Introduction

Accurately and efficiently estimating crop parameters during a season is essential for predicting yield, assessing drought stress, and determining nitrogen (N) requirements for a more resource-efficient and sustainable agriculture based on precision farming (Li et al., 2017; Lukina et al., 2001). In this context, the key plant traits are the green area index (GAI), the driving force in radiation uptake and transpiration, and the above-ground N uptake of the plant biomass (total N). In interaction they represent the conversion of water, radiation, and nitrogen into harvestable plant biomass. Remote sensing of total N and GAI by measuring spectral reflectance via sensors based on unmanned aerial vehicles (UAV) in various crops has gained interest in recent years, due to the possibility of generating high spatial and temporal resolution data at relatively low cost and effort (Barmeier & Schmidhalter, 2017; Bukowiecki et al., 2020; Prey et al., 2020; Rose & Kage, 2019).

In parallel with advances in remote sensing technology, scientists have developed methods to interpret spectral data by calibrating it to crop parameters. The concept of summarizing spectral data using vegetation indices (VI) was first proposed by Jordan (1969), where one spectral band is standardized by dividing it by a second band. Currently, it is widely accepted that this approach results in more stable calibrations under varying irradiance conditions and VI estimation errors of crop parameters can be reduced (Bukowiecki et al., 2020; Verger et al., 2014). However, it is necessary to identify the optimal wavelengths and combinations of vegetation indices (VIs) for calibrating to both GAI and total N. The results of calibration studies aiming to estimate crop parameters from VIs vary due to differences in considered crops, growth stages, locations, and crop parameters (Baret & Guyot, 1991; Barmeier & Schmidhalter, 2017; Fu et al., 2021; Gitelson et al., 2005). For example, Rose et al. (2023b) demonstrated that a uniform GAI calibration can be applied to four different crops, but emphasized the need for crop-specific parameters for total N estimation due to crop-specific relationships for GAI and total N. Furthermore, they estimated GAI of winter oilseed rape with one single calibration for the whole-season, but provided two different calibrations for total N before and after winter. In contrast, Müller et al. (2008) proposed a single calibration for GAI and for N in winter oilseed rape that is applicable from sowing until the start of flowering, whereas Ma et al. (2019) concluded that growth stage specific calibrations are required due to variations in plant architecture and leaf coverage throughout the growing season.

Another challenge in remote sensing is the diversity of spectral sensors and that sensors relevant in practice are generally limited to multispectral sensors due to weight and size limitations. As a result, VIs recommended for one sensor may not be computable by another due to missing bands, or calibration models may not be transferrable due to differences in available wavelengths or bandwidths. For example, Rose et al. (2023b) used a hyperspectral ground-based sensor and developed a calibration (total N calibration) that can be applied to various crops using a simple ratio of 720 and 710 nm. However, most UAV-based sensors provide at maximum only one band with a bandwidth of at least 10 nm in the red edge (RE) reflectance range. RE is defined as the wavelengths of sharp increase of crop reflection between 680 and 750 nm (Horler et al., 1983). Thenkabail et al. (2000) recommend the use of sensors with at least 12 different bands to accurately describe some crop characteristics for various crops. However, currently, most UAV-based multispectral sensors freely available on the market have no more than five bands. Therefore, it is necessary to conduct thorough assessments of the lower-cost alternatives available on the market and evaluate their performance in estimating crop traits.

To calibrate spectral data to crop parameters and assess its precise and reliable applicability, it is essential to acquire destructive reference data at various growth stages of the crops, spanning multiple seasons and locations. However, this involves a high working effort and delays the application of new remote sensors. As a consequence, it is common practice to integrate spectral data directly as VI in yield forecasts or fertilization recommendations (Piikki et al., 2022; Jahromi et al., 2023; Karimli et al., 2023; Petersen et al., 2023). The use of the VI has clear drawbacks as it heavily relies on the specific wavelengths and bandwidths of the sensor, as well as the informative value of the VI for the particular crop and canopy parameter under consideration. To save time and effort on destructive sampling for each new spectral sensor, studies attempt to bypass data requirements by using radiative transfer models to simulate reference measurements (Kooistra & Clevers, 2016; Richter et al., 2012). Others use leave-one-out procedures to overcome the problems of too small data sets (Clevers et al., 2017). Both approaches have drawbacks as the data only partly represents the real measurement conditions, and the evaluation is not clearly independent. Furthermore, the calibration procedures and resulting calibrations are rather difficult to comprehend and to apply. Bukowiecki et al. (2021) have demonstrated that by an UAV spectral sensor calibrated to the GAI of winter wheat and a large number of UAV-flights parallel to Sentinel-2 image recording, it was possible to calibrate the satellites spectral sensor without any further destructive sampling. However, this concept still requires extensive calibration of at least one remote spectral sensor for crops and crop parameters of interest.

Therefore, as a follow-up to the work of Bukowiecki et al. (2020) to calibrate a UAV-based spectral sensor to the GAI of winter wheat, this study aims to establish a calibration for GAI and total N for several commonly grown crops in Germany at once, namely: silage maize (Zea mays L.), winter barley (Hordeum vulgare L.), winter oilseed rape (Brassica napus L.), and winter wheat (Triticum aestivum L.) (Statistisches Bundesamt, 2023). To achieve this goal, a comprehensible and practicable calibration procedure is introduced. Whether a crop-specific whole-season calibration to GAI and total N is possible is tested and also if a uniform GAI calibration can be used for all four crops. The applicability of these calibrations is evaluated with data of different years and locations and demonstrated on an evaluation data set collected in three different seasons.

