˅**3DQLFOHLQLWLDWLRQ**

*0.01*

*0.02*

*0.18*

*0.05*

*0.15*

*0.08*


**Figure 2.** *Cont.*

*0.01*

*0.05*

*0.04*

*0.03*

*0.08*

*0.08*

*0.00*

*0.00*

*0.04*

*0.01*

**GFEDGFEDGFEDGFEDGFEDGFEDGF%DGFED**


˅**+HDGLQJ**

**Figure 2.** Mean value comparisons for each Multiplex index at the three growth stages and each N application rate (kg N ha−1) for variety KY 131 in 2013. Means and standard errors are shown in each cell for N rate treatments with each measurement mode (Above Canopy, AC; On-the-go, OG; Leaf Scale, LS). Different lowercase letters at the bottom of the plot at each growth stage indicate significant differences according to the least significant difference test at P ≤ 0.05.

#### *3.2. Changes in Multiplex Indices ("On-The-Go" Mode) over Growth Stages under Di*ff*erent N Supplies*

The SFR\_G, SFR\_R, NBI\_G, and NBI\_R indices demonstrated an increasing trend as N rate increased, while a decreasing trend was shown for FLAV (Figure 2). Comparatively, the ANTH and BRR\_FRF values were less sensitive to the changes in N rates. The values of these SFR\_G, SFR\_R, NBI\_G, and NBI\_R indices increased from the PI to SE stage but decreased slightly from the SE to HE stage, because the panicle formation decreased the chlorophyll/N concentration in the upper layer at the HE stage. The opposite was true for the BRR\_FRF, FLAV, and ANTH. NBI\_G and NBI\_R could differentiate different N application rates the best regardless of the growth stages, followed by SFR\_G, SFR\_R, BRR\_FRF, and FLAV. The performance of ANTH was the worst (Figure 2).

#### *3.3. Correlations between Multiplex Indices ("On-The-Go" Mode) and N Status Indicators*

The linear regression results of the seven Multiplex indices and the five N status indicators at three growth stages across the two rice varieties are shown in Table 4. The SFR\_G, SFR\_R, NBI\_G, and NBI\_R indices were positively correlated with the N indicators whereas the BRR\_FRF, ANTH, and FLAV were inversely associated with them. The R<sup>2</sup> of the regression models based on these indices varied from 0.03 to 0.78. The best performing index varied at different stages, but NBI\_G and NBI\_R showed steady high correlations with all five N status indicators. The second-best performing indices were BRR\_FRF and FLAV. The SFR\_G and SFR\_R indices displayed high or moderate correlations with the N indicators during the PI or HE stage, respectively.


**Table 4.** The coefficients of determination (R2) for the linear relationships between standard and normalized Multiplex indices and N status indicators (leaf N

Compared to the counterpart of the standard indices, the normalized su fficiency indices SFR\_GNSI, SFR\_RNSI, and ANTHNSI exhibited better linear relationships with LNC, PNC, and NNI in most of the cases, especially at the SE and HE stages. The NBI\_GNSI and NBI\_RNSI displayed enhanced relationships with the LNC and PNC at the PI and HE stages, and with the NNI at the HE stage. The BRR\_FRFNSI, FLAVNSI showed improved associations with PNC at the PI and HE stages, and with NNI at the SE and HE stages. All the standard indices showed moderate–high relationships with the AGB and PNU during the PI and HE stages, while at the HE stage, greatly improved R<sup>2</sup> values were obtained using the normalized indices.

#### *3.4. Validation of the Estimation Models for N Status Indicators*

In order to diagnose rice N status, linear regression models between the Multiplex indices and N indicators were established. The regression models varied across growth stages. Table 5 lists the best performing models at the PI, SE, and HE stages. The best performing indices di ffered across the stages. However, the relationships of NBI\_G and NBI\_R with N indicators were relatively more stable. After normalization, the NBI\_RNSI showed an absolute advantage for N status estimation at the PI and SE growth stages, while the FLAVNSI demonstrated to be optimal for estimating most of the N indicators.

**Table 5.** Equations and coe fficients of determination of linear regression models (n = 40) at di fferent growth stages based on the best performing Multiplex index and crop N indicators (LNC, PNC, NNI, PNU, and AGB).


