Estimation of Chlorophyll Content in Wheat Based on Optimal Spectral Index

Abstract

Chlorophyll content in wheat leaves reflects its growth and nutritional status, which can be used as a health index for field management. In order to evaluate the potential of hyperspectral data to estimate the chlorophyll content in wheat leaves, this study focused on the leaves of wheat at the flag-picking stage, flowering stage, grain-filling stage, and maturity stage. Based on the framework of five vegetation indexes, the spectral index was constructed by using the combination of 400–1000 nm bands. The correlation between the constructed spectral index and the measured chlorophyll value was analyzed, and the optimal spectral index was screened using the correlation coefficient. Based on the optimal spectral index, polynomial regression, random forest, decision tree, and artificial neural network were used to establish the estimation model for chlorophyll value, and the optimal model for estimating the chlorophyll value of wheat leaves was selected through model evaluation. The results showed that the five optimal spectral indices at the four growth stages were primarily composed of the red band, red edge band, and near-infrared band. The five optimal spectral indices during the grain-filling stage had the highest correlation with the chlorophyll value, and the absolute value of the correlation coefficient was greater than 0.73. The accuracy of the estimation model established in the four growth stages was different, with the estimation accuracy of the flag stage being the best, showing an R² and RMSE of 0.79 and 2.63, respectively. These results indicate that hyperspectral data are suitable for estimating the chlorophyll value of wheat leaves, and the polynomial regression model of the flag-picking period can be used as the optimal model for estimating the chlorophyll value of wheat leaves.

1. Introduction

As one of the three major cereals, wheat is also the primary grain in China, making it critical to monitor the growth status of wheat [1]. As the most important pigment in the process of photosynthesis, chlorophyll is directly related to the light energy utilization of wheat. The variation in the chlorophyll index reflects the growth status of wheat at different stages and can serve for agricultural production and field activities [2]. Therefore, accurate estimation of the chlorophyll content is an essential prerequisite for high-yield and high-quality wheat production [3].

Traditional methods for measuring chlorophyll content may require manual sampling in the field combined with laboratory quantitative testing, which is time-consuming. Another limitation is that the method is not suitable for large-scale crops [4]. In recent decades, remote sensing hyperspectral technology has been widely used in crop growth monitoring, yield estimation, and pest prediction [5], and also provides a new method for estimating the chlorophyll content in crop leaves. Red light and near-infrared are sensitive to green crops in the remote sensing spectral band, which contains a large amount of spectral information related to leaf chlorophyll content, and many scholars have carried out a lot of research. For example, as early as 1983, Horler et al. [6] studied the correlation between the spectral characteristics of vegetation and chlorophyll concentration and found that the “red edge” position of the spectrum (i.e., the wavelength value at the maximum value of the derivative spectrum of vegetation near 700 nm) has a great contribution to the estimation of vegetation chlorophyll concentration. Curran (1989) [7] used data to establish a regression equation between the location of the red edge and the chlorophyll content: CHL = 0.05 × REP − 32.1. Wu et al. [8] (2000) found that the spectral reflectance characteristics of vegetation populations at the 762 nm wavelength band and their derivative spectral data had a high correlation with chlorophyll density and developed a regression model based on the correlation. After that, Schlemmer et al. [9] proposed a calculated normalized difference vegetation index (NDVI) by processing the hyperspectral reflectance of maize leaves as a derivative and realized an efficient estimation of maize chlorophyll content based on multivariate linear regression. Shen et al. [10] built a partial least squares regression model to estimate the chlorophyll content in wheat leaves.

