*2.1. Experiments and Materials*

The experiments were conducted in Hengshui City, Hebei Province, China. Seventy-two sampling areas were present in the test field, as shown in Figure 1, with six fertilization levels to ensure the gradient of chlorophyll content. The nitrogen fertilizer was pure nitrogen, and the phosphorus fertilizer was P2O5. The spectral data were collected in three growth periods on the basis of the growth time and conditions, namely, G1 (6 leaf stage), G2 (9 leaf stage), and G3 (12 leaf stage). During growth periods, one leaf sample was collected in each sampling area, so that two hundred and sixteen leaf samples were collected.

**Figure 1.** Locations and treatments of the experiment.

The overall process of analyzing reflection spectrum data and chlorophyll content is shown in Figure 2; this mainly included the collection of spectrum data, preprocessing of the spectral data, selection of the characteristic variables, and establishment of a detection model for the chlorophyll content. The selection of the characteristic variables was conducted on the basis of comparison of the characteristic wavelengths and wavelet features selected by maximum correlation coefficients and local extrema of the correlation coefficients.

**Figure 2.** Data analysis flow chart.

#### *2.2. Field Spectrum Data Collection and Chlorophyll Content Measurement*

The ASD FieldSpec® HandHeld 2 was used in the field to measure the canopy reflectance of corn. This tool is a hand-held spectrometer with a wavelength range of 325–1075 nm, wavelength accuracy of 1 nm, and spectral resolution of <3.0 nm at 700 nm [40]. Sample leaves were randomly chosen in each sample area to measure the spectral reflectance. Sample leaves were then sealed for subsequent chlorophyll extraction experiment. Spectrum reflectance data were collected three times above the leaf during the spectrum measurement. The averaged reflectance was taken as an original spectral datum.

The chlorophyll content of the sample leaves was measured in the laboratory via SHIMADZU UV2450 spectrophotometry. The spectrophotometry measurement wavelength range was 190–900 nm and the band width was 0.1–5 nm. The spectral resolution was 0.1 nm and the stray light was lower than 0.015%. The main stems of corn leaves were removed, and the leaves were shredded and evenly mixed. Approximately 0.4 g crushed leaf samples were soaked in 25 mL acetone and absolute ethanol mixture for 24 hours, and the mixture ratio was 2:1. During soaking, the mixed solutions were shaken three times to accelerate the chlorophyll extraction. The absorbances of extract solution at 645 and 663 nm were measured with a UV spectrophotometer. The concentrations of chlorophyll *a* (*Chla*) and chlorophyll *b* (*Chlb*) were calculated using the following equations:

$$\text{Cll}\_4 = 12.72 \times A\_{663} - 2.59 \times A\_{645\prime} \tag{1}$$

$$\text{Chl}\_b = 22.88 \times A\_{645} - 4.67 \times A\_{663} \tag{2}$$

$$\text{Cbl}\dagger = \text{Cbl}\_{\text{d}}\text{(mg L}^{-1}) + \text{Cbl}\_{\text{b}}\text{(mg L}^{-1}),\tag{3}$$

where *A*<sup>645</sup> and *A*<sup>633</sup> are the absorbances of the extract solution at 645 and 663 nm, respectively, and *ChlT* is the total chlorophyll [41].

#### *2.3. Spectrum Data Preprocessing*

The corn canopy spectrum collected in the field environment contained noise information due to the uneven surface of the sample, random noise, different optical paths, and light scattering. First, a Savitzky–Golay (S-G) filter was used to smoothen the reflection spectrum, and the smoothing window was set to 13 [42]. The S-G filter is based on the principle of least squares. Multiple fitting was performed to the original signal in the correction window and the final conversion result was calculated by the multiplicity of the fitting. Using *m* (*m* is odd) continuous wavelength points as the window, the data points inside the smooth window were fitted by *p*-order polynomial function, and the polynomial equation combination was obtained. The smoothing coefficient was obtained using least-square fitting, and the corrected spectral value of the center point of the window was calculated. By successively moving the position of the smoothing window and repeating the above polynomial fitting steps, the spectra after S-G filtering were obtained.

Second, the standard normal variable (SNV) method was used to process the smoothed spectral curve to reduce the influence of the scattering effect [43]. Standard normal variable correction is often used to eliminate the effects of different particle sizes, surface scattering, and optical path differences in NIR diffuse reflectance spectra. The SNV correction of sample spectra were independent of each other and did not involve the spectral information related to the sample set. First, the sample spectrum was centralized, which means that the mean value of spectral reflectance of each spectral data was subtracted from the sample. The standard deviation of the sample reflectance was then used to scale up. After SNV correction, the spectral mean of each sample became 0 and the variance became 1. The SNV correction spectrum of sample *j* is as follows:

$$A\_{j,SNV} = \left(A\_j - \overline{A\_j}\right) / \sigma\_{j,\prime} \tag{4}$$

where *Aj* is the mean value of the spectrum of sample *j* and σ*<sup>j</sup>* is the standard deviation of the spectrum of sample *j*.

Wavelet analysis is one of the potential technologies used in the extraction of weak hyperspectral information. Wavelet transform is a function combination that decomposes a complex signal into simple subsignal components. The spectral signal can be decomposed into subsignals of different frequencies when applied to the analysis of crop spectral data. We effectively used the overall structural characteristics of spectral information and extracted the weak information hidden in the spectral signal. Moreover, we searched for the optimal combination of the subsignal components to detect the chlorophyll content of the crop canopy.