Figure 3 shows the RE values of the validation models for six Multiplex indices (SFR\_G, SFR\_R, BRR\_FRF, FLAV, NBI\_G, and NBI\_R) and the N status indicators. The RE values for AGB and PNU estimations based on these indices decreased steadily with advancing growth stages, while a slightly increasing trend was observed for the LNC and PNC estimation models from the SE to HE stage. The RE values for LNC (4.50%–10.24%) and PNC (5.87%–10.87%) models were much smaller than those for AGB (15.49%–30.18%) and PNU (19.31%–31.25%), while the REs of NNI remained similar during the three growth stages. At the earlier to middle growth stages, NBI\_R and NBI\_G presented a lower RE than the other four indices for all the five N indicators. At the HE stage, however, the prediction accuracies of the six indices were similar.

**Figure 3.** The relative error (RE) values of the validation analysis based on the regression models of the six Multiplex indices and the N status indicators for (**a**) above-ground biomass (AGB), (**b**) plant N uptake (PNU), (**c**) leaf N concentration (LNC), (**d**) plant N concentration (PNC), and (**e**) nitrogen nutrition index (NNI) at the panicle initiation (PI), stem elongation (SE), and heading (HE) stages.

#### *3.5. Rice N Status Diagnosis*

The best performing indices including SFR\_G, BRR\_FRF, NBI\_G, NBI\_R, NBI\_GNSI, NBI\_RNSI and FLAVNSI were validated using independent data sets (Table 5). Moderate–high model performance with R<sup>2</sup> ranging from 0.34 to 0.82 was observed, especially for the LNC, PNC, and NNI estimations. The areal agreemen<sup>t</sup> and Kappa statistics were compared at the critical N fertilizer application stages (SE and HE) to evaluate the N diagnostic accuracies of the indices. Results confirmed that the NNI models based on NBI\_R and NBI\_G performed consistently well at the SE and HE growth stages, and their corresponding NSI indices further improved the results (Table 6). At the SE stage, the NBI\_RNSI achieved the highest diagnostic accuracy (areal agreemen<sup>t</sup> = 90%; Kappa = 0.84), while the best accuracy was achieved by FLAV at the HE stage (areal agreemen<sup>t</sup> = 90%; Kappa = 0.76). In addition, across the two growth stages, the NBI\_RNSI showed the highest diagnostic consistency, followed by the BRR\_FRFNSI.


**Table 6.** Agreement and Kappa statistics for different indices (SFR\_G, SFR\_R, BRR\_FRF, FLAV, ANTH, NBI\_G, and NBI\_R) and corresponding normalized indices (SFR\_GNSI, SFR\_RNSI, BRR\_FRFNSI, FLAVNSI, ANTHNSI, NBI\_GNSI, and NBI\_RNSI) regarding diagnostic results (Nitrogen Nutrition Index) at different growth stages.

\*\*\*Significantatthe0.001level;\*\*Significantatthe 0.01 level;\*Significantatthe0.05level;NSNotsignificant.

## **4. Discussion**

#### *4.1. Multiplex Measurement Modes and Estimation of Crop N Indicators by Fluorescence Indices*

The N treatment effects were more significant for the readings obtained in the OG mode than those collected in the AC mode (Table 3, Figure 2), which is different from the finding by Diago et al. [39] who reported a 20% loss of information occurred when using the Multiplex on-the-go (compared to the AC mode) for N assessment of grapevine. This is because the OG measurements in this study were taken manually by placing the Multiplex sensor right on the top of the rice rather than a small distance above the rice canopy while passing through the rice paddy. In contrast, in the study by Diago et al. [39], the Multiplex sensor was mounted onto an all-terrain vehicle and placed 1.5 m above the ground so that the leaves on the mid-part of the canopy were automatically measured at a 20 cm distance, the same measuring distance as their AC mode. In addition, this study revealed that measurements made using the LS mode were the least sensitive to N supply, contrasting to the result by Zhang et al. [41] who found Multiplex measurements made from corn leaves were more capable of distinguishing plant N status than those made from above the plants. While the leaf scale measurements made in this study were collected in the laboratory by taking ten leaves in the second position from the top, the leaf-borne measurements by Zhang et al. [41] were made on 20 representative plants in the center two rows of each plot in the field, which is more similar to the OG rather than the LS method in this study. Another possible reason for their better results with the LS method is that the OG measurements do not give much time to choose leaves and result in more random leaf choosing than the LS method, which may have an unwanted tendency to choose "good" leaves. This is particularly true for maize, because systematic use of a representative leaf is easier, as it is well known which is the most representative leaf for each growth stage, given its determined growth. For rice, the individual leaves are quite small and the signal obtained during the measurements taken on a leaf is relatively weak and can be easily affected by other factors. One advantage the "measuring in motion" or OG mode has is efficiency, especially when the sensor is mounted on a vehicle or other automatic devices, which might make practical applications of such non-destructive technology over large areas possible. Bringing the sensor close to or even touching the leaves of the crop in OG mode may help reduce information loss. However, further well-designed studies are needed to confirm this finding.

