Spectral-Frequency Conversion Derived from Hyperspectral Data Combined with Deep Learning for Estimating Chlorophyll Content in Rice

Abstract

As a vital pigment for photosynthesis in rice, chlorophyll content is closely correlated with growth status and photosynthetic capacity. The estimation of chlorophyll content allows for the monitoring of rice growth and facilitates precise management in the field, such as the application of fertilizers and irrigation. The advancement of hyperspectral remote sensing technology has made it possible to estimate chlorophyll content non-destructively, quickly, and effectively, offering technical support for managing and monitoring rice growth across wide areas. Although hyperspectral data have a fine spectral resolution, they also cause a large amount of information redundancy and noise. This study focuses on the issues of unstable input variables and the estimation model’s poor applicability to various periods when predicting rice chlorophyll content. By introducing the theory of harmonic analysis and the time-frequency conversion method, a deep neural network (DNN) model framework based on wavelet packet transform-first order differential-harmonic analysis (WPT-FD-HA) was proposed, which avoids the uncertainty in the calculation of spectral parameters. The accuracy of estimating rice chlorophyll content based on WPT-FD and WPT-FD-HA variables was compared at seedling, tillering, jointing, heading, grain filling, milk, and complete periods to evaluate the validity and generalizability of the suggested framework. The results demonstrated that all of the WPT-FD-HA models’ single-period validation accuracy had coefficients of determination (R²) values greater than 0.9 and RMSE values less than 1. The multi-period validation model had a root mean square error (RMSE) of 1.664 and an R² of 0.971. Even with independent data splitting validation, the multi-period model accuracy can still achieve R² = 0.95 and RMSE = 1.4. The WPT-FD-HA-based deep learning framework exhibited strong stability. The outcome of this study deserves to be used to monitor rice growth on a broad scale using hyperspectral data.

1. Introduction

Currently, nearly half of the world’s population consumes rice as a staple food, and with the worldwide population growth, the demand for rice is expected to continue to grow in the future [1]. As the original rice-producing region, China ranks second in the world in terms of planted area and first in terms of production, so it is evident that China’s regional rice production plays a crucial role in the world’s food security [2]. The traditional approach to increasing rice yield has relied mainly on breeding and genetic improvement. China has achieved fruitful research results in this area, and rice varieties grown in different regions have gradually achieved a stable level [3]. The quality of development and final yield of rice are closely related to its cropping practices, growing conditions, and growing environment, such as planting density, chlorophyll content, and weed stress [4]. Chlorophyll is the most important indicator pigment in photosynthesis and the power source for plants to convert light energy into chemical energy. From the perspective of precision agriculture, since chlorophyll content is directly related to plant stress and senescence, the real-time monitoring of chlorophyll content is of great significance in grasping the growth status of crops in real time. For example, the chlorophyll content was rapidly and non-destructively estimated to evaluate the wheat drought stress [5].

Conventional sampling survey methods require a large number of samples from rice crops and are destructive. In addition, the poor operating environment of paddy fields can also lead to inefficient data collection, and there is even the possibility of contamination of the samples for testing. These factors greatly increase the degree of sampling difficulty in manual surveys and reduce the credibility of sample data information [6]. At the same time, traditional crop management is highly subjective, relying mainly on the manager’s empirical judgment as the basis for decision-making. Such subjective judgments are often less accurate and prone to the excessive application of pesticides or fertilizers, which negatively impacts crop growth and the soil environment. Therefore, there is an urgent need for a comprehensive application of modern information technology, which can effectively enhance the economic benefits of crops, and implement locally adapted field management strategies for crops grown on a wide scale.

The non-destructive monitoring of chlorophyll content in rice has been made possible with the development of spectroscopic techniques [7]. However, spectral monitoring of rice chlorophyll content is not only affected by changes in canopy structure but also by the growth period and soil background factors [8]. Hyperspectral remote sensing technology has shown great advantages in plant remote sensing research and application due to its characteristics of having large spectral information, high resolution, and strong continuity, which can directly quantify the weak spectral differences in crops [9]. Many studies have shown that the use of hyperspectral data to obtain plant biochemical parameters quickly and accurately has become an important means of plant growth monitoring, providing an effective pathway for the rapid, efficient, and non-destructive detection of crop chlorophyll content [10].

