Sensitivity of Vegetation Indices for Estimating Vegetative N Status in Winter Wheat

Precise sensor-based non-destructive estimation of crop nitrogen (N) status is essential for low-cost, objective optimization of N fertilization, as well as for early estimation of yield potential and N use efficiency. Several studies assessed the performance of spectral vegetation indices (SVI) for winter wheat (Triticum aestivum L.), often either for conditions of low N status or across a wide range of the target traits N uptake (Nup), N concentration (NC), dry matter biomass (DM), and N nutrition index (NNI). This study aimed at a critical assessment of the estimation ability depending on the level of the target traits. It included seven years’ data with nine measurement dates from early stem elongation until flowering in eight N regimes (0–420 kg N ha−1) for selected SVIs. Tested across years, a pronounced date-specific clustering was found particularly for DM and NC. While for DM, only the R900_970 gave moderate but saturated relationships (R2 = 0.47, p < 0.001) and no index was useful for NC across dates, NNI and Nup could be better estimated (REIP: R2 = 0.59, p < 0.001 for both traits). Tested within growth stages across N levels, the order of the estimation of the traits was mostly Nup ≈ NNI > NC ≈ DM. Depending on the number (n = 1–3) and characteristic of cultivars included, the relationships improved when testing within instead of across cultivars, with the relatively lowest cultivar effect on the estimation of DM and the strongest on NC. For assessing the trait estimation under conditions of high–excessive N fertilization, the range of the target traits was divided into two intervals with NNI values < 0.8 (interval 1: low N status) and with NNI values > 0.8 (interval 2: high N status). Although better estimations were found in interval 1, useful relationships were also obtained in interval 2 from the best indices (DM: R780_740: average R2 = 0.35, RMSE = 567 kg ha−1; NC: REIP: average R2 = 0.40, RMSE = 0.25%; NNI: REIP: average R2 = 0.46, RMSE = 0.10; Nup: REIP: average R2 = 0.48, RMSE = 21 kg N ha−1). While in interval 1, all indices performed rather similarly, the three red edge-based indices were clearly better suited for the three N-related traits. The results are promising for applying SVIs also under conditions of high N status, aiming at detecting and avoiding excessive N use. While in canopies of lower N status, the use of simple NIR/VIS indices may be sufficient without losing much precision, the red edge information appears crucial for conditions of higher N status. These findings can be transferred to the configuration and use of simpler multispectral sensors under conditions of contrasting N status in precision farming.


Introduction
Optimizing nitrogen (N) supply in wheat production is essential, both economically and for reducing environmental side effects from excessive N supply [1][2][3][4]. Unlike under West European conditions, optimum N fertilization levels are often largely exceeded in China [5][6][7] where only about 30% of the applied N is removed through the produce [8]. In contrast, the prevailing N intensities are without discussing the level and range of the target traits, which are relevant in practice. The latter are however crucial for estimating the N status in high-yielding environments and for potentially identifying effects of over-fertilization within a limited data range [35]. In addition, the disturbing effect of cultivar-specific morphology and phenology can affect the use of spectral methods [28]. However, the effect was rarely compared, while mostly the relationships were tested either for one cultivar or across cultivars. Comparing index relationships both within and across morphologically differing German and Chinese cultivars, often weaker relationships were observed across cultivar groups [28]. Due to the complex effects, further empirical testing under varying conditions is required.
Using seven years' data of an N fertilization trial conducted in high-yielding West-European conditions, this study aimed at comparing selected vegetation indices for estimating wheat traits related to the vegetative and pre-flowering canopy N status. Traits included NNI, Nup and their contributing traits DM and NC. The objective of the study was to investigate empirically the index rankings and the trait estimability as influenced by (i) years and growth stages, (ii) cultivars within years and (iii) by the N status and the absolute level of the target traits.

Field Trials
The experiment was conducted over seven years in 2009-2011, 2013, 2014, 2016, and 2018 in southeast Germany (48.406 N, 11.692 E) under rain-fed conditions to evaluate the N response in biomass and nitrogen accumulation, and sensor-based detection of over-fertilization. The soil was mostly homogeneous Cambisol of silty clay loam with a pH of 6.4, K 2 O-content of 12 mg 100 g −1 , P 2 O 5 -content of 12 mg 100 g −1 and C org of 1.2%. The annual precipitation in this region is approximately 800 mm with an average temperature of 8 • C. In all years, N fertilization was differentiated in eight N levels, ranging from 0 to 420 kg N ha −1 , incrementally increasing by 60 kg N ha −1 . According to the local practice, N application was split into three doses. N was applied at the beginning of vegetation in spring, at early stem elongation, and at booting-flowering (Table 1) . Cultivars were arranged in rows of eight plots within the replicates ('main plots') for technical reasons. In 2014, chemical straw-shortening growth regulator (GR) was included as an additional factor (treated/non-treated). In the other years, all plots were treated with Chlormequate-based straw-shortening according to local practice. Depending on the disease incidence, adapted fungicide treatments were applied. Sowing density was 300-350 kernels m −2 . Adequate amounts of P, K, Mg, and S were supplied.

