saturation phenomena are well-documented in crops with overlapping leaves, underscoring the need for phenology-specific calibration.

Plot-scale UAV imaging faced unique tradeoffs between spatial resolution and environmental interference. At 50 m altitude, bidirectional reflectance under solar noon conditions necessitated intensive image stitching to mitigate angular artifacts. Conversely, 150 m altitude introduced motion blur and chromatic aberrations in consumer-grade cameras, degrading spectral fidelity. The 100 m altitude optimally balanced resolution and stability, yielding consistent accuracy (NRMSE< 16%). These observations corroborate Rasmussen et al. (2016), who identified solar angle variations as a key uncertainty in UAV-based vegetation indexing. Post-segmentation GMR thresholding further enhanced accuracy by suppressing soil/water interference, validating its utility in operational settings.

Collectively, these results establish a scale-dependent hierarchy: leaf-scale precision, canopy-scale field adaptability, and plot-scale scalability. Each scale addresses distinct agricultural needs, from high-stakes breeding trials to regional fertilizer management.

To assess the sensitivity of SMLR models to GMR segmentation thresholds in rice LNC estimation, we systematically evaluated threshold values from 0 to 20 in 5-unit increments. This analysis excluded leaf-scale imagery due to its uniform white background, which permits reliable segmentation via Otsu’s method without requiring advanced techniques. Wang et al. (2013) established the efficacy of GMR segmentation for rice-background isolation in digital imagery, providing a methodological foundation for our canopy- and plot-scale analyses.

At the canopy scale, GMR thresholds > 5 during the booting stage effectively differentiated rice canopies from complex backgrounds (soil, water residues) in intact images ( Figure 3 ). Paradoxically, post-segmentation SMLR models exhibited reduced accuracy compared to pre-segmentation models ( Table 6 ), suggesting that canopy-level segmentation may filter ecologically relevant spectral variation. This observation aligns with Lee and Lee (2013), who demonstrated stronger correlations between VIs and crop parameters in non-segmented datasets, implying that segmentation disrupts natural reflectance gradients associated with canopy density and microtopography. The challenge intensified during flowering and filling stages, where panicles and senescing leaves exhibited spectral similarities ( Figure 3 ), complicating pixel classification. However, combined-stage models demonstrated notable accuracy improvements post-segmentation (R² increase ≤ 9.26%, Table 6 ), likely because multi-temporal calibration compensates for illumination variations and structural heterogeneity. This compensation mechanism aligns with the principle that temporal integration buffers single-stage segmentation errors (Zheng et al., 2018).

At the plot scale, it is notable that most researchers have not incorporated UAV-based image segmentation into their models for crop parameter estimation (Li et al., 2016). This oversight may lead to inaccuracies in parameter estimation. In this study, we adopted specific GMR threshold values to segment images obtained from different flight altitudes, aiming to enhance the accuracy of LNC estimation. This approach allowed us to address the variations in lighting and perspective associated with different altitudes. During the reproductive stage, rice achieves nearly full coverage, creating significant shading effects. The darker regions in the plot images were primarily due to the lower parts of the rice canopy being shaded by the upper parts (Wang et al., 2013). This shading effect poses challenges for accurate image segmentation and analysis. When the GMR threshold value exceeded 0, darker areas were classified as background. After segmentation, the optimal SMLR models showed some enhancement in the reliability of the assessments ( Table 6 ), indicating the potential benefits of segmentation in improving model accuracy. However, ANOVA revealed no significant differences in the R² values obtained from various altitudes at the same growth stage after GMR segmentation (p > 0.05). This suggests that while segmentation can enhance model reliability, its impact may be limited by other factors such as altitude and growth stage.

In summary, our findings indicate that while GMR segmentation can enhance the reliability of SMLR models for LNC estimation, its effectiveness varies across different spatial scales and growth stages. This variability underscores the need for tailored segmentation approaches when monitoring rice nitrogen status.