Estimating foliar anthocyanin content of purple corn via hyperspectral model

Abstract To date, the foliar anthocyanin content was either determined via the pH differential or HPLC methods, both of which are slow and destructive. Here, a hyperspectral model was established to estimate the foliar anthocyanin content of purple corn (Zea mays L. var. Jingzi No. 1). The reflectivity (P) of the foliar hyperspectral was inverted to 1/P, lg P, 1/lg P, P′, 1/P′, lgP′, and 1/lgP′. The correlation coefficient between these inversions and the foliar anthocyanin content was plotted against the hyperspectral wavelength. The wavelength of inversions around 650 nm was sensitive to the foliar anthocyanin content. The hyperspectral model was fitted via linear, polynomial, power, exponential, and logarithmic functions with the sensitive band as independent variable and the anthocyanin content as function. The hyperspectral model (y = 3,000,000,000 × W 685 4.5896) fitted via inversion of lgP′ showed the highest determination coefficients (0.768) among all models. The hyperspectral model was well validated with a determination coefficient of 0.932 and an RMSE of 0.0065. Moreover, the accuracy and stability of the hyperspectral model were further enhanced with a determination coefficient of 0.954 and RMSE of 0.0047 when the anthocyanin content of the sample was below 20 mg/g. Hence, the hyperspectral model estimated the foliar anthocyanin content of purple corn quickly and nondestructively.

In contrast, a hyperspectrum technique is a nondestructive, quick, and simple method to measure the pigment of plants (Palmer et al., 2006;Qin & Lu, 2008). The technique measures the foliar nitrogen content (He et al., 2016), chlorophyll content (Croft et al., 2015;Zhang, Chen, Miller, & Noland, 2008), and moisture content (Clevers, Kooistra, & Schaepman, 2010). However, the hyperspectral model of the anthocyanin content was not plotted due to a weak relationship between the foliar anthocyanin content and the hyperspectrum. One possible reason for this was the overlap between the chlorophyll and anthocyanin absorption peaks (Sims & Gamon, 2002). The other reason was the influence of the moisture and leaf scattering absorption (Hatfield, Gitelson, Schepers, & Walthall, 2008).
Therefore, a hyperspectral model was developed to evaluate the foliar anthocyanin content of Jingzi No. 1. To reduce the influence of chlorophyll and moisture as well as leaf scattering, a sensitive band was selected via multiple linear regressions. Then, the hyperspectral model was inversed via linear, polynomial, power, exponential, and logarithmic functions with the sensitive band as independent variable and the anthocyanin content as function. The hyperspectral model will provide nondestructive and quick support for the harvest of purple corn. Moreover, the further application of the hyperspectral model based on the geographic information will provide instant and reliable support for the harvest decision at large scale.

| Foliar sampling
Jingzi No. 1 (Zea mays L.) was harvested 85 days after sowing in September 2014 and 2015 in Yanqing Farmer. The whole leaf was dark red with a moisture content of 6%-12%. A piece of leaf was randomly picked from a strain of the corn and was cut into an 8 cm × 6 cm blade in the middle. The cut leaf was immediately scanned via mobile hyperspectral radiometer (ASD Field Spec Pro FR, US), which was coupled with a LI-Cor1800-12S external integrating sphere. After calibration with the white board, the hyperspectrum was collected from 350 to 2,500 nm at an interval of 1 nm. The hyperspectrum was scanned with a distance between sample and radiometer of 5 cm and a view angle of 25° on a sunny and clear day from 10:00 a.m. to 2:00 p.m. The foliar spectral was repeated 10 times, and the obtained results were averaged. The external integrating sphere was used to ensure the repeatability of the spectra. The data were processed with the software ViewSpec Pro, Version 2.14 (Analytical Spectral Device, Inc5335 Sterling Drive Suite A, Boulder, CO 80301). The scanned leaf was numbered and sealed in a polyethylene bag for anthocyanin content determination in the laboratory. A total of 500 pieces of leaves were collected.