Materials and methods

Study site and field trials

The majority of the spectral reflection data from a UAV-based commercial camera and destructive sampling data of silage maize, winter barley, winter oilseed rape, and winter wheat were collected between 2017 and 2022 in different field trials at the Hohenschulen experimental farm of the Kiel University (10.0° E, 54.3° N, 21 m a.s.l.) in the federal state of Schleswig–Holstein in the North of Germany (Fig. 1, Table 1). The study site has a long-term mean temperature of 8.8 °C and an annual precipitation sum of 806 mm. Regarding winter wheat, a share of the data was already presented by Bukowiecki et al. (2021). Additional winter wheat and winter oilseed rape data was obtained for different commercial fields in Schleswig–Holstein in 2018 and 2019 (Fig. 1). In 2021 and 2022 silage maize data was not only collected in Schleswig–Holstein, but also at Wehnen (experimental station of the Chamber of Agriculture of the federal state of Lower Saxony; 8.12° E, 53.2° N, 8 m a.s.l.) and at Bad Hersfeld (experimental station of the Hesse Department of Agricultural Affairs; 9.67° E, 50.8° N, 204 m a.s.l) (Fig. 1). In the long-term, the mean temperature is higher at these sites (experimental station in Lower Saxony: 9.2 °C, experimental station in Hesse: 9.12 °C), with less precipitation in Hesse (633 mm) and higher precipitation in Lower Saxony (817 mm).

Fig. 1
figure 1

Map of the study sites. Points indicate experimental stations (Exp. Station), whereas crosses mark commercial farms (Com. Farm). Colors represent which crop was sampled. Only at the experimental station in Schleswig–Holstein (SH, white point) data of all four crops (silage maize, winter barley, winter oilseed rape and winter wheat) was collected. From the experimental stations in Hesse (HE) and in Lower Saxony (LS) only silage maize data was used and at the commercial farms in Schleswig–Holstein only winter oilseed rape and winter wheat were sampled

Full size image
Table 1 Overview of the data set (GAI: green area index, total N: total amount of nitrogen, HE: Hesse, LS: Lower Saxony, SH: Schleswig–Holstein, Exp. Station: experimental station, Com. Farm: commercial farm). As no calibration model was finalized with data after mid of flowering, the number of sampling data is only given as total sum, but not split in calibration and evaluation data (ncalibration, nevaluation)
Full size table

Data collection

Plant sampling

At every sampling date, aboveground plant biomass was sampled at 3–84 spots of the site and its BBCH stage (Lancashire et al., 1991) was scored. The number of sampling spots depended on the type of trial, the variation within the crop and the experimental question. The sampling area was 0.25 m2 for winter cereals and 1 m2 for winter oilseed rape. For silage maize, 8–15 plants were sampled at each sampling spot and the sampling area was recorded. To determine GAI, fresh matter was fractionized into leaves, senescent leaves (if available), stems, ears, pods, cobs, husks, and panicles. The one-sided area of all fractions was determined with a LiCor 3100 leaf area meter (LiCor Inc., NE, USA) and total green area was calculated as the sum of the area of green leaves, stems, pods, and ears. Total green area divided by the sampling area represents the GAI. Afterwards, each fraction was dried and its nitrogen concentration was determined by near infrared spectroscopy (NIRS). All silage maize data from the trials in Hesse and in Lower Saxony and parts of the silage maize data from Schleswig–Holstein in 2021 and 2022 were measured with the Phoenix 5000 (Bluesun Sci, USA) whereas all winter crop data and the other share of the silage maize data were analyzed with the NIRSystems 5000 scanning monochromator (FOSS GmbH, Rellingen, Germany).

As the destructive determination of GAI during senescence is prone to error (Bukowiecki et al., 2020), no winter crop GAI data after flowering was included in the analysis. Additionally, the GAI data for silage maize from the last two sampling dates in 2022 had to be excluded due to noticeable senescence caused by the unusually warm and dry summer. The data after the start of flowering was excluded for winter oilseed rape as the distinctive flowering of oilseed rape significantly shifts the spectral signal. In some cases, nitrogen concentration and aboveground dry matter amount were determined, but GAI was not measured. This resulted in differing data sets per crop species for the calibration of UAV spectral data to GAI and total N (Table 1). However, it should be noted that the data set for the calibration of each crop to both crop parameters was large, representing a broad range of crop variation and sampling conditions. The dataset includes at least five different cultivars, eight level of N fertilization, 15 growth stages, 13 sampling dates and three growing seasons. In addition, the data from the commercial farms in Schleswig–Holstein, and from the experimental stations in Lower Saxony and Hesse could be used to test the applicability of the calibration models to different locations.

Spectral reflectance measurements

Multispectral data were collected with a Sequoia multispectral camera (Parrot Drones SAS, France, Paris), carried by a fixed-wing drone (eBee, senseFly). For each sampling date, spectral and destructive data collection was done within at most three days to ensure a negligible plant development between both dates. The Sequoia camera measures simultaneously incoming solar radiation and surface reflection in four bands: green (550 nm, bandwidth: 40 nm), red (660 nm, bandwidth: 40 nm), RE (735 nm, bandwidth: 10 nm) and near infrared (NIR, 790 nm, bandwidth: 40 nm). Picture resolution was controlled by flight height and ranged between 8 × 8 and 12 × 12 cm. After the flight, pictures were georeferenced using the software eMotion3 (eMotion3, SenseFly SA, Switzerland). Then they were corrected by incoming sunshine and stitched together in the software Pix4Dmapper (Pix4D SA., Switzerland). On the resulting four maps, one for every band, sampling spots were identified, and the reflectance data was extracted using QGIS version 3.30 (QGIS Development Team, 2018). For all crops but silage maize, all pixels of a sampling spot were summarized as median to be more robust towards weeds. For silage maize, pixels were summarized as arithmetic means. This was based on the consideration, that during early growth of silage maize, most pixels of the sampling spots belonged to the soil between the maize rows. Hence, the median would not represent plant growth during this part of the season.