Strong relationships between the Multiplex indices (SFR\_G, SFR\_R, BRR\_FRF, FLAV, NBI\_G, and NBI\_R) measured in the OG mode and the five N indicators were achieved with low RE and high R<sup>2</sup> values (Table 4, Figure 3). This finding conforms to previous research results in this field [26,42,55,56]. Many studies confirmed SFR was a good fluorescence index for chlorophyll content monitoring [24,38,57]. However, in this study, it was found that the R<sup>2</sup> of the SFR\_G, SFR\_R for LNC, PNC, and NNI estimation decreased steadily from the early-stage to later stages, while an opposite trend was observed for FLAV. Padilla et al. [55] found that the relationships between the NNI and SFR\_G changed with the phenological stages of cucumber (*Cucumis sativus* L.). Firstly, the consistency of the relationship between chlorophyll content and N concentration varied with crop development,

leading to different performances of SFR for N concentration estimation. For example, the linear correlation between LNC and chlorophyll meter readings of rice was weaker at the SE stage than at other growth stages [58]. Secondly, the weaker differentiation ability of SFR under the unlimited N conditions may also be a reason [55]. The performance of FLAV increased from the PI to HE stages, which was confirmed by Padilla et al. [59], who found the relationship of FLAV and NNI increased at the middle to late growth stages. The better performance of FLAV at the later stage may be attributed to the accumulation of the flavonols content in leaves under light radiation [44,60]. The NBI\_G and NBI\_R indices were shown as the best indices for estimating the N indicators (Table 5). Many studies have proven that the NBI indices appeared to be the most efficient in estimating the N status [47,56,59,61]. This is because the NBI is a ratio of SFR and FLAV, which makes it more robust than using FLAV or SFR alone to reduce the effects of leaf age or other factors [34,36,47]. The NBI\_G and NBI\_R had similar performance in this study, as demonstrated by Longchamps and Khosla [27]. Moreover, Longchamps and Khosl [27] found that SFR was less sensitive to N application than NBI, which conforms to our results, as shown in Figure 2. In most cases, the SFR\_G and SFR\_R indices could not distinguish between the 100 and 130 kg N ha−<sup>1</sup> treatments, but NBI could. The BRR\_FRF index was significantly correlated with the N nutritional status and was especially sensitive to N deficiency in this research. When there is N stress, the fluorescence ratio of blue–green/far-red will increase after exposure to elevated UV radiation to avoid or alleviate the damage of the photosynthetic apparatus [62]. Generally, the UV-protection response takes place before the chlorophyll damage can be seen, so the BRR\_FRF can also be considered as a potential index that can realize early N deficiency detection [63]. The BRR\_FRF was also very sensitive to environmental stresses, such as disease and drought [38,63,64]. The ANTH index provided by Multiplex is commonly used to reflect anthocyanin content, which corresponds to the maturation degree of fruit [49,65]. In this research, the low values of ANTH were due to the low anthocyanin content in the rapid vegetative growth phase for rice [66]. Nevertheless, ANTH was also found to be closely related to the leaf chlorophyll concentration in some studies [38,41]. This study revealed that ANTH was significantly related to N status indicators in PI and HE growth stages with moderate R<sup>2</sup> values (Table 4).

