Analysis of the Effects of Different Spectral Transformation Methods on the Estimation of Chlorophyll Content of Reclaimed Vegetation in Rare Earth Mining Areas

Abstract

Ion adsorption rare earths are an important strategic resource, but their leach mining causes post-mining wastelands and tailings to suffer from soil sanding, acidification, and heavy metal contamination. This makes natural vegetation recovery difficult, relying mainly on artificial reclamation; however, the reclaimed vegetation grows poorly due to environmental stress. Hyperspectral remote sensing technology, with its high efficiency, non-destructive nature, and wide-range monitoring capability, can accurately estimate the physiological parameters of reclaimed vegetation. This provides support for environmental regulation in mining areas. In this study, three typical types of reclaimed vegetation in the Lingbei Rare Earth Mining Area, Dingnan County, Ganzhou City, were analyzed. Hyperspectral data and the corresponding chlorophyll content were collected to compare the spectral differences between reclaimed and normal vegetation. The spectral data were processed using mathematical transformation, fractional order differentiation, discrete wavelet transform, and continuous wavelet transform. Sensitive bands were extracted, and multispectral transformed feature bands were integrated. Linear and machine learning regression models were used to estimate chlorophyll content. The effects of different spectral processing methods on chlorophyll estimation were then analyzed. The results showed that reclaimed vegetation had higher spectral reflectance than normal vegetation, with the red valley shifting towards the long-wave direction and a steeper red edge slope. Different spectral transformation methods impact the accuracy of chlorophyll content estimation. Using appropriate methods can improve estimation accuracy. Fusing multi-spectral transformation features can achieve relatively good results. Among the models, the random forest regression model provides the best performance in estimating the chlorophyll content of reclaimed vegetation. This study provides a scientific basis for rapid and accurate monitoring of reclaimed vegetation growth in rare earth mining areas, supporting environmental management and decision-making and contributing to ecological restoration.

1. Introduction

Rare earth (RE) elements are vital for national security and advanced technologies. They are used in aerospace, renewable energy, electronics, and defense industries [1]. In the past few decades, RE mining has mainly used three methods: pool leaching, heap leaching, and in situ leaching. However, these methods, along with the use of acidic solutions, have caused significant environmental damage. This includes large amounts of wasteland and tailings. Soil desertification, acidification, and heavy metal contamination have worsened. These impacts hinder vegetation growth and threaten the health and livelihoods of local communities [2]. Due to the challenges of vegetation recovery, ecological restoration now depends mainly on artificial reclamation efforts [3].

Chlorophyll is essential for life on Earth. It plays a key role in plant growth and development, affecting photosynthesis in green plants. It is also an important indicator for assessing plant health [4]. However, traditional monitoring methods have several drawbacks. They are time-consuming, labor-intensive, lack timeliness, and may damage plant structures. In contrast, hyperspectral remote sensing technology offers several advantages. It has a wide range of bands and high spectral resolution, allowing for the precise estimation of chlorophyll content [5]. This technology supports large-scale, non-contact monitoring and enables non-destructive detection, making it an efficient and reliable tool for assessing vegetation health.

In recent years, researchers worldwide have conducted numerous studies on estimating plant chlorophyll content, yielding significant results. To improve the accuracy of chlorophyll content estimation, many have applied mathematical transformations (MTs) to raw spectral data. For example, Zhang et al. processed apple tree spectra using Savitzky–Golay (SG) smoothing, continuum removal (CR), and second-order derivatives (SDs). They found that the SD-transformed spectra had the highest correlation with chlorophyll content, improving estimation accuracy [6]. Nigela Tuerxun et al. applied different orders of derivatives to jujube leaf spectra, combining these transformations with spectral dimensionality reduction and a support vector regression model. Their results showed that the processed spectral data enhanced model performance [7]. Li et al. used fractional order differentiation (FOD) for hyperspectral analysis of grape leaves across different varieties and growth stages, combining it with the random forest (RF) algorithm to estimate chlorophyll content. Their method demonstrated high monitoring accuracy and application potential [8]. Some studies have also combined multiple MTs with signal-processing techniques to evaluate their effects on chlorophyll estimation accuracy. For example, Li et al. applied continuous wavelet transform (CWT) and traditional spectral transformations to process cotton spectral data. They used three regression methods to construct an estimation model, and the CWT-based model demonstrated better predictive accuracy than the traditional spectral transformation model [9]. Xiao et al. utilized FOD and CWT to process citrus spectral data, improving the correlation between spectral reflectance and chlorophyll content [10]. Wang et al. evaluated four spectral processing algorithms for their sensitivity and accuracy in estimating chlorophyll content in dragon fruit. They found that the discrete wavelet-differential (DWT) algorithm significantly improved estimation accuracy [11]. Previous studies have shown that applying MTs and signal processing techniques can effectively improve the correlation between spectral data and chlorophyll content, thereby enhancing the accuracy of chlorophyll estimation. Additionally, incorporating advanced machine learning algorithms can improve the model’s predictive power and robustness, especially when dealing with complex and high-dimensional spectral data. These methods have demonstrated strong adaptability across various crop types, helping to account for differences in spectral features and reduce the impact of environmental interference on estimation accuracy. However, most existing research has focused on common economic crops, often studying only a single crop type. In contrast, research on chlorophyll content estimation in vegetation from reclaimed land in RE mining areas is limited. Furthermore, most studies have used a limited set of spectral processing methods, failing to fully explore the potential of various spectral transformation techniques in this context. Therefore, exploring multiple spectral transformation methods for estimating chlorophyll content in reclaimed vegetation is crucial, as it could provide more accurate support for monitoring vegetation growth in mining reclamation areas.

Hyperspectral data have great potential for ecological monitoring, but their high dimensionality and sensitivity to noise make them prone to environmental interference. In RE mining areas, the surface soil and waste materials can affect the spectral characteristics of vegetation. This leads to a large amount of irrelevant information in the spectral data, which reduces the accuracy of chlorophyll content estimation [12]. At the same time, the presence of water vapor and gases in the atmosphere further increases noise, reducing the accuracy of chlorophyll content estimation. To address these challenges, this study focused on three typical reclaimed vegetation species in RE mining areas: Eucalyptus globulus, Photinia serrulata, and Vernicia fordii. Various spectral transformation methods were applied to their spectral data to explore how these methods reduce environmental interference, improve spectral sensitivity to chlorophyll content, and enhance estimation accuracy. Furthermore, a chlorophyll content estimation model was constructed by integrating features from different spectral transformations applicable to various reclaimed vegetation types.

Given the diversity of reclaimed vegetation types, the large volume of spectral data, and significant differences in their characteristics, the random forest regression model was chosen for chlorophyll content estimation in this study. The random forest regression model is effective at handling high-dimensional data. It captures linear relationships well and can also adapt to complex nonlinear relationships. Additionally, it has strong noise resistance. To verify the effectiveness of different spectral transformation methods for estimating chlorophyll content in vegetation from RE mining areas, this study also introduced partial least squares regression (PLSR) and support vector regression (SVR) models for comparison. This study aims to provide a scientific foundation for vegetation monitoring and ecological restoration in RE mining areas. It also lays the groundwork for applying hyperspectral remote sensing technology in large-scale ecological monitoring, contributing to environmental protection and sustainable development.