Data processing and statistical analysis

All subsequent data processing and statistical analyses were conducted using the R statistical environment, version 4.1.0 (R Core Team, 2021). The statistical measures used were the mean absolute error (MAE), the relative mean error (rMAE; the MAE standardized by the measured mean value) and the coefficient of determination (R2).

All possible combinations of two bands (simple ratios) and their squares were calculated. Squares of the simple ratios were used considering the known non-linear behavior of VIs above a certain level of soil cover, respectively above a certain GAI (Haboudane et al., 2004). This yielded 12 potential terms to characterize the GAI of all four crops and the total N for winter wheat and silage maize, all of which were tested.

In winter oilseed rape, a distinctly different GAI/ total N-ratio before and after winter was found (data not shown). Additionally, the GAI/ total N-ratio of winter barley shifted noticeably at booting. Therefore, a binary factor was included in the total N calibration models for these two crops (for winter oilseed rape 0: before, 1: after winter, and for winter barley 0: until booting, 1: afterwards). Consequently, 25 different terms had to be analyzed for the total N models for winter barley and winter oilseed rape, testing each squared and non-squared term with and without the influence of the binary term and the individual influence of the binary component.

The package “ExhaustiveSearch” (Jagdhuber et al., 2021) was used to assess the predictive capability of all combinations of these terms for GAI and for total N. No interactions were tested. The relative importance of the components in the linear models was analyzed in terms of their R2 contribution by averaging over order among all selected components using the package “relaimpo” (Grömping, 2006).

To ensure a robust evaluation, data was split for each crop and crop parameter into an independent calibration and evaluation data set. For this purpose, the arithmetic mean of the crop parameter was calculated crop-wise for each sampling date. In the next step, sampling dates was sorted from lowest to highest average value. Then, sampling dates were assigned alternating, to calibration and evaluation data set. The sampling date with the highest and lowest mean value was allocated in the calibration data set. As sampling number was not equal at each sampling date, the trial, sizes and proportion of evaluation and calibration data set differed and ranged between 42 and 58% for calibration and for evaluation (Table 1). For all crops and crop parameters, the calibration data set contained at least 191 samplings at six different sampling dates and for two different growing seasons.

The possibility of using one single calibration for all four crops was only tested for GAI as the results of Rose et al. (2023b) negate the possibility of a uniform total N calibration for these four crops. Furthermore, the integration of the growth stage dependent factor for winter barley and the pre-/post-winter factor for winter oilseed rape prevented a uniform model for total N. The data was likewise split into a calibration and an evaluation data set as for the crop-specific calibration (ncalibration = 1352, nevaluation = 1386). Each data set contained at least data for six different sampling dates per crop.

For total N a whole-season calibration was not possible. Hence, it was analyzed until which point in season a total N calibration is possible. For this purpose, the calibration and evaluation procedure were repeated stepwise, leaving aside in each step the highest BBCH stage still included in the data set. The phenological inflection point, when the error terms started to increase, was detected as last phenological phase in which total N can be predicted.

Application of the calibrations

For a further evaluation of the calibration models, the best calibrations were applied to Sequoia multispectral data of a large crop rotation trial during the growing seasons of 2019/20, 2020/21 and 2021/22. The obtained GAI and total N data at plot level were used to analyze the correlation between N uptake at flowering and total N at harvest and between the GAI-based calculated radiation uptake and finally harvested aboveground biomass. According to Monteith (1977), intercepted radiation and final biomass should correlate closely, as dry matter is the product of intercepted radiation and radiation use efficiency (RUE).

The field trial was located at the Hohenschulen experimental station in Schleswig–Holstein and is described in more detail by Rose et al. (2023a). Each year silage maize, winter barley, winter oilseed rape and winter wheat were cultivated at five different N fertilization levels with four replications. Spectral reflectance was measured bi-weekly from March until the harvest of silage maize in October. No data was collected between the sowing of the winter crops in autumn and the next spring (March) due to the problems of applying the Sequoia camera during winter irradiance conditions in Northern Germany (Holzhauser et al., 2022). The number of spectral measurement dates per crop and season varies between 10 and 14. In the growing season 2021, and particularly in 2020, data were sparse from the beginning of August onwards and therefore during a large part of the silage maize growing season. For winter oilseed rape, reflectance data during flowering were excluded from the analysis because of a strong masking effect of yellow petals. The total N calibration was only applied until the last flight before flowering. The start of flowering of winter oilseed rape was distinguished via UAV pictures. For the other crops, flowering was assigned according to the BBCH scale as assessed by surrounding field trials (data not shown). Silage maize flowered after July 20, winter wheat flowered in the last three days of May, and winter barley flowered after May 20. The stem elongation of winter barley was estimated to start on April 20.

Yield data at plot level were obtained by a combine harvester or with a maize chopper for silage maize. The nitrogen concentration of the harvested product was determined by NIRS (NIRSystems 5000 scanning monochromator, FOSS GmbH, Rellingen, Germany). For the estimation of total N at harvest, average nitrogen harvest indices (NHI) per crop in the seasons 2015/16, 2016/17 and 2017/18 (Rose et al., 2023a) were used (0.9 for winter barley, 0.86 for winter wheat and 0.75 for winter oilseed rape). For the calculation of total biomass, mean harvest indices (HI) as assessed at the same trial in earlier seasons were applied (data not shown): HIwinter barley = 0.55, HIwinter oilseed rape = 0.35, HIwinter wheat = 0.5. For silage maize, NHI and HI equal 1. Daily weather data were obtained from a local weather station.