Data Acquisition and Calculations
The total above-ground biomass was sampled at differing growth stages and at differing frequencies: Zadok's growth stage (ZGS) 32 Table 2).The sampling area per plot differed between years from 0.13 m −2 in 2010 to 2.7 m −2 in 2011. After drying at 60 • C, dry matter weight (DM [kg ha −1 ]) was determined through weighing. Nitrogen concentration (NC [%]) was determined using a ratio mass spectrometer with an ANCA SL 20-20 preparation unit (Europe Scientific, Crewe, UK) from 2009 to 2014. In the following years, near-infrared spectroscopy (NIRS) using a FOSS NIRS 6500 (NIRSystem, Silver Spring, Md.) and an FT-NIRS (Bruker, MPA, Germany) was used instead. DM and NC values were combined for the NNI as NNI = NC/(5.35 × DM (−0.442) ) [16] and Nup [kg ha −1 ] as Nup = DM × NC. Spectral measurements were conducted for each sampling date using the mobile sensor platform Phenotrac (I-IV). It was equipped with the same hyperspectral passive bidirectional sensor (tec5, Oberursel, Germany) [22], which measures at a nominal resolution of 3.3 nm and was used from 350 to 1000 nm. In 2018, a similarly measuring handheld hyperspectral Handyspec (tec5, Oberursel, Germany) sensor with a nominal resolution of 2.0 nm was used instead. Index comparison indicated a very close agreement between both sensors.
Spectral vegetation indices (SVI) were selected based on previous evaluation on similar traits under comparable growth conditions [22,29,30] (Table 3). The selection was limited to the REIP and simple ratio indices which were found to be more sensitive than normalized indices in dense canopies [31]. Table 3. List of vegetation indices tested in this study.

Data Analysis
SVIs and reference data were submitted to analysis of variance and treatment means were compared using Tukey's HSD test. SVIs were tested in regression analysis with the target traits. The following approaches were applied using R version 3.4.2 [40]: (i) Regression across the full data from different years and growth stages testing quadratic indext rait relationships. The minimum trait value with a non-positive slope of the index~trait relationship was extracted as a 'saturation point'. In addition, fitted values with negative slopes were replaced with the fitted value reached for the saturation point, resulting into quadratic + plateau curves ( Figure 2) to avoid counterintuitive overfitting with declining index values for high reference trait values. The noise equivalent (NE) was calculated as root mean squared error (RMSE) divided by the 1 st derivative of the index values over the reference traits [28,41].
(ii) Linear or quadratic trait~index relationships selected based on the Akaike information criterion (AIC) within dates across main plot treatments (1-3 cultivars or two levels of growth regulators in 2014 only).
(iii) Linear and quadratic trait~index relationships selected based on the AIC within dates within main plot treatments.
(iv) Linear relationships within two intervals of the target traits, referring to low N status of plots with NNI < 0.8 and high N status of plots with NNI > 0.8.
Relationships were compared by coefficients of determination (R 2 ), RMSE, and mean-and range-normalized RMSE.

Treatment Effects on Target Traits and Indices
Owing to the different growing conditions together with the different phenology of the cultivars, growth stages differed markedly across years. In addition, the levels of reference plant traits differed substantially by growth stages by years ( Figure 1; Supplementary Table S1; Supplementary Table S2). DM ranged from an average across all treatments of 1. Quadratic response of reference traits to incremental N fertilization (N levels) by sampling dates (year/month/day) and main plot treatments. Nitrogen was applied incrementally increasing by 60 kg ha −1 from 0-420 kg N ha -1 in three doses (Table 1). For nitrogen nutrition index (NNI), the threshold used (NNI </> 0.8) for dividing the data into two intervals is drawn as a horizontal line. Interval 1 (NNI < 0.8) and interval 2 (NNI > 0.8) are indicated as circles and rectangles, respectively. See Supplementary Figure S1 for the N-response of the index R760_730. Figure 1. Quadratic response of reference traits to incremental N fertilization (N levels) by sampling dates (year/month/day) and main plot treatments. Nitrogen was applied incrementally increasing by 60 kg ha −1 from 0-420 kg N ha −1 in three doses (Table 1). For nitrogen nutrition index (NNI), the threshold used (NNI </> 0.8) for dividing the data into two intervals is drawn as a horizontal line. Interval 1 (NNI < 0.8) and interval 2 (NNI > 0.8) are indicated as circles and rectangles, respectively. See Supplementary Figure S1 for the N-response of the index R760_730.