The results show that remote sensing monitoring based on vegetation index (VI) has spatiotemporal continuity, which is convenient for long-term monitoring of crop growth information and has a high application value in chlorophyll estimation [11]. Qiao et al. [12] studied the response characteristics of leaf chlorophyll content (LCC) and the vegetation index of field maize under different coverage conditions. They constructed a canopy LCC estimation model using random forest (RF) and partial least squares methods. Jia et al. [13] found that a new combination of spectral indices aided in determining sensitive bands, enhanced the correlation of leaf nitrogen content, and performed reliably in the regression model of nitrogen content in flue-cured tobacco leaves. With the rapid development of machine learning algorithms, multiple linear regression (MLR) [14], partial least squares regression (PLSR) [15], random forest regression (RFR) [16], support vector machine regression (SVR) [17], and other methods have achieved good results in estimating leaf chlorophyll content. Liu et al. [18] utilized continuous wavelet transform to analyze the chlorophyll content in potatoes. The results showed that the continuous wavelet transform method could effectively extract the wavelet characteristics that were sensitive to the chlorophyll content in potatoes and enhance the correlation between the spectrum and chlorophyll content. Scholars have extensively used various machine learning algorithms and methods to conduct research on the quantitative monitoring of crop attributes, achieving remarkable results. Therefore, the choice of appropriate spectral pretreatment methods, vegetation index methods, and machine learning methods is crucial for inverting the chlorophyll content in crop leaves using spectroscopy. Therefore, it is particularly important to choose appropriate spectral pretreatment methods, vegetation index methods, and machine learning methods to invert the chlorophyll content in crop leaves by spectroscopy.

The analysis of the dynamic changes in chlorophyll content and vegetation index at different growth stages can directly or indirectly reflect the growth status of wheat. In this study, the chlorophyll content in wheat at four different growth stages, namely the flag-picking stage, flowering stage, grain-filling stage, and maturity stage, was taken as the research object. A five-point smoothing pretreatment was performed on the original hyperspectral reflectance. The correlation matrix method was used to screen out the optimal combination bands of chlorophyll content and five vegetation indices in the range of 400–1000 nm, and the chlorophyll content in wheat was estimated by polynomial regression, random forest, decision tree, and artificial neural network algorithm. This study explored the optimal vegetation index combination and the best estimation model at different growth stages, providing a theoretical basis and technical support for real-time monitoring of wheat growth and development.

2. Materials and Methods

2.1. Study Area

The study area is located at the Xinxiang site of the Chinese Academy of Agricultural Sciences (35.2° N, 113.8° E), with a plot area of 3.0 m × 1.4 m, as shown in Figure 1. Each plot has a separator of 0.2 m in the east–west direction and 1.0 m in the north–south direction during the winter wheat growing season between 2020 and 2021. The wheat variety was Zhengmai 369 [19], and the soil type was fluvial soil. The natural conditions of all plots in the experiment were completely the same; the same standard field operations and management, such as fertilization, watering, insect control, weeding, seedling leveling, etc., were implemented on each plot. Xinxiang City experiences a mild continental monsoon climate with four distinct seasons, and the seasonal characteristics are obvious. The spring is dry and windy, the summer is hot and rainy, and the average temperature difference between day and night in autumn is large. The winter is cold, with scarce rain and snowfall. The average precipitation and the inequality of distribution are mainly concentrated from June to September, and the annual precipitation generally accounts for about 75% of the national precipitation, which is suitable for the growth of wheat (Figure 1).

2.2. Hyperspectral Data Acquisition

Hyperspectral data of the wheat canopy were acquired using the FieldSpec^®4 Hi-Res portable object spectrometer manufactured by ASD (Analytica Spectra Devices, Inc., Longmont, CO, USA) from the United States [20]. The instrument collects the spectral information in a view of 25° at the 350 nm~2500 nm band [21]. The sampling interval is 1.4 nm in the band range of 350 nm to 1000 nm, and the sampling interval is 2 nm in the band range of 1000 nm to 2500 nm, and the spectral resolution is 3 nm [22]. Spectral reflectance data of the wheat canopy were covered over the flag-picking period (2 May), flowering stage (12 May), grain-filling stage (17 May), and maturity period (24 May) of wheat, including the key growth period of wheat. The experiment was carried out between 11:00 local time (LT) and 12:00 LT in sunny or cloudy weather. During the collection, the cloud coverage was less than 20%. The hyperspectral detector was in a vertical state, and the heights of the detector and the canopy were approximately 30 cm for the measurement of each study area. A 40 cm × 40 cm whiteboard was utilized to reduce the influence of natural light on the spectrum [23]. We measured the spectral curve 10 times for each cell. After measurements, data were processed using View Spec Pro 6.0 software to extract the spectral reflectance, then the spectral reflectance value of each cell is averaged over the 10 spectral curves.