To calculate radiation uptake, GAI was interpolated from sowing until harvest above growing degree days (base temperature = 0 °C) by locally weighted scatterplot smoothing (LOESS). The LOESS algorithm is implemented in base R (R Core Team, 2023) and was used with a constant smoothing parameter of 0.25. For the interpolation through the whole growing season, the spectral data were complemented by assuming a GAI of 0 at sowing. Between sowing and the first measurement date, GAI was interpolated linearly above growing degree days. For winter crops, a GAI of 0 was assumed at harvest as well. For silage maize, GAI measurement of the last flight was applied to the harvest date. To account for the temperature dependency of photosynthesis (Rose et al., 2017), incoming photosynthetic active radiation (PAR) was weighted by a trapezoid function with crop-specific transition points as proposed by Rose et al. (2023b). Next, the daily weighted intercepted PAR was computed using weighted PAR, GAI, and crop-specific extinction coefficients from Rose et al. (2023b) (0.654 for silage maize, 0.729 for winter cereals, and 0.7 for winter oilseed rape) as input for the Beer-Lambert law (Monsi & Saeki, 1953). In a last step, weighted intercepted PAR was cumulated from sowing until harvest and correlated with final biomass.

Results

Calibration models for GAI

For each crop a large range of GAI values was represented in the dataset (0–6 m2 m−2). Yet, GAI values exceeding 4 m2 m−2 for winter oilseed rape were sampled only on one date, restricting the ability to assess the performance of GAI models above this threshold for this particular crop. The best GAI calibrations provided R2 values of 0.86–0.95 for the calibration data and of 0.82–0.92 for the evaluation data, with the closest correlations between measured and predicted GAI for winter wheat and the least close for winter oilseed rape (Fig. 2). The calibration models exhibited consistent performance across all combinations of trial locations and years. For each crop there was a slight tendency to underestimate high GAI values (Fig. 2). The MAE for calibration and evaluation data did not exceed 0.48 m2 m−2 (Table 2). Comparing the prediction accuracy for the different crops in terms of rMAE, the GAI of silage maize was predicted with the highest precision (rMAE = 17%) and the GAI of winter oilseed rape with the least precision (rMAE = 24%).

Fig. 2
figure 2

Correlation of measured and predicted GAI for calibration and evaluation for silage maize, winter barley, winter oilseed rape and winter wheat. Colors illustrate the different growing seasons and trial sites. The linear regression of measured vs. predicted GAI and the coefficient of correlation is indicated for each data set

Full size image
Table 2 Calibration models and statistical measures for calibration and evaluation (GAI: green area index, total N: total amount of nitrogen, MAE: mean absolute error, rMAE: relative mean absolute error). For the calculation of total N of winter barley and winter oilseed rape, additional terms are added to predict values until/after BBCH30 or before/after winter
Full size table

The calibration models for the four crops showed notable differences in the selection and weighting of simple ratios: Each of the GAI models included all four spectral bands provided by the Sequoia camera (Table 2). At least two of these calibration components were squared ratios. None of the ratios (squared or non-squared) was used in all four crops. However, certain patterns in the R2 contribution can be discerned (Fig. 3): In all crop-specific calibrations the ratios including NIR provided the most information. Only for silage maize, two major components did not include NIR; Red/RE and (Red/RE)2. In contrast, (NIR/RE)2 was an important part in each GAI calibration for the narrow row crops but was not included in the GAI calibration for silage maize. In all calibrations for the graminoids (silage maize, winter barley, and winter wheat), the ratios NIR/Green and (NIR/Green)2 contributed significantly to explaining the variation in GAI. Comparing the calibration of both winter cereals, the one for winter wheat used two of the three ratios used by winter barley; the two squared components (NIR/Green)2 and (NIR/RE)2.

Fig. 3
figure 3

The R2 contribution of the components in the calibration models (light: low, dark: high). The contribution of the components including the binary factor for pre/post-winter dates of winter oilseed rape and pre/post-booting of winter barley is not shown

Full size image

Calibrating one uniform GAI model for the four crops resulted in high values for R2 for calibration and evaluation (Fig. 4). However, the rMAE was distinctly higher compared to the crop-specific models. In the evaluation, the rMAE was 24% for silage maize, 26% for winter barley, 29% for winter oilseed rape, and 31% for winter wheat. Hence, the rMAE was 5–12% lower with a crop-specific GAI model (Table 2). In the calibration, the GAI of winter wheat was distinctly overestimated while in the evaluation, only low values were not well depicted. For winter barley, the GAI was underestimated in terms of the calibration data set but in the evaluation, there was the same tendency of overestimating low GAIs and underestimating high GAIs as for winter wheat. For winter oilseed rape and silage maize high GAI values were noticeably underestimated in the calibration and the evaluation. The uniform GAI calibration included all four bands of the Sequoia, composed in seven components, two of them were squared terms. The pattern of R2 contributions for the components of the uniform GAI model did not resemble that of the crop-specific models (Fig. 3).

Fig. 4
figure 4

Correlation of measured and predicted GAI for calibration and evaluation with one calibration model for all crops. Colors indicate the different crops. The linear regression of measured vs. predicted GAI and the coefficient of correlation is indicated for each data set

Full size image

Calibration models for total N

The analysis revealed a distinct increase of the total N prediction error if the data set included growth stages after mid—flowering (≥ BBCH67) (Fig. 5). This increase was noticeably sharper for winter wheat than for silage maize. For silage maize, the minimal prediction error was obtained including data from sowing until mid—flowering. In contrast, for winter wheat, rMAE was low as long as only data until the end of booting (< BBCH50) were included. For both crops, rMAE was larger if the data set was reduced to data until end of stem elongation (< BBCH40). As a consequence of this analysis, data after mid—flowering were excluded for the further calibration work for total N of all four crops (Table 1).