#### *4.2. Normalized Nitrogen Su*ffi*ciency Fluorescence Indices*

Our research involved two years and two varieties of experiments. In these experiments, N fertilization rate is the main variable, and the variation is so high that it will probably override any other source of variation. In a commercial field, many factors can influence N availability, N and Chlorophyll relationship, or Chlorophyll (+FLAV and ANTH) relations to fluorescence indices, including biotic or abiotic stresses at the moment of measurement or in the history of the crop or even the field. The normalized N sufficiency index approach has been suggested to reduce the influence of the varieties, developmental stages, and other variables on SPAD values or spectral data [3,11,67]. From the results of this study, in most cases, the normalized NSIs were better associated with the LNC, PNC, and NNI (Table 4). The R<sup>2</sup> of the ANTHNSI was improved the most, followed by the NBI\_GNSI and NBI\_RNSI. However, the improvement in R<sup>2</sup> for BRR\_FRFNSI was minimal. The variance analysis of this study showed consistent results, which demonstrated that NSI indices could reduce the influence of inter-annual and growth stage differences. Since NNI itself is a diagnostic criterion, it represents an optimal N status when NNI is equal to one [68]. Most of the NSI indices greatly improved the NNI diagnostic accuracy at the critical topdressing (SE and HE) stages (Table 6). Similarly, Lu et al. [43] observed that the NNI inversion through the normalized vegetation indices further improved the N nutrition diagnostic results of rice.

Hussain et al. [69] proposed a critical NSI value of 0.90 for rice. However, in this study, when the NSI indices were 0.90, different optimal NNI values, ranging from 0.85 to 2.14, were derived by different indices at different developmental stages. Only the corresponding optimal NNI values for the NBI\_GNSI and NBI\_RNSI indices were close to one (ranging from 0.91 to 1.19). Therefore, to avoid the risk of misdiagnosis, the NSI threshold was not used as a diagnostic criterion directly. A possible

reason for this is that the N fertilizer application rate in this study was only 1.3–1.6 times higher than the optimal amount instead of 1.8–2.0 times higher than recommended for the well-fertilized N plot as Hussain et al. [69] suggested. Furthermore, all of the Multiplex indices were divided by the readings of the N rate with the largest shoot dry matter at each sampling date to obtain a sufficiency index. However, Varvel et al. [70] suggested that the maximum readings within each cropping system, variety, and year should be considered as the normalized criterion. Obviously, with different normalization criteria, different sufficiency indices will be obtained, which will affect the corresponding analysis results. Another limitation of the NSI approach is that well-fertilized reference plots need to be established in each farmer's field for practical application purposes, and some farmers may not be willing to do this. More in-depth and systematic research is expected in the future.

#### *4.3. The Application Potential of the Multiplex Sensor*

The Multiplex indices presented good R<sup>2</sup> values for LNC and PNC estimations at the earlier growth stages (Table 4). In particular, the validation data showed that the RE values for LNC and PNC estimations were as low as 6%–7% (Figure 3c,d). This is consistent with the results of Cerovic et al. [71] and Agati et al. [56], who have shown a high correlation between the fluorescence index and LNC. NBI and LNC had a fairly linear relationship. Therefore, the NBI indices can be used to more accurately estimate a wider range of LNC. Agati et al. [56] also found the results based on reflectance imaging (camera picture) are less sensitive to N application than fluorescence-based indices. Research by Stroppiana et al. [72] and Yu et al. [19] on rice showed unsatisfactory results for the estimation of LNC and PNC based on reflectance spectroscopy. This is possibly due to the fact that the effect of N on the leaf area index and biomass is much greater than its effect on chlorophyll content. Second, near-infrared radiation is hardly absorbed in the canopy and is highly transmissive, so its correlation with leaf area index or biomass is extremely high; while visible light, especially the blue and red radiation, is easily absorbed by chlorophyll and its transmittance is low, so it is highly correlated with chlorophyll content [15,72]. On the other hand, changes in plant metabolism indicators are fast or slow due to changes in response to the environment. However, the sensitivity of reflectance-based parameters does not always provide satisfactory monitoring results [73]. Demotes-Mainard et al. [74] observed that changes in N concentration took precedence over changes in biomass. Thus, fluorescence-based techniques that are highly sensitive to plant N status information may address the limitation of reflectance-based methods [27,73]. Similarly, the Multiplex indices, especially the NBI\_G and NBI\_R, presented accurate estimation for NNI, with R<sup>2</sup> reaching a maximum of 0.72–0.78, and the validation results also showed a low inversion error for NNI (RMSE ≤ 0.16, RE ≤ 15%) (Table 4). Many studies have confirmed that NBI has a strong estimation potential for NNI [47,55,59]. This is because NBI is the ratio of SFR to FLAV. The SFR index was considered to be an important parameter for estimating chlorophyll concentration, which was often used as an index of surface-based N [75], while the FLAV parameter directly reflects flavonol content, which is controlled by light as well as leaf mass per area, and has a strong correlation with leaf mass [76]. Therefore, NBI as the SFR/FLAV ratio is the best N nutrition diagnostic index.