2.3. Determination of Chlorophyll Content

Chlorophyll content was measured with a portable chlorophyll meter, and three samples with uniform growth representing the growth of the whole plot were selected from each experimental area. The chlorophyll content in the top leaves for each sample was estimated by averaging values over the three samples. In this study, SPSS25 is mainly used for data analysis. Under the same study conditions,

Table 1. Statistics of chlorophyll content.

Child-Bearing Period	Minimum Value	Maximum Value	Mean Value	Variance
flagging stage	47.8	58.8	54.5	6.6
flowering period	45.1	57.8	51.4	10.8
pustulation period	37.8	53.2	47.4	16.4
maturity	5.1	38.4	17.2	69.8

shows the chlorophyll content at different growth stages. It is clear to note that the statistic value gradually decreased from the flagging stage to the maturity stage in terms of minimum, maximum, and mean values. However, the variance shows an inverse trend during the whole growing cycle.

2.4. Construction of Vegetation Index

Since the vegetation spectrum is mainly water absorption after 1100 nm, the reflectance of the spectral curve of wheat in the water absorption zone will decrease. Moreover, it is irregular, so spectral reflectance data from 400 to 1000 nm were used to analyze and construct an estimate of the chlorophyll content model [24].

Based on the spectral reflectance curves of wheat, the general trend of spectral reflectance of wheat in different growth stages is basically the same, and the results are shown in Figure 2. The overall trend is low reflectivity in the visible band (400–700 nm) and high reflectivity in the near-infrared band (800–1000 nm). It can be seen that the spectral curve characteristics are closely related to the chlorophyll content, and the chlorophyll content can be estimated using the reflectance spectrum of wheat.

The correlation between the chlorophyll content and the spectral reflectance of wheat was obtained by analyzing the correlation between the denoised hyperspectral reflectance of wheat at the flag-picking stage, flowering stage, grain-filling stage, and maturity stage and the corresponding chlorophyll content. Figure 3 shows that there were significant differences in the correlation between spectral reflectance and chlorophyll content in leaves at different growth stages. However, the absolute value of the correlation coefficient is at most 0.32, which indicates that the correlation between chlorophyll content and spectral reflectance is weak. On the whole, the correlation between the original spectral reflectance and chlorophyll content in the four growth stages was low.

In order to solve the problem that the spectral characteristics are easily influenced by the differences in the physiological information of the crops themselves [25,26], the correlation matrix method was used to construct a spectral index sensitive to wheat information to optimize the estimation of chlorophyll in wheat, and the following spectral index was selected, which is detailed in

Table 2. Vegetation index formula.

Vegetation Index	Computing Formula	Explanation
Normalized Difference Vegetation Index (NDVI)	${(R}_i-R_j)/{(R}_i+R_j)$	$i,j$ are any band in the spectrum; $R_i,R_j$ are reflectivity
Difference Vegetation Index (DVI)	${(R}_i-R_j)$
Ratio Vegetation Index (RVI)	${(R}_i/R_j)$
Renormalized Difference Vegetation Index (RDVI)	${(R}_i-R_j)/sqrt{(R}_i+R_j)$
Green Chlorophyll Index (GCI)	${(R}_i-R_j)-1$

. In the spectral wavelength range of 400–1000 nm, the spectral index was calculated for the combination of all wavelengths of the original hyperspectral [27], and then the correlation between all spectral indices and chlorophyll content was analyzed. The wavelength positions of i and j, where the maximum correlation coefficient was located, were used as the optimal wavelength combination, and the calculated spectral index value was the optimal spectral index with the highest correlation with the chlorophyll index of wheat.