Fig. 5
figure 5

Relative mean error (rMAE) in the evaluation of calibration models for total N of silage maize and winter wheat. The data set size increases from left to right by including stepwise data of higher growth stages (< BBCH30: sowing - end of tillering, < BBCH40: sowing - end of stem elongation, < BBCH50: sowing - end of booting, < BBCH60: sowing - end of heading, < BBCH67: sowing - mid of flowering, < BBCH75: sowing - medium milk, all data: sowing - harvest)

Full size image

This condensed dataset comprised N uptake values of up to 20 g m−2 for each of the four crops. Again, high values were sparse for winter oilseed rape (see section `Calibration models for GAI`). As for GAI, there was a tendency to underestimate high amounts of total N (Fig. 6). This was especially evident in predicting winter oilseed rape, but it also held true for silage maize with N uptakes exceeding 20 g m−2. Regarding the rMAE of the evaluation (Table 2), the prediction accuracy for the total N of winter oilseed rape (27%) was lowest, followed by the one of silage maize (23%). For winter cereals, rMAEs were noticeably lower. The range of R2 in the calibration was 0.88–0.94 and in the evaluation 0.76–0.9 (Fig. 6). rMAEs were relatively similar and slightly higher for the estimation of total N as for GAI (Table 2: for GAI 17–24%, for total N 17–27%).

Fig. 6
figure 6

Correlation of measured and predicted amount of total N for calibration and evaluation for silage maize, winter barley, winter oilseed rape and winter wheat. Colors illustrate the different growing seasons and trial sites. The linear regression of measured vs. predicted total N and the coefficient of correlation is indicated for each data set

Full size image

Each calibration for total N included all bands of the Sequoia. Again some ratios were integrated more frequently in the calibration models and contributed more to the R2 of the model: (NIR/RE)2 was a major part in the total N calibrations for all narrow row crops (Fig. 3). (Green/RE)2 was included in the calibration model for the three graminoids (Table 2). Silage maize and winter oilseed rape shared only one common component. Each calibration included at least two squared terms. For winter oilseed rape and winter barley, two and three bands respectively were differently weighted in the phases until/after BBCH30 or before/after winter due to the additional binary factor (see section `Data processing and statistical analysis`).

Application of the calibrations

The weather conditions of the growing seasons 2019/20, 2020/21 and 2021/22 differed noticeably, with distinct effects on plant growth: The springs of 2020 and 2021 were relatively cold, which particularly affected the development and yield of winter oilseed rape (2020: 270 g m−2, in comparison to 2021: 360 g m−2, 2022: 370 g m−2). In contrast, the summer of 2021 was warmer than the average resulting in very good growing conditions for silage maize (2021: 2110 g m−2, in comparison to 2020: 1840 g m−2). However, the temperatures were too high during grain filling of winter wheat (2021: 60 g m−2, in comparison to 2020: 70 g m−2, 2022: 78 g m−2). For 2021 it is also worth mentioning that in some silage maize plots large damage was caused by game animals early in the season. In 2022 below-average precipitations in the month after silage maize sowing caused low yields (average 1650 g m−2). The winter barley yields were the least variable between the years (2020: 740 g m−2, 2021: 820 g m−2, 2022: 880 g m−2).

The GAI calibrations provided stable and plausible values through the three seasons for each of the four crops (Fig. 7). A distinct growth in spring in the winter crops with the highest GAI values reached in winter barley, followed by winter wheat and the lowest maximal GAIs in winter oilseed rape. The GAI of silage maize begins to increase at the onset of winter crop senescence and attains GAI values within the same range as winter cereals. The hampered growth of winter oilseed rape in the spring of 2020 and the drought-related stronger senescence of silage maize in 2022 were reproduced by the spectral measurements.

Fig. 7
figure 7

Interpolated mean GAI courses of silage maize, winter barley, winter oilseed rape and winter wheat. Points indicate the days of UAV-based spectral reflactance measurements

Full size image

The UAV-based estimated cumulative weighted intercepted PAR correlated closely with final aboveground biomass (Fig. 8): 93–97% of the variation in total dry matter was explained in each year for winter cereals. For winter oilseed rape, both total dry matter and weighted intercepted PAR showed less variation in the different years and the R2 of radiation uptake-explained yield variation was somewhat lower, yet still between 0.75 and 0.88. The cumulative weighted intercepted PAR varied the least in silage maize, with a maximal difference of 107–158 MJ m−2 between unfertilized and highly N fertilized plots. R2s were the smallest for silage maize as well; 0.54 in 2020, 0.24 in 2021 and 0.33 in 2022 (Fig. 8).

Fig. 8
figure 8

Correlation of final aboveground biomass with intercepted weighted photosynthetic active radiation (PAR) of silage maize, winter barley, winter oilseed rape, and winter wheat in three growing seasons. Intercepted PAR was calculated with UAV-based estimated GAI data

Full size image

The RUE was always lower for winter oilseed rape (2020: 1.17 g MJ−1, 2021: 1.62 g MJ−1, 2022: 1.86 g MJ−1) than for winter barley (2020: 1.93 g MJ−1, 2021: 2.44 g MJ−1, 2022: 2.7 g MJ−1) and winter wheat (2020: 2.25 g MJ−1, 2021: 2.12 g MJ−1, 2022: 2.98 g MJ−1). Hence, RUE of winter wheat was in two of the three years higher than of winter barley. The RUE of silage maize was especially high in 2021, and only in 2022 below the RUE of winter cereals (2020: 2.75 g MJ−1, 2021: 3.28 g MJ−1, 2022: 2.65 g MJ−1).