The main methods used in modeling the estimation of growth parameters are multiple stepwise regression (MSR), principal component regression, BP neural networks, and random forest regression (RFR) [11,12,13,14]. Machine learning and deep learning algorithms have shown strong potential in crop growth monitoring and feature fusion. For instance, the original reflectance, spectral index, and wavelet features were combined by RFR to estimate the chlorophyll content of winter wheat under nitrogen deficiency and powdery mildew stress, and the color features, spectral indices, and chlorophyll fluorescence intensity were fused using multiple linear regression (MLR) and partial least squares regression (PLSR) models to predict sorghum leaf chlorophyll content [15,16]. Existing studies have shown that the spectral indices calculated based on two or more bands correlate well with crop chlorophyll content [17,18]. However, these spectral indices were calculated by utilizing only a few bands of information, which could not comprehensively and effectively extract the hidden weak information and did not reflect the advantages of high spectral resolution of hyperspectral remote sensing data. In terms of hyperspectral data preprocessing techniques, first-order differential, logarithmic, inverse, and wavelet packet transform (WPT) can improve the accuracy of crop growth parameter estimation under specific scenarios (nitrogen fertilizer control and water stress control), such as chlorophyll content prediction at the leaf scale in viroid-inoculated oil palm seedlings, potato leaf area index monitoring under water stress conditions, and carya illinoensis chlorophyll content valuation [19,20,21,22]. However, it is difficult to maintain generalization in estimating chlorophyll content at different stages.

Wavelet analysis is one of the promising techniques in hyperspectral weak information extraction, which is often used to solve the problems of spectral noise and unstable inversion parameters by decomposing a complex signal into a functional combination of simple sub-signal components [21]. In order to make full use of the advantage of the informativeness of hyperspectral data, combined with spectral signal preprocessing techniques and the deep neural network (DNN), the harmonic analysis theory was introduced in this paper to perform time-frequency spatial transformation of the traditionally processed spectral data [23]. The data compression effect was achieved by suppressing or eliminating the background noise of the ground feature through harmonic decomposition. The main research objectives are (1) to analyze the correlation between the parameters extracted by different hyperspectral preprocessing techniques and the chlorophyll content of rice; (2) to compare the accuracy of the parameters obtained by different hyperspectral preprocessing techniques using DNN for estimating the rice chlorophyll content; and (3) to analyze the generalized performance of the framework combining spectral preprocessing techniques and DNN for chlorophyll content estimation in different rice periods.

2. Materials and Methods

2.1. Experimental Design and Chlorophyll Content Measurement

Two experiments were conducted from February to May and July to September at the rice experimental bases in the Lingshui Lizu autonomous county, Hainan Province, and Ezhou City, Hubei Province, respectively. In the 42-plot experiment, a total of 42 rice varieties were laid out in the study area, each with an area of 100 m². The rice in all plots was transplanted on 8 January and harvested on 20 May. On 4 February (seedling stage), 25 February (tillering stage), 9 March and 19 March (jointing stage), 31 March (heading stage), 17 April (grain filling stage), and 28 April (milk stage), four uniformly distributed points within each plot were selected for chlorophyll content measurements (expressed as SPAD values). The top three leaves were selected on each rice plant, and each point was measured 10 times, and the average of the measurements was taken as the SPAD value of the plot. This experiment was used for the initial validation of the models for estimating chlorophyll content in rice. The 48-plot experiment contained a total of 48 plots planted with different rice varieties, each with an area of about 36 m². Rice chlorophyll content was collected on 2 July (seedling stage), 11 July (tillering stage), 21 July (jointing stage), 6 August (heading stage), 21 August (grain filling stage), and 2 September (milk stage), respectively. The collection methods were the same as in the 42-plot experiment. This experiment was used to perform a transportability or robustness validation of the models for estimating chlorophyll content in rice. Except for the different varieties, the other management practices of the two experiments were kept the same. The plots of the different experiments were distributed as shown in Figure 1. In both of the above experiments, the varieties designed were common indica rice. The genotypes of rice in the experiments are not specified because rice genotypes were not a key concern of this study and there is a large number of rice varieties.