2.5. Modeling Methods

Polynomial Regression (PR) is a special case of linear regression in which the dimensionality of a variable is converted from nonlinear to linear. This is a way of increasing the number of times an independent variable, and when we set the number (greater than 1) on an independent variable, we attain the result of projecting these data into a higher power space.

Random Forest (RF) is a bagging ensemble method based on a set of decision trees [28]. The basic principle of random forest is to utilize the self-service method resampling technique to extract a plurality of self-service sample sets from the original sample and carry out decision tree modeling for each self-service sample set. Each node variable of the tree is generated. Node segmentation in a plurality of predictors is randomly selected, and the average value of all decision tree predictions is taken as the final prediction result. In general, a random forest will randomly generate hundreds or thousands of classification trees and then select the tree with the highest repetition as the final result, which improves the prediction accuracy of the model by summarizing a large number of classification trees [29]. When using the random forest regression algorithm to construct the model, the overall error of different data sets can be balanced, and the operation performance can be improved by adjusting a few parameters, and the performance is higher.

Decision Tree (DT) is a predictive model used to represent the mapping relationship between the attributes of entities and their corresponding value. Each node in the tree represents an object, and the path of each branch represents a possible property, with each leaf node corresponding to the path that this leaf node travels through. When constructing a decision tree, the characteristics to be divided are typically chosen based on the principle of maximizing purity after partitioning. Additionally, if it is a tree, pruning is often necessary to eliminate nodes that could result in increased validation errors, thus preventing overfitting [30].

An Artificial Neural Network (ANN) is a computational model inspired by the human brain’s function, designed to solve practical problems by mimicking the brain’s information processing capabilities. It is an algorithmic model that can judge the problems of the human brain from the perspective of information processing. The ability to approximate and the information is used to complete the simplest abstract calculations [31]. The neural network is composed of an input layer, an output layer, and one or more hidden layers, which are given corresponding weights through the connection between neurons, and the weights are continuously adjusted through training and learning algorithms to ultimately reach an optimum.

2.6. Accuracy Evaluation Index

Three indicators were selected for the comprehensive evaluation of different models. They are the deterministic coefficient (R²), root mean square error (RMSE), and relative error (RE). Among the three indicators, the closer the R² value of the coefficient of determination is to 1, the smaller the RMSE and RE are, which indicates the higher accuracy of the model and the more consistent the prediction effect is with the actual situation. The calculation formula is as follows:

$${\mathrm{R}}^2=1-\frac{{\sum }_{i=1}^n{\left(\widehat{y_i}-y_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}$$

(1)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(\widehat{y_i}-\overline{y}\right)}^2}$$

(2)

$$\mathrm{RE}=\frac{1}{n}{\sum }_{i=1}^n\frac{\left|{\widehat{y}}_i-y_i\right|}{y_i}$$

(3)

where $y_i$ represents the observed value of the test sample; $\widehat{y_i}$ represents the estimate of the test sample; $\overline{y}$ represents the average of the observations of the test sample; n sample number.

3. Results

Python software 2020.1 was used to calculate NDVI, DVI, RVI, RDVI, and GCI at any two wavelength combinations at the flag stage, flowering stage, grain-filling stage, and maturity stage of wheat, and the correlation analysis with the chlorophyll content in wheat was carried out. The correlation coefficients between NDVI, DVI, RVI, RDVI, and GCI and canopy chlorophyll content were distributed diagonally at (400, 400) and (1000, 1000), and the two-dimensional distribution map of the correlation coefficient R was generated for each growth period, and different colors were assigned according to the size of R [32]. As shown in Figure 4, the horizontal and vertical coordinates are the hyperspectral wavelengths of the wheat canopy, and the wavelength range is 400–1000 nm. In the figure, each point is the value of the correlation coefficient between the vegetation index and the chlorophyll content in the canopy constructed by the combination of the two wavelengths of the horizontal and vertical axes corresponding to the point. The shift in chlorophyll content and spectral index color from blue to red in wheat was indicated by a change from a negative correlation to a positive correlation.