| Measurement of pH differential method
The anthocyanin content of samples was determined via pH differential method previously described (Zhao et al., 2008). The leaf (10 g) was smashed and stirred in 50 ml liquid (a solution of 60% (v/v) ethanol acidified with citric acid (1%, w/v)) at 60°C for 120 min. The ethanol extracts were centrifuged at 9,000 rpm and 20°C for 10 min. The supernatants were evaporated to dryness at 46°C with a rotary evaporator Büchi R-3000 (Büchi Labortechnik AG, Switzerland). Then, the concentrate was freeze-dried. An aliquot of the dried concentrate (1 mg) was placed into a 25-ml volumetric flask and filled to the final volume with pH 1.0 buffer.
Another 1 mg of the sample was placed into a 25-ml volumetric flask and filled to a final volume with pH 4.5 buffer. Absorbance was measured via spectrophotometer (UV-1800, Shimadzu, Japan) at 510 and 700 nm, respectively. Absorbance was calculated as Abs = (A 510 nm − A 700 nm ) pH 1.0 − (A 510 nm − A 700 nm ) pH 4.5 with the molar extinction coefficient for cyanidin 3-glucoside of 26,900.
Total anthocyanin content was calculated using the following equation and expressed as grams of cyanidin 3-glucoside equivalents per 1 g sample (Equation 1).
where Abs represents the absorbance, e represents the cyanidin 3-glucoside molar absorbance [26,900 ml/(mmol·cm)], L represents the cell path length (1 cm), MW represents the molecular weight of anthocyanin (449.2 Da), D is a dilution factor, V represents the final volume (ml), and G represents the dry material (mg).

| Screening of the sensitive band
The reflectivity of the hyperspectrum (P) was inverted to the reciprocal of the reflectivity (1/P), the logarithm of the reflectivity (lg P), the reciprocal of the logarithm of the reflectivity (1/lg P), the firstorder differential of the reflectivity P ′ , the first-order differential of the reciprocal of the reflectivity 1∕P � , the first-order differential of the logarithm of the reflectivity ( lg P ′ ), and the first-order differential of the reciprocal of the logarithm of the reflectivity 1∕ lg P � .
The differential inversion of spectral reflectance was calculated via where P represents the reflectance of a band of λ and Δλ represents the interval from λ i to λ i−1 .
The correlation coefficient between the foliar anthocyanin content and hyperspectral vectors or the inverted vectors was evaluated via

| Establishment and validation of the hyperspectral model
The hyperspectral model for the anthocyanin content was fitted via linear, polynomial, power, exponential, and logarithmic functions with the sensitive band as independent variable and the anthocyanin content as function. Specifically, the reflectivity value of the sensitive band was plotted against the foliar anthocyanin content via linear (Y = a × X + b), where LAC i and PLAC i represent the anthocyanin content and predicted anthocyanin content of the purple corn leaf, respectively; N represents the number of the validation.

| Screening of sensitive bands
Hyperspectrals usually contain noise due to atmospheric, instrumental, and geometric distortions (Gomez, Oltra-Carrió, Bacha, Lagacherie, & Briottet, 2015). Consequently, reducing the atmospheric influences and shortening the hyperspectrum range reduced the noises of the hyperspectral. Specifically, the LI-C or 1800-12S external integrating sphere was coupled with the hyperspectral radiometer. The external integrating sphere provided stable illumination and appropriate reflection for the sample, thus reducing the noise of the reflectivity.
Moreover, moisture is another factor that enhanced the noise of the reflectivity (Croft et al., 2015;Zhang et al., 2008). The hyperspectrum of 1,400~2,500 nm is sensitive to the moisture content, especially to the bands of 1,450 and 1940 nm (Clevers et al., 2010). Consequently, only the hyperspectrum of 400~1,400 nm was used in the following inversions to reduce noise.
Prof. Zhou, a reviewer, suggested that the reflectivity of chlorophyll will overlap with that of anthocyanin, and the moisture and some The 1/P of the hyperspectral was positively correlated with the foliar anthocyanin content of purple corn. The correlation coefficient of the visible light band was higher than that of both the near-infrared and middle-infrared bands. The correlation coefficient of 1/P was

| Modeling
The hyperspectral model was fitted via linear, polynomial, power, exponential, and logarithmic functions with the sensitive band as the independent variable and the anthocyanin content as the function (Table 1).
Each model was randomly trained by a total of 400 samples to ensure  " was used to estimate the foliar anthocyanin content of purple corn.

| Validation of the hyperspectral model
The hyperspectral model based on the inversion of the lg P ′ was validated via the remaining 100 samples (Figure 3)