Spectral Variations of Reclamation Vegetation in Rare Earth Mining Areas Using Continuous–Discrete Wavelets and Their Impact on Chlorophyll Estimation

Abstract

Ion-adsorption rare earth mining areas are primarily situated in the hilly regions of southern China. However, mining activities have led to extensive deforestation of the original vegetation. The reclamation vegetation planted for ecological restoration faces significant challenges in surviving under environmental stresses, including heavy metal pollution, ammonia nitrogen contamination, and soil drought. To rapidly and accurately monitor the growth of reclamation vegetation, this study investigates the spectral variations and their impact on the accuracy of chlorophyll estimation, utilizing hyperspectral data and relative chlorophyll content (SPAD). Specifically, continuous–discrete wavelet transforms were applied, along with the original spectra and first derivative spectra, to enhance spectral anomalies in the reclamation vegetation and identify chlorophyll-sensitive spectral features. Additionally, multiple linear stepwise regression and backpropagation neural network models were employed to estimate chlorophyll content. The results revealed the following: (1) the d5 and d6 scales of the discrete wavelet effectively highlighted spectral anomalies in the reclamation vegetation; (2) Salix japonica (Salix fragilis L.), among typical reclamation species, exhibited poor adaptability to the environmental conditions of the rare earth mining area; (3) the backpropagation neural network model demonstrated superior performance in chlorophyll estimation, with the spectral features Fir, Fir_d4, Fir_d5, and Fir_d6 significantly enhancing the accuracy of the model, achieving an R² of 0.93 for Photinia glabra (Photinia glabra (Thunb.) Maxim.). The application of continuous–discrete wavelet transforms to hyperspectral data significantly improves the precision of chlorophyll estimation, underscoring the potential of this method for the rapid monitoring of reclamation vegetation growth.

1. Introduction

Ion-adsorption rare earth deposits, primarily located in the hilly regions of southern China, represent a unique category of rare earth minerals. What distinguishes these deposits is that rare earth elements, which are essential to modern technological fields such as petrochemicals, military, and defense, are adsorbed in ionic form onto the surface of clay minerals [1,2,3]. Extracting these elements requires specialized mining techniques such as pond leaching, heap leaching, and in situ leaching. While these methods effectively recover rare earth elements, they also cause severe deforestation, soil degradation, desertification, and acidification, posing significant threats to the ecological security of the mining areas. Various reclamation measures, including the planting of multiple typical species of reclamation vegetation (the species of plants used for reclamation, RV), have been implemented to restore the environment. However, the growth of these species has been suboptimal due to the severe environmental stresses they encounter. Therefore, accurately assessing the adaptability of typical RV species to environmental disturbances and monitoring their growth is critical for providing more constructive recommendations for future ecological restoration efforts.

The spectral characteristics and biochemical parameters, such as chlorophyll, of vegetation can effectively represent its growth conditions and health status. When vegetation growth is affected by environmental factors like heavy metal pollution, variations occur in the cellular tissue and biochemical structure of the leaves. This leads to changes in the reflectance intensity and wavelength position of the spectral characteristics, as well as anomalies in biochemical parameters like chlorophyll [4,5,6,7]. Utilization of the high spectral resolution, narrow bands, and continuous nature of hyperspectral data can effectively identify these changes and anomalies, enabling the monitoring of vegetation growth under environmental stress [7,8]. This has been confirmed by many researchers [9,10]. Additionally, to more accurately monitor vegetation growth, in addition to fully utilizing hyperspectral data, the input data for biochemical parameter inversion models can be used because they vary depending on the different subjects and environmental disturbance backgrounds. For example, Huazhe Li et al. [11] studied dominant mangrove species capable of tolerating high salinity. They used measured hyperspectral data to invert relative chlorophyll content (SPAD), providing a technical reference for large-scale mangrove monitoring. Xiaozhe Cheng et al. [12] measured the hyperspectral reflectance of healthy and powdery mildew-infected rubber tree leaves, extracted wavelet features using continuous wavelet transform (CWT), and constructed an inversion model for leaf powdery mildew based on these features. This supported the monitoring and control of rubber tree powdery mildew. While these studies have explored input data for biochemical parameter inversion models suitable for different vegetation under various growth conditions, providing strong support for their monitoring, most of them have focused on relatively common vegetation with more singular environmental disturbance backgrounds.

The ion-adsorption rare earth mining areas are affected by mineral extraction, leading to multiple environmental stresses, including ammonia nitrogen pollution, heavy metal pollution, and drought stress. Consequently, the spectral characteristics of RV in these areas exhibit diversity. The commonly used input data for biochemical parameter inversion models, such as raw hyperspectral data and spectra after ordinary mathematical transformations, do not effectively adapt to the RV in mining areas. Therefore, more complex signal processing methods should be adopted, and their impact on the inversion models of biochemical parameters, such as chlorophyll, should be explored. Wavelet transform, which accommodates both time and frequency domains, includes continuous wavelet transform (CWT) and discrete wavelet transform (DWT), and has gained increasing attention from researchers in recent years. Its advantages in local feature extraction of non-stationary signals and effective capture of instantaneous frequency [7,8,9] have been widely applied in spectral transformation. For instance, Xiaoyu Huang et al. [13] used CWT to process soil hyperspectral data and constructed a soil organic carbon estimation model based on the processed results, achieving good results. Similarly, Juanjuan Zhang et al. [14] applied DWT to decompose the first-order derivative transformation results of raw soil hyperspectral data, building a total nitrogen content estimation model that significantly improved estimation accuracy. However, the application of wavelet transform in spectral processing has primarily focused on subjects not affected by environmental stresses. The effectiveness of spectral transformation and its impact on biochemical parameter inversion models for vegetation under multiple environmental stresses remain to be explored.

