UAV Mapping of the Chlorophyll Content in a Tidal Flat Wetland Using a Combination of Spectral and Frequency Indices

Abstract

The chlorophyll content of leaves is an important indicator of plant environmental stress, photosynthetic capacity, and is widely used to diagnose the growth and health status of vegetation. Traditional chlorophyll content inversion is based on the vegetation index under pure species, which rarely considers the impact of interspecific competition and species mixture on the inversion accuracy. To solve these limitations, the harmonic analysis (HA) and the Hilbert–Huang transform (HHT) were introduced to obtain the frequency index, which were combined with spectral index as the input parameters to estimate chlorophyll content based on the unmanned aerial vehicle (UAV) image. The research results indicated that: (1) Based on a comparison of the model accuracy for three different types of indices in the same period, the estimation accuracy of the pure spectral index was the lowest, followed by that of the frequency index, whereas the mixed index estimation effect was the best. (2) The estimation accuracy in November was lower than that in other months; the pure spectral index coefficient of determination (R²) was only 0.5208, and the root–mean–square error (RMSE) was 4.2144. The estimation effect in September was the best. The model R² under the mixed index reached 0.8283, and the RMSE was 2.0907. (3) The canopy chlorophyll content (CCC) estimation under the frequency domain index was generally better than that of the pure spectral index, indicating that the frequency information was more sensitive to subtle differences in the spectrum of mixed vegetation. These research results show that the combination of spectral and frequency information can effectively improve the mapping accuracy of the chlorophyll content, and provid a theoretical basis and technology for monitoring the chlorophyll content of mixed vegetation in wetlands.

1. Introduction

Wetland vegetation is the most important part of a wetland ecosystem and a key indicator to measure the health of a wetland ecosystem. Wetland vegetation can comprehensively reflect the ecological characteristics of wetlands and plays an extremely important role in the material exchange and energy conversion of wetlands [1]. The chlorophyll content of wetland vegetation is the most important factor affecting the spectral variability of wetland vegetation. The chlorophyll content of leaves is widely used to diagnose vegetation growth and health and is an indicator of vegetation environmental stress, photosynthetic capacity, and the plant development stage [2]. Wetland ecological monitoring is an important measure to study the dynamics of wetland ecosystems and protect the ecosystem services of wetlands. At the same time, it is also an important link in developing the mixed competition mechanism between alien invasive species and local species, and the research results can provide methodological support for the protection measures of the local species. The estimation of the vegetation chlorophyll content through remote sensing plays an important role in wetland ecological monitoring research. In recent years, with the rapid development of hyperspectral remote sensing, the level of wetland ecological monitoring has further improved, especially in the quantitative inversion of vegetation biochemical components by remote sensing. The retrieval of the canopy chlorophyll content (CCC) based on remote sensing technology is the key link to establish the conversion between spectral signals and photosynthesis, which has been realized through quantitative ecological monitoring, and the research results can provide important basic data for studying the evolution of wetland ecosystems and identifying ecological safeguards [3,4,5]. At present, a large number of chlorophyll content estimation studies are mainly concentrated on crops and forestry, while there are few studies on wetlands plants such as Phragmites australis and Spartina alterniflora. As we all know, S. alterniflora is one of the typical invasive species in the wetland of the Yangtze River Estuary of Chongming Island, and S. alterniflora has a strong impact on the growth of local wetland species [6]. The invasion of S. alterniflora encroaches on the living space of local organisms to varying degrees, forms a single S. alterniflora plant community, which affects the coastal biological habitat, and leads to the decline of biodiversity in the original wetland ecosystem. Therefore, rapid monitoring and evaluation of the growth of S. alterniflora in Chongming Island have important practical significance for understanding population dynamics and distribution of this species.