2. Data and Methods

2.1. Study Area Overview

The Lingbei RE mining area is located in the northern part of Dingnan County, Ganzhou City, Jiangxi Province. The terrain consists of mountains and hills, with red soil as the dominant type, which has strong adsorption properties, aiding the accumulation of RE elements. The elevation is approximately 300 to 500 m, and the dominant rocks are granite and sandstone. The region has a complex geological structure and a subtropical monsoon humid climate. Its geographic coordinates are approximately 114°58′12″ E to 115°11′24″ E and 24°51′00″ N to 25°03′00″ N, covering about 213 km². RE mining has been ongoing for over 30 years, using methods like pool leaching, heap leaching, and in situ leaching. The combination of long-term mining and the area’s high temperatures and rainfall has severely impacted the ecosystem, causing soil desertification, water and soil erosion, and heavy metal contamination. For this study, we selected the Jiazi Bei reclamation mining site as the research area. The site is located at specific coordinates of 115°2′22″–115°3′16″ E and 24°58′21″–24°59′15″ N, covering an area of about 1.6 km². Since 2004, ecological restoration efforts have been made, with large-scale planting of reclamation vegetation. However, the harsh conditions have led to poor plant growth overall. To monitor vegetation growth quickly, non-destructively, and accurately, we selected three typical reclamation species: Eucalyptus globulus, Photinia serrulata, and Vernicia fordii. The goal was to explore methods for estimating chlorophyll content under the environmental stress of mining to effectively monitor the growth of reclamation vegetation in the mining area. An overview map of the study area is shown in Figure 1.

2.2. Data Acquisition and Preprocessing

The spectral reflectance of reclaimed vegetation leaves in the RE mining area was measured using an ASD FieldSpec4 spectrometer, covering a wavelength range of 350–2500 nm. The number of spectral bands collected during each acquisition was 2150. In the 350–1000 nm range, the spectral resolution was 3 nm with a sampling interval of 1.4 nm, whereas in the 1000–2500 nm range, the spectral resolution was 10 nm with a sampling interval of 2 nm. The spectral measurements were conducted from 11:00 AM to 2:00 PM on 14–15 August 2020 under clear and windless weather conditions. During the measurements, the spectrometer had a field of view of 7.5°, and the probe was positioned vertically above the leaves at a distance of approximately 0.6 m. A whiteboard calibration was performed before data collection, and 10 to 15 spectral measurements were taken for each leaf, with their average value used to generate the spectral reflectance curve for the sample [13]. In this study, samples of three representative types of reclaimed vegetation were collected from the Jiazi Bei mining site. These plants exhibited poor growth, low coverage, and a single vegetation type. At the same time, corresponding samples of three types of normal vegetation were also collected from an area more than 1 km upstream of the RE mine site, characterized by healthy growth, high coverage, and diverse vegetation types.

Immediately after each spectral measurement, the SPAD-502 chlorophyll meter was used to measure the SPAD value of the leaves, obtaining chlorophyll content data. During the measurements, leaf veins were avoided, and 10 to 15 measurements were uniformly taken around each leaf, with their average value used as the sample’s chlorophyll content [14]. The SPAD chlorophyll meter operates based on the principle of transmission, using two LEDs to emit red light (650 nm peak wavelength) and near-infrared light (940 nm peak wavelength) onto a specific part of the leaf. By measuring the light density difference at these two wavelengths, the relative chlorophyll content can be estimated. The SPAD value is dimensionless and has a high correlation with chlorophyll content, making it a common parameter for characterizing chlorophyll levels [15].

To minimize the influence of environmental factors and human error during the acquisition of hyperspectral data in the field, data preprocessing is necessary. This includes removing anomalous spectra, reducing noise, and calculating mean values. The spectral mean value of each vegetation type is used as the center, with 1.5 times the standard deviation set as the threshold to define the spectral interval. Spectra outside this interval are considered anomalous and excluded. During field collection, environmental factors can affect the chlorophyll content, cell structure, and water content of plants, altering their spectra in the visible and near-infrared regions. Additionally, the short-wave infrared region may exhibit noise due to strong absorption bands from water vapor and carbon dioxide [16]. Based on this, this study focused on the 350–1350 nm band for analysis, excluding the bands above 1350 nm. After preprocessing, the final valid datasets consisted of 132 groups of reclaimed Eucalyptus globulus, 54 groups of reclaimed Photinia serrulata, 86 groups of reclaimed Vernicia fordii, 43 groups of normal Eucalyptus globulus, 33 groups of normal Photinia serrulata, and 35 groups of normal Vernicia fordii.

2.3. Vegetation Spectral Data Processing Methods

2.3.1. Mathematical Transformation Methods for Vegetation Spectra

The spectral data collected from the RE mining area were influenced by environmental stress, resulting in more noise and weakening the accuracy of leaf chlorophyll estimation. To reduce background noise and improve the correlation between spectra and chlorophyll content, various mathematical transformations (MTs) were applied to the spectral data [17]. These transformations were performed on the original spectral ($\mathrm{R}$) data and their first derivative (FD, $\mathrm{R}\prime$), including logarithmic first derivative (LFD, lg$\mathrm{R}\prime$), radical first derivative (RFD, $\sqrt[2]{\mathrm{R}}\prime$), cube root first derivative (CRFD, $\sqrt[3]{\mathrm{R}}\prime$), logarithmic opposite-number first derivative (LOFD, −lg$\mathrm{R}\prime$), reciprocal transformation (RT, 1/$\mathrm{R}$), reciprocal first derivative (RTFD, 1/$\mathrm{R}\prime$), absorbance transformation and first derivative (ATFD, 1/lg$\mathrm{R}\prime$), opposite-number logarithmic first derivative (OLFD, $-\mathrm{R}\prime$), and continuum removal (CR). The FD and CR are calculated as follows:

$${\mathrm{R}}^{\prime }({\mathsf{\lambda }}_{\mathrm{i}})={\displaystyle \frac{\mathrm{R}\left({\mathsf{\lambda }}_{\mathrm{i}+1}\right)-\mathrm{R}\left({\mathsf{\lambda }}_{\mathrm{i}-1}\right)}{∆{\mathsf{\lambda }}_{\mathrm{i}}}}$$

(1)

In Equation (1), $\mathrm{R}\prime ({\mathsf{\lambda }}_{\mathrm{i}})$ represents the FD spectrum, ${\mathsf{\lambda }}_{\mathrm{i}}$ denotes the wavelength of each band, and $∆{\mathsf{\lambda }}_{\mathrm{i}}$ is the distance between two wavelengths.

$$\mathrm{D}=1-{\displaystyle \frac{{\mathrm{R}}_{\mathrm{a}}}{{\mathrm{R}}_{\mathrm{b}}}}$$

(2)

Equation (2) defines $\mathrm{D}$ as the depth of the absorption band, ${\mathrm{R}}_{\mathrm{a}}$ as the reflectance at the center of the band, and ${\mathrm{R}}_{\mathrm{b}}$ as the reflectance of a continuous medium at the same wavelength as ${\mathrm{R}}_{\mathrm{a}}$.