2.2. Measurement of Rice Canopy Hyperspectral Data

In this experiment, an ASD FieldSpec 4 spectrometer (Analytical Spectral Devices Inc., Boulder, CO, USA) was used to collect hyperspectral data from the rice canopy. The measurement range of the spectrometer was 350–2500 nm, with a spectral resolution of 3 nm for 350–1000 nm and 8 nm for 1000–2500 nm. The field of view of the spectrometer was 25 degrees. Spectral measurements were performed at six stages: seedling, tillering, jointing, heading, grain filling, and milk stages. At the seedling stage, the rice leaf area index (LAI) was around 1 and the proportion of soil background exposed was high. At the tillering stage, LAI began to exceed 2, and the proportion of soil background occupying the canopy decreased significantly. After the jointing stage, the rice canopy gradually closed (LAI exceeded 5), and the proportion of the canopy occupied by soil background was negligible. During the measurements, the sensor probe was vertically downward, with a vertical height of about 0.5 m from the rice canopy (coverage of approximately 0.04 m²), and all measurements were corrected with a whiteboard before measurement (the effects of soil background is largely negligible). Four target points were randomly selected in each plot and 10 spectra were collected from each target sampling point, and finally, the average value of the spectra was taken as the spectral measurement value of the target point.

2.3. Spectral Preprocessing Techniques

2.3.1. First-Order Differential Processing

The effect of noise on the target can be eliminated by transforming the spectra. A large number of existing studies have shown that differential spectra in characteristic bands have a higher correlation with crop biochemical parameters [24,25]. For vegetation, the spectral measurement is easily affected by the observation angle and illumination, which makes the signal-to-noise ratio of spectral data relatively low. After differential processing, the influence of changing illumination conditions can be reduced, and the effects of background and atmospheric scattering can also be partially or completely eliminated. The biochemical absorption properties of the crop are emphasized by enhancing the spectral differences. The first-order differentiation of the reflectance spectra is calculated as follows:

$$R^{\prime }({\lambda }_i)=\left[R({\lambda }_{i+1})-R({\lambda }_{i-1})\right]/({\lambda }_{i+1}-{\lambda }_{i-1}),$$

(1)

where λ_i−1, λ_i, and λ_i+1 are neighboring wavelengths and $R^{\prime }\left({\lambda }_i\right)$ is the first-order differential spectral value corresponding to wavelength λ_i.

2.3.2. Wavelet Packet Transform

Wavelet packet decomposition is often used in signal denoising. The process of decomposing a spectral signal produces information in both low and high frequency parts. Spectral noise is frequently in the high-frequency part. In order to remove the noise in the high-frequency information, the wavelet packet transform (WPT) was adopted in this paper to decompose and reconstruct the high-frequency information more finely. WPT was introduced based on wavelet transform. Compared with wavelet transform, WPT is superior in decomposing and reconstructing high-frequency information, and the information processing result has no redundancy nor omission, so it is more conducive to spectral denoising and original information retention. Daubechies wavelet is a wavelet function constructed by Ingrid Daubechies, a world-renowned scholar of wavelet analysis, and we generally abbreviate it as dbN, where N is the order of the wavelet. The wavelet mother function used in this study was Db10 [26]. Spectral reconstruction was performed based on the optimal wavelet packet basis and the quantized optimal wavelet packet decomposition coefficients to obtain the spectra after WPT noise reduction processing. After WPT processing, the spectra were first-order differentiated to obtain WPT-FD data.