At present, the analysis of research data shows that the remote sensing monitoring method for invasive wetland vegetation is mainly through the application of remote sensing images of different time phases to the spatial distribution and dynamic change monitoring of S. alterniflora. The use of remote sensing images of different time phases can effectively monitor the spatial distribution and dynamic changes of an invasive species [7], which can provide basic data support for ecological protection and biodiversity research. However, the current use of remote sensing images to monitor S. alterniflora has certain limitations [8,9,10]. Limited by the spectral resolution of remote sensing images, the phenomenon of the same material with different spectra and the same spectrum with foreign material is serious, which can result in statistical errors in the distribution and quantity of S. alterniflora. At the same time, in some study areas, S. alterniflora is mixed with other vegetation, and there may be numerous mixed pixels in an image, which makes it difficult to monitor the spatial distribution pattern and diffusion dynamics of S. alterniflora. Therefore, it is difficult to accurately estimate the chlorophyll content of mixed vegetation in an area. With the rapid development of computer technology and the continuous emergence of various digital, lightweight, small-size, and high-precision sensors, the performance of UAV (unmanned aerial vehicle) remote sensing systems continues to improve, and the scope of application continues to expand [11,12,13,14]. At present, UAV remote sensing systems have been applied to vegetation monitoring [15,16,17], precision agriculture [14,18,19,20], and other applications. UAV hyperspectral technology can obtain high-spatial, high-time resolution, and multiscale remote sensing data. Compared with traditional remote sensing technology, UAV hyperspectral technology has the characteristics of multiple bands, a narrow spectral range, continuous bands, and a large amount of auxiliary data [21,22,23]. At present, most studies mainly use UAV images to extract ground object information and to monitor changes in plant height and aboveground biomass [24,25], whereas research on the CCC quantitative inversion of S. alterniflora and P. australis based on UAV images is lacking. Concerning the inversion of the chlorophyll content, past researchers mainly selected the traditional spectral index to construct the model, resulting in a large number of chlorophyll content estimation vegetation indices; among them, the ratio vegetation index and the normalized vegetation index are the two most commonly used forms for constructing the vegetation index. For example, Wu et al. [26] established an estimation model for the chlorophyll content of winter wheat based on the hyperspectral vegetation index, and the results showed that the two improved comprehensive vegetation indices, MCARI/OSAVI [705,750] and TCARI/OSAVI [705,750], were the most suitable chlorophyll content inversions. Main et al. (2011) researched the robustness of 73 vegetation indices that have been reported to estimate the leaf chlorophyll content, and the results showed that the index established based on the derivative spectrum of the red edge position and the derivative spectrum of the red edge inflection point were suitable for estimating the leaf chlorophyll content [27]. Liu et al. (2017) selected the most suitable spectral index from 79 conventional spectral indices through correlation analysis for chlorophyll content inversion of S. alterniflora and found that MNDVI8 and TCARI resulted in the best chlorophyll content estimation models [28]. Zhuo et al. (2020) used principal component analysis to build an estimation model of the chlorophyll content of S. alterniflora leaves based on harmonic decomposition parameters, and the research results proved that the frequency domain parameters were suitable for the estimation of the chlorophyll content of wetland vegetation [29]. The above results show that chlorophyll remote sensing estimation technology is relatively advanced. At present, most existing research results are based on pure species, in which the spectral morphology is relatively similar and the spectral difference information is not obvious. However, when plants are in a mixed growth mode, interspecific competition will exist between different species, and the internal parameters and canopy structure of the vegetation will change accordingly. The small changes in the internal parameters of the vegetation do not differ greatly in the spectrum, but a sudden change may exist in some bands. In this case, the spectral index will be invalid, and the accuracy of the model will also be greatly affected.

In summary, in view of the shortcomings in previous wetland vegetation chlorophyll content inversion research, the frequency transformation theory of harmonic analysis and Hilbert transform were used in this study to convert the spectral dimension to the frequency dimension. Moreover, the chlorophyll inversion method under a mixed growth state of P. australis and S. alterniflora is discussed in the manuscript. The key points of this study are to establish an appropriate chlorophyll inversion model using different input parameters under a mixed growth status and to provide effective technical means for remote sensing monitoring of wetland vegetation growth at a regional scale.

2. Materials and Methods

2.1. Introduction to the Study Area

In this study, the typical coastal salt marsh wetland in the Yangtze River Estuary—Beiliuyao of Chongming Island, Shanghai was selected as the UAV experimental research area; it is a mixed growth zone with P. australis and S. alterniflora. The geographical location of this area is shown in the figure below. The north side of the study area is the estuary of the north branch of the Yangtze River estuary, and the south side is an artificial levee. The study area is located in the middle; the northern edge of the subtropical zone has a maritime monsoon climate, with a mild and humid climate, four distinct seasons, and abundant rainfall [30]. It is surrounded by the Yangtze River estuary and the East Sea, and the water has a large heat capacity and a good regulation effect on the temperature in the area. The central position of the study area is shown in point A of Figure 1 (121°45′ E, 31°37′ N).

2.2. Data Acquisition and Preprocessing

2.2.1. Field Experimental Design

The location of the UAV field experimental design research area in this experiment is show at position of point A in Figure 1. The size of the research area was approximately 300 × 200 m². It is a typical estuarine wetland vegetation type, in which P. australis and S. alterniflora are in a mixed growth mode. To set ground control points, a total of 80 ground control points were selected as sample squares, and the size of the sampling square was set up as 1 × 1 m² in this area before the acquisition of UAV image data. Each sampling point was surrounded by four bamboo poles, and the bamboo poles were marked with labels for data registration and splicing. The points were set according to the appropriate distance, and the azimuths were evenly distributed in the center of the entire flight path of the UAV (Figure 2).