2.3.2. Fractional Order Differentiation of Vegetation Spectra

Fractional-order differentiation (FOD), an extension of traditional calculus, is widely used in fields such as image processing and signal enhancement [18]. Compared to traditional integer-order differentiation, FOD can capture finer spectral features related to chlorophyll content, thereby improving the accuracy of chlorophyll estimation models. FOD has a smaller variation interval, resulting in finer spectral resolution. It can effectively remove noise while retaining detailed information related to chlorophyll [19]. In this paper, the Grünwald–Letnikov differential method is applied to process the leaf hyperspectral data, calculated as follows:

$${\displaystyle \frac{{\mathrm{d}}^{\mathsf{\alpha }}\mathrm{f}\left(\mathsf{\lambda }\right)}{\mathrm{d}{\mathsf{\lambda }}^{\mathsf{\alpha }}}}\approx \mathrm{f}\left(\mathsf{\lambda }\right)+\left(-\mathsf{\alpha }\right)\mathrm{f}\left(\mathsf{\lambda }-1\right)+{\displaystyle \frac{-\mathsf{\alpha }(\begin{array}{c}-\mathsf{\alpha }+1)\end{array}}{2}}\mathrm{f}(\begin{array}{c}\mathsf{\lambda }-2)\end{array}+\cdots +{\displaystyle \frac{\mathsf{\Gamma }\left(-\mathsf{\alpha }+1\right)}{\mathrm{m}!\mathsf{\Gamma }\left(-\mathsf{\alpha }+1\right)}}\mathrm{f}\left(\mathsf{\lambda }-\mathrm{m}\right)$$

(3)

In Equation (3), λ denotes the value at the corresponding point, Γ(⋅) is the Gamma function, α is an arbitrary order, and m is the difference between the upper and lower limits of differentiation. In this study, α ranges from 0 to 2 in steps of 0.2. When α is an integer (0, 1, or 2), it corresponds to the original spectrum, FD, and SD, respectively [20].

2.3.3. Discrete Wavelet Transform of Vegetation Spectra

Discrete wavelet transform (DWT) is a method used to decompose a signal on a discrete scale. It extracts and separates the signal’s local features through processing in both the time and frequency domains, thereby enhancing the aggregation and sensitivity of spectral information [21]. This technique is commonly used for estimating and analyzing chlorophyll and heavy metal content. In this paper, the ‘Db5’ wavelet function is applied to process the plant hyperspectral data. The data are decomposed into low-frequency and high-frequency components, with the low-frequency part retained and the high-frequency part thresholded. The low-frequency part is then further decomposed to extract new low-frequency and high-frequency information. After multi-layer decomposition, the original signal is transformed into several sub-signals. The calculation formula is as follows:

$$f\left(\lambda \right)=l_j\left(\lambda \right)+{\sum }_{k=1}^j{d}_k\left(\lambda \right)$$

(4)

In Equation (4), $f\left(\lambda \right)$ represents the spectral signal, $j$ is the decomposition layer, $l_j$ is the low-frequency component, and $d_k$ is the high-frequency component. The extracted detail coefficients are denoted as d1–d8.

2.3.4. Continuous Wavelet Transform of Vegetation Spectra

The continuous wavelet transform (CWT) analyzes a signal in the continuous domain by adjusting both the scale (scaling) and position (translation) of the wavelet, capturing its features at various scales and positions. This method is widely used in signal and image processing. In vegetation spectral data processing, the CWT effectively extracts features, suppresses noise, highlights details, and enhances spectral reflectance [22]. In this paper, the ‘gaus4’ mother wavelet is used to decompose the leaf hyperspectral data and capture subtle features in the vegetation spectrum. The calculation is as follows:

$$\mathrm{W}\mathrm{f}\left(\mathrm{a},\mathrm{b}\right)\le \mathrm{f};{\psi }_{a,b}\ge {\int }_{-\mathrm{\infty }}^{+\mathrm{\infty }}\mathrm{f}\left(\mathrm{t}\right){\psi }_{a,b}\left(\mathrm{t}\right)\mathrm{d}\mathrm{t}$$

(5)

where $\mathrm{W}\mathrm{f}\left(\mathrm{a},\mathrm{b}\right)$ represents the wavelet transform coefficients. The wavelet mother function can be scaled and translated to obtain the wavelet basis function, ${\psi }_{a,b}$, as in Equation (6).

$${\psi }_{a,b}\left(t\right)={\displaystyle \frac{1}{\sqrt{a}}}\psi \left({\displaystyle \frac{t-b}{a}}\right)$$

(6)

where $\mathrm{f}\left(\mathrm{t}\right)$ denotes vegetation reflectance spectral albedo; $t$ denotes spectral band; $a$ denotes scale factor; and $b$ denotes the translation factor. Due to the large amount of decomposed data, 10 specific scales ranging from 2¹ to 2¹⁰ were selected for analysis. These extracted scales are denoted as C1–C10 [23].

2.4. Estimation Model Construction and Testing

2.4.1. Random Forest Regression Model

Random forest (RF) regression is an ensemble learning method used to perform regression tasks [24]. It combines multiple decision trees, each trained on a different subset of the data. The final prediction is made through a voting mechanism. RF can handle high-dimensional data effectively, capturing linear relationships and adapting to complex nonlinear ones. It is also robust to noise. In chlorophyll content estimation, RF automatically extracts features from the data, identifies complex relationships between spectral features and chlorophyll content, and evaluates the importance of each feature in the prediction.

2.4.2. Partial Least Squares Regression Model

Partial least squares regression (PLSR) is a widely used method for regression analysis. It works by mapping high-dimensional spectral data into a low-dimensional latent variable space, extracting the components that best represent the variation in the data, and performing regression analysis based on these components. PLSR effectively handles multicollinearity by analyzing the common variation patterns between input variables and target variables [25]. While PLSR performs well with linear relationships, its performance may be limited when dealing with complex nonlinear relationships as it relies on linear combinations of latent variables.

2.4.3. Support Vector Regression Model

Support vector regression (SVR) is a regression method based on statistical learning theory developed by Vladimir Vapnik. It is widely used for classification and regression tasks [26]. In this study, a nonlinear radial basis function (RBF) kernel was used to identify the complex relationship between spectral data and chlorophyll content. The RBF kernel maps the data into a high-dimensional space by calculating the distance between data points and hyperplanes, effectively capturing nonlinear features. SVR improves the regression model by adjusting the kernel function’s parameters to enhance prediction accuracy.

2.4.4. Model Validation

During model training and testing, RF, PLSR, and SVR were evaluated using the same metrics. The training and testing samples were split in a 7:3 ratio, and the optimal model was selected through cross-validation, with the model having the smallest RMSE chosen as the final model. During testing, model performance was assessed by comparing the actual and predicted values in the test set using R², RMSE, and MRE. A model with R² close to 1, a small RMSE, and a low MRE indicates strong estimation ability and results that closely match the actual values. The specific formulae are as follows:

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

(7)

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

(8)

$$\mathrm{M}\mathrm{R}\mathrm{E}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\displaystyle \frac{\left|y_i-y_i^{\prime }\right|}{y_i}}\times 100\%$$

(9)

where $y_i$ and ${y^{\prime }}_i$ denote the actual and predicted values of chlorophyll content, respectively, and n is the number of samples.