2.3.3. Harmonic Decomposition

Even with the aforementioned data preparation, there are still several noise and redundancy issues in the spectral information. Harmonic analysis (HA) was presented in this work to solve these issues by performing a time-frequency conversion of the spectral data processing as a sequence signal [27]. After harmonic decomposition, the spectra energy characteristic parameters can be found. Any time series function f(t) with respect to time t can be written as a superposition of multiple sine (cosine) waves. This is known as harmonic theory, and it can be defined as fitting the period function f(t) of a digital or time series signal by the superposition of sine (cosine) waves. The spectral profile with N bands can be seen as a function of period N when processing spectral data with HA. The spectral profile of every sample can be understood as the superposition of a sequence of sine (cosine) waves, comprising the harmonic residual term (A₀/2), amplitude (A_h, B_h, C_h), and phase (φ), which are the distinctive elements of the energy spectra, according to the HA decomposition. If a set of spectra consisting of N bands is denoted as V(k) = (v₁, v₂,…, v_N), the expansion equation obtained from the h time harmonic decomposition is as follows:

$$\begin{array}{c}V\left(k\right)={\displaystyle \frac{A_0}{2}}+{\displaystyle \sum _{h=1}^{\infty }}[A_h\mathit{cos}(2\pi hk/N)+B_h\mathit{sin}(2\pi hk/N\left)\right]\\ ={\displaystyle \frac{A_0}{2}}+{\sum }_{h=1}^{\infty }{C}_h\mathit{sin}(2\pi hk/N+{\phi }_h),\end{array}$$

(2)

The characteristic components of the h-time harmonic decomposition are calculated as follows:

$$A_0/2={\displaystyle \frac{1}{N}}{\sum }_{k=1}^N{v}_k,$$

(3)

$$A_h={\displaystyle \frac{2}{N}}{\sum }_{k=1}^N{v}_k\mathit{cos}(2\pi hk/N),$$

(4)

$$B_h={\displaystyle \frac{2}{N}}{\sum }_{k=1}^N{v}_k\mathit{sin}(2\pi hk/N),$$

(5)

$$C_h=\sqrt{({A_h}^2+{B_h}^2)},$$

(6)

$${\phi }_h={\mathit{tan}}^{-1}(A_h/B_h)$$

(7)

where A₀/2 is the harmonic residual term; C_hsin(2πhk/N + φ_h) is the harmonic component with h time; and A_h, B_h, C_h, and φ_h are the cosine amplitude, sine amplitude, amplitude of the harmonic component, and the phase of the harmonic component, respectively. Among them, A₀/2, C_h, and φ_h reflect the average value of the energy of the spectra in each band, the variation of the energy in different bands, and the position of the band in which the energy appears in amplitude, respectively.

2.4. Deep Learning Model and Validation Methods

The statistical regression modeling-based methods for estimating crop chlorophyll content include both linear and nonlinear models. As agroecosystems are complicated and many of the processes involved are nonlinear, the mechanism of crop growth formation usually behaves nonlinearly [28]. Therefore, nonlinear models are commonly used to monitor crop growth, such as support vector regression, random forest, and partial least squares [29]. However, such traditional machine learning methods are limited in their ability to capture complex nonlinear relationships among data, and deep learning is capable of extracting multi-scale and multi-level features and abstracting combinations of these features into high-level features. Existing studies have shown that deep learning models outperform traditional machine learning and show great potential for growth monitoring using remotely sensed data [30].

Neural networks are based on the extension of perceptual machines, and deep neural network (DNN) can be understood as a neural network with many hidden layers. The neural network layers within DNN are categorized into input, hidden, and output layers by stacking multiple hidden layers to achieve the extraction of higher order feature information, which helps to improve the performance in regression prediction. The DNN model has now become an effective method for solving a variety of classification and regression problems [8]. The network structure designed in this paper contains one input layer, three hidden layers, and one output layer; the layers are directly fully connected to each other, and the neurons between the same layers are not connected (the network structure is shown in Figure 2). The variables obtained based on WPT-FD and WPT-FD-HA screening were converted to one-dimensional vectors corresponding to the measured chlorophyll content of rice, respectively, and used as input layers. The final output was the predicted chlorophyll content. ReLU was selected as the activation function in the DNN model, the root mean square error (RMSE) was used as the optimization function, and the mean square error (MSE) was selected as the loss function for 150 iterations. In this study, Python 3.7 was adopted to build this network structure based on the Keras 2.3.0 framework.