As for GAI, reasonable time courses of total N were obtained by applying the calibrations to the multispectral data of the different crops (not shown). The correlation between UAV-based estimated total N until flowering and total N at harvest was nearly as close as between total dry matter and intercepted PAR (Figs. 8, 9). For winter cereals, 83–94% of the variation of total N at the end of the season was described by the UAV-based measured total N until flowering, for winter oilseed rape a little less (76–84%) and for silage maize the explained variation did not exceed 50% in 2020 and 2022. In 2021 the explained variance was only 4%. It is worth mentioning that in 2020 and 2021 in winter barley and winter oilseed rape at all N levels, the UAV-based estimated total N until flowering nearly equaled total N at harvest (94–106%). The same applies for winter wheat in both unfertilized and low nitrogen-fertilized plots. However, the ratio between N uptake at harvest and until flowering increased with fertilization level by about 50%. Silage maize nearly doubled its total N from flowering until harvest (2020: + 84%, 2021: + 91%). In 2022, the correlation between post-anthesis and pre-anthesis N uptake underwent a shift in all crops and at all N levels, increasing by approximately 50% compared to 2020 and 2021 (Fig. 9).

Fig. 9
figure 9

Correlation of N yield at harvest with UAV-based assessed total aboveground N uptake until flowering of silage maize, winter barley, winter oilseed rape, and winter wheat in three growing seasons. Values above the black 1:1-line imply that the N amount harvested is higher than the total N measured at flowering. Shapes indicate different N levels (0 kg N ha−1: unfertilized, up to 150 kg N ha−1: fertilized low, over 150 kg N ha−1: highly fertilized)

Full size image

Discussion

Calibration models for GAI and total N

Whole-season applicability

The main objective of this study was to establish a calibration for GAI and total N for silage maize, winter barley, winter oilseed rape, and winter wheat. This was possible for GAI, but only up to mid-flowering for the total N calibration. Afterwards, a significant decrease in prediction accuracy was observed and this decline was more pronounced for winter wheat than for silage maize. This can be traced back to the fact that spectral reflection is primarily driven by GAI (Rose et al., 2023b). Consequently, total N is only accurately described as long as it correlates with GAI (Rose et al., 2023b). As the GAI/ total N ratio changes constantly from the end of the vegetative growth onwards—due to the storage of high amounts of N in non-green plant organs (Lemaire et al., 2007)—the estimation of total N is no longer possible once vegetative growth ceases. As senescence and N retranslocation is far more pronounced in winter wheat than in silage maize, the rMAE increased were correspondingly sharper. This argumentation is additionally underlined by the slightly higher accuracy in the estimation of GAI (rMAE = 17–24%) than of total N (rMAE = 18–27%). Furthermore, it matches with the requirement of including an additional factor in the total N calibration of winter barley and winter oilseed rape, which differentiates phases of the season with distinctly different GAI/ total N ratios. A noticeably changing GAI/ N ratio after winter in winter oilseed rape was also observed by Rose et al. (2023b). However, to the best of our knowledge, it has not been described for winter barley after booting.

Regarding the latest suitable stage for applying a uniform calibration for total N, Mistele and Schmidhalter (2010) described a close correlation between VI and total N until flowering as well. In contrast, Rose et al. (2023b) determined the end of the correlation for silage maize, winter wheat and winter barley already at BBCH39. This might be traced back to the relative short time period between BBCH39 and BBCH65. For example, in silage maize BBCH40–50 does not exist and BBCH51–59 takes in general not more than a few days under the given mild-humid temperate climate conditions. Moreover, these developmental phases are not as relevant as leaf or stem development for considerations of radiation uptake. They are also of less importance for fertilization management since applications in winter cereals are commonly done at tillering, the start of stem elongation and/or just before start of flowering. Hence, these growth stages might have been underrepresented in the data set of Rose et al. (2023b).

The decreasing accuracy of the VI-based estimation of crop parameters with increasing crop density is a well-known phenomenon (Haboudane et al., 2004; Viña et al., 2011). This problem can be reduced by including the RE band (Delegido et al., 2013), which allows whole season GAI calibrations even at high canopy densities (Bukowiecki et al., 2020). Nevertheless, it is common to observe slight underestimations of high GAIs even when RE is integrated into the calibration (Bukowiecki et al., 2020; Nguy-Robertson et al., 2014; Rose et al., 2023b). The saturation of total N calibrations for data until flowering above a distinct N uptake is consistent with the findings of Mistele et al. (2004). They concluded that besides the growth stage, high biomass limits the applicability of spectral data to describe N uptake.

Despite this saturation effect, the accuracy of the presented GAI and total N calibrations for all four crops is high: For the GAI of winter wheat, the relative prediction error is 2% less than that of Rose et al. (2023b) and in the same range as the results of Broge and Mortensen (2002). The calibration model of Thenkabail et al. (2000) describes 78% of the GAI variation in silage maize, which is 11% less than the approach proposed here. In winter oilseed rape, the rMAE of Rose et al. (2023b) is 6% higher and the R2 of Müller et al. (2008) is slightly lower. Regarding the prediction of total N in winter wheat until flowering, our results are very close to those of Lelong et al. (2008) and Mistele and Schmidhalter (2010), but distinctly better than those of Hansen and Schjoerring (2003). Even the less accurate estimate of total N in winter oilseed rape is more accurate than the predictions of Rose et al. (2023b) and comparable to the results of Müller et al. (2008). Rose et al. (2023b) have found a much better description of total N in winter wheat than in winter oilseed rape. These consistent findings could be attributed to the difficulties in sampling high values of total N (respectively of GAI), as for winter oilseed rape the correlation between crop parameter and spectral signal breaks earlier than for winter wheat, already at the beginning of winter oilseed rape flowering.

Crop-specific calibration

In each total N and GAI calibration, the NIR, Green, and RE bands are included. This matches the findings of several other studies (Gitelson et al., 2005; Viña et al., 2011; Nguy-Roberstson, 2014). The additional use of the Red band is in line with the results of Haboudane et al. (2004) and Thenkabail et al. (2000), who used Red, Green and NIR for the estimation of the leaf area index. In line with Müller et al. (2008), Mistele and Schmidhalter (2010), Prey et al. (2020) and Rose et al. (2023b), the ratio of NIR/RE was identified as the key component for most calibrations. The exception were the models for silage maize, which instead included the RE band in the ratios Green/RE or Red/RE. The high importance of NIR-standardized components matches the findings of Müller et al. (2008), Bukowiecki et al. (2020) and Rose et al. (2023b). The frequent use of RE in a simple ratio either with NIR, or with Red or Green, underlines its importance for the estimation of crop parameters (Hansen & Schjoerring, 2003). In accordance with Rose et al. (2023b), squared and non-squared simple ratios had to be included in the calibrations.

Due to the high workload of plant sampling and the fast-paced development of multispectral sensors for remote sensing, most calibration work focuses on data for one crop (Bukowiecki et al., 2021; Piikki et al., 2022), one location (Bukowiecki et al., 2020) or consists of only one growing season (Lelong et al., 2008). Nevertheless, it is common practice to apply the same calibration model to crop parameters of different crops or for different field sites (Verrelst et al., 2015). The presented data set is characterized by its comprehensiveness in terms of crops, range of GAI and total N values, sampling dates, growing seasons and trial sites and allows a more detailed view. The applicability of the presented calibrations to data for the different locations and years is an important finding regarding the application of remotely sensed data for agricultural applications.

However, our findings propose crop-specific calibrations for each crop parameter. The prediction error for GAI was distinctly higher for the three graminoids with a uniform GAI model and there was a strong tendency for underestimating high GAIs. Furthermore, low GAIs were not well depicted. This contrasts to the results of Rose et al. (2023b) who used a ground based hyperspectral sensor. In that latter study it was stated that a uniform GAI calibration with a simple ratio of NIR (810 nm)/RE (710 nm) was possible. However, NIR/RE is not included in the GAI calibration presented here for silage maize and, in addition, the GAI models for the four studied crops vary distinctly.

This might arise upon the different wavelengths and bandwidths of the multispectral sensors in the NIR and RE band. In the bands provided by the Sequoia sensor, the distinctly different crop characteristics could have influenced the correlation between crop parameters and spectral reflection significantly, as found for example by Kang et al. (2016) between narrow row and wide row crops. Nguy-Robertson et al. (2014) identified resemblances in GAI calibration models across various crops, attributing these similarities to shared characteristics in canopy architecture and leaf chlorophyll contribution. This is consistent with the finding that the components in the calibration models are more similar for winter barley and winter wheat than for winter oilseed rape, and that silage maize contrasts the most. This is also consistent with the observation that, with a uniform GAI model, there is a greater similarity in the performance of the prediction for winter cereals than for silage maize and winter oilseed rape. Assuming that the bands of the Sequoia are not optimally placed, the inclusion of the four bands in each GAI and total N calibration in most cases could be a compensating effect. Following Gitelson et al. (2005) and Nguy-Robertson et al. (2012) that the inclusion of a green band results in a species-specific calibration, the suboptimal positioned bands of the Sequoia camera may have induced the requirement of crop-specific calibrations. However, Richter et al. (2012) described a leaf area estimation with two spectral bands as possible, but suggested the use of 6–8 bands for higher accuracy. Furthermore, Thenkabail et al. (2000) calibrated crop-specifically to leaf area index, although they worked with a hyperspectral sensor. They stated that for the optimal estimation of leaf area index of most crops a multispectral sensor of 12 bands would be required. For their crop-specific calibrations, they have combined four of these 12 different bands, corresponding to the number of bands chosen in the calibration models presented here. Therefore, the truth may lie somewhere in between; there is a need to utilize four bands for an accurate estimation of crop parameters. However, dealing with the challenge of adjusting optimal band widths for various crops and crop parameters proves to be challenging with a multispectral sensor.

Application of the calibrations

Applying the calibrations to the multispectral data obtained from UAVs results in consistent and meaningful trends for GAI and total N in silage maize, winter barley, winter oilseed rape, and winter wheat across three different years. Seasonal GAI estimates reflected well the expected variation due to different weather conditions. Due to the difficulties in measuring GAI during senescence (Bukowiecki et al., 2020), only silage maize data after flowering were used for GAI calibration. However, following the argumentation of Bukowiecki et al. (2020), the GAI calibrations can nevertheless be applied to spectral data during senescence. The high R2 between calculated whole-season radiation uptake and final aboveground dry matter of the winter crops support this assumption. In this context, the low explained variation of silage maize yield by radiation uptake stands out. In 2020 and 2021 this might be traced back to fewer UAV flights after flowering and therefore less accurate GAI courses, but R2 in 2022 is low as well. In each of the three years the low variation in radiation uptake in silage maize between the different N levels compared to the winter crop is striking. This is in accordance with a low response of leaf area index to varying N supply as reported by Massignam et al. (2012) and Vos et al. (2005). Instead of GAI, silage maize changes the amount of N per unit leaf area (specific leaf nitrogen, SLN). Hereby, silage maize reduces the photosynthetic capacity of its leaf, an important aspect for the correlation between radiation uptake and yield formation, which is not considered by the GAI calibration and the approach used for the calculation of radiation uptake. Considering varying SLN due to N fertilization levels (Massignam et al., 2009) might be a possible solution. Particularly low correlations between radiation uptake and final dry matter yield as well as between total N at flowering and N yield in 2021 may be due to damage caused by animal game.

The quality of UAV-based derived whole-season GAI data and thereby calculated radiation uptake is underlined by the calculated RUEs, matching to values reported in literature for silage maize (Andrade et al., 1992, 1993), winter barley (Fletcher et al., 2013; Jamieson et al., 1995; Sieling et al., 2016), winter oilseed rape (Fletcher et al., 2013), and winter wheat (Wilson & Jamieson, 1985; Sinclair & Muchow, 1999; Shearman et al., 2005; Sieling et al., 2016). Furthermore, higher RUEs in winter cereals than in winter oilseed rape are in accordance with the findings of Sieling et al. (2016), and mostly higher RUEs in silage maize than in the three winter crops to the C3/C4-differentation match to the results of Fletcher et al. (2013). Variation in RUE between crops and years were found by Sieling et al. (2016) as well. However, the variability might be overestimated by the application of a crop-specific uniform harvest index which is known to differ between N levels and cultivars (Rose & Kage, 2019).

The explained variation of total N at harvest by the UAV-based estimated N uptake at flowering of 76–94% is very high for the winter crops (low R2 for silage maize have already been discussed) compared to the R2 = 0.46–0.61 obtained by Papakosta and Gagianas (1991) in winter wheat. This can be traced back to the larger variation of N fertilization levels and therefore N yield of the presented trial; five levels from 0 to 320 kg N ha−1 in comparison to the two 150 kg N ha−1 variants of Papakosta and Gagianas (1991).

The marked crop- and year specific shifts in the relationship between UAV-measured total N uptake until flowering and total N at harvest merit further discussion. A significant effect of N fertilization on the ratio of pre-flowering to total N uptake in winter wheat has also been found by Delogu et al. (1998). In accordance with the UAV-based findings of 2020 and 2021 presented here, they stated that in unfertilized plots no N uptake after flowering occurred, in contrast to the high N uptake in fertilized plots. Their results for post-anthesis N uptake and the relationship between pre-anthesis and post-anthesis N uptake measured by Papakosta and Gagianas (1991) of 73% and by Van Sanford and MacKown (1987) of 83% in winter wheat, match to the results presented here. In contrast to winter wheat, the proportion of N uptake after flowering to pre-anthesis N uptake in winter oilseed rape and winter barley does not increase with N fertilization (Schjoerring et al., 1995). Rossato et al. (2001) describe no significant N uptake after flowering in winter oilseed rape. This matches to our results in 2020 and 2021, whilst the non-fertilizer sensitive increase of total N found by Schjoerring et al. (1995) corresponds more to the results in 2022. Gombert et al. (2010) attributed the cessation of N uptake in winter oilseed rape at flowering to a depleted N supply. They also demonstrated that, in the presence of high soil N supply, post-anthesis N uptake takes place. Among others, Gammelvind et al. (1996) have described increasing leaf area as the main effect of increased N fertilization in winter oilseed rape. Also Guo et al. (2021) describe the maintenance of GAI after flowering (stay green) as decisive factor for post-anthesis N uptake. Hence, the increased post-anthesis N uptake in 2022 could be due to higher soil N reserves at flowering, inducing a prolonged phase of higher GAIs. Slowed down uptake of mineral fertilizer in the dry spring of 2022 does seem realistic. The assumption of a stay green effect could be analyzed further by the UAV data and the developed GAI calibration for winter oilseed rape. The ratio of post-anthesis to pre-anthesis N uptake is the highest in silage maize in all considered years as its senescence is slower than in winter crops. In line with our findings, Ciampitti and Vyn (2011) describe a ratio of 51–71%, depending on N level, cultivar, and location. Cultivar- and year-specific effects in the relation between pre- and post-flowering N uptake were also reported by Papakosta and Gagianas (1991) and Guo et al. (2021).

These shifts in the correlation between N uptake at flowering and total N at harvest, and similarly between radiation uptake and final biomass should be kept in mind when discussing possible applications of UAV-based crop parameter data. By the use of the presented calibrations, it is now possible to derive fast and precise estimations of whole-season GAI and total N until flowering. The importance of these data is shown and discussed with respect to radiation uptake, radiation use efficiencies, the ratio between N uptake pre- and post-anthesis and crop response to N fertilization, such as prolonged stay green. Yet, to gain the full benefit of these data in terms of yield prediction or N fertilization recommendations, they should be integrated into appropriate models considering effects of altering weather conditions, varying locations, and different cultivars.

Conclusion

Models for the whole-season estimation of GAI and the assessment of total N uptake until flowering by UAV-based spectral measurements were calibrated for four important crops: silage maize, winter barley, winter oilseed rape, and winter wheat. NIR-based ratios were identified as key components for all calibrations. A robust and easy-to-use methodological framework for calibrating spectral sensors to crop parameters is provided. Applied to an independent evaluation dataset of different years and locations, the calibrations produced accurate predictions for both crop parameters and for each crop. With the bands provided by the UAV-based sensor, it was not possible to establish a consistent calibration for accurate GAI estimation across all four crops. Distinct changes of the GAI/ N ratio within the vegetative phase of winter oilseed rape and winter barley required the inclusion of corresponding components in the calibration models for the precise estimation of total N. The applicability of the UAV-based crop data, which can now be derived with little effort, with regard to various yield-relevant processes such as radiation and N uptake has been demonstrated in a large field trial. Restrictions in the transferability of calibration models between different sensors and crops are discussed, just as the dependence of the benefits of the UAV-based crop data on their interpretation by appropriate crop models. The presented results support the use of UAV-borne remote sensing data for precision agriculture and provide a comprehensive set of calibrations for GAI and total N in several crops.