3. Results

3.1. Original Spectral Characterisation

After pre-processing the hyperspectral data collected from normal and reclaimed vegetation leaves in the field, the mean spectral reflectance for each normal and reclaimed vegetation leaf was calculated to obtain the spectral reflectance data of the three reclaimed vegetation types. The comparison of spectral reflectance between normal and reclaimed vegetation is presented in Figure 2.

From Figure 2, it is clear that the spectral curves of the three types of vegetation, after mean value processing, show the same trend, which is consistent with typical vegetation spectral characteristics. Both normal and reclaimed vegetation in the RE mining area exhibit the ‘two valleys and one peak’ characteristic, indicating similar spectral trends. However, compared to normal vegetation, reclaimed vegetation has higher spectral reflectance in the visible and near-infrared wavelengths, which suggests clear signs of environmental stress. Specifically, near 660 nm, the red valley of reclaimed vegetation shifts towards the long-wave direction, with increased reflectance near 680 nm, a steeper slope in the red-edge reflectance, and an elevated green peak. These changes are closely related to the effects of environmental stress, pests, and diseases in RE mining areas, leading to damaged leaf structures, reduced chlorophyll content, and decreased water content [27].

Table 1. Statistical information on spectral characteristics.

Vegetation Type	Green Peak		Red Valley		Red Edge
	Position	Reflectance	Position	Reflectance	Position	Slope
Normal Eucalyptus globulus	557	0.1389	672	0.0762	706	0.0133
Reclaimed Eucalyptus globulus	555	0.1926	673	0.0729	701	0.0172
Normal Photinia serrulata	554	0.1234	671	0.0501	711	0.0125
Reclaimed Photinia serrulata	556	0.1679	675	0.0629	701	0.0154
Normal Vernicia fordii	553	0.0936	669	0.0567	718	0.0133
Reclaimed Vernicia fordii	555	0.1796	673	0.0653	700	0.0182

presents the spectral statistics information for three types of normal and reclaimed vegetation. From a physiological perspective, the reduction in chlorophyll in reclaimed vegetation significantly decreases its ability to absorb red and blue light, especially in the red-light region around 660 nm. This leads to an increase in reflectance and a shallower, shifted absorption valley towards the long-wave direction. The shift in the ‘red edge’ position reflects the changes in chlorophyll’s spectral absorption properties [28]. The red edge region (680–750 nm), which is the transition zone from visible to near-infrared (NIR), typically shows a rapid increase in reflectance in healthy vegetation. However, in reclaimed vegetation, reduced chlorophyll and altered leaf structure result in a decreased ability to absorb red light and an increased NIR reflectance, leading to a steeper slope of the red edge and a noticeable change in reflectance. Additionally, the ‘green peak’ of reclaimed vegetation is elevated, further indicating limited photosynthesis. The reduction in chlorophyll decreases the absorption of green light, enhancing its reflection and increasing spectral reflectance at the ‘green peak’—a feature particularly prominent in reclaimed vegetation. Consequently, reclaimed vegetation in RE mining areas often exhibits symptoms such as leaf spotting and yellowing, reflecting poor growth conditions.

These subtle changes in spectral reflectance provide a scientific basis for monitoring and evaluating the health of vegetation under environmental stress. The shift in the absorption valley near 660 nm, the increase in the slope of the red edge, and the enhancement of the green peak reflectance reflect the adaptive physiological responses of reclaimed vegetation to stress. These findings are important for understanding the ecological restoration process of reclaimed vegetation in RE mining areas and provide valuable support for spectral data analysis and chlorophyll estimation model construction.

3.2. Effect of Different Spectral Transformations on the Correlation Between SPAD Values and Spectral Parameters

In this study, the original spectra and spectra processed using MT, FOD, DWT, and CWT methods were analyzed for their Pearson correlation with the chlorophyll content of the three types of reclaimed vegetation. The correlation coefficient, r, was used to quantify the level of correlation. The resulting correlation heat map is shown in Figure 3. Based on significance analysis, sensitive bands were selected, with bands exhibiting extreme significance (p < 0.01) chosen as the sensitive bands for this study. These bands form the data basis for subsequent chlorophyll content estimation. The number of sensitive bands selected by each method is presented in Figure 4. The preliminary results indicate that different spectral transformation methods have significant effects on both the correlation and extraction of sensitive bands.

When analyzing the raw spectral data (R or OR), the correlation with chlorophyll content was weak. For Eucalyptus globulus, Photinia serrulata, and Vernicia fordii, 125, 241, and 247 sensitive bands were extracted, respectively, indicating that unprocessed data had limited capacity to improve SPAD correlation. However, after applying MTs, correlations were significantly improved. For example, LFD, RFD, and CRFD increased the number of sensitive bands, with RFD showing the best performance, extracting 352 sensitive bands in Photinia serrulata. RFD enhanced spectral features and increased SPAD correlation through root-based processing and smoothing. The RFD transform effectively enhanced spectral features and improved the correlation with SPAD values through root sign processing and smoothing differentiation. In addition, the FOD method significantly influenced the correlation between spectral data and chlorophyll content. As the differential order increased from 0.2 to 2.0, the correlation improved, and the number of sensitive bands first increased and then decreased. This indicates that middle to high-order fractional differentiation can better capture subtle changes in the spectral data, thus improving the correlation with SPAD values. The DWT extracts different levels of detail coefficients by decomposing spectral signals, strengthening the relationship between spectral data and SPAD values. Medium-level detail coefficients perform best for sensitive wavelet extraction, as they effectively capture key information in the spectral signals. The CWT analyzes spectral features at different scales, with the mesoscale (C4–C6) bands showing particularly good performance. These bands significantly increase the number of sensitive bands, indicating that the mesoscale CWT is highly effective in extracting spectral features and improving the correlation with SPAD values.

Overall, different spectral transform methods significantly affect the correlation between SPAD values and spectral parameters. Root-first-order differentiation (RFD) and fractional-order differentiation (1.0–1.8 orders) effectively improved the correlation of the spectral data. The DWT worked best with mid-level detail coefficients (d3–d6), while the CWT performed best in the mesoscale range (C4–C6). These findings suggest that selecting appropriate spectral transform methods can greatly enhance the correlation between spectral data and SPAD values, extract more sensitive bands, and provide valuable data for chlorophyll content estimation.

3.3. Impact of Different Spectral Transformations on Chlorophyll Content Estimation in Reclaimed Vegetation

In this study, the sensitive bands and their corresponding chlorophyll content data, selected by four spectral processing methods, were divided into training and test sets in a 7:3 ratio. Three regression models were used to analyze the effects of different spectral transformations on chlorophyll content in reclaimed vegetation. The evaluation metrics of the test set for chlorophyll content estimation under different spectral transformations are shown in

Table 2. Random forest regression model estimation results.