2.2.2. Ground Data Acquisition

In this study, ground control verification points were set up from south to north, and a total of 80 ground sample points were set up in the experimental area. The SPAD value of leaves at each sample point was measured with a SPAD-502 handheld chlorophyll meter, The SPAD value was a parameter to measure the relative content of chlorophyll or represent the green degree of a plant. The chlorophyll content measurement included the following steps. First, five vegetation pieces were selected for each sample point. Second, five SPAD values were measured at different parts of each leaf, and the average value was taken as the SPAD value of the leaf. Finally, the average value of five leaves was taken as the SPAD value of the canopy chlorophyll content of the sample point. At the same time, the leaf area index (LAI) of the vegetation canopy was measured by an LAI-2200 and the geographic location of the sample point was recorded.

2.2.3. UAV Data Acquisition and Preprocessing

(1) 

Introductionofthe UAV sensor parameters

The drone used in this experiment was a Matrice 600 Pro produced by Dajiang Company of China. The drone includes the aircraft itself, a remote control, and supporting DJI GO App. The hyperspectral camera on the machine is a GaiaSky-mini2 imaging system, a cost-effective airborne imaging system that was developed for small rotor drones. The GaiaSky-mini2 imaging system has a built-in scanning system and a stabilization system. The imaging system successfully overcomes the problem of poor imaging quality caused by the vibration of the UAV system. Figure 3 and

Table 1. UAV sensor parameters.

Spectral Range	Frame Pixels	Spectral Resolution	Sensor	Lens Parameter	Carrying Platform
400–1000 nm	1392 × 1040	3.5 nm ± 0.5 nm	CCD ICX285	23 mm	Load > 3 kg

show the GaiaSky-mini2 imaging system and sensor parameters.

(2) 

UAV hyperspectral data acquisition and preprocessing

(a) Data acquisition

The UAV hyperspectral image data were acquired by a GaiaSky-mini2 hyperspectral imager. The spectral range was 400–1000 nm, and the spectral resolution was 3.5 nm. To obtain effective canopy hyperspectral reflectance data, data collection was carried out when the weather was clear, with no wind or breeze. The collection test time was the same as the ground sample collection date, the flight time was 10:00–14:00, and the flying height was set to 150 m; the ground resolution was 7 cm. The entire data acquisition was synced with the ground measurement data, and a total of four phases of image data, from August–November 2020, were acquired.

(b) Data calibration and splicing

The original UAV hyperspectral images acquired in the field are affected by atmospheric water vapor and some systematic errors of the sensor itself during the acquisition process. The image must be calibrated before use. Human–machine data calibration requires three processing steps: lens calibration, reflectance calibration, and atmospheric correction, in which lens calibration adopts the lens calibration parameter file of UAV which can correct the influence of image internal distortion caused by built-in push scan on splicing, reflectance calibration is completed by observing the whiteboard and lens background file in advance, atmospheric correction is based on the spectral reflectance of gray cloth calibrated by the National Institute of Metrology, and the standard reflectance data are obtained after these three steps. The final experimental research area is obtained through image registration, stitching, and cropping, the data preprocessing process is shown in Figure 4, and the processed image is shown in Figure 5.

2.3. Harmonic Analysis

Harmonic analysis (HA) performs time-frequency space conversion on the processed spectral data as a sequence signal to solve the problems of noise and information redundancy in the preprocessing to obtain the inversion factor. The HA theory was first proposed by Jakubauskas [31]. Tis harmonic theory is to express the time t series f(t) through the superposition of sine (co)sine waves (harmonics); that is, it transforms the time series function from the time domain to the frequency domain. In the frequency space, a number of sine (and co) sine curves of different frequencies are superimposed on each other to represent any time curve in the time domain. Any time series function f(t) about time t can be represented by several positive (the Cosine sum is represented by the superimposed waveforms [31]). The spectral data collected by the spectrometer have high spectral resolution, and the spectral data of each sample can be expressed as a continuous curve. Therefore, when using HA to process spectral data, the spectral curve composed of L bands can be regarded as a function of period L. HA decomposition expresses the spectral curve of each sample as a series of harmonic residuals (A₀/2), amplitudes (A_h, B_h, C_h), phases (${\phi }_h$), and other energy spectrum characteristic components composed on the sine (co) sine wave superposition sum. If a group of spectra composed of N bands is expressed as $F(t)=(f_1,f_2\dots ,f_s,\dots ,f_t)$, the spectral value of each band is recorded as follows: t is the band sequence number (t = 1, 2 (spectrum L), L is the total number of bands (period), and the harmonic decomposition expansion formula obtained by the h–th harmonic analysis is shown in Equations (1) and (2):

$$\begin{array}{l}f(t)=\frac{A_0}{2}+{\displaystyle \sum _{h=1}^{\infty }[A_h\cos (2\pi ht/L)+B_h\sin (2\pi ht/L)]}\\ =\frac{A_0}{2}+{\displaystyle \sum _{h=1}^{\infty }{C}_h\sin (2\pi ht/L+{\phi }_h)}\end{array}$$

(1)

In Formula (1)

$$\{\begin{array}{l}{A}_h=\frac{2}{N}{\displaystyle \sum _{t=1}^{L}f(t)\cos (2h\pi t/L)} h=0,1,2,\cdots \\ B_h=\frac{2}{N}{\displaystyle \sum _{t=1}^{L}f(t)\sin (2h\pi t/L)} h=1,2,\cdots \\ C_h=\sqrt{A_h^2+B_h^2}\\ {\phi }_h=\arctan (A_h/B_h)\end{array}$$

(2)

In Formulas (1) and (2), t is the band number, A₀/2 represents the harmonic remainder, L is the total number of bands, h indicates the number of harmonic analyses, $C_h\sin (2\pi ht/L+{\phi }_h)$ represents the h th harmonic component, A_h represents the cosine harmonic energy spectrum and B_h represents the sinusoidal harmonic energy spectrum, C_h represents the amplitude, and ${\phi }_h$ represents the phase.

2.4. Hilbert Transformation Theory

The Hilbert–Huang Transform (HHT) is a new signal time-frequency joint analysis method proposed by HUANG et al. 1998 [32], which is composed of empirical mode decomposition (EMD) method and Hilbert transform. As a signal processing method, the EMD method is essentially smoothing processing. A series of intrinsic mode functions (IMFs) can be produced in the process of EMD decomposition of the original signal, where each IMF component contains components of different frequency ranges from high to low. EMD eliminates the riding waves of the signal through multiple movement processes and has a smoothing effect on the sequence signal. Each modal function must meet the following two conditions: (a) The number of extreme points and zero-crossing points must be equal or differ by only one in the entire time range; (b) At any time, the lower envelope and local minimum, the average value of the upper envelope sum of the maximum value, must be zero. This method can adaptively decompose nonlinear and non-stationary signals according to signal characteristics and obtain a high-resolution spectral structure. The frequency resolution contained in each frequency band varies with the signal, and has the characteristics of adaptive multi-resolution analysis. The signal is empirically decomposed to obtain the IMF; the Hilbert transform can be performed on each IMF to obtain the meaningful instantaneous frequency, and the frequency change with time can be accurately expressed. The specific calculation method is defined in Formulas (3)–(7).

$$H(c_i(t))=\frac{1}{\pi }{\displaystyle \int {}_{-\infty }^{+\infty }}\frac{c_i(\tau )}{t-\tau }{d}_{\tau }$$

(3)

In Formula (3), H(c_i(t)) is the Hilbert transform function of the i-th IMF component c_i(t). When c_i(t) and H(c_i(t)) are conjugate complex numbers, the analytical signal Z_i(t) can be constructed as follows:

$$Z_i(t)=c_i(t)+jH(c_i(t))=a_i(t)e^{{\phi }_i(t)}$$

(4)

$$a_i(t)=\sqrt{c_i{}^2(t)+H^2{c}_i(t)}$$

(5)

$${\phi }_i(t)=\arctan \frac{H(c_i(t))}{c_i(t)}$$

(6)

$$\begin{array}{l}W{c}_i(t)=\frac{d{\phi }_i(t)}{dt}\\ c_i(t)=\mathrm{Re}(a_{ci}(t)e^{{\displaystyle \int W{c}_i(t)dt}}\\ x(t)=\mathrm{Re}({\displaystyle \sum _{i=1}^n{a}_{ci}(t)e^{{\displaystyle \int W{c}_i(t)dt}})}\end{array}$$

(7)

where a_i(t) is the amplitude or energy of the analytical signal, φ_i(t) is the phase of the analytical signal, Wc_i(t) is the instantaneous frequency of the component, x(t) is the original signal, and n is the number of IMF components.

The envelope Z_i(t) of the signal is obtained by modulating the analytical signal E_i(t), $E_i(t)=\sqrt{c_i{}^2(t)+H^2{c}_i(t)}$, and then the Hilbert envelope spectrum can be obtained by performing spectrum analysis on the envelope signal of formula x(t).

To quantitatively describe the characteristics of the HHT envelope spectrum of the vegetation canopy under different competition ratios, the HHT envelope spectrum characteristic parameters are calculated as envelope spectrum peak index (ESP1) and envelope spectrum kurtosis coefficient (ESP2). ESP1 is defined as the ratio of the maximum value of the envelope spectrum amplitude to its root mean square. ESP2 is the normalized fourth-order central moment, which can be used to reflect the distribution characteristics of the envelope spectrum signal. The expressions of ESP1 and ESP2 are as follows, (8) and (9), respectively.

$$ESP1=\frac{Max(x_f)}{\left|{\varphi }_f\right|}$$

(8)

$$ESP2=\frac{E{(x-u)}^4}{{\sigma }^4}=\frac{1}{N}{\displaystyle \sum _{j=1}^N\frac{{(x-u)}^4}{\sigma }}$$

(9)

where N represents the length of the envelope spectrum signal, μ represents the mean value of the envelope spectrum signal x, σ is the standard deviation of the signal x, and x_i is the i-th corresponding frequency amplitude.

2.5. Spectral Index and Frequency Index Calculation

2.5.1. Spectral Index Calculation

The vegetation index refers to a calculation method that mainly uses the different linear combinations of the ratio between the visible spectrum and the near-infrared spectrum to reflect the vegetation spectrum information. The main reason is that vegetation has strong sensitivity in the visible and near-infrared bands, and the combination of bands between them enhances the hidden vegetation information. The reported hyperspectral index is used to estimate the photosynthetic pigment content of green plant leaves at the canopy level, and by referring to some of existing documents [33,34,35,36,37,38,39,40,41,42], a total of 10 related spectral indexes were selected for the CCC. The abbreviations and specific calculation formulas are shown in

Table 2. Definition and calculation of the spectral indices.

Spectral Index Definition	Abbreviation	Calculation	Ref.
Chlorophyll Red-edge Index (Chlred-edge)	Chlred-edge	(R780/R705) − 1	[33]
Normalized Difference Vegetation Index	NDVI	(R800 − R670)/(R800 + R670)	[34]
Double Difference Index (DDN)	DDN	2 R710 − R660 − R760	[35]
Modified Chlorophyll Absorption in Reflectance Index (MCARI)	MCARI	[(R701 − R671) − 0.2(R701 − R549)]/(R701/R671)	[36]
Normalized Pigment Chlorophyll Index	NPCI	(R680 − R430)/(R680 + R430)	[37]
Red edge Vegetation Stress Index (RVSI)	RVSI	(R712 + R752)/2 − R732	[38]
Structure-insensitive Pigment Index (SIPI)	SIPI	(R800 − R445)/(R800 − R680)	[39]
Simple Ratio (SR)	SR	R800/R670	[40]
Modified Red-edge Normalized Difference Vegetation Index	MNDVIred	(R750 − R705)/(R750 + R705 − R445)	[41]
Meris Terrestrial Chlorophyll Index	MTCI	(R754 − R709)/(R709 − R681)	[42]

. Different spectral index calculation formula is different in order to keep each index at the same quantitative level without affecting the original index properties; each index was normalized between 0 and 1. The normalization formula is as follows:

$$x_{\mathrm{Norm}}=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}$$

(10)

In the above formula, x_i is each index, x_min is the minimum value of all sample sets of the index, x_max is the maximum value of all sample sets of the index, and x_Norm is the normalized index.

2.5.2. Frequency Index Calculation

When the original spectral image data were analyzed by Fourier transform, the spectral information was converted into frequency information, and the pixels in each band were traversed. Finally, the original hyperspectral image was decomposed into four harmonic components, A_h, B_h, C_h, and φ_h, and the image of the harmonic characteristic components obtained by the decomposition was used to establish a mapping model of the chlorophyll content on a regional scale. Figure 6 shows the harmonic decomposition process of the UAV hyperspectral imagery.

2.6. Partial Least Squares Regression Theory

Partial least squares regression (PLSR) analysis is a commonly used multivariate data analysis method in the field of mathematics. At present, it has been widely used in various fields [43]. This method is mainly suitable for linear regression model analysis with a large number of independent variables and high autocorrelation. The core idea of PLSR is to continuously extract the principal component information and then use the dependent variables of the extracted components for regression model analysis. In this manuscript, the PLSR model was established based on the calculated spectral parameters and frequency parameters.

2.7. Research Technical Route

The experiment was conducted using hyperspectral data of the UAV and ground-measured data. First, the image was preprocessed to obtain effective research data, and then the spectral characteristics and frequency characteristics were analyzed to calculate the spectral index and frequency indices. After correlation analysis with the chlorophyll content, 10 spectral indices and 10 frequency indices were selected as model input parameters. In the process of model construction, the indices were divided into the pure spectral indices, pure frequency indices, and mixed index. Based on these three indices, chlorophyll content estimation models were established. Finally, the constructed estimation model was applied to the UAV images to construct the spatial distribution of the chlorophyll content. The detailed flow of the experiment is shown in Figure 7.

3. Results and Analysis

3.1. Correlation Analysis between the Indices and the Chlorophyll Content

The sensitivity of each index to the chlorophyll content was analyzed, and the correlation analysis results of the various indices calculated in different months and the measured chlorophyll contents were obtained (Figure 8). There were differences in the results between different growth periods. Among them, the correlation coefficient of the chlorophyll estimation model was relatively low in August; at this time, the two species were in the vigorous growth period, and the impact of interspecies competition was the greatest. With respect to the process of chlorophyll inversion, the estimation accuracy of the model was the lowest. The accuracy of the model in September was better than that in other months, possibly because the growth of S. alterniflora continued, the chlorophyll contents of the plants gradually increased, S. alterniflora accounted for a large proportion of the canopy spectral information, and the correlation between the spectral index and chlorophyll content was improved compared with months. In October, when the P. australis leaves turned yellow, the canopy spectrum was more affected by the background of the underlying surface, and the estimation accuracy of the model began to decline. In November, the two species were almost completely withered, and the leaf chlorophyll content and the canopy LAI decreased sharply, which caused the canopy spectrum to be greatly affected by the underlying surface, influencing the estimation accuracy of the model.

3.2. Construction of the Chlorophyll Content Estimation Model

The chlorophyll content estimation model was constructed based on the correlation analysis between the chlorophyll content and the indices. The 80 training samples and verification samples of the model were divided into 8 groups, each with 10 data points. Among them, four groups were randomly selected for chlorophyll content modeling, two groups were used as ground modeling sample verification set 1, and the remaining two groups were used as image chlorophyll mapping error verification set 2. The input parameters of the chlorophyll estimation model included the pure spectral, pure frequency, and mixed indices. The chlorophyll content PLSR estimation model was constructed based on these three indices, and the accuracy verification results of the chlorophyll estimation model in different periods are shown in Figure 9, Figure 10 and Figure 11.

Figure 9, Figure 10 and Figure 11 show the model accuracy verification results obtained based on the ground verification data. Analysis of the results indicated that the chlorophyll inversion results of different months had the same characteristics; the mixed index chlorophyll estimation model had the highest accuracy, followed by the frequency domain, and the pure spectral index had the lowest accuracy. This is mainly due to the following two reasons. On the one hand, in the process of UAV image acquisition, it is difficult to ignore the influence of the underlying surface. Therefore, the accuracy of the spectral curve is affected. To obtain a better noise removal effect, the calculated frequency domain parameters successfully avoided the calculation uncertainty of the spectral parameters, and this process avoided the error caused by the selection of the spectral characteristic band. On the other hand, the growth ratio of the ground vegetation in the wild is complex, and mixed growth is prevalent. Mixed growth can lead to the change in canopy morphology and canopy spectral reflectance, which ultimately affects the remote sensing inversion of vegetation chlorophyll content. Since the frequency analysis is based on all bands, it can accurately reflect the spectral differences caused by vegetation morphological changes. These two species were still in the growing period in August and September, and the whole image appeared green. Since the leaf thickness of S. alterniflora is greater than that of P. australis, and the growth period of S. alterniflora was later than that of P. australis, the chlorophyll content of S. alterniflora was generally higher than that of P. australis, while the spectral difference was relatively weak. Therefore, using only the spectral index to build a model cannot meet the chlorophyll inversion requirements under mixed growth conditions in the wild. When the frequency parameter is added, this parameter can amplify the difference information while maintaining the characteristics of the vegetation itself. There will be large differences in the frequency parameters between two different species, so the frequency domain parameters are more suitable for estimating the chlorophyll content of mixed vegetation than that of spectral index. August is the period of vigorous growth of P. australis and S. alterniflora, so it is most strongly affected by interspecies competition, and the spectral difference was more obvious than other months. In the process of chlorophyll inversion, the estimation accuracy of the model was the lowest. In September and October, with the invasion influence of S. alterniflora, the growth of S. alterniflora was still luxuriant but the leaves of P. australis began to turn yellow. The leaf chlorophyll content difference of between P. australis and S. alterniflora was increased gradually, and the spectral difference in the vegetation also increased. At this time, the accuracy of the model established by the spectral index improved. However, affected by the background of the underlying surface, the highest estimation accuracy was observed for the mixed index model. In November, the two types of vegetation withered, the chlorophyll content declined overall, the spectrum was more affected by the underlying surface background, and the inversion accuracy of the model was also the lowest. Therefore, in the case of mixed growth of P. australis and S. alterniflora, the chlorophyll content estimation from UAV images needs to consider interspecies competition and the invasion of vegetation together and to use a combination of the spectral and frequency indices to improve the estimation accuracy.

3.3. Chlorophyll Content Mapping Based on UAV Images

To determine the spatial distribution characteristics of the chlorophyll content of wetland vegetation at the regional scale, the chlorophyll content estimation model established by ground data was applied to the UAV hyperspectral image, and the chlorophyll distribution mapping research was carried out in different periods. The spatial distribution of the wetland vegetation chlorophyll content and the influence of interspecies competition on the accuracy of chlorophyll mapping in different periods were analyzed, and the results are shown in Figure 12.