In deep learning models, a common training-test dataset segmentation method is the non-independent data split. For example, within the same experimental region, the data is partitioned into a training set and a test set according to 2:1 or 3:1 [30,31]. If the test set is spatially randomly sampled, spatial autocorrelation can lead to model overfitting. If the test set is spatially blocked, spatial autocorrelation leads to a narrower data distribution in the test set than in the training set. In addition, it does not test the transferability of the model. In this study, based on the traditional validation method, the accuracy and transferability of the model were further tested by combining the independent data split method. The schematic diagram of the two validation methods is shown in Figure 3. It can be seen that a non-independent data split is most likely to have homogeneity of variance. Tests of such homogeneous data do not provide a true indication of the model’s transportability [7]. In contrast, an independent data split generally has heterogeneity of variance. The combination of these two methods can improve the effectiveness of the validation results. In this study, the 42-plot experimental data were used for the construction and initial validation of the rice chlorophyll content model (non-independent data split). The dataset was randomly divided into 2:1 as the model calibration and validation sets, respectively. The 48-plot experimental data were used for model validation across time and location (independent data split). All model accuracy was assessed using the coefficient of determination (R²) and RMSE.

3. Results

3.1. Analysis of Changes in Rice Canopy Spectra with Chlorophyll Content and Growth Stage

Figure 4a shows the spectral reflectance corresponding to different chlorophyll contents in rice. It can be seen that in the visible bands (about 350–750 nm), the spectral reflectance decreased with increasing chlorophyll content, showing an overall negative correlation between reflectance and SPAD values. In the near-infrared bands (about 750–1750 nm), spectral reflectance was positively correlated with SPAD values. Figure 4b shows the characteristics of rice canopy reflectance changes with growth stages. It is observed that in the visible bands, the spectral reflectance has been increasing from the tillering stage to the grain filling stage, and the increase in magnitude continues to decrease. During the grain tillering stage, rice grew rapidly, the chlorophyll content increased dramatically, and the leaves absorbed lights strongly at red and blue bands, resulting in lower spectral reflectance in these bands. The visible bands showed the largest change in spectral reflectance from the grain filling stage to the milk stage. Figure 4c demonstrates the correlation between the original canopy spectra and the chlorophyll content of rice at different growth stages. It might be noticed that the correlations between the raw spectra and the chlorophyll content of rice at different growth stages varied considerably, with most of the correlations at the jointing stage. The correlation between the first-order derivative spectra and chlorophyll content reached a maximum value near 700 nm during the six growth periods observed.

3.2. Selection of Characteristic Bands

The original spectral data, its first-order differential spectra (FD), and the first-order differential spectral data after wavelet packet reconstruction (WPT-FD) were correlated with the chlorophyll contents of rice for the whole period, and the results are shown in Figure 5. It is evident that the overall negative correlation between the raw spectral reflectance and the rice chlorophyll contents was found. The FD and WPT-FD data can significantly improve the correlation with the rice chlorophyll contents and showed a positive and negative alternating trend. In the process of selecting the characteristic bands, in order to reduce the total number of bands for the estimation of rice chlorophyll content and to achieve the effect of data compression, the principles of selecting the characteristic bands were finally determined as follows: (i) the correlation coefficient was more than 0.8, (ii) the correlation coefficient was selected according to the absolute value from the largest to the smallest, and (iii) the number of bands was moderate. Two hundred and seventy characteristic bands were selected from the FD and WPT-FD data of the whole and each growth period of rice, respectively.

3.3. Harmonic Decomposition

The WPT-FD data extracted from the rice canopy spectra were decomposed harmonically to obtain the harmonic energy spectra characteristic components A₀/2, A_h, B_h, C_h, and φ_h, and their correlation coefficients with leaf chlorophyll content were calculated. The total number of bands in the harmonic decomposition was 270, and, considering the period of the sine (cosine) function, the number of harmonic decompositions was determined to be 270. Figure 6 shows the correlation coefficients between the different harmonic energy spectral characteristic components and the chlorophyll content. It was observed that the correlation between the harmonic components at the beginning and the end of the decomposition number and the chlorophyll content was significantly higher. Finally, the components with correlation coefficients greater than 0.8 were selected from the seedling, tillering, jointing, heading, grain filling, and milk stages of rice, respectively.