3.1. Chlorophyll Correlation Analysis at Different Growth Stages

According to the correlation between the vegetation index and chlorophyll content at the flag-picking stage, the two-dimensional matrix was analyzed to determine the optimal combination band. The optimal combination bands of NDVI, DVI, RVI, RDVI, and GCI are, respectively, $R_{679}{R}_{683}=-0.581,R_{984}{R}_{982}=-0.469,R_{637}{R}_{431}=-0.368,R_{681}{R}_{682}=-0.528,$ and $R_{683}{R}_{679}=-0.582;$ It can be seen that the correlation coefficients of NDVI, RNDVI, and GCI are all greater than 0.5. Therefore, we chose the combination of NDVI and GCI with a high correlation coefficient with chlorophyll content during the flag-picking period, which can be used to establish the model in the later stage (Figure 4).

At the flowering stage, the correlation coefficients of NDVI and RDVI can reach more than 0.6, while the correlation of RVI is lower than 0.5, which is not suitable for modeling. NDVI and RDVI were selected as input variables for modeling the chlorophyll content in wheat during the flowering stage.

At the grain-filling stage, the correlation coefficients of NDVI, RVI, and GCI all reached more than 0.6, so we selected these three vegetation indices of NDVI, RVI, and GCI as the optimal vegetation indices of wheat chlorophyll content at the grain-filling stage, which were used to establish the model.

At maturity, the correlation coefficients of NDVI (989, 996), DVI (403, 401), RVI (576, 709), RDVI (493, 401), and GCI (989, 996) were −0.588, −0.532, −0.344, −0.550, and −0.588, respectively. Through the analysis, it can be seen that the correlation coefficients of NDVI, DVI, RDVI, and GCI all reached 0.5, so the two vegetation indices of NDVI and GCI were selected for the construction of the later model at the maturity stage. According to the correlation analysis, the vegetation index with a high correlation was selected to construct the model.

Table 3. Selection of independent variables at different growth stages.

Child-Bearing Period	Variable
flagging stage	$\mathrm{NDVI}(R_{679}{R}_{683}$$),\mathrm{GCI}(R_{683}{R}_{679}$)
flowering period	$\mathrm{NDVI}{(R}_{407}{R}_{412})$$,\mathrm{RDVI}(R_{991}{R}_{989}$)
pustulation period	$\mathrm{NDVI}(R_{587}{R}_{720}$$),\mathrm{RVI}(R_{587}{R}_{934}$$),\mathrm{GCI}(R_{587}{R}_{724}$)
maturity	$\mathrm{NDVI}(R_{989}{R}_{996}$$),\mathrm{GCI}(R_{989}{R}_{996}$)

shows the selection of independent variables.

3.2. Modeling Results and Analysis

Through the above correlation analysis, the optimal vegetation index with a good correlation with chlorophyll content was determined. The input variable of this model is the optimized spectral index with a high correlation with chlorophyll content, in which 80% of data were used for modeling, 20% of data were used for verification, and the modeling accuracy verification set models R², RMSE, and RE are shown in

Table 4. Wheat model results.

Child-Bearing Period	Modeling	Modeling Set	Validation Set
		R2	R2	RMSE	RE
flagging stage	PR	0.353	0.792	2.632	3.977
	RF	0.810	0.320	4.027	3.512
	DT	0.705	0.518	3.913	3.268
	ANN	0.526	0.537	2.060	2.948
flowering period	PR	0.376	0.682	3.236	5.855
	RF	0.906	0.784	3.858	3.064
	DT	0.908	0.220	2.477	1.837
	ANN	0.659	0.474	2.526	4.048
pustulation period	PR	0.674	0.668	4.288	8.243
	RF	0.872	0.677	4.651	3.829
	DT	0.794	0.376	5.803	4.331
	ANN	0.706	0.616	2.846	4.917
maturity	PR	0.401	0.371	8.376	10.764
	RF	0.838	0.249	8.926	7.589
	DT	0.447	0.683	8.662	6.070
	ANN	0.847	0.773	7.230	9.785

.