The fluorescence-based indices are more sensitive to chlorophyll or N content than the reflectance-based indices and can detect the difference in N nutrition status earlier. However, the difference of the stage-based models between the indices and the N nutrition indicators based on the canopy reflectance instrument is smaller than that based on fluorescence [59]. The surface area of the crop involved in each test when using the canopy reflectance spectroscopy sensors is larger than the fluorescence sensors [59]. Therefore, canopy reflectance measurements are more representative, while fluorescence instruments require increasing the number of tests to obtain sufficiently representative data. Although the performance of the fluorescence sensor was quite good for estimating LNC, PNC, and NNI, 22%–60% of their variability was still not explained. In addition, the fluorescence sensor did not perform very well for estimating plant biomass in the middle to late stages (Table 4). It has

been suggested to combine the fluorescence and reflectance data to improve the estimation of plant N status [3,73]. This may be one of the important research directions in the future.

## **5. Conclusions**

This research compared the LS, OG, and AC measurement modes of the fluorescence instrument Multiplex ®3 and determined that the OG mode was best suited for this rice N status study. Using the OG mode, stable test results and crop growth information were derived. The results revealed that the fluorescence indices of NBI, SFR, BRR\_FRF, and FLAV were significantly correlated to all five N status indicators from the PI through HE growth stages. Among them, NBI\_G and NBI\_R were the best performing indices and highly correlated to LNC (R<sup>2</sup> = 0.52–0.68), PNC (R<sup>2</sup> = 0.52–0.71), NNI (R<sup>2</sup> = 0.69–0.78), AGB (R<sup>2</sup> = 0.47–0.64), and PNU (R<sup>2</sup> = 0.68–0.72) at the three growth stages. The normalized su fficiency indices of the Multiplex parameters could greatly improve the LNC, PNC, and NNI estimation ability, especially at the HE stage. The N diagnostic results indicated that the NBI\_RNSI and FLAV achieved the highest diagnostic accuracy rate (90%) at the SE and HE stage, respectively, while NBI\_RNSI showed the highest analytical consistency across growth stages. The results sugges<sup>t</sup> that the Multiplex sensor can be used to reliably estimate N nutritional status for rice in cold regions, especially for the estimation of LNC, PNC, and NNI. The normalized su fficiency indices based on Multiplex indices may further improve the accuracy of N nutrition diagnosis by reducing the di fferences between years and varieties.

**Author Contributions:** Y.M. and G.B. conceived and guide the study. S.H. and Q.C. conducted the field experiments. S.H. and H.Y. performed the data analysis. S.H. wrote the original manuscript. Y.M., F.Y. and V.I.S.L.-W. revised the manuscript. G.B. and H.Y. reviewed and edited the manuscript. All authors read and approved the final manuscript.

**Funding:** This research was financially supported by the National Key Research and Development Program of China (2016YFD0200600, 2016YFD0200602), National Basic Research Program (2015CB150405), and the Norwegian Ministry of Foreign A ffairs (SINOGRAIN II, CHN-17/0019).

**Acknowledgments:** We would like to thank the supports by Wen Yang, Huamin Zhu, and Fengyan Liu at the Jiansanjiang Institute of Agricultural Sciences. We also would like to thank Jianning Shen, Weifeng Yu, and Shanshan Cheng for their fieldwork and contributions in spectral data collection.

**Conflicts of Interest:** The authors declare no conflict of interest.