3.4. Estimation and Validation of the Chlorophyll Content of Rice at Different Stages Based on Diverse Variables

In this paper, in order to reduce the number of model input variables and improve the estimation performance of the model when estimating the chlorophyll content of rice at different growth stages using DNN, the selected components at six growth stages of rice were treated as the input factors of the deep learning model, respectively. The input variables were randomly divided into calibration and validation sets according to 2:1 to construct the rice chlorophyll prediction model based on WPT-FD and WPT-FD-HA, and the model accuracy is shown in

Table 1. The results of the DNN models for predicting rice chlorophyll content based on different variables and growth stages.

Growth Stages	Variables	Calibration		Validation
		R2	RMSE	R2	RMSE
Seedling stage (Cal:Val = 28:14)	WPT-FD	0.895	2.467	0.766	2.865
	WPT-FD-HA	0.921	1.596	0.911	0.905
Tillering stage (Cal:Val = 28:14)	WPT-FD	0.946	1.305	0.832	2.304
	WPT-FD-HA	0.992	0.946	0.985	0.975
Jointing stage (Cal:Val = 56:28)	WPT-FD	0.984	1.065	0.863	2.674
	WPT-FD-HA	0.997	0.938	0.994	0.953
Heading stage (Cal:Val = 28:14)	WPT-FD	0.954	1.379	0.845	2.283
	WPT-FD-HA	0.997	0.934	0.991	0.958
Grain filling stage (Cal:Val = 28:14)	WPT-FD	0.941	1.298	0.723	2.984
	WPT-FD-HA	0.990	0.936	0.983	0.962
Milk stage (Cal:Val = 28:14)	WPT-FD	0.946	0.338	0.733	2.863
	WPT-FD-HA	0.990	0.934	0.986	0.958
Whole growth stage (Cal:Val = 200:94)	WPT-FD	0.924	1.406	0.707	3.334
	WPT-FD-HA	0.982	0.947	0.971	1.664

. It can be found that the validation set accuracies based on WPT-FD variables were significantly lower than those based on the WPT-FD-HA framework when predicting rice chlorophyll content in a single period. As described in more detail, the R² of the WPT-FD models was generally lower than 0.8 and the RMSE was greater than 3, while the R² of the models based on the WPT-FD-HA framework were all more than 0.9, and the RMSE were all less than 1. When estimating the chlorophyll content of rice in multiple periods, the WPT-FD-HA framework showed comparable accuracies to the single-period models and demonstrated generalization across different periods and the whole period. It was significantly higher than the accuracy of the estimation model based on WPT-FD variables (R² = 0.971 vs. 0.707, RMSE = 1.664 vs. 3.334).

Figure 7 and Figure 8 illustrate the comparison of measured and predicted chlorophyll content values in the single period validation set. From the single-period model, it could be noticed that the chlorophyll estimation error was the largest at the seedling stage (points corresponding to different chlorophyll content values were far away from the 1:1 line). The WPT-FD model showed that the predicted line of rice chlorophyll content at different periods differed from the 1:1 line, whereas the predicted line almost overlapped with the 1:1 line in the WPT-FD-HA model. Excellent estimation was observed at both high and low values when using the WPT-FD-HA framework. As can be seen from the scatter plot of the multi-period model (Figure 9), at low values of chlorophyll content, the WPT-FD-based model showed saturation (a large number of points clustered in the same range). In comparison, no significant aggregation was shown in the WPT-FD-HA model. Combined with the characteristics of rice cultivation during the different periods, it indicated that this framework possessed a certain degree of resistance to soil effects. From the comparison of a single period and full period, the accuracy of the WPT-FD model fluctuated greatly in different periods, and the accuracy in the full period was relatively low. However, the accuracy based on the WPT-FD-HA model showed consistently high accuracy in both the single period and the full period, indicating that the framework has the stability of estimating rice chlorophyll content in different periods.

To further prove the stability of the modeling framework proposed in this study for estimating the chlorophyll content of rice in different periods, the method was validated based on the 48-plot experiment, and the results are shown in Figure 10. It was easily seen that the WPT-FD model was less stable in cross-regional experiments, and the fluctuation of model accuracy in a single period was obvious. From the WPT-FD-HA validation model, it can be observed that both the single period and the full period showed high accuracy. The accuracy of the generalized rice chlorophyll model for the full period had an R² of 0.95 and an RMSE of 1.4 (N = 288).

4. Discussion

In this paper, the chlorophyll content of rice was estimated for single and multiple periods by spectral data preprocessing techniques, including FD, WPT, and HA decomposition, combined with a deep learning model. The significance of frequency-domain predictors was highlighted. The whole methodology system was applied to hyperspectral data in terms of data processing, parameter selection, and fusion. Furthermore, the technique was generalized to meet the prediction of rice chlorophyll content for different periods as well as across regions.

Most of the current studies on monitoring crop growth parameters are based on the ground, unmanned aerial vehicles (UAVs), or satellite visible and multispectral data [32,33]. Many optical data cannot cover the entire period of crop growth due to cloud cover and revisit cycles. Moreover, the low spatial resolution of satellite imagery is often not applicable to plot-level applications, limiting its use in crop growth monitoring [34]. A ground-based fixed platform can realize crop growth monitoring at the individual and small-plot scale in the field. Compared with satellite platforms, this platform can provide data with higher spatial, temporal, and spectral resolution, but its monitoring scale is small and inefficient, which is only suitable for small-scale applications, constraining its application in large-area crop growth monitoring [35]. Drones can obtain different kinds of images by flying at low altitudes, making them ideal for agricultural applications, while overcoming the shortcomings of satellites and ground platforms, and providing practical tools for precision agriculture. Different sensors carried by UAVs have been used in various scenarios. For example, hyperspectral cameras were utilized for the estimation of leaf area index [36] and chlorophyll content [37]. However, there is still a lack of relevant research. Thermal infrared and LiDAR sensors have been employed in measuring vegetation canopy temperature, water content [38], biomass [39], and yield [40]. However, these sensors are quite heavy for UAVs. They are also relatively expensive, and the data acquired is difficult to process and analyze. Relatively inexpensive RGB cameras are also frequently used in agriculture. However, the lack of highly sensitive red-edge and near-infrared bands that can be used for vegetation monitoring makes it impossible to acquire stable spectral data. Therefore, it is insufficient for in-depth quantitative studies and applications [41]. Multi-spectral sensors only provide information in a small number of bands, which makes it difficult to obtain satisfactory results in complex field scenarios [42]. Concerning crop growth monitoring techniques, there has been a gradual transition from traditional statistical methods to machine learning methods. For example, partial least squares regression (PLS), principal component regression (PCR), stepwise multiple linear regression, and random forest regression were used to construct models for estimating leaf physiological parameters of different crops [43]. In addition, the rapid development of deep learning techniques has also provided effective tools for the prediction of crop growth parameters [31,44]. Therefore, based on the above background, in situ measurement of hyperspectral data was utilized in this study, which was expected to apply the modeling framework to UAV and satellite hyperspectral data.

Vegetation canopy spectra are closely related to vegetation growth, since plant canopy reflectance carries valuable information about canopy interactions with solar radiation, such as vegetation absorption and scattering [45]. In the visible range, vegetation strongly absorbs sunlight and exhibits low reflectance due to the presence of pigments, and in the near-infrared wavelengths, vegetation exhibits relatively high reflectance influenced by thick plant tissues and canopy structures [46]. Currently, optical vegetation indices based on combinations of reflectance in different wavelength bands are widely used for crop growth parameter extraction in the field and regionally [47,48]. For example, the ratio vegetation index has been shown to be the most sensitive in leaf area index estimation, while the red-edge vegetation index has better results in chlorophyll content estimation [47]. However, a single vegetation index makes it difficult to obtain high accuracy in many complex scenarios. For instance, most vegetation indices fail in rice growth monitoring in scenarios with high soil background influence and dense vegetation growth [42,47]. In addition to vegetation indices, height [49,50], texture [51,52], and temperature [40] information have good auxiliary effects in crop growth parameter extraction, but the extraction of this information will inevitably increase the complexity and cost of the UAVs. Furthermore, machine learning or deep learning was used in most of the studies to fuse different types of parameters, which is rarely validated across regions and over time, making it difficult to apply them directly.