During the flag-picking period, combined with polynomial regression, random forest, decision tree, and artificial neural network modeling. The R², RMSE, and RE of the polynomial regression model were 0.792, 2.632, and 3.977, respectively. Compared with random forest, decision tree, and artificial neural network, the R² is increased by 0.472, 0.274, and 0.255, respectively. The best model validation results are shown in Figure 5a.

The R², RMSE, and RE of the random forest model at the flowering stage were 0.784, 3.858, and 3.064, respectively. When compared with polynomial regression, decision tree, and artificial neural network model, the R² increased by 0.102, 0.564, and 0.310, respectively. Additionally, the RE decreased by 2.791 and 0.984 compared with the polynomial regression and artificial neural network models, respectively. The model validation results are shown in Figure 5b.

During the filling period, the R², RMSE, and RE of the random forest model were 0.677, 4.651, and 3.829, respectively. The values were 0.061 higher than the R² of artificial neural networks, and RE decreased by 4.414, 0.520, and 1.088 for multinomial regression, decision trees, and artificial neural networks, respectively. The model validation results are shown in Figure 5c.

At the maturity stage, the R², RMSE, and RE of the artificial neural network model for estimating chlorophyll content were 0.773, 7.230, and 9.785, respectively. The R² of polynomial regression, random forest, and decision tree increased by 0.402, 0.524, and 0.09, respectively, while the RMSE decreased by 1.146, 1.696, and 1.432, respectively. The model validation results are depicted in Figure 5d.

4. Discussion

Accurately obtaining crop chlorophyll value is of great significance for crop growth monitoring and yield estimation [33]. With the rapid development of artificial intelligence technology, remote sensing inversion technology, which combines spectroscopy technology with machine learning, is likely to become an important means of chlorophyll remote sensing inversion in the future [34]. Compared with the existing hyperspectral vegetation index, the optimized spectral index is superior, which can solve the problems of low regional applicability of some conventional hyperspectral vegetation indices and “oversaturation” under different growth conditions. It can be seen that the band combination and model algorithms are key to restricting the prediction effect of chlorophyll.

Wu et al. [35] demonstrated the feasibility of inverting chlorophyll content from a hyperspectral vegetation index composed of reflectance in specific bands. Wu et al. [36] used any combination of bands to construct four vegetation indices, and the chlorophyll estimation model established by the random forest method could accurately estimate the SPAD of the rice canopy. Previous studies have shown that the construction of an optimized spectral index can effectively reduce the interference of external environmental factors, accurately identify the sensitive bands of crop growth parameters, and significantly improve the monitoring accuracy and robustness of crop nitrogen, chlorophyll, and leaf area indices [37]. Therefore, in this study, preferred spectral indices and machine learning models were used to estimate the chlorophyll value of wheat leaves, which offset the noise caused by environmental changes to varying degrees, improved the accuracy of model estimation, and the constructed model had strong universality in different regions. The results of this study showed that the optimization of the spectral index could improve the correlation between vegetation index and chlorophyll content. Through the correlation analysis between wheat spectral index and chlorophyll content at different growth stages, it was found that the optimal vegetation index of the correlation between wheat vegetation index and chlorophyll at flag-picking stage, flowering stage, grain-filling stage, and maturity stage had NDVI. Similar to the results of Chen Shengbo’s [38] study on the estimation of chlorophyll in maize, indicating that the NDVI vegetation index could better reflect the growth status of crops. The optimal sensitive bands of NDVI in the four growth periods are quite different from the previously proposed NDVI (680, 800 nm) spectral sensitive bands. The may be due to the fact that the original NDVI (680, 800 nm) is based on the chlorophyll content in maize, while wheat and maize have significant differences in growth period [39], so crop type and growth period will affect the estimation ability of spectral index. Therefore, it is of great significance to optimize the bands for specific crops to figure out the best sensitive bands to improve the estimation of chlorophyll content.

At present, many studies have been conducted on chlorophyll estimation, but the estimation results of different studies are quite different and uncertain. Combined with the scatter diagram, it is found that there are many underestimated points in this study. They may be related to the selection of sampling location and local terrain differences. Overall, the uncertainty of wheat chlorophyll estimation is quite large, which may be mainly due to differences in chlorophyll measurement data. The chlorophyll value is the basic datum used for estimation, and the number of samples and the quality of samples directly affect the accuracy of chlorophyll estimation. Sample points with regional representativeness can improve the accuracy of estimation, and the differences in sampling technology, sampling process, and experimental measurement process will also lead to different data measurement results [40]. In Figure 5c, the coordinates of the maximum error point are (41.70, 48.12), indicating that the difference between the actual value and the predicted value is 6.42. This may be due to the error in the chlorophyll measurement process or the error of measurement equipment or operator. Therefore, it is necessary to ensure the unification and standardization of sampling and experimental processes as much as possible to reduce errors.

The growth stage is also an important factor that affects the ability of the spectral index to estimate the chlorophyll content [41]. Different models have different bands of optimal vegetation index combinations at different growth stages. According to the verification effect of the test set, the PR model performed the best at the flag-picking stage, with the best vegetation index of NDVI (679, 683) and GCI (683, 679), and the best performance of the RF model at the flowering stage and maturity stage, with NDVI (407, 412), RDVI (991, 989), and NDVI (587, 720), RVI (587, 934), and GCI (587, 724), respectively. The ANN model at the maturity stage performed the best, and the optimal number of vegetations were NDVI (989, 996) and GCI (989, 996), among which the model built at the flag-picking stage performed the best.

At the same time, it also generates new ideas and methods for research in related fields. In the future, we can further explore how to translate this research into agronomic practice, such as it can provide a theoretical basis for the design of active sensors based on specific bands so as to reduce the burden of large-scale hyperspectral remote sensing data processing and meet the real-time needs of precision agriculture. In addition, the relationship between other vegetation indices and wheat growth indicators can be investigated to improve the accuracy and practicability of monitoring further. Data in this study are based on samples of different growth stages, but the constructed optimized spectral index has not been validated in different environments and years. In order to improve the robustness and generalization ability of the model, the applicability of different wheat varieties needs to be further tested.

5. Conclusions

In this study, based on the measured hyperspectral data and four types of machine learning models, the correlation matrix method was used to optimize the spectral index of the 400~1000 nm band. We discovered that the optimal spectral index was highly correlated with the chlorophyll content in wheat at different growth stages, but the optimal vegetation index with a high correlation between chlorophyll content at different growth stages was different. The GCI vegetation index had the highest correlation with chlorophyll during the flag-picking period. The correlation coefficients between NDVI and RDVI vegetation index and wheat chlorophyll were greater than 0.6 at the flowering stage, and the chlorophyll content had the strongest correlation with GCI at the grain-filling stage, and NDVI and GCI showed a higher correlation with chlorophyll at maturity stage. Among them, the chlorophyll content in wheat at the grain-filling stage was the most correlated. Compared with the estimation models of chlorophyll content at different growth stages, the RF model based on NDVI, RDVI, and NDVI, the RVI and GCI vegetation indices achieved good estimation results during both the flowering and irrigation stages of wheat. The optimal vegetation indices were NDVI and GCI at both the flag-picking stage and the maturity stage, among which PR was found to be the best in estimating chlorophyll content at the flag-picking stage, and the accuracy R² of the validation set was 0.792. The ANN model was the most ideal for estimating chlorophyll content at the maturity stage.