Method	Form	Eucalyptus globulus				Photinia serrulata				Vernicia fordii
		Top 3 Bands	R2	RMSE	MRE	Top 3 Bands	R2	RMSE	MRE	Top 3 Bands	R2	RMSE	MRE
OR	R	704, 705, 698	0.538	4.539	9.29%	706, 698, 703	0.814	4.788	13.15%	704, 700, 699	0.767	3.061	10.48%
MT	LFD	752, 753, 747	0.722	3.520	11.25%	676, 482, 745	0.921	3.124	9.63%	666, 652, 667	0.807	2.781	10.14%
	RFD	704, 705, 698	0.829	2.757	9.27%	646, 753, 631	0.934	2.858	6.84%	656, 666, 487	0.770	3.037	10.15%
	CRFD	753, 631, 747	0.726	3.494	9.64%	753, 747, 756	0.907	3.376	7.27%	652, 666, 657	0.798	2.849	9.68%
	LOFD	753, 747, 752	0.738	3.419	11.30%	756, 748, 744	0.929	2.963	8.90%	666, 751, 744	0.803	2.815	11.17%
	RT	707, 699, 704	0.530	4.580	9.27%	699, 698, 706	0.744	5.609	13.84%	699, 608, 703	0.755	3.135	9.32%
	RTFD	590, 716, 722	0.716	3.558	7.86%	675, 756, 751	0.842	4.402	11.37%	715, 725, 716	0.835	2.572	9.17%
	ATFD	752, 658, 589	0.668	3.847	10.41%	755, 744, 743	0.931	2.920	8.94%	665, 680, 667	0.820	2.692	12.36%
	OFD	645, 646, 647	0.811	2.902	6.76%	756, 755, 485	0.891	3.665	7.98%	665, 656, 553	0.771	3.034	12.56%
	CR	746, 714, 708	0.685	3.746	11.50%	752, 743, 751	0.925	3.043	8.29%	744, 745, 743	0.835	2.577	7.35%
FOD	0.2	693, 705, 702	0.588	4.285	9.55%	693, 695, 696	0.745	5.592	13.72%	695, 697, 699	0.783	2.955	10.26%
	0.4	704, 699, 703	0.725	3.504	7.56%	699, 568, 515	0.794	5.032	13.01%	694, 696, 519	0.795	2.872	11.55%
	0.6	689, 700, 688	0.722	3.522	7.61%	647, 554, 708	0.810	4.829	11.65%	671, 663, 693	0.768	3.053	11.14%
	0.8	647, 648, 698	0.824	2.798	7.00%	510, 512, 491	0.800	4.958	11.81%	669, 693, 658	0.778	2.984	11.35%
	1.0	645, 646, 697	0.815	2.875	6.96%	756, 755, 753	0.876	3.902	8.96%	486, 660, 665	0.786	2.934	11.89%
	1.2	644, 645, 646	0.833	2.729	6.56%	746, 745, 736	0.897	3.557	10.56%	663, 501, 693	0.831	2.604	11.56%
	1.4	644, 646, 695	0.750	3.341	7.40%	619, 730, 596	0.917	3.200	8.88%	659, 658, 604	0.835	2.575	10.33%
	1.6	644, 651, 643	0.786	3.092	7.87%	713, 712, 633	0.909	3.339	9.48%	603, 644, 637	0.840	2.533	10.23%
	1.8	650, 644, 649	0.795	3.026	7.60%	503, 632, 508	0.902	3.476	8.95%	667, 603, 625	0.813	2.742	10.81%
	2.0	650, 638, 654	0.798	3.002	7.22%	706, 624, 508	0.884	3.779	10.58%	667, 524, 639	0.856	2.406	8.62%
DWT	d1	683, 699, 687	0.655	3.922	8.46%	517, 694, 697	0.694	6.135	13.07%	694, 668, 693	0.816	2.721	9.22%
	d2	652, 660, 699	0.730	3.472	9.01%	675, 539, 662	0.823	4.667	12.52%	699, 650, 665	0.850	2.454	6.81%
	d3	630, 655, 631	0.838	2.692	6.32%	704, 601, 625	0.930	2.930	10.53%	527, 544, 521	0.821	2.683	10.96%
	d4	621, 655, 1051	0.712	3.584	8.37%	747, 748, 539	0.761	5.417	12.49%	540, 654, 511	0.861	2.364	8.59%
	d5	825, 754, 725	0.795	3.025	8.36%	666, 755, 737	0.842	4.412	10.52%	565, 664, 629	0.822	2.673	10.41%
	d6	710, 672, 627	0.645	3.977	7.71%	765, 764, 589	0.916	3.217	8.91%	520, 581, 519	0.795	2.872	11.17%
	d7	423, 575, 426	0.618	4.125	8.50%	801, 685, 693	0.835	4.503	10.50%	704, 584, 803	0.773	3.021	12.35%
	d8	740, 1212, 746	0.261	5.739	14.53%	558, 557, 556	0.789	5.093	10.12%	564, 566, 577	0.564	4.186	13.05%
CWT	C1	697, 649, 681	0.719	3.537	8.49%	665, 540, 684	0.841	4.422	14.23%	679, 678, 656	0.814	2.736	7.36%
	C2	648, 647, 621	0.766	3.233	7.00%	632, 577, 622	0.829	4.587	13.92%	508, 651, 656	0.865	2.327	8.27%
	C3	652, 672, 651	0.792	3.044	6.90%	653, 700, 701	0.887	3.729	10.65%	629, 618, 506	0.860	2.373	10.35%
	C4	623, 639, 566	0.815	2.875	7.16%	469, 707, 472	0.828	4.592	10.47%	656, 501, 524	0.794	2.874	11.24%
	C5	622, 620, 621	0.714	3.570	7.13%	757, 501, 664	0.904	3.428	9.19%	483, 482, 673	0.840	2.538	10.51%
	C6	514, 723, 717	0.584	4.307	9.76%	796, 806, 803	0.836	4.494	12.17%	719, 716, 350	0.760	3.105	12.38%
	C7	351, 352, 873	0.518	4.635	12.12%	876, 867, 718	0.831	4.562	9.65%	447, 864, 445	0.799	2.841	10.56%
	C8	356, 408, 709	0.584	4.308	10.00%	361, 351, 357	0.845	4.367	8.08%	412, 361, 436	0.843	2.510	9.26%
	C9	580, 613, 575	0.142	6.185	22.99%	699, 728, 1340	0.632	6.729	14.51%	714, 713, 709	0.699	3.479	11.35%
	C10	355, 350, 351	0.056	6.489	19.32%	435, 434, 415	0.394	8.631	17.82%	485, 494, 495	0.660	3.697	14.92%

,

Table 3. Partial least squares regression model estimation results.