Relationships across Measurement Dates: Noise Equivalent, R 2 and Saturation Points
The index~trait relationships across all dates revealed distinct differences (Table 4). For DM, only the R900_970 provided useful relationships (R 2 = 0.47), followed by the red edge indices. In contrast, the NIR_green and NIR_red indices scattered strongly between dates (not shown). For NC, none of the indices showed a unique relationship without decreasing index values for high NC values (max. R 2 = 0.24 of the NIR_green). The different dates were less distortive for Nup and NNI, where for both traits the REIP (R 2 = 0.59) and both red edge simple ratio indices performed best. Table 4. Coefficients of determination (R 2 ) with significance levels (***: p < 0.001; n.s.: p > 0.05) calculated across the total data range from all dates, with non-negative slopes (

Index Comparisons within and across Cultivars by Measurement Dates
Within dates, regression analysis was conducted across N levels. Trait ~ index relationships were tested both across and within the main plot treatments (i.e., cultivars; growth regulator in 2014). The range with plateau-like values was considered to be non-distinguishable by the index. Thus, it was excluded for calculating the noise equivalent (NE). Indices differed in reaching this plateau, indicating differing saturation thresholds (Supplementary Figure S4). Low saturation points were found for NC with a maximum saturation point of 3.4% (NIR_red), accompanied by steeply increasing NE values (Figure 3). In accordance with the R 2 -based ranking, only the R900_970 resolved almost the full data range (relative saturation point 0.98; 12.5 t DM ha −1 ) for DM. Correspondingly, the R900_970 showed the lowest NE values with the relatively flattest increase across the data range ( Figure 3). It was followed by the similar red edge indices and higher NE values of the NIR_green and NIR_red indices. In contrast, for NNI, the index ranking differed in R 2 , NE, and saturation points. In spite of higher saturation points of the NIR_red and NIR_green indices and relatively lower NE ( Figure 3) values in the upper data range, these indices showed stronger scattering in the lower data range with higher NE values and overall lower R 2 -values.
For Nup, most indices reached lower relative saturation points (0.62-0.75) than for NNI, corresponding to 223 kg N ha −1 (REIP, NIR_green) and 265 kg N ha −1 (R780_740). Clearly lowest NE values were observed for the red edge indices.

Index Comparisons within and across Cultivars by Measurement Dates
Within dates, regression analysis was conducted across N levels. Trait~index relationships were tested both across and within the main plot treatments (i.e., cultivars; growth regulator in 2014).
Generally, the influence of measurement dates was larger than the differences by main plot treatments (Supplementary Table S3, Supplementary Table S4).
For all traits, the best relationships were observed in 2014 and 2016 both within and across main plots. 2011 was characterized by non-significant relationships for NC, weak relationships for NNI (max. R 2 = 0.15) and Nup (max. R 2 = 0.31), but medium relationships for DM especially within the main plots. Compared across measurement dates, coefficients of determination (R 2 ) differed overall little between the regressions across main plots and within main plots for DM. For Nup, followed by NNI and especially NC, R 2 -values were clearly lower from regressions across than within main plots    Table S3 and  Supplementary Table S4 for original results.
Generally, the influence of measurement dates was larger than the differences by main plot treatments (Supplementary Table S3, Supplementary Table S4).
For all traits, the best relationships were observed in 2014 and 2016 both within and across main plots. 2011 was characterized by non-significant relationships for NC, weak relationships for NNI (max. R 2 = 0.15) and Nup (max. R 2 = 0.31), but medium relationships for DM especially within the  For the three N-related traits, the index ranking was comparable within and across main plots. Either the REIP or the R760_730 ranked best but both indices performed very similarly and were closely followed by the R780_740 index. The other indices were clearly less suited and the explained variance was approximately 4-10% lower for the N traits. For DM, however, the NIR_red index gave on average best (R 2 = 0.62) and the REIP weakest (R 2 = 0.55) relationships across main plots with a similar performance of the other indices. Compared within the main plots, however, the red edge indices slightly outperformed the others for DM. Linear relationships may be preferable because they are less likely to hide saturation issues. However, for most trait*date*index combinations, the differences in R 2 values between linear and quadratic fits were low (Supplementary Figure S3).
RMSE-values normalized to the mean and the range of the target traits generally resulted in similar index rankings (Figure 4d; Supplementary Table S5) as from the coefficients of determination (Figure 4a). However, the higher mean-normalized RMSE values for Nup than for the other traits were not reflected in lower R 2 -values. In contrast, R 2 values were rather low for NC but mean-normalized RMSE similar as for DM and NNI.

Regressions by NNI-Based Intervals
The NNI was used for dividing the reference trait range into an interval of sufficient/excessive N supply (NNI > 0.8) and an interval with N deficiency (NNI < 0.8) for assessing the specific ability to estimate the target traits under conditions of high N status, within date*main plot combinations ( Figure 5).
between the regressions across main plots and within main plots for DM. For Nup, followed by NNI and especially NC, R 2 -values were clearly lower from regressions across than within main plots (Figure 4), although all target traits differed by cultivars on most dates (Supplementary Table S1). Thus, for NC, R 2 -values decreased by approximately 25% when calculated across main plots. All vegetation indices showed a similar decrease for the N-related traits. In contrast, the NIR_red, NIR_green and R90_970 indices achieved relatively higher R 2 -values for DM across than within main plots.
For the three N-related traits, the index ranking was comparable within and across main plots. Either the REIP or the R760_730 ranked best but both indices performed very similarly and were closely followed by the R780_740 index. The other indices were clearly less suited and the explained variance was approximately 4-10% lower for the N traits. For DM, however, the NIR_red index gave on average best (R 2 = 0.62) and the REIP weakest (R 2 = 0.55) relationships across main plots with a similar performance of the other indices. Compared within the main plots, however, the red edge indices slightly outperformed the others for DM. Linear relationships may be preferable because they are less likely to hide saturation issues. However, for most trait*date*index combinations, the differences in R 2 values between linear and quadratic fits were low (Supplementary Figure S3). RMSE-values normalized to the mean and the range of the target traits generally resulted in similar index rankings (Figure 4d; Supplementary Table S5) as from the coefficients of determination ( Figure  4a). However, the higher mean-normalized RMSE values for Nup than for the other traits were not reflected in lower R 2 -values. In contrast, R 2 values were rather low for NC but mean-normalized RMSE similar as for DM and NNI.

Regressions by NNI-Based Intervals
The NNI was used for dividing the reference trait range into an interval of sufficient/excessive N supply (NNI > 0.8) and an interval with N deficiency (NNI < 0.8) for assessing the specific ability to estimate the target traits under conditions of high N status, within date*main plot combinations ( Figure 5).  Compared by R 2 , RMSE as well as mean-and range-normalized RMSE-values, all traits were better estimated in the lower (interval 1) than in the upper (interval 2) NNI-based trait interval ( Figure 6). In general, the order of the R 2 -values among traits was DM < NC ≈ NNI ≈ Nup in interval 1 and DM < NC < NNI < Nup in interval 2 (Figure 6a). Accordingly, the order by range-normalized RMSE was DM > NC > NNI ≈ Nup in interval 1 but DM > NC ≈ Nup > NNI in interval 2. However, the order by mean normalized RMSE was DM ≈ Nup > NNI ≈ NC in interval 1 and Nup > DM > NNI > NC in interval 2. For DM, all indices performed on average similarly in interval 1 (R 2 = 0.45-0.47, RMSE = 480-520 kg ha −1 ; mean-normalized RMSE ≈ 12%) but differed more in interval 2 with the R900_970 being relatively less suited mainly due to the poor relationships during stem elongation in 2018 ( Supplementary Figures S5 and S6). For all N traits, the REIP and R760_730 ranked highest (average R 2 of the REIP = 0.40 for NNI, 0.46 for NC and 0.47 for Nup) in interval 1. In interval 2, the R780_740 reached slightly higher R 2 -values and lower absolute and normalized RMSE values than the R760_730. While the index discrimination was minor in interval 1, the R 2 , especially of the NIR_red and R900_970 in interval 2, was about 1/3 lower than that of the best-performing indices in interval 2. In spite of the overall weaker estimation in interval 2, the contrary was found for some date*main plot combinations, mainly in 2009 (Supplementary Figure S5). Compared by R 2 , RMSE as well as mean-and range-normalized RMSE-values, all traits were better estimated in the lower (interval 1) than in the upper (interval 2) NNI-based trait interval ( Figure  6). In general, the order of the R 2 -values among traits was DM < NC ≈ NNI ≈ Nup in interval 1 and DM < NC < NNI < Nup in interval 2 (Figure 6 a). Accordingly, the order by range-normalized RMSE was DM > NC > NNI ≈ Nup in interval 1 but DM > NC ≈ Nup > NNI in interval 2. However, the order by mean normalized RMSE was DM ≈ Nup > NNI ≈ NC in interval 1 and Nup > DM > NNI > NC in interval 2. For DM, all indices performed on average similarly in interval 1 (R 2 = 0.45-0.47, RMSE = 480-520 kg ha -1 ; mean-normalized RMSE ≈ 12%) but differed more in interval 2 with the R900_970 being relatively less suited mainly due to the poor relationships during stem elongation in 2018 ( Supplementary Figures S5 and S6). For all N traits, the REIP and R760_730 ranked highest (average R 2 of the REIP = 0.40 for NNI, 0.46 for NC and 0.47 for Nup) in interval 1. In interval 2, the R780_740

Discussion
Saturation of vegetation indices is a frequently discussed limitation for the application of spectral methods in dense canopies [42,43]. This study aimed at assessing the estimation of DM, NC, NNI, and Nup under varying conditions. Dominated by the effect of N fertilization, positive trait~index relationships were found on all dates. An asymptotical negative NC~DM relationship was confirmed across dates, which was converted into the NNI [13,15].

The Effect of Years, Growth Stages and Cultivars
Despite the strong scattering between dates in the regression approach across dates, relevant index~trait relationships were found for NNI and Nup. However, unlike to Cao et al. (2018) [23] but similar to Cao et al. (2013) [24] for rice, more robust relationships across growth-stages for NNI than for Nup were not confirmed. The fact that across years, the lowest R 2 -values were found for NC in spite of mostly close relationships within dates, may indicate that these observed relationships were driven indirectly by the NC's close relationship with DM and Nup ( [25]; Supplementary Figure S2). The strong influence of growth stages for NC estimation is in line with [25]. Due to the year-and stage-specific shifts especially for DM and NC, a global model would still have resulted in substantial absolute errors. Possibly, the canopy water mass as detected by the R900_970 is a more integrative signal compared to structural NIR/VIS and NIR/red edge indices. This may explain that the latter indices were not suited across measurement dates for DM due to the concomitant influence of NC and leaf morphology, while the R900_970 was the least sensitive to NC across dates. For DM, the strongest index differentiation with lower NE of the R900_970 is in line with the previous recommendation of the water band for overcoming saturation in dense canopies [44,45]. Still, phenological shifts, which remain relevant also for the other traits and relationships across growth stages, allow only a relative assessment of the N status if growth differs too strongly [23,46,47].
For the DM and NC, a higher saturation point was associated with a higher coefficient of determination and lower NE whereas for the NNI and Nup, the differing index rankings by saturation point, NE and R 2 -values indicate that the NE alone may be misleading. The higher relative saturation points for DM and NNI than for NC and Nup should not be over-interpreted due to the strong scattering for medium trait values and the few data points contained in the upper DM and Nup range. In contrast, the low index values, which were associated with high NC values, indicate that all indices were dominantly responsive to the low DM/Nup values of these plots. The use of less structure-sensitive indices would have to be tested under such conditions [48].
The predominantly small difference between linear and quadratic relationships (Supplementary Figure S3) within growth stages for red edge indices is in line with [30]. However, it implies that the clearly curvilinear across-dates models may not be suited especially for dense canopies due to the distorting phenological effect. Moreover, the substantial different index rankings in the within-date regressions indicate that especially the across-stage robustness of the R900_970 counteracts with the performance within date*main plot conditions. This is in line with the relatively good performance of the R900_970 across main plots in contrast to within main plots for DM, possibly indicating being less sensitive to morphological differences of the cultivars. Possibly due to the more different growth stages and the different environmental conditions in the trial years, the NE level across dates and growths stages was higher than previously reported for the same target traits [46] and for Nup and NC in [28].
Testing across main plot treatments, notably cultivars, may cause 'stretching' of the data. However, it combines differences in morphology and phenology as well, which explains that the effect differed between traits and measurement dates. The stronger negative effect of main plots on the estimation of N traits than of DM may relate to the influence of differing chlorophyll content between cultivars [49]. The predominantly weaker relationships across than within cultivars are in line with previous findings [27,28,47]. While this aspect is relevant for discriminating genotypes in phenotyping or for the use on a landscape scale, the distortive influence is less of a problem for precision farming with commonly only one cultivar. Still, absolute errors would occur if the models were not adapted for the cultivar-specific influence.

The Estimation Potential Compared by Traits and Statistical Measures
Weaker index-relationships for DM and NC compared to NNI and especially Nup in all approaches (across dates, within dates, and by intervals) are in line with previous findings [22,[28][29][30], and may be caused by the effect of N fertilization, which increased both DM and NC in the same 'direction'. With both traits contributing to the Nup and NNI calculation, this may have additionally 'stretched' these traits. Thus, DM (but not NC) was generally less distinguishable in the post hoc test than Nup and NNI. As the coefficient of determination, the range-normalized RMSE indicated a weaker estimation of DM as well, but a similar estimation of NC and Nup in interval 2 in spite of higher R 2 -values for Nup, as well as better estimation of NNI than that of Nup regardless of a similar R 2 -level. The mean-normalized RMSE resulted in identical index rankings as the absolute and range-normalized RMSE but may be less useful to compare the trait estimability due to differing absolute levels of the traits. Despite the overall useful relationships, the estimation failed for some date*main plot combinations (Supplementary Figure S5)-mainly when the differentiation in the target traits was weak as for the very flat N response of NC in 2011 (Figure 1). Differing sampling areas and occasional temporal shifts between plant samplings and spectral measurements constitute further error sources.
The strong stage-specific clustering in the NC~index relationships could be addressed through combining a structural index as the NDVI with red edge information [28]. On the other hand, the NNI showed the least clustering across stages due to its immanent correction by DM, explaining its better estimation than that of NC. Thus, it may also be preferred for guiding fertilization strategies [15].
Comparing datasets and statistical measures (Supplementary Table S5), R 2 and NE ranked the indices similarly for all traits for the whole data ('across all') and within date*main plots for NC, NNI, and Nup, but not for DM. The similar index rankings for NNI and Nup from all approaches did not go along with constantly weaker performance of the NIR_red and R900_970. However, due to the relatively poorer performance of the R900_970 within the main plots, the rankings within dates differed as tested across and within the main plots for DM. Still, using the whole data is not appropriate for assessing the actual index precision in the relevant data range of sufficient to excessive N supply. Besides the similar index rankings between datasets, the comparable rankings for the three N-related traits are promising for transferring algorithms between traits.

The Influence of the N Status
The results corroborate better trait estimation and a weaker index differentiation in sparse canopies. Although the index ranking was comparable in intervals 1 and 2, the relative advantage of the red edge indices especially over the NIR_red and R900_970 indices was clearly stronger in interval 2 and stronger for the N-related traits than for DM. The limitation of NIR_red as for the similar NDVI is well established for dense canopies [22,50,51]. The limitation of the water band information for N traits may be due to its insensitivity to NC regardless of being less prone to saturation [45]. Mostly ranking higher than the NIR_red, the NIR_green confirmed its advantage for dense canopies [32,41,52,53]. Still, red edge information appears inevitable for canopies under conditions of high-excessive N supply due to being less prone to saturation [31,41,54,55]. With the similar performance of the three red edge indices, no clear ranking could be established. The previously reported advantage of NIR/red edge simple ratio indices [22,56] and the similar CI red-edge [57,58] over the REIP was not confirmed, similar as in [30]. However, a slight trade-off was found with the R760_730 ranking better than the R780_740 in interval 1 and vice versa. This may be related to a shift in the red edge position at high N status [30,58], favoring a red edge band further right. The position of the red edge band between 720 and 740 nm agrees with optimization of the related CI red-edge index for chlorophyll estimation [58], and similar optimum combinations were found for in-season estimation of grain Nup and NC [44].

Conclusions
In this paper, the results obtained support the theory that spectral proximal sensing can be applied under conditions of high N status, aiming at avoiding overfertilization. While the index rankings were mostly consistent between datasets and statistical measures, the comparison of the trait estimability differed between RMSE and R 2 -values. Optimized index selection only slightly improved the estimation under conditions of low N status, indicating that simple NIR/red two-band multispectral sensors may be sufficient. In contrast, the relatively better performance of the red edge indices suggests that red edge information is crucial under conditions of high N status and in dense canopies.