(1) Exp1: different N application rates

Four inbred lines with great differences in nitrogen use efficiency were selected as the test varieties, including CIMBL123, CML422, 526018, and CIMBL78. The sensitivities of these varieties were as follows: CIMBL123 has a low soil and plant analyzer development (SPAD) value and yield with low nitrogen fertilizer. CML422 has a high SPAD value and yield with low nitrogen fertilizer. 526,018 has a low SPAD value and yield with high nitrogen fertilizer. CIMBL78 has a high SPAD value and yield with high nitrogen fertilizer. Maize seedlings were cultured in a complete nutrient solution with major vault protein (MVP) stone in the greenhouse until they had two outward leaves and one heart leaf. Then, three N fertilizer application rates were set up as follows: complete N treatment (N concentration was 5 mmol/L), 1/2 N treatment (N concentration was 2.5 mmol/L), and 1/4 N treatment (N concentration was 1.25 mmol/L). Before the V7 stage, 1.5 L nutrient solution was poured three times. A volume of 1 L nutrient solution was poured at the jointing stage and the male powder dispersing stage.

We collected leaf samples at the V6 stage and the filling stage. The 6th fully unfolded leaf and the leaf under the ear were cut off, and the hyperspectral images were obtained immediately. The veins and yellow areas of the leaves were then removed, and the remaining leaves were placed in tinfoil bags, frozen in liquid nitrogen, and stored in the refrigerator at –80°C for the amino acid content determinations. Six replicates were taken for the different N application experiments. A total of 144 samples were collected in Exp1.

(2) Exp2: N starvation treatment

Two inbred lines, namely, CIMBL123 and CML422, were selected as the test varieties. The Center for Crop Functional Genomics and Molecular Breeding of China Agricultural University provided all the test varieties. The maize seedlings were cultured in deionized water. In the early stage, the seedlings were cultured with a complete nutrient solution. The seedlings were treated with a low N treatment (0.05 mmol/L) when they had two leaves and one heart.

Leaf samples were collected every 3 days for a total of 13 times. The second fully expanded leaf was cut off from top to bottom, and the hyperspectral images were obtained immediately. The veins and yellow areas of the leaves were then removed, and the remaining leaves were placed in tinfoil bags, frozen in liquid nitrogen, and stored in the refrigerator at –80°C for the various amino acid content determinations. Six replicates were taken for the different N application experiments. A total of 146 samples were collected in Exp2.

The hyperspectral images of maize leaves were collected after each sampling. The Pika-L imaging spectrometer was used to obtain the hyperspectral images. Pika-L images provide the band range of 400–1,000 nm with a total of 300 spectral channels and 900 spatial channels. The spectral resolution was 2.1 nm. The pixel size is 5.86 μm with a field of view of 17.6°. This equipment has the advantages of low astigmatism, low distortion, and a high signal-to-noise ratio.

A hyperspectral image acquisition system was designed and is shown in Figure 1. The system was mainly composed of Pika-L, a personal computer (PC), a halogen lamp, a mobile carrier platform, a stepper motor, a speed controller, and a blade flattening device. A halogen lamp provided stable light similar to sunlight to obtain a stable hyper-spectrum of leaves. The power of the halogen lamp is 220 W. To reduce the influence of the external environment on image quality, the hyperspectral image acquisition of maize leaves was carried out in a relatively stable dark room. Each leaf was spread flat on the platform. The hyperspectral image of the leaf was obtained directly above the leaf using the Pika-L spectrometer. Before the experiment, the hyperspectral imaging system was turned on and preheated for 30 min. The parameters of this system were set as follows: exposure time was 4.35 ms, and the speed of the electronically mobile carrier platform was 6 mm/s.

The hyperspectral images obtained include green leaves and the background. The normalized difference vegetation index (NDVI) can be used to separate green leaves from the background. NDVI is calculated by the reflectance of the near-infrared band and the red band (Formula 1) (Thenkabail et al., 2000). This study set a threshold (NDVI > 0) to distinguish the leaf pixels from the background pixels. The average hyperspectral reflectance of green leaf pixels was obtained to estimate the content of amino acids in maize leaves.

where r_nir and r_red are the reflectance of 780 nm and 660 nm, respectively.

The amino acid content was determined using LC-MS. The liquid chromatography used was ACQUITY UPLC I-Class (Waters, United States). Mass Spectrometer adopted the Q Exactive Focus system (Thermo Fisher, United States). Thermo Xcalibur 4.0 was used for data analysis. The measurement process includes the following processes: (1) Sample processing. The leaf samples were ground into powder and freeze-dried. The 20-mg freeze-dried powder was weighed as a subsample, adding 1 ml of water. Then the subsample was shaken by an ultrasonic crusher for 30 min. The subsample was centrifugally rotated for 10 min at 14,000 rpm/min.

(2) Sample derivatization. A volume of 10 μl of supernatant was taken, 50 μl of borate buffer solution and 20 μl derivative solution were added, the resultant solution was placed at room temperature for 1 min and then derived on an oscillator at 55°C for 10 min.

(3) Suction and filtration. The derived sample was cooled to room temperature and then filtered using a 1-ml syringe and filter membrane. (4) Bottling and measuring sample. The filtered sample was transferred to the glass bottle, the sample on the machine was tested, and the data were exported. (5) Drawing the standard curve of amino acids. The standard sample of amino acids was diluted to different concentrations. The peak values of molecular ions varied gradually with the increase of solution concentration, showing a linear relationship. (6) Calculating the reference value of amino acids. The Thermo Xcalibur4.0 software was used to process the mass spectrogram. The types of amino acids were determined according to the retention time and mass-charge ratio, and the peak values of molecular ions were recorded. Finally, the contents of various amino acids were obtained by putting the ion peak value into the equation of the standard curve of various amino acids.

There were 24 amino acids in maize leaves, including alanine (Ala), γ-aminobutyric acid (GABA), β-aminobutyric acid (BABA), arginine (Arg), aspartic acid (Asp), citrulline (Cit), glutamic acid (Glu), glycine (Gly), histidine (His), isoleucine (Ile), leucine (Leu), lysine (Lys), methionine (Met), ornithine (Orn), phenylalanine (Phe), proline (Pro), sarcosine (Sar), serine (Ser), threonine (Thr), tryptophan (Trp), tyrosine (Tyr), glutamine (Gln), valine (Val), and asparagine (Asn).

Savitzky-Golay filter was used to remove noise from the hyperspectral reflectance. Savitzky-Golay filter is one of the commonly used filtering methods in spectral preprocessing and can improve the smoothness of the spectrum and reduce the noise interference (Dai et al., 2017). Due to the different magnitude of various amino acid contents, z-score standardization was used to deal with the amino acid content.

The estimation models of the 24 amino acid contents were constructed based on the reflectance of all bands, sensitive band range, and sensitive bands with the PLSR method. The PLSR, proposed by Herman Wold in the 1970s, cannot only reduce the dimension of the data but also solve the collinearity between the bands (Wu and He, 2014). In this study, the leave-one-out cross-validation was used to determine the number of principal components. We calculated the predicted residual error sum of squares (PRESS) of the predicted value of n–1 principal component and selected the principal components with the lowest PRESS for regression modeling. For all models, 70% (203) of the samples were used as the training set to construct the model, and the remaining 30% (87) were used as the test set to evaluate the model’s accuracy. To eliminate the random error, the modeling process was repeated 100 times, and the average result of the 100 repetitions was taken as the final result.

Hyperspectral data contain hundreds of bands. Data redundancy and multicollinearity need to be addressed. Studies have shown that only using sensitive bands to establish the model can not only reduce the computational burden but also improve the accuracy and stability of the model (Wan et al., 2020). In this study, the reflectance of maize leaves was obtained at 400–1,000 nm. The greater the reflectivity variability of this band, the more sensitive it is to amino acids. The CV (Equation 2) was used to determine the sensitive band range of each amino acid.

where SD and mean represent the standard deviation and mean value, respectively.

Using the selected sensitive band range, we constructed the PLSR model of each amino acid and performed the regression coefficient test of the model. When screening sensitive bands, we referred to the study by Meng et al. (2013). Taking the band reflectance of the two regions as input variables, the estimation models of amino acids in maize leaves were established based on PLS regression. The regression coefficient was used to quantify the correlation between the band and the model. The larger the absolute value of the regression coefficient, the stronger the correlation between the band and the model. The absolute values of the regression coefficients of each band were sorted from small to large. The bands were removed one by one, and the model was then reconstructed. The reconstructed model was evaluated according to the PRESS. The band was counted when the PRESS value of the model was at its minimum. The above process was repeated 100 times. The bands with frequencies greater than 80 Hz used in modeling with the minimum model PRESS were taken as the sensitive bands of that amino acid.

The evaluation indices of the model include the determination coefficient (R²), root mean square error (RMSE), relative root mean square error (RE), and relative percent deviation (RPD). The average values of R², RMSE, RE, and RPD with test set for 100 times were used to evaluate the performance and stability of models.

where m is the number of samples, y_i, yi¯, yi^ are the measured and the predicted values of various amino acid contents of sample i, and SD represents standard deviation.

The descriptive statistics for the entire sample are reported in Table 1. The descriptive statistics of the data included range, standard deviation (SD), and CV. The mean values of Sar, Ala, Glu, and Ser were relatively large, indicating that these amino acid contents in the samples were relatively high. The CV of Gln, Asn, Ser, and Gly was larger than the other amino acids, which may be that these amino acids were more sensitive to N treatment.

Figure 2 shows the comparison of various amino acid contents in maize leaves of two inbred lines sampled at the early and later stages of the nitrogen starvation experiment. In Figure 2, the early and later stages refer to the first three and the last three samples in the nitrogen starvation experiment, respectively. It can be seen that the contents of various amino acids of the two inbred lines in the later stage were lower than those in the early stage. The contents of alanine, γ-aminobutyric acid, arginine, glutamic acid, proline, sarcosine, threonine, and tyrosine in the later stage were significantly lower than those in the early stage.

With the spectral reflectance of all bands as the independent variable and the amino acid contents as the dependent variable, we established the PLSR model of 24 amino acid contents. The validation results of the model using the test set are shown in Table 2. The estimation accuracies of β-aminobutyric acid, ornithine, citrulline, methionine, and histidine were the best, with R², RE, and RPD of the test set in the range of 0.84–0.95, 9.68%–20.38%, and 2.52–4.95. The models of sarcosine, alanine, glutamic acid, proline, threonine, aspartic acid, and leucine had relatively good estimation accuracy, with R², RE, and RPD of the test set in the range of 0.57–0.73, 23.23%–39.75%, and 1.53–1.95. The performance of the other amino acid models was relatively poor.

Figure 3 shows the CV value of the spectral reflectivity of each sample (A) and all samples (B). The CV and variation range of the samples were large in the range of 400–717.08 nm and small in the range of 717.08–1,100 nm. We further constructed and validated the PLSR model of each amino acid based on the reflectance in the ranges of 400–717.08 nm and 717.08–1,100 nm, respectively. The results are shown in Table 3. The estimation model of citrulline was relatively good when using the reflectance in the range of 717.08–1,100 nm, while the estimation models of most other amino acids performed well when using the reflectance in the range of 400–717.08 nm. Therefore, the bands in the range of 400–717.008 nm were more sensitive to various amino acids than those in the range of 717.08–1,100 nm.

The specific sensitive bands of various amino acids were further screened in the range of 400–717.08 nm. We established the PLSR model of each amino acid using the reflectance in the range of 400–717.08 nm and performed the regression coefficient test of the model. Figure 3 shows the usage frequency of each band in 100 times modeling. The dark colors indicate the more times the band appeared and the more important the band was. As seen in Figure 4, the sensitive bands of most amino acids were mainly concentrated in the ranges of 505.39–604.95 nm and 651.21–714.10 nm.

Table 4 shows the validation results of PLSR model using test set for each amino acid based on the sensitive bands. The estimation accuracies of methionine, ornithine, sarcosine, alanine, and asparagine were improved compared with the models constructed with the reflectance in the range of 400–717.08 nm. However, the estimation accuracies of alanine, histidine, threonine, tryptophan, citrulline, β-aminobutyric acid, and aspartic acid were almost unchanged, and those of other amino acids were relatively decreased. In summary, when modeling with the reflectance of the sensitive bands, the estimation accuracies of 11 amino acids by test set were improved or equivalent to that of the model using the reflectance of band range.

The evaluation results of the optimal model for each amino acid and the bands used are summarized in Table 5. It generally suggests that the model estimation accuracies of β-aminobutyric acid, ornithine, citrulline, methionine, histidine, and sarcosine using test set were relatively high, with R² more than 0.7. Among the 24 amino acids, five amino acids obtained the best estimation accuracy based on the reflectance of sensitive bands. A total of 15 amino acids obtained the best estimation accuracy based on the reflectance of the sensitive band range, of which 14 amino acids used the reflectance in the range of 400–717.08 nm.

Figure 5 shows the results of testing one model randomly selected from 100 PLS regression models by the optimal estimation method. The predicted values of histidine, sarcosine, glutamic acid, and alanine were close to the measured values. The measured and predicted values of threonine, proline, leucine, and aspartic acid also matched well.