Method	Form	Eucalyptus globulus			Photinia serrulata			Vernicia fordii
		R2	RMSE	MRE	R2	RMSE	MRE	R2	RMSE	MRE
OR	R	0.401	8.877	15.55%	0.744	4.845	11.35%	0.690	4.180	9.64%
MT	LFD	0.649	4.286	8.47%	0.804	3.996	10.48%	0.783	4.664	11.88%
	RFD	0.617	4.068	8.11%	0.751	4.550	11.17%	0.765	3.166	8.42%
	CRFD	0.788	3.202	6.42%	0.806	4.183	10.70%	0.720	4.749	11.61%
	LOFD	0.532	5.033	10.19%	0.790	4.816	9.25%	0.806	3.932	9.74%
	RT	0.403	5.376	10.41%	0.627	5.308	12.52%	0.648	4.973	12.84%
	RTFD	0.481	4.776	9.52%	0.807	5.394	11.08%	0.746	4.379	11.62%
	ATFD	0.599	4.397	8.55%	0.816	3.785	9.55%	0.703	5.410	11.40%
	OFD	0.635	4.060	8.18%	0.825	5.367	10.43%	0.745	3.085	8.33%
	CR	0.520	4.551	9.08%	0.807	4.086	9.54%	0.724	4.298	11.05%
FOD	0.2	0.529	7.779	14.92%	0.574	7.136	14.69%	0.745	3.555	8.45%
	0.4	0.573	5.644	11.15%	0.753	6.809	14.55%	0.721	3.459	8.91%
	0.6	0.729	3.561	7.55%	0.826	5.006	9.75%	0.710	3.326	8.20%
	0.8	0.807	3.132	6.54%	0.851	4.343	7.47%	0.761	3.179	7.51%
	1	0.768	3.198	6.45%	0.836	4.570	8.86%	0.835	2.736	7.70%
	1.2	0.714	3.748	7.11%	0.791	5.239	10.69%	0.823	4.236	9.37%
	1.4	0.723	4.110	7.90%	0.662	7.510	13.79%	0.668	6.072	10.25%
	1.6	0.557	5.090	9.86%	0.560	7.457	12.84%	0.736	6.029	11.00%
	1.8	0.377	6.572	13.00%	0.380	8.420	16.30%	0.809	6.808	7.23%
	2	0.313	7.167	14.52%	0.570	7.239	15.11%	0.710	3.559	9.48%
DWT	d1	0.399	7.452	14.63%	0.765	5.975	17.46%	0.697	5.804	14.72%
	d2	0.336	6.590	13.07%	0.749	6.720	13.77%	0.755	3.776	9.55%
	d3	0.751	3.493	6.92%	0.832	4.474	8.48%	0.710	4.838	12.09%
	d4	0.796	4.438	9.69%	0.821	4.765	8.69%	0.843	2.748	6.63%
	d5	0.658	5.391	10.33%	0.822	4.595	7.63%	0.761	3.432	8.44%
	d6	0.554	5.215	9.86%	0.764	6.397	12.95%	0.798	2.935	7.29%
	d7	0.734	3.766	7.64%	0.799	6.075	12.90%	0.763	5.068	13.77%
	d8	0.557	4.613	9.32%	0.734	6.857	11.20%	0.677	6.168	16.58%
CWT	C1	0.697	4.325	8.70%	0.623	8.570	14.66%	0.623	4.762	12.96%
	C2	0.680	4.411	8.61%	0.610	8.217	14.22%	0.652	4.477	12.04%
	C3	0.750	3.977	7.87%	0.734	6.768	13.88%	0.785	3.020	7.73%
	C4	0.777	3.496	7.36%	0.753	4.419	9.22%	0.710	3.294	9.28%
	C5	0.671	4.438	9.14%	0.779	4.123	9.27%	0.659	4.569	10.85%
	C6	0.766	4.082	8.55%	0.758	4.316	9.23%	0.726	3.689	9.46%
	C7	0.364	7.291	14.81%	0.742	4.713	10.12%	0.757	3.240	9.12%
	C8	0.548	8.427	17.21%	0.648	5.014	14.49%	0.634	4.950	13.43%
	C9	0.227	9.727	20.25%	0.431	8.595	15.07%	0.679	5.182	12.25%
	C10	0.234	6.632	14.08%	0.422	8.493	16.73%	0.502	7.717	15.83%

and

Table 4. Support vector regression model estimation results.

Method	Form	Eucalyptus globulus			Photinia serrulata			Vernicia fordii
		R2	RMSE	MRE	R2	RMSE	MRE	R2	RMSE	MRE
OR	R	0.571	5.008	10.58%	0.637	7.851	12.94%	0.721	4.687	11.19%
MT	LFD	0.708	4.956	9.73%	0.651	6.603	12.18%	0.784	2.919	10.50%
	RFD	0.754	3.621	7.59%	0.723	5.017	10.88%	0.778	3.012	11.05%
	CRFD	0.619	5.245	11.43%	0.706	6.342	12.91%	0.785	2.942	11.15%
	LOFD	0.704	4.931	10.66%	0.656	6.565	12.47%	0.812	3.796	8.50%
	RT	0.524	5.194	10.81%	0.603	8.282	13.66%	0.598	4.486	11.60%
	RTFD	0.735	3.787	8.59%	0.696	6.503	8.37%	0.734	3.218	7.59%
	ATFD	0.710	4.792	10.24%	0.635	6.584	11.80%	0.760	3.032	10.09%
	OFD	0.618	4.910	10.47%	0.766	5.092	9.95%	0.770	3.066	9.88%
	CR	0.626	4.451	9.59%	0.804	5.588	10.83%	0.772	3.301	9.16%
FOD	0.2	0.610	4.146	8.84%	0.739	6.253	12.99%	0.744	5.096	13.01%
	0.4	0.638	4.345	8.88%	0.714	6.220	13.14%	0.795	5.130	12.67%
	0.6	0.669	4.615	9.49%	0.795	6.111	12.55%	0.819	4.662	11.72%
	0.8	0.627	5.009	10.12%	0.775	6.063	12.01%	0.783	4.431	11.84%
	1	0.610	4.918	10.47%	0.776	6.092	11.25%	0.804	4.386	11.26%
	1.2	0.630	4.670	9.89%	0.787	5.980	12.05%	0.768	4.016	10.40%
	1.4	0.629	4.689	9.94%	0.765	6.443	12.65%	0.791	4.092	10.53%
	1.6	0.591	4.669	9.88%	0.722	6.735	13.57%	0.813	3.499	8.74%
	1.8	0.587	4.899	10.24%	0.715	6.673	12.83%	0.791	3.788	9.35%
	2	0.588	4.940	10.31%	0.702	7.137	14.27%	0.833	2.534	6.71%
DWT	d1	0.388	5.737	12.38%	0.570	8.580	17.60%	0.672	3.815	10.14%
	d2	0.450	5.267	11.38%	0.577	8.136	16.60%	0.709	3.410	8.34%
	d3	0.735	3.727	7.85%	0.714	7.038	14.80%	0.809	3.499	8.95%
	d4	0.602	4.513	9.22%	0.690	7.293	14.81%	0.817	3.333	8.51%
	d5	0.590	4.771	10.26%	0.712	6.348	12.49%	0.767	4.366	10.87%
	d6	0.528	4.835	10.33%	0.558	7.166	12.71%	0.507	4.547	11.80%
	d7	0.497	4.878	10.11%	0.509	7.389	13.24%	0.646	5.585	14.74%
	d8	0.271	6.149	13.21%	0.542	7.656	15.85%	0.591	5.318	12.54%
CWT	C1	0.505	5.201	11.27%	0.609	8.285	17.16%	0.751	3.198	8.07%
	C2	0.470	5.292	11.19%	0.519	7.445	15.60%	0.776	3.012	7.70%
	C3	0.604	4.568	9.76%	0.465	8.116	15.76%	0.736	3.724	9.66%
	C4	0.749	4.266	9.09%	0.510	7.767	13.85%	0.721	4.643	11.89%
	C5	0.400	5.358	11.63%	0.764	5.926	10.46%	0.736	5.400	14.66%
	C6	0.349	5.377	11.47%	0.784	6.735	13.86%	0.665	6.020	15.49%
	C7	0.371	5.333	11.70%	0.712	8.778	18.34%	0.729	4.703	11.93%
	C8	0.605	5.370	10.33%	0.613	8.392	17.65%	0.769	3.447	9.18%
	C9	0.007	8.755	18.16%	0.534	8.932	18.31%	0.676	6.499	13.25%
	C10	0.017	7.817	16.75%	0.370	8.430	18.51%	0.687	4.804	11.15%

. In Figure 5, we present the best estimation results for three types of reclaimed vegetation from different models as the optimal examples. These examples are more representative and highlight the differences in the best spectral transformation methods applicable to chlorophyll content estimation models for different reclaimed vegetation.