This study aims to utilize continuous–discrete wavelet transform and other mathematical methods to enhance the spectral variation details of typical RV in ion-adsorption rare earth mining areas under multiple environmental stresses. Additionally, it explores the impact of different spectral feature inputs on the accuracy of chlorophyll inversion models (CIMs), which effectively monitor the growth of RV and provide constructive recommendations for ecological restoration efforts. To achieve this, a series of hyperspectral data processing steps was applied to typical RV in the mining areas, including first derivative transformations and continuous–discrete wavelet transforms, to enhance the spectral details and identify chlorophyll-sensitive features. Based on these sensitive features, the CIM of the multiple linear stepwise regression (MLSR) and the backpropagation neural network (BPNN) were developed. Through this analysis, the study investigates the adaptability of RV to environmental stresses in the mining area, and determines the appropriate CIM and input parameters for each vegetation type.

2. Materials and Methods

2.1. Study Area Overview

The Lingbei rare earth mining area is located approximately 2 km north of Dingnan County in Jiangxi Province, China. It is situated between latitudes 24°51′24″ N and 25°02′56″ N, and longitudes 114°58′04″ E and 115°10′56″ E, covering an area of about 213 square kilometers. This mining area has a history of over 50 years of extraction history. Initially, “pool leaching” and “heap leaching” techniques were used, causing surface stripping and altering the physical structure of the land. Later, “in situ leaching” was adopted, which further disrupted the soil structure. These mining techniques have led to the degradation of the original vegetation in the area and resulting slow growth of RV. This study selected the vegetation reclamation area at the Jiazi Bei mining site as the research area. Salix japonica (D222014080), Tung tree (Vernicia fordii (Hemsl.) Airy Shaw, D099059001), Masson pine (Pinus massoniana Lamb., C007008018), Photinia glabra (D217028017), and Blue gum (Eucalyptus globulus Labill., D161004009) were chosen as typical RV species for study [15], focusing on the spectral characteristic changes in RV under environmental stress (Figure 1).

2.2. Data Acquisition and Preprocessing

The hyperspectral data required for the experiment were collected on-site. During data collection, the weather was clear, with no wind or clouds, and the vegetation leaves were at a mature stage. The leaf spectra were measured in the field using an ASD (Analytical Spectral Devices) Field Spec 4 spectroradiometer (Malvern PANalytical, Longmont, Colorado, United States of America) from the USA, with sampling intervals of 1.4 nm (350–1000 nm) and 2 nm (1001–2500 nm). Before collecting the spectral curves of the vegetation, the spectrometer was warmed up for about 5 min. Following this, the instrument’s dark current was measured, optimized, and calibrated with a white reference to achieve calibration of the hyperspectral instrument before measurements were taken. During the measurement process, the distance between the sample leaf and the spectrometer lens was maintained at 15 cm. To study the differences between the spectra of reclaimed vegetation and normal vegetation, leaf spectra were collected under normal environmental growth conditions in the upstream, unmined area within 1 km of the reclamation site, following the principle of uniform distribution in the same environmental context. The collected hyperspectral data were preprocessed to remove anomalous spectra, denoise, and average, to mitigate the impact of soil conditions and human interference on the leaf spectra of vegetation in the rare earth mining area [2]. During the preprocessing, 13 groups of reclaimed Salix japonica, 1 group of reclaimed Tung tree, 5 groups of reclaimed Masson pine, 3 groups of Photinia glabra, and 19 groups of Blue gum were excluded. Ultimately, the usable sample sizes for reclaimed Salix japonica, reclaimed Tung tree, reclaimed Masson pine, reclaimed Photinia glabra, and reclaimed Blue gum were 79, 86, 74, 54, and 132, respectively. Additionally, because there is an obvious noise band in the short-wave infrared spectrum, spectral bands above 1350 nm were excluded. The mean of the spectral data from all qualified leaf samples was used as the final spectral reflectance (Figure 2).

In addition to collecting leaf spectral data, leaf samples were also measured for chlorophyll relative content (SPAD). The instrument used for this measurement was the SPAD-502 chlorophyll meter (Konica Minolta, Tokyo, Japan) [16]. Measurements were taken uniformly around the leaf veins 10 to 15 times, and the average was calculated as the final chlorophyll content for each sample. The IQR method was used to eliminate the relevant outliers (Figure 3).

$$m={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{R}_i},$$

(1)

where m represents the final spectral reflectance, $R_i$ represents the spectral reflectance for each qualified leaf sample, and n represents the total number of qualified leaf samples.

2.3. Spectral Feature Transform Method

Leaf spectral reflectance exhibits abnormal changes when vegetation is subjected to environmental stress. Therefore, mathematical transformations of spectral reflectance can refine spectral information and amplify anomalies, further reflecting vegetation growth. In this study, we not only used simple mathematical transformation methods such as first-order derivatives, but also employed the wavelet transform method, which can simultaneously decompose the signal in both time and frequency directions, for spectral transformation. DWT is a form of wavelet transform. It uses discrete scale and translation parameters to extract and analyze the details of the signal at discrete scales. Its advantages, including multiscale analysis and the ability to highlight local features, have been effectively applied in variance analysis of signals [10]. The corresponding “discretization” can be realized in the form of a power series, i.e.,

$$a=a_0^j,b=k{a}_0^j{b}_0(j,k\in Z),$$

(2)

where $a_0$ > 1 is usually assumed when $a_0$ ≠ 1, k is a coefficient, and $b_0$ is a change factor. If the discrete wavelet representation of the signal f(t) is

$${\overline{\Psi }}_{j,k}\left(t\right)=a_0^{-{\displaystyle \frac{j}{2}}}\Psi \left(a_0^{-j}t-k{b}_0\right),$$

(3)

where t denotes time and ψ denotes the wavelet level, then the DWT coefficient C_j,k of the signal f(t) is denoted as

$$C_{j,k}={\displaystyle {\int }_{-\infty }^{+\infty }f\left(t\right){\overline{\Psi }}_{j,k}\left(t\right)dt},$$

(4)

where C_j,k includes high-frequency detail coefficients and low-frequency approximation coefficients under different scale decompositions, the high-frequency detail coefficients denote the fine signals in the original information, and the low-frequency approximation coefficients represent the macroscopic signals in the original information [17].

In this study, DWT was used to decompose the hyperspectral data of each type of RV in the mining area, to obtain spectral local features and variation features at different scales, allowing for a detailed analysis of the local variation in RV spectra and their adaptive ability to environmental stress. The multilayer decomposition wavelet function used in the study is Db5, and the number of decomposition layers is 8.

2.4. Spectral and Chlorophyll-Sensitive Feature Extraction

The measured hyperspectral data contain a large number of cases with a strong correlation between some bands, so the spectral data were correlated with the chlorophyll content to screen sensitive features for use as input data in the CIM and to reduce the large number of cases with a strong correlation possibility of data redundancy [18]. Continuous wavelet transform (CWT) is outstanding in terms of localizing changes in spectral signals from the continuous time–frequency direction [19], which is compatible with the chlorophyll content in the spectral signals as a change in a specific frequency, and it is suitable for mining deeper spectral features that are sensitive to chlorophyll to provide accurate information for chlorophyll inversion. Therefore, in order to further explore the decomposition ability of wavelet transform on the spectra of vegetation under multiple environmental stresses, this study used continuous wavelets to further multiscale-process the aforementioned spectral transform results and screen the spectral features that are more sensitive to chlorophyll content.

The principle of the continuous wavelet transform is to integrate the wavelet basis functions over the scale–translation plane to obtain the wavelet coefficients with continuous scale and translation parameters. Its calculation formula is

$$W_f\left(c,d\right)={\displaystyle {\int }_{-\infty }^{+\infty }f\left(t\right){\Psi }_{c,d}\left(t\right)dt},$$

(5)

$${\Psi }_{c,d}\left(t\right)={\displaystyle \frac{1}{\sqrt{c}}}\Psi \left({\displaystyle \frac{t-d}{c}}\right),$$

(6)

where W_f(c, d) represents the wavelet coefficients of f(t), f(t) denotes the spectral signal, Ψ(t) denotes the wavelet mother function, and c and d denote the scale and translation factors, respectively. The study uses gaus4 as the wavelet mother function for the continuous wavelet transform.

In addition, the correlation screening indices selected in this study include the correlation coefficient r and significance p-value, which characterizes the degree of correlation between variables and measures whether the correlation is significant or not, respectively. Generally speaking, the threshold of significance p-value needs to be controlled below 0.05 to indicate a potential correlation, while the threshold of correlation coefficient r should be analyzed on a case-by-case basis.

2.5. Construction of Chlorophyll Inversion Model

The RV in rare earth mining areas faces multiple environmental stresses, and there is limited research on this topic. Exploring the complex relationships between the physiological parameters of vegetation and measured hyperspectral data remains a challenge. This study used the spectral transformation-sensitive features of different types of RV, identified in the above work, as independent variables, and the chlorophyll content of RV as the dependent variable. We established linear and nonlinear CIMs for different types of RV based on MLSR and BPNN algorithms. The performance of these models is evaluated from the perspectives of different input data and different modeling approaches.

MLSR is capable of evaluating the input variables step by step, adding or eliminating variables based on their significance, and is able to solve the problem of high dimensionality and high covariance caused by highly covariate variables [20]. Based on this feature, the basic idea of MLSR is to gradually introduce feature parameters into the model and test the significance of each introduced feature parameter. The significant feature parameters are retained in the model; however, if the introduction of new features makes the original feature parameters no longer significant, then the original feature parameters are rejected. This process continues until the model no longer introduces new feature parameters and also does not reject the original feature parameters, thereby achieving the optimal model. The expression of its equation is

$$Y=a_0+a_1{X}_1+a_2{X}_2+a_3{X}_3+\dots +a_{F}a{X}_F,$$

(7)

where Y represents the chlorophyll content of the vegetation in the mining area, X_i (i = 1 … p) is the leaf spectral reflectance, a₀ is a constant for the stepwise regression model, and a_i (i = 1 … p) is the coefficients obtained sequentially through regression calculation.

BPNN is a widely used classical artificial neural network characterized by strong nonlinear mapping capabilities and self-organization, as well as adaptive self-learning abilities, making it highly suitable for complex logical operations and nonlinear problems [21]. A typical BP neural network consists of three components: the input layer, hidden layer, and output layer. Its learning process involves both forward and backward transmission. In the forward transmission process, signals enter the hidden layer from the input layer, where they undergo layer-by-layer processing at the hidden layer nodes before being transmitted to the output layer, which evaluates the results at each processing layer. If the output layer does not yield the desired output, the network sends an error signal back along the original connection path, and each node in the hidden layer adjusts the weights and thresholds of the neural nodes through feedback of the error information, aiming to minimize the error and achieve a clear mapping between the input and output layers [22]. In this study, the measured hyperspectral reflectance data of leaves and measured chlorophyll data were used as input factors of the input layer (x₁, x₂, ..., x_n), whose structure is shown in Figure 4.

In this study, we employed a five-fold cross-validation method to partition the dataset. Each training set contained 80% of the entire dataset samples, while each test set contained 20%. The numbers of training samples for reclaimed Salix japonica, Tung tree, Masson pine, Photinia glabra, and blue gum were 63, 68, 59, 43, and 105, respectively; the corresponding test samples were 16, 18, 15, 11, and 27.

2.6. Precision Analysis

To effectively compare the inversion effect of the two models, this study consistently utilizes the coefficient of determination (R²) and the root mean square error (RMSE) to assess model accuracy: R² indicates the relationship between the calculated prediction value and the mean value; the closer it is to 1, the better the match between the predicted and actual values; RMSE describes the degree of fluctuation in the data; a lower value indicates a more accurate model. The formulas for these evaluation metrics are as follows:

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{({\stackrel{\wedge }{y}}_i-y_i)}^2}}{{\displaystyle \sum _{i=1}^n{({\stackrel{\wedge }{y}}_i-\overline{y})}^2}}},$$

(8)

$$RMSE=\sqrt{{\displaystyle \frac{1}{m}}{\displaystyle \sum _{i=1}^m{\left(y_i-\stackrel{\wedge }{y_i}\right)}^2}},$$

(9)

where m represents the number of input-sensitive bands characterizing the RV of the mine site, y_i denotes the predicted value of chlorophyll content, $\widehat{y_i}$ indicates the measured value of chlorophyll content, and $\overline{y}$ represents the mean of the actual observed chlorophyll values.

3. Results

3.1. Statistical Characteristics of Chlorophyll in RV

The SPAD measurements of five typical RV leaves in the mining area were statistically analyzed using one-way ANOVA, with the results presented in

Table 1. SPAD statistics of different RV.

Reclamation Vegetation	Number of Measurements	Minimum Value	Maximum Value	Average Value	Standard Deviation	Coefficient of Variation	Significant Difference
Salix japonica	79	29.4	52.2	41.3	5.1	12.3%	0 (1.22 × 10−6)
Tung tree	86	16.2	55.6	32.3	7.2	22.3%
Masson pine	74	19.7	79.8	59.7	13.8	23.1%
Photinia glabra	54	24.8	73.8	44.6	10.6	23.7%
Blue gum	132	25.4	60.8	42.6	6.3	14.8%

. It can be observed that the standard deviations and coefficients of variation of the SPAD measurements of the five types of RV leaves exhibited a similar trend, with coefficients of variation exceeding 10% and reaching a maximum of 23.7%. This indicates that the five types of RV were affected by environmental disturbances in the mining area to varying degrees. Additionally, the SPAD measurements of the leaves of the five RV species showed significant differences at the p < 0.05 level. These findings suggest that different types of typical RV in mining areas are affected by environmental disturbances and respond distinctly to environmental disturbances. Independent analysis of the hyperspectral measurement data of different vegetation can provide more accurate information for chlorophyll estimation and enhance our understanding of the growth status of different species as well as the impacts of environmental stresses.

3.2. Leaf Spectral Response of RV

To better understand the influence of environmental disturbances on RV and emphasize spectral variations, we compared the original hyperspectral data of both RV and normal vegetation, as well as their spectral transformations using first-order derivatives and DWT. The comparison maps between the original spectra and the first-order derivative spectra, as well as between the original spectra and the first-order derivative spectra with DWT for the vegetation in different environments, are shown in Figure 5 and Figure 6.

As shown in Figure 5, the spectral fluctuation trends of RV generally align with those of normal vegetation. However, in the original spectra, the “reflection peak” in the green light band and the “absorption valley” in the red light band for RV shift towards longer wavelengths, and the spectral reflectance is consistently higher than that of normal vegetation, highlighting a distinctive change pattern. Additionally, the changes in RV are more pronounced in regions where the reflectance of the first-order derivative spectra sharply rises and falls. In the visible light band, these changes primarily occur at the “three edge parameters,” with specific values shown in

Table 2. Statistics of “trilateral parameters” for different vegetation.

Type of Vegetation	${\mathit{D}}_{\mathit{b}}$	${\mathit{\lambda }}_{\mathit{b}}$	${\mathit{D}}_{\mathit{y}}$	${\mathit{\lambda }}_{\mathit{y}}$	${\mathit{D}}_{\mathit{r}}$	${\mathit{\lambda }}_{\mathit{r}}$
Normal Salix japonica	0.0011	525	−0.00004	628	0.0080	724
Reclamation Salix japonica	0.0033	520	−0.00012	629	0.011	703
Normal Tung tree	0.0013	523	−0.00002	627	0.013	718
Reclamation Tung tree	0.0036	520	−0.00008	629	0.018	701
Normal Masson pine	0.0031	518	−0.00007	628	0.013	718
Reclamation Masson pine	0.0034	520	−0.00006	629	0.012	717
Normal Photinia glabra	0.0025	523	−0.00008	629	0.015	719
Reclamation Photinia glabra	0.0032	520	−0.00010	597	0.015	702
Normal Blue gum	0.0018	523	0.00008	629	0.013	706
Reclamation Blue gum	0.0019	524	−0.000004	629	0.014	708

${\mathit{D}}_{\mathit{b}}$ and ${\mathit{\lambda }}_{\mathit{b}}$ are the amplitude and position of the blue edge parameter; ${\mathit{D}}_{\mathit{y}}$ and ${\mathit{\lambda }}_{\mathit{y}}$ are the amplitude and position of the yellow edge parameter; ${\mathit{D}}_{\mathit{r}}$ and ${\mathit{\lambda }}_{\mathit{r}}$ are the amplitude and position of the red edge parameter.

. From Table 2, it can be seen that compared to normal vegetation, the blue edge, yellow edge, and red edge positions for RV have shifted to varying degrees, with the red edge showing the largest shift, frequently moving towards shorter-wavelength blue light, and the corresponding spectral values reaching the highest. These spectral anomalies are typical characteristics of vegetation under environmental stress.

Figure 6 shows the spectral curves of RV and normal vegetation at different decomposition levels after performing DWT on different spectral bases. It can be observed that, for the same vegetation type in different environmental contexts, the signal waveforms at each decomposition level are generally similar but exhibit subtle differences. Furthermore, d5 and d6, based on the original spectra and first-order derivative spectra, respectively, better represent the spectral differences between RV and normal vegetation than other wavelet coefficients. This difference is primarily reflected in the crossover points of the peak–valley features of the vegetation spectral curves in the red edge region, ranging from 620 nm to 750 nm, which effectively compensate for the indistinct peak–valley features or overlaps in spectral curves at other scales. In addition to capturing spectral anomalies, the comparison of spectral curves of different RV types and their normal vegetation counterparts at these two scales reveals that the spectral signals of reclaimed Blue gum and reclaimed Masson pine are relatively consistent with those of their corresponding normal vegetation. In contrast, reclaimed Salix japonica, reclaimed Tung tree, and reclaimed Photinia glabra exhibit significant differences, which gradually increase.

3.3. Spectral and Chlorophyll-Sensitive Feature Extraction of RV

The study first conducted a correlation analysis of the original spectra, first-order derivative spectra, and chlorophyll content, selecting sensitive bands based on the correlation coefficient (r) and the significance p-value. Next, continuous wavelet transform (CWT) was applied to both the original and first-order derivative spectra of each RV, followed by a correlation analysis with chlorophyll content. This analysis provided quantitative insights into the relationship between vegetation spectra and chlorophyll content across 10 scales, as shown in Figure 5.

From Figure 7, it is evident that, whether CWT was applied to the original spectra or the first-order derivative spectra, the overall trend for each vegetation type remained consistent across different scales. Certain scales also exhibited specific band ranges with notably higher correlations. To identify the scales with the most pronounced correlations and facilitate scale selection, the study conducted sensitivity feature screening based on the significance p-value and correlation coefficient. This was done across the 10 scales of both the original and first-order derivative spectra CWT for each vegetation type, and the results were statistically analyzed and are presented in

Table 3. Statistical analysis of 10 scale sensitive features of CWT spectra based on different spectra of different RV.

Type of Vegetation		Reclamation Salix japonica		Reclamation Tung tree		Reclamation Masson pine		Reclamation Photinia glabra		Reclamation Blue Gum
		Max	Num	Max	Num	Max	Num	Max	Num	Max	Num
Ori_CWT	d1	0.6	169	0.9	180	0.7	320	0.7	194	0.8	231
	d2	0.7	328	0.9	273	0.7	253	0.8	268	0.8	280
	d3	0.7	389	0.9	339	0.6	229	0.8	289	0.8	354
	d4	0.8	440	0.9	377	0.7	140	0.8	303	0.8	391
	d5	0.7	464	0.9	235	0.6	184	0.8	330	0.8	435
	d6	0.8	652	0.9	480	0.6	188	0.8	403	0.7	454
	d7	0.6	738	0.7	599	0.3	156	0.8	372	0.5	724
	d8	0.7	892	0.9	731	0.3	99	0.8	603	0.6	644
	d9	0.6	911	0.8	559	0.3	158	0.8	603	0.4	776
	d10	0.6	701	0.7	561	0.3	229	0.7	353	0.3	292
Fir_CWT	d1	0.6	145	0.8	141	0.6	328	0.7	175	0.7	190
	d2	0.6	287	0.8	234	0.7	280	0.7	228	0.8	293
	d3	0.7	335	0.8	309	0.6	237	0.8	275	0.8	306
	d4	0.8	392	0.9	373	0.7	139	0.8	323	0.8	372
	d5	0.7	472	0.9	387	0.6	172	0.8	377	0.8	499
	d6	0.7	614	0.9	504	0.7	170	0.8	427	0.8	497
	d7	0.5	569	0.8	506	0.3	291	0.7	332	0.5	694
	d8	0.7	887	0.9	848	0.4	118	0.9	640	0.7	684
	d9	0.6	803	0.8	758	0.3	240	0.7	620	0.5	724
	d10	0.6	841	0.6	634	0.3	138	0.7	558	0.3	534

.

Table 3 reveals that, for most RV, scales d4, d5, d6, and d8 contain many sensitive features and show high correlations with chlorophyll content. Therefore, these four scales were selected to capture the chlorophyll-sensitive bands.

Table 4. Each input data of the chlorophyll inversion model.

Input Data Scheme	Input Data Content	Scheme Abbreviation
Scheme 1	Original spectral sensitive band (Ori)	Ori
Scheme 2	First-order derivative spectral sensitive band (Fir)	Fir
Scheme 3	d4 scale sensitive band after CWT of the original spectrum (Ori_CWT_d4)	Ori_CWT
Scheme 4	d5 scale sensitive band after CWT of the original spectrum (Ori_CWT_d5)
Scheme 5	d6 scale sensitive band after CWT of the original spectrum (Ori_CWT_d6)
Scheme 6	d8 scale sensitive band after CWT of the original spectrum (Ori_CWT_d8)
Scheme 7	d4 scale sensitive band after CWT of the first-order derivative spectra (Fir_CWT_d4)	Fir_CWT
Scheme 8	d5 scale sensitive band after CWT of the first-order derivative spectra (Fir_CWT_d5)
Scheme 9	d6 scale sensitive band after CWT of the first-order derivative spectra (Fir_CWT_d6)
Scheme 10	d8 scale sensitive band after CWT of the first-order derivative spectra (Fir_CWT_d8)

lists the input data for the CIM.

3.4. Comparison and Evaluation of RV Chlorophyll Estimation Models

By utilizing the sensitive features extracted from various spectral processing methods as independent variables and the measured relative chlorophyll content as the dependent variable, MLSR and BPNN were employed to construct the CIM. This approach allows for examining the impact of different input data on model accuracy and identifying which model type is more suitable for chlorophyll inversion in RV. This provides valuable insights for accurately monitoring various types of RV in mining areas.

3.4.1. Effect of Ori on the Chlorophyll Estimation Model of RV

The results of the CIM in MLSR and the BPNN for RV, which were constructed using sensitive bands derived from the original spectra, are shown in

Table 5. Accuracy of different input data schemes in CIM of MLSR.

Input Data Scheme	Type of Vegetation	R2	RMSE	Input Data Scheme	Type of Vegetation	R2	RMSE
1	Reclamation Salix japonica	0.71	2.76	2	Reclamation Salix japonica	0.72	2.69
	Reclamation Tung tree	0.86	2.75		Reclamation Tung tree	0.92	2.01
	Reclamation Masson pine	0.85	13.29		Reclamation Masson pine	0.65	8.20
	Reclamation Photinia glabra	0.70	5.82		Reclamation Photinia glabra	0.87	3.86
	Reclamation Blue gum	0.74	3.25		Reclamation Blue gum	0.83	2.59
3	Reclamation Salix japonica	0.73	2.66	4	Reclamation Salix japonica	0.69	2.88
	Reclamation Tung tree	0.88	2.47		Reclamation Tung tree	0.87	2.57
	Reclamation Masson pine	0.57	9.15		Reclamation Masson pine	0.30	11.64
	Reclamation Photinia glabra	0.76	5.18		Reclamation Photinia glabra	0.74	5.40
	Reclamation Blue gum	0.79	2.91		Reclamation Blue gum	0.80	2.85
5	Reclamation Salix japonica	0.68	2.92	6	Reclamation Salix japonica	0.54	3.46
	Reclamation Tung tree	0.84	2.87		Reclamation Tung tree	0.79	3.31
	Reclamation Masson pine	0.12	13.00		Reclamation Masson pine	0.08	13.34
	Reclamation Photinia glabra	0.75	5.30		Reclamation Photinia glabra	0.75	5.30
	Reclamation Blue gum	0.57	4.16		Reclamation Blue gum	0.62	3.87
7	Reclamation Salix japonica	0.75	2.55	8	Reclamation Salix japonica	0.63	3.10
	Reclamation Tung tree	0.88	2.55		Reclamation Tung tree	0.86	2.76
	Reclamation Masson pine	0.59	8.95		Reclamation Masson pine	0.39	10.83
	Reclamation Photinia glabra	0.74	5.42		Reclamation Photinia glabra	0.78	4.96
	Reclamation Blue gum	0.74	3.23		Reclamation Blue gum	0.75	3.15
9	Reclamation Salix japonica	0.69	2.86	10	Reclamation Salix japonica	0.51	3.58
	Reclamation Tung tree	0.85	2.83		Reclamation Tung tree	0.83	2.95
	Reclamation Masson pine	0.13	12.98		Reclamation Masson pine	0.23	12.10
	Reclamation Photinia glabra	0.74	5.43		Reclamation Photinia glabra	0.78	4.98
	Reclamation Blue gum	0.61	3.97		Reclamation Blue gum	0.71	3.40

and Figure 8. When using the original spectra as input parameters, the multiple linear regression models for chlorophyll inversion exhibited consistent and satisfactory accuracy across all RV types, with the highest R² of 0.85 observed in the reclamation Masson pines. On the other hand, while the BP neural network model achieved an R² of 0.83 for the reclamation Tung tree, the overall R² values exhibited less stability, particularly showing lower accuracy for the reclamation Masson pines.

3.4.2. Effect of Fir on the Chlorophyll Estimation Model of RV

Taking the sensitive bands of the first derivative spectra as independent variables, the MLSR and BPNN models for chlorophyll inversion of each type of RV were constructed, and the results are shown in Table 5 and Figure 8. In general, in the MLSR chlorophyll inversion model constructed in the sensitive band of the first derivative spectrum, the model accuracy across the RV types was relatively consistent, with the highest R² reaching 0.92, while the performance for reclaimed Masson pine was notably lower, at only 0.65. Similarly, for the BPNN chlorophyll inversion model, the inversion accuracy of the reclaimed oil tree was 0.9, while that of reclaimed Masson pine was only 0.3. By comparing the inversion effects of the two inversion models on different vegetation, the inversion accuracy of the BPNN on other vegetation is shown to be slightly lower than that of MLSR, except that the inversion accuracy R² of the BPNN on reclaimed bamboo and Salix japonica is 0.16 higher than that of MLSR.

3.4.3. Effect of Ori_CWT on the Chlorophyll Estimation Model of RV

The sensitive bands of different scales obtained from Ori_CWT were used as inputs for the MLSR and BPNN chlorophyll inversion models for each type of RV, and the results are presented in Table 5 and Figure 8 and Figure 9. It can be observed that the accuracy (R²) of the d4 scale for chlorophyll inversion across various vegetation types is particularly notable, with the MLSR and BPNN accuracies for the reclaimed Tung tree reaching their highest values of 0.88 and 0.89, respectively. In all scales, vegetation types that exhibit high inversion accuracy with the BPNN generally perform similarly with MLSR, but not vice versa, and most vegetation can achieve the optimal chlorophyll inversion accuracy on the BPNN.