Based on the analysis of the mapping results of the different models in Figure 12, the results established by the pure spectral indices differed from the other two indices, especially the mixed indices, and this difference is related to the mixed growth of P. australis and S. alterniflora; the area marked with a red circle is the mixed area of reed and S. alterniflora. When P. australis and S. alterniflora are under mixed growth conditions, the canopy spectrum is affected by interspecies competition, which reduces the estimation accuracy, and it is impossible to accurately map the chlorophyll content based on the pure spectrum index. In August, the density of vegetation mixing and the degree of mixing in the spectrum were relatively high. Analysis of the mapping results in August showed that in areas with a low chlorophyll content, the mapping difference between the pure spectral index and frequency index and the mixed index was small. Therefore, based on field sampling and analysis, the proportion of S. alterniflora in the high-value area was larger, the growth density was high, and the mixed area was dominant. According to the ground-measured data, the pure spectral index was effective in the inversion of the chlorophyll content for a single species, but in the mixed scenario, some spectral index monitoring effects were invalid, and the weak difference in the spectrum caused by the change in the mixing density could not be detected; consequently, obvious regional differences were observed in the high SPAD value. The mapping results were relatively similar for September and October, which is related to the decline in the mixed density of vegetation. Due to the continuous competition and rapid growth of S. alterniflora, the mixing degree of vegetation and the number of mixed pixels decreased, which improved the correlation between the spectrum and chlorophyll content. The accuracy of the pure spectral index model in the November mapping results was very low because P. australis withered during this month. Another major influential factor is the decline in vegetation coverage and canopy spectrum. Affected by the underlying surface, the spectral information is complex. The November verification results were the worst, especially the pure spectral index estimation accuracy, which was very low. The frequency index can remove part of the noise during the decomposition and conversion of the image, especially when the spectrum is affected by the underlying surface. The difference between the frequency information of the underlying soil and that of the vegetation will be amplified in the frequency conversion, so the estimation accuracy of the chlorophyll content can be improved.

3.4. Verification of the Accuracy of the Chlorophyll Content Estimation Model

Based on verification dataset 2, the results of UAV image chlorophyll mapping and the ground-measured data were compared and verified. The statistical results of the estimation accuracy are shown in Figure 13 and

Table 3. Accuracy verification statistics of UAV chlorophyll mapping.

Index	August		September		October		November
	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE
Pure Spectral Indices	0.6459	2.9618	0.7246	2.5768	0.6784	2.3675	0.5143	4.4254
Pure Spectral Indices	0.6817	2.6835	0.7817	2.3876	0.7421	1.9879	0.5767	3.7585
Mixed Indices	0.7525	2.1757	0.8073	2.0943	0.7684	1.5685	0.6438	3.1548

. Analysis of the model estimation accuracy in Table 3 indicated consistency with the verification results in verification set 1, and the accuracy of the chlorophyll mapping of images in different periods was compared. The chlorophyll estimation results in September were the best, and the results in November were the worst. Based on a comparison of the chlorophyll mapping results of images from the same period, the estimation model constructed based on the pure spectral index had the lowest accuracy, and the accuracy of the mixed index model was significantly higher than that of the spectral index model. The mixed index model had the highest accuracy in September, with a coefficient of determination (R²) of 0.8073 and a root–mean–square error (RMSE) of 2.0943. The research results showed that the effect of the mixed index was significantly better than that of the pure spectral index.

3.5. Analysis of SPAD Estimation in Time Series

Two ground sample points of mixed vegetation areas on UAV images from August to November were randomly selected to verify the accuracy of the model (Figure 14). The accuracy of the chlorophyll content estimation model was analyzed from the time series based on these two sample points. The estimation results of the chlorophyll content from different growth periods were compared, which showed that the August and November results were worse than those obtained for September and October. Due to the interspecies competition between the two species, the estimation accuracy of the model was low, and the SPAD value of the leaves gradually decreased in October and November. P. australis and S. alterniflora began to wither in November, and the impact of interspecific competition decreased. However, as the vegetation coverage decreased, the yellowing and wilting of leaves led to a decrease in LAI, and the influence of the underlying soil background increased during the spectrum acquisition process, which seriously affected the vegetation index. This resulted in a severe decrease in the accuracy of the model. Analysis of the three index models indicated that in different periods, the estimation accuracy of the mixed index was the best. Using only the spectral index or the frequency index will lead to a larger error between the estimated value of the chlorophyll content and the actual measured value because interspecies competition affects the pure spectral index and cannot be effectively improved. On the one hand, it can be concluded from the above analysis results that the chlorophyll content estimated by the mixed index was more suitable for each phenological phase in the mixed growth mode. On the other hand, the estimation results indicated that the interspecific competition have a greater impact on the estimation of the chlorophyll content, and frequency information can be taken into consideration under mixed growth conditions.

4. Discussion

4.1. Effect of Mixed Growth on the Accuracy of Chlorophyll Content Estimation

The mixed growth of wetland vegetation leads to different plant morphologies, which has a certain impact on vegetation chlorophyll content estimation. Changes in plant morphology affect the radiation transmission process inside the vegetation and influence the spectral response law. Comparing the inversion results of different index models, it can be found that the PLSR estimation model established by the pure spectral indices had a poor ability to accurately estimate the chlorophyll content, and the estimation accuracy of the model was significantly lower than that of the frequency indices and the mixed indices. The best model R² of the pure spectral index ground verification data was 0.7246 and the RMSE was 2.5768; in contrast, the R² of the mixed index model was 0.8073 and the RMSE was 2.0943. The research results showed that when estimating the chlorophyll content of vegetation in the mixed mode, the estimation accuracy of the pure spectral index model is more vulnerable to the interspecific competition caused by mixed growth, and the mixed indices can effectively improve the estimation accuracy of the model [4]. Analysis of the mapping accuracy verification results for the four different periods indicated that the model estimation in September was the best. As the mixed growth density of P. australis and S. alterniflora decreased gradually, S. alterniflora became dominant, the intensity of interspecific competition and the degree of spectral mixing decreased, and the correlation between the chlorophyll content and vegetation index increased, which resulted in an improvement in the chlorophyll inversion accuracy in September. After October, P. australis began to flower and wither, the LAI decreased, and S. alterniflora also began to wither in November. The spectrum in these two periods was more affected by the underlying soil background; thus, the estimation accuracy of the model decreased. The above analysis shows that mixed vegetation growth not only changes the plant morphology but also affects its internal biochemical parameters and spectral radiation transmission. Some conventional spectral indices have difficulty accurately monitoring weak differences. Therefore, to improve the inversion accuracy of the CCC in the mixed growth area of wetland vegetation, it is necessary to select the appropriate growth period and find new wetland parameters.

4.2. The Influence of the Input Parameters on the Model Accuracy

According to the analysis of the research results in Figure 9, Figure 10, Figure 11 and Figure 12 and Table 2, the different input indices had a large impact on the accuracy of UAV canopy chlorophyll content inversion, and index screening was particularly important. An analysis of the research results of the related literature [28] indicated that traditional spectral indices are suitable for estimating the chlorophyll content of pure species, but when P. australis and S. alterniflora are under mixed growth conditions, the canopy spectral response regulation of vegetation changes with mixed growth. It is difficult to obtain a good estimation based solely on pure spectral indices, which indicates that chlorophyll inversion based on pure spectral indices has some limitations [29]. The shortcomings of using a spectral index to the estimate chlorophyll content under mixed growth conditions include the following two aspects. On the one hand, spectral information cannot accurately detect the minor changes in plant internal parameters caused by mixed growth, and the correlation between the spectral indices and chlorophyll content decreases, which could affect the inversion accuracy of the model. On the other hand, the spectral indices are calculated by limited band information, and the selection of characteristic bands is a key step in the calculation of the spectral indices. Due to the difference in the vegetation canopy structure under different mixing ratios, the spectral indices cannot fully reflect changes in plant internal biochemical components. Therefore, to improve the accuracy of chlorophyll inversion, it is necessary to explore new wetland inversion parameters. Coincidentally, the frequency domain information can reflect changes in the vegetation canopy structure, and the frequency indices can be combined with the spectral indices to collaboratively invert the canopy chlorophyll content under mixed vegetation. Through the above analysis, to improve the estimation accuracy of mixed vegetation based on UAV images, appropriate frequency domain information can be added to the chlorophyll content estimation.

5. Conclusions

This study took the typical vegetation, P. australis and S. alterniflora, in the wetlands of the Yangtze River Estuary as the research objects to estimate the chlorophyll content under mixed vegetation at the regional scale. The research data included UAV hyperspectral images and SPAD values of P. australis and S. alterniflora leaves. The traditional spectral index was based on spectral feature analysis. In addition, harmonic decomposition and Hilbert transform theory were introduced, and the frequency index was obtained through frequency domain analysis, the research based on the PLSR model with the pure spectral index, pure frequency index, and mixed index as the input parameters to construct the chlorophyll content estimation model. A chlorophyll content distribution map was constructed for the regional scale based on the UAV hyperspectral image. This result illustrates showed that: (1) A comparison of the accuracy of the estimation models under the three types of input indices for the same period indicated that the estimation accuracy of the pure spectral index was the lowest, followed by the frequency index, whereas the mixed index estimation effect was the best. This can illustrate that frequency combinations effectively reduced the influence of mixed competition on the accuracy of chlorophyll estimation. (2) The estimation accuracy of November was lower than that for the other months; the R² of the pure spectral index was only 0.5143, and its RMSE was 4.4254, whereas the mixed index estimation effect was the best, while the estimated R² for September reached 0.8073, and the RMSE was only 2.0943. (3) The frequency index calculation process removes part of the spectral noise, so the combination of the frequency domain indices and the spectral index can effectively improve the UAV mapping accuracy of the chlorophyll content. The above research results show that the frequency domain index can more accurately reflect the changes in the chlorophyll content of vegetation under a mixed growth environment and can enhance the identification of the spectral information of P. australis and S. alterniflora. Therefore, the frequency domain index can be included to improve the chlorophyll content estimation accuracy under a mixed vegetation growth mode. The results of this experiment provide a methodological basis and technical support for the estimation of UAV hyperspectral chlorophyll content under the condition of mixed vegetation growth in wetlands at a regional scale and are of great significance for the monitoring of wetland vegetation growth.