3.3.1. Comparison of Chlorophyll Content Estimation Methods for Reclaimed Vegetation Under Different Spectral Transformations

The study focused on analyzing the effect of data processed using different spectral transformation methods. The RF regression model was used to estimate the chlorophyll content. From Table 2, estimation performance is poor with untransformed raw spectral data, particularly for Eucalyptus globulus, where R² is 0.538 and RMSE is 4.539. With the exception of RT, MT improves estimation accuracy, especially with RFD, achieving R² values of 0.829, 0.934, and 0.835 for Eucalyptus globulus, Photinia serrulata, and Vernicia fordii, respectively. This indicates that MT effectively suppresses background noise and enhances spectral features. FOD further refines spectral features, with estimation accuracy increasing as the differentiation order rises from 0.2 to 2.0. Mid- to high-order FOD (1.0–2.0) shows the best effect, yielding R² values of 0.833, 0.917, and 0.856 for Eucalyptus globulus, Photinia serrulata, and Vernicia fordii, respectively, demonstrating its capability to capture spectral details and improve accuracy. For DWT, spectral–chlorophyll correlation is strengthened by decomposing the spectral data into multi-level detail coefficients (d1–d8). Mid-level detail coefficients (d3–d5) are most effective, balancing sensitive band extraction with key spectral information. For instance, Photinia serrulata achieves an R² of 0.930 at the d3 level with an RMSE of 2.930. CWT extracts spectral features at varying scales (C1–C10), with low to mid-level scales (C2–C5) performing best. For example, Vernicia fordii at C2 reaches an R² of 0.865 and RMSE of 2.327, Eucalyptus globulus at C4 achieves an R² of 0.815 and RMSE of 2.875, and Photinia serrulata at C5 has an R² of 0.904 and RMSE of 3.428. However, the performance decreases at higher scales (C6–C10), suggesting that the C2–C5 scales are more effective in capturing spectral features and thus improving estimation accuracy. In the above results with better chlorophyll content estimation, the MRE was mostly kept within 10%, indicating that the deviation between predicted and actual values remained within acceptable limits, allowing for more accurate estimation in most cases.

RF can output feature importance. We analyzed the bands with the highest contributions and found that the bands with higher contributions varied across different spectral ranges under different spectral transformation methods. The original spectral data and MT bands mainly fall within the 700–750 nm range, which is closely linked to changes in chlorophyll content. The bands in FOD are more dispersed, with the most effective bands around the ’red valley’ (600–680 nm). Bands from the DWT and CWT are concentrated in the red edge region (680–750 nm), reflecting chlorophyll’s absorption and reflection characteristics. This indicates that, under the influence of the RE mining area environment, the sensitive bands of reclaimed vegetation are primarily concentrated near the ’red valley’ and ’red edge’.

3.3.2. Comparison of Chlorophyll Estimation Methods for Reclaimed Vegetation in Different Regression Models

To validate the effectiveness of different spectral transformation methods for estimating chlorophyll content in reclaimed vegetation, we compared the results of three regression models: random forest (RF), partial least squares regression (PLSR), and support vector regression (SVR) (see Table 2, Table 3 and Table 4, and Figure 5). The comparison shows that the RF model outperforms both PLSR and SVR in terms of overall estimation accuracy.

In the PLSR and SVR models, the chlorophyll content estimated from the raw spectra had lower accuracy. However, after applying spectral transformation methods, the estimation accuracy improved significantly. The results following spectral transformation showed a similar trend to RF, indicating that the appropriate spectral transformation can enhance chlorophyll content estimation accuracy. PLSR, being a linear regression model, is best suited for regression problems with linear relationships. However, chlorophyll content estimation involves complex nonlinear relationships, which limits PLSR’s performance compared to RF. Specifically, PLSR produced lower R² and higher RMSE and MRE values than RF. While SVR is effective for handling nonlinear relationships, it struggles with linear relationships, leading to less accurate results than RF. In contrast, RF combines multiple decision trees to address both linear and nonlinear relationships, demonstrating strong noise immunity and achieving superior results.

In summary, these findings confirm the superiority of the random forest model for estimating chlorophyll content in reclaimed vegetation in RE mining areas. Additionally, they highlight the importance of selecting both the appropriate spectral transformation method and the suitable regression model to achieve better estimation results.

3.3.3. Estimation of Chlorophyll Content of Reclaimed Vegetation Incorporating Multispectral Transform Features

To further explore suitable methods for estimating chlorophyll content in different reclaimed vegetation of RE mining areas, this study integrates features from multiple spectral transformations and inputs the extracted sensitive bands along with chlorophyll content data into an RF model. The model outputs the top five contributing sensitive bands and evaluation metrics, as shown in Figure 6a–c and Figure 7. Compared to the single-spectral transformation model, the accuracy of the fused model decreased slightly, with R² values of 0.820, 0.883, and 0.841 for Eucalyptus globulus, Photinia serrulata, and Vernicia fordii, respectively. However, the fused model significantly improved estimation accuracy compared to the original spectral data, with R² values increasing by 28.15%, 6.94%, and 7.47% for Eucalyptus globulus, Photinia serrulata, and Vernicia fordii, respectively, with the largest improvement observed for Eucalyptus globulus. The five most sensitive bands showed that the primary sensitive bands for Eucalyptus globulus were concentrated in the visible region, where they had a strong correlation with chlorophyll content, resulting in more accurate estimations.

Compared to the RF model, both PLSR and SVR showed lower estimation accuracy (Figure 6). PLSR, as a linear regression model, struggles with the nonlinear relationship between chlorophyll content and spectral data, which lowers its accuracy. Although SVR can handle nonlinear relationships, it does not capture the complex connection between spectra and chlorophyll, preventing it from surpassing RF. This confirms the superiority of the RF model.

The results show that integrating multispectral transformation features improves the accuracy of chlorophyll content estimation in reclaimed vegetation of RE mining areas. This provides a solid foundation for vegetation growth monitoring.

4. Discussion

This study analyzed spectral differences of vegetation in RE mining areas and evaluated the effects of various spectral transformation methods on chlorophyll content estimation. The results show that environmental factors led to spectral variations in reclaimed vegetation, which differed from that of normal vegetation. Applying methods like MT, FOD, DWT, and CWT improved chlorophyll estimation accuracy, with optimal methods identified for different reclaimed vegetation types. By fusing multiple spectral features, the accuracy of chlorophyll estimation improved, and the R² value increased significantly.