3.4.4. Effect of Fir_CWT on the Chlorophyll Estimation Model of RV

The sensitive bands of different scales obtained from Fir_CWT were utilized as inputs for the MLSR and BPNN inversion models of chlorophyll for each type of RV, and the results are presented in Table 5 and Figure 9. Similar to the findings from the model using sensitive bands obtained from Ori_CWT as input data, the chlorophyll inversion accuracy R² of Fir_CWT_d4 for each vegetation type in both MLSR and BPNN models is superior to that of other vegetation, and the optimal accuracy for most vegetation types is also achieved at this scale. In addition, the inversion accuracy of the BPNN is higher than that of MLSR on most vegetation scales.

3.4.5. Effects of Different Input Data on Chlorophyll Retrieval Model of RV

Figure 10 illustrates the impact of different spectral input data on the performance of multiple linear stepwise regression (MLSR) and BP neural network (BPNN) models for chlorophyll content in RV. For the MLSR models, most vegetation types achieved good accuracy when Ori, Fir, Ori_CWT_d4, and Fir_CWT_d4 were used as input parameters. Notably, the reclamation Tung tree achieved the highest R² of 0.92 on Fir, marking the best accuracy among all vegetation types using the MLSR model. Reclamation Masson pine, Photinia glabra, and Blue gum also reached R² values of 0.8 with Ori and Fir. For the BPNN models, Fir, Ori_CWT at scale d4, and Fir_CWT at scales d4, d5, and d6 provided effective chlorophyll inversion results for most vegetation types. Reclamation Photinia glabra achieved the highest R² of 0.93 on Fir_CWT_d4, ranking first in accuracy across all vegetation types and models. Reclamation Salix japonica and Tung tree also performed well on Fir, with R² values of 0.88 and 0.9, respectively. Overall, the accuracy of CIM varied across different input parameters and methods. While the maximum achievable accuracy for chlorophyll inversion was comparable between the MLSR and BPNN models, the BPNN models generally outperformed the MLSR models in terms of R² across most input data. This was especially evident for reclamation Masson pine, where the accuracy on Ori_CWT_d6 improved by 0.75. These results suggest that nonlinear regression is better suited to capturing the complex relationship between spectral data and chlorophyll content.

Comparing the RMSE of the MLSR and BPNN models reveals that the MLSR models consistently show better stability across all vegetation types compared to the BPNN models. This may be related to the relatively small sample size used in the study conducted. While nonlinear regression can achieve higher accuracy with smaller sample sizes, it tends to lack stability.

4. Discussion

The ecological environment of rare earth mining areas is in a critical situation due to long-term uncontrolled mining. Planting RV in these areas helps alleviate ecological pressure and promotes restoration. However, mineral mining has introduced multiple environmental stresses, including heavy metal pollution, ammonia nitrogen pollution, and soil drought, which hinder the healthy growth of RV. Accurate monitoring of five typical RV types in the mining area is essential to provide valuable input for future ecological restoration efforts. To achieve this, this study utilized measured hyperspectral data of RV in the mine area. The spectral data were further transformed and decomposed using first-order derivatives and continuous wavelets to enhance the spectral information. By analyzing the spectral anomalies of the RV, the adaptive capacity of each species to the various environmental stresses was assessed. Additionally, a chlorophyll inversion model for RV was constructed using MLSR and BPNN models, with the results of different spectral transformations as independent variables. The study also examined the impact of different input data on the accuracy of the model, comparing the performance of linear and nonlinear models.

4.1. Analysis of Typical Characteristics of Spectral Variation in RV

By comparing the results of the original spectra, first-order derivative spectra, and the spectra obtained from discrete wavelet transform based on both RV and normal vegetation, typical spectral variation features exhibited by RV under multiple environmental stressors can be observed. These features include a significant increase in near-infrared spectral reflectance, noticeable shifts in the positions of the green peak band, red valley band, and red edge position in the first-order derivative spectra, as well as the appearance of crossover peaks and valleys between RV and normal vegetation in the discrete wavelet transform [22]. These phenomena primarily occur due to the vegetation’s response to environmental stress, including changes in leaf structure and water content. This not only leads to decreased chlorophyll and light absorption rates but also increases the likelihood of the “green loss” phenomenon, resulting in changes in reflectance and shifts in spectral positions [10,23].

Quantitative analysis of these variability features can provide insights into how vegetation adapts to environmental conditions. In many similar studies, researchers typically use six characteristic spectral parameters from the original spectra—such as peak reflectance and wavelength of the green band, and trough reflectance and wavelength of the blue and red bands [24]—or the “three-edge parameters” from first-order derivative spectra [25] to represent the specific changes in spectral features under environmental stress. However, for vegetation experiencing single or multiple environmental stresses, the differences in reflectance and wavelength are often subtle, and there is no standard threshold for determining these differences. This makes it challenging to quickly and accurately analyze the specific growth conditions of the vegetation. In this study, discrete wavelet transformation at scales d5 and d6 effectively decomposes and clearly represents the spectral information of vegetation. By comparing the spectral curves under normal and stressed environmental conditions, the vegetation’s growth status can be directly assessed and provide a basis for analyzing its environmental adaptability.

4.2. Environmental Adaptability of Different RV

As previously mentioned, the study investigates the environmental adaptability of five typical RV species in rare earth mining areas by analyzing the spectral curve differences between reclaimed and normal vegetation based on DWT. The species’ adaptability, ranked from highest to lowest, is as follows: reclaimed Blue gum, reclaimed Masson pine, reclaimed Photinia glabra, reclaimed Tung tree, and reclaimed Salix japonica. It is worth noting that Salix japonica, as a kind of vegetation that supports growth under severe conditions, can retain soil moisture and requires minimal maintenance, and has been widely used in the ecological remediation of environments contaminated by heavy metals and organic pollutants [26]. Raushan Kumar et al. [27] explored the role of Salix japonica in heavy metal exclusion and enrichment in fly ash dumpsites in India in the context of trace element pollution and achieved good results; Hao Ma et al. [28] investigated the excellent cultivation of Salix japonica in the desert and saline soil under the background of drought and sandy winds; Wei Chen et al. [29] found that Salix japonica had excellent uranium enrichment properties in uranium-contaminated soils compared to other vegetation. However, in this study, the growth performance of reclaimed Salix japonica was poor, and its environmental adaptability was weak.

This issue has also been observed in other studies. For instance, Hu Hongling [30] found that Salix japonica had a low survival rate under drought conditions, while Wu Xiang [31] discovered that Salix japonica exhibited poor osmotic regulation under salt stress. However, the reasons for Salix japonica’s weaker adaptability differ under various environmental stresses. Considering the complex vegetation growth environment in rare earth mining areas, we reviewed a substantial amount of literature and identified two primary factors contributing to Salix japonica’s poor environmental adaptability in the context of this study. On the one hand, Salix japonica’s capacity for environmental remediation mainly manifests in its absorption of heavy metal elements [32]. However, the environmental stressors in rare earth mining areas extend beyond heavy metal pollution. Influenced by the discharge of wastewater from heap leaching, heap stacking, and in situ leaching mining methods, the ammonia nitrogen pollution in the soil of mining areas is equally severe. Moreover, mining activities result in a large area of exposed ground surface in mining areas, with phenomena such as drought coexisting, which may limit the growth of Salix japonica due to additional environmental pressures. On the other hand, Salix japonica is easily outcompeted by weeds and other vegetation. Many mining areas are predominantly bare ground with abundant weeds, and the growth rate of Salix japonica is slower than that of herbaceous species, resulting in inadequate access to resources such as water and sunlight [33]. Therefore, for the ecological restoration of ion-adsorption-type rare earth mining areas, planting Salix japonica may not be the best choice. Therefore, for ecological restoration in areas affected by multiple environmental stresses such as ammonia nitrogen pollution and heavy metal contamination, like those in ion-adsorbed rare earth mining regions, planting Salix japonica may not be the most effective choice.

4.3. The Indicative Role of Biochemical Parameter Inversion in Assessing the Environmental Adaptability of RV

Key biochemical parameters, such as chlorophyll, are not only directly linked to the physiological state of vegetation but also serve as important indicators of environmental stress levels. In this study, we applied continuous–discrete wavelet transform to decompose and filter the hyperspectral data from reclaimed vegetation in mining areas. This method enabled the accurate extraction of chlorophyll-sensitive spectral features and significantly improved the precision of chlorophyll inversion. Across different wavelet decomposition scales, the optimal inversion accuracy for all reclaimed vegetation exceeded 0.9, with a maximum accuracy of 0.93. In contrast, previous chlorophyll inversion studies have largely focused on crops in farmland, forests, and saline–alkali vegetation. For example, Wang et al. [34] achieved a chlorophyll inversion accuracy of 0.87 for maize using stepwise regression feature extraction combined with ensemble regression. Shen et al. [35] employed first-order derivatives and principal component analysis on hyperspectral winter wheat data, combined with partial least squares regression, reaching an inversion accuracy above 0.5. Zheng et al. [36] used original and first-order differential spectra to invert the chlorophyll fluorescence parameters of Suaeda glauca, achieving an accuracy of 0.81. Zarco-Tejada et al. [37] used optical indices to estimate chlorophyll content in coniferous forests, with an accuracy of 0.40.

While these studies have made progress, most of them reported optimal chlorophyll inversion accuracies below 0.9, and limited research specifically targets vegetation in mining areas, especially reclaimed species. These findings demonstrate that the continuous–discrete wavelet transform proposed in this study shows strong potential for biochemical parameter inversion under multiple environmental stresses. It not only facilitates the precise extraction and analysis of chlorophyll changes but also helps assess the ecological adaptability of vegetation and rapidly evaluates its growth status in complex environments. This has significant implications for optimizing ecological restoration measures.

Although this study provides an analysis of the environmental adaptability of reclaimed vegetation in areas affected by multiple environmental stresses and offers a theoretically effective monitoring approach, the reasons behind the varying adaptability between different vegetation species require further exploration. This is closely related to the internal mechanisms of vegetation, such as the extent of heavy metal and ammonia nitrogen pollution they are exposed to, as well as the biomass of their growth environment and soil properties. Therefore, future research will focus on further analyzing the spectral responses to these mechanisms, to deepen our understanding of the growth differences between vegetation species and to assist in future ecological restoration efforts.

5. Conclusions

The study found that complex mathematical transformation methods, such as Discrete Wavelet Transform (DWT), effectively amplify spectral information of vegetation under various environmental stresses and highlight detailed variations. Compared to the original spectra and first-order derivatives, DWT at scales d5 and d6 more effectively enhances spectral variation information and underscores the differences between RV and normal vegetation. This supports accurate monitoring of vegetation’s environmental adaptability under multiple stresses. Notably, Salix japonica has shown poor environmental adaptability in areas subjected to multiple environmental stresses, such as rare earth mining regions.

In modeling the relationship between spectral information of different RV types and chlorophyll content, the BP neural network (BPNN) has demonstrated superior performance compared to multiple linear stepwise regression (MLSR). Among the models, chlorophyll-sensitive bands at scales d4, d5, and d6 of Fir and Fir_CWT have shown excellent results across various vegetation types. This indicates that the combination of these input data with BPNN is well suited for chlorophyll inversion in vegetation subjected to multiple environmental stresses, such as ammonia nitrogen and heavy metal pollution. In other words, using the proposed continuous–discrete wavelet transform to decompose and filter hyperspectral data can help improve the accuracy of chlorophyll estimation in vegetation under multiple environmental stresses. This approach enables fast and precise monitoring of vegetation growth.