In this study, hyperspectral remote sensing technology was applied to the estimation of chlorophyll content in rice, and the theory of harmonic decomposition was invoked to carry out the research by transforming from the spectral domain to the frequency domain. The research method broadens the current research field of remote sensing to diagnose the chlorophyll content of rice and provides a new idea for the estimation of crop chlorophyll content. The innovation of the whole study is that the frequency-domain parameters obtained from harmonic decomposition were substituted for the spectral factors as the final input factors of the model, thus effectively reducing the human errors brought about by the calculation process of the conventional vegetation parameters. The frequency-domain signal is capable of displaying the essential characteristics and different information of the signal more clearly.

The wavelet packet theory was used to decompose the original spectral signals, and the spectra were reconstructed according to the optimal wavelet packet basis and the quantized optimal wavelet packet decomposition coefficients. From Figure 4, it can be seen that the correlation between the wavelet packet reconstructed spectra and the chlorophyll content of rice is better than that of the original spectra near the infrared and near-infrared bands, because wavelet packet decomposition yields a series of high-frequency and low-frequency signals, and some Gaussian noises tend to be contained in the high-frequency components [53]. In wavelet packet reconstruction, the noise contained in the original spectra was effectively reduced by choosing the appropriate wavelet packet basis to calculate the decomposition coefficients and then recombined into a new set of spectral signals.

Chlorophyll content estimation errors come from spectral noise and model input parameters. Vegetation indices were mostly adopted as model input parameters in conventional chlorophyll content estimation models. Different scholars calculated the vegetation index using various wavelengths and could not precisely select a specific wavelength, thus bringing some errors to the estimation model [54,55]. The purpose of introducing the harmonic decomposition is to eliminate the calculation error of the vegetation index due to the non-uniqueness of wavebands. Harmonic decomposition dealt with the whole spectra without involving a specific band reflectance, and finally, a series of harmonic decomposition parameters were obtained. From these parameters, the variables with a higher correlation with chlorophyll content were filtered out to realize the transformation of the input parameters from the spectral domain to the frequency domain.

By comparing the model accuracy of the single-period and full-period estimations of rice chlorophyll based on WPT-FD and WPT-FD-HA variables, it was finally found that the latter exhibited high stability in single-period and full-period estimations and high validation accuracy in experiments at different times and locations. This modeling framework of spectral preprocessing technique combined with DNN is expected to be applied in large-scale hyperspectral crop growth monitoring. Of course, it should be noted that the shortcomings of this study still exist. The ASD spectrometer for measuring rice canopy hyperspectra requires complex optical configurations, which are expensive and difficult to realize for high-throughput data access. However, the method proposed in this study for estimating rice chlorophyll content using hyperspectral data has practical applications. Utilizing airborne or spaceborne hyperspectral sensors for data access will become a future trend and can provide a simple and high-throughput data source for the application of this method [56,57].

5. Conclusions

This paper focuses on the non-destructive monitoring of chlorophyll content in rice at seedling, tillering, jointing, heading, grain filling, and milk stages. By analyzing the correlation between variables obtained by relying on spectral preprocessing techniques and chlorophyll content at each period, two estimation models, WPT-FD and WPT-FD-HA, were constructed based on the DNN algorithm. Through the comparative analysis of the accuracy of the models for a single growth period as well as for the whole period, it can be found that the spectral noise was effectively reduced by WPT reconstruction, and the correlation between rice canopy spectra and chlorophyll content was improved. Moreover, the harmonic decomposition further removes the noise in the first-order differential spectra and has the effect of data compression. After the harmonic decomposition, the correlation between the obtained components and the chlorophyll content of rice was significantly improved. Meanwhile, the harmonic decomposition converted the spectral domain to the frequency domain, which avoided the uncertainty in the calculation of spectral parameters and played a role in reducing parameter calculation errors. It should be emphasized that the rice chlorophyll content estimation model based on DNN and WPT-FD-HA has good adaptability in important individual and entire periods. This suggests that the WPT-FD-HA framework is a chlorophyll content prediction method applicable to the full rice period and has some generalizability. The unified estimation model of rice chlorophyll content proposed in this study is expected to be useful for rice growth monitoring at a large scale using hyperspectral data.