It has been shown that vegetation under environmental stress exhibits different spectral characteristics compared to normal vegetation. For instance, Li et al. found that the spectra of winter wheat displayed a ‘red edge’ under low temperatures [29]. Similarly, this study also found that, under the influence of the mining area’s unique environment, the spectral characteristics of reclaimed vegetation differed significantly from those of normal vegetation. Specifically, the red valley became shallower, and the slope of the red edge increased. These changes reflect the spectral adaptation of vegetation and the optimization of resource use during photosynthesis under environmental stress. This phenomenon is closely related to the environmental conditions of heavy metal pollution and the restricted water and nutrient supply in the soil of RE mining areas. Regarding chlorophyll content estimation, previous studies have demonstrated the effectiveness of spectral transform methods in improving estimation accuracy [30,31,32]. In this paper, we explore the applicability of various spectral transform methods for reclaimed vegetation. Interestingly, we found that not all spectra processed by MT are suitable for estimating chlorophyll content in reclaimed vegetation. For example, inverse transforms did not perform well. FOD proved highly effective in revealing subtle variations at middle and high orders, particularly in the red-valley and red-edge regions, significantly enhancing the correlation between spectra and chlorophyll content. Additionally, DWT and CWT effectively maintained a high correlation between spectra and chlorophyll content through multi-scale and multi-level analyses. However, excessively large scales or levels in these transforms may lead to the loss of spectral information, affecting estimation accuracy. In a novel approach, this study integrates multiple spectral transform features for chlorophyll content estimation. The results indicate that this method significantly improves estimation accuracy compared to using the original spectra alone. This integration provides a new perspective and methodology for chlorophyll content estimation in reclaimed vegetation within RE mining areas.

This paper applies various spectral processing methods to enhance the accuracy of hyperspectral data for chlorophyll content estimation. This improvement provides strong support for ecological restoration in RE mining areas. Hyperspectral remote sensing technology enables quick and accurate assessment of reclamation effects without disturbing the ecosystem, offering a scientific basis for ecological restoration. Chlorophyll content serves as an indicator of vegetation growth and health. It provides real-time feedback on the progress of ecological restoration in mining areas, making it an effective indirect method for evaluating reclamation success. This approach not only aids in assessing the restoration effects of reclaimed vegetation but also lays the groundwork for future management strategies. For instance, it can guide adjustments to restoration strategies, optimization of planting schemes, or targeted conservation efforts in specific areas, thus ensuring the effectiveness of ecological restoration. Beyond its application in RE mining areas, the enhanced chlorophyll content estimation method used in this study is also applicable to other ecosystems under environmental stress. It can help identify changes in vegetation health, supporting the assessment and restoration of diverse ecosystems. In the ecological restoration of other types of mines and abandoned lands, mining activities and human disturbances often result in severe vegetation destruction and soil degradation. In such contexts, reclamation efforts face significant challenges. Hyperspectral technology can accurately monitor chlorophyll content in vegetation, assess its health in real-time, and provide data to inform the adjustment and optimization of reclamation efforts. Hyperspectral data can also reveal the spatial distribution of vegetation growth and chlorophyll content, identifying potential problems in the restoration process. This allows for targeted restoration measures that ensure the accuracy of the restoration work. In decertified areas, water scarcity and soil degradation are major barriers to vegetation restoration. Hyperspectral technology can effectively monitor chlorophyll changes, revealing how vegetation responds to environmental stress. This information supports the optimization of irrigation, water management, and vegetation planting patterns, promoting the sustainable development of desertification control. In addition, hyperspectral technology plays a crucial role in the restoration of degraded grasslands and abandoned arable land. Vegetation degradation due to overgrazing, climate change, and improper farming practices can be monitored through hyperspectral remote sensing, particularly changes in chlorophyll content. This helps assess restoration progress and provides a scientific foundation for grassland restoration and land management. In wetland and forest restoration, hyperspectral technology can track water quality, ecological health, and vegetation growth by monitoring chlorophyll changes in real-time. This helps detect ecological imbalances promptly and supports wetland protection and forest ecosystem restoration efforts. Therefore, Improving the accuracy of chlorophyll content estimation not only provides important support for ecological restoration in RE mining areas but also offers valuable insights for restoration in other ecosystems, helping contribute to the sustainable development of the global environment.

The current spectral transformation methods have shown progress in chlorophyll estimation, but they still require improvement and further validation in terms of computational efficiency and applicability across a broader range of real-world scenarios. Firstly, since the samples in this study are primarily from RE mining areas, they are relatively homogeneous. In the future, the sample range should be expanded, and more diverse data should be collected to improve the model’s generalization ability and estimation accuracy. Secondly, while multispectral fusion improves chlorophyll estimation accuracy, high-dimensional data and computational complexity remain challenges. Large-scale hyperspectral data increase storage and processing requirements, leading to longer processing times or wasted resources. Redundant information also adds complexity, potentially affecting the stability and accuracy of the estimation. For large-scale monitoring tasks, such as environmental monitoring or agricultural remote sensing, this computational burden can slow system response and hinder real-time or near-real-time analyses. To address this, dimensionality reduction techniques (e.g., principal component analysis, feature selection) can help reduce redundancy. These can be combined with optimization algorithms (e.g., genetic algorithms, particle swarm optimization) to select relevant features for chlorophyll, thereby improving computational efficiency, shortening processing times, and enhancing estimation accuracy. Lastly, the study did not fully account for environmental factors such as soil composition, heavy metal concentration, and moisture, which could introduce bias in chlorophyll estimations. Subsequent work should incorporate these environmental variables into the model and optimize it using deep learning algorithms. By including these factors as input features and employing methods such as multiple regression or deep learning techniques (e.g., convolutional neural networks, recurrent neural networks), a nonlinear relationship between chlorophyll content and environmental factors can be established. Through these strategies, the multiple factors influencing chlorophyll content can be more comprehensively addressed, improving the model’s adaptability and accuracy and providing more reliable support for ecological restoration assessment and decision-making in mining areas.

In the future, combining hyperspectral satellite data with ground-based hyperspectral data for ecological assessment and monitoring will be a viable direction. Satellite data provide wide coverage but are limited by low spatial resolution and atmospheric effects. Ground data offer high precision but cover a smaller area. By merging the two, errors in satellite data can be corrected, improving the accuracy of vegetation physiological parameter inversion. This combination can also promote the development of applications in agriculture, ecological monitoring, and environmental assessment.

5. Conclusions

This study focused on reclaimed Eucalyptus globulus, Photinia serrulata, and Vernicia fordii in the Lingbei RE Mining Area, Dingnan County. It used hyperspectral data of vegetation and chlorophyll content data for analysis. Three regression models were used to build a chlorophyll content estimation model. The study compared the original spectra of normal and reclaimed vegetation and applied spectral processing to the reclaimed vegetation spectra. It then analyzed how different spectral processing methods affected chlorophyll estimation accuracy. The results of the study show the following:

(1)

The spectral characteristics of the three reclaimed vegetation types varied due to environmental stress in the RE mining area. These changes reflect how the reclaimed vegetation adapted to the environment.

(2)

Spectral transformation methods effectively reduce background noise, enhance spectral data, and improve the correlation between spectral parameters and chlorophyll content in reclaimed vegetation.

(3)

The effectiveness of spectral transformation methods in improving chlorophyll content estimation accuracy varies, and suitable spectral transformation methods can significantly improve the estimation accuracy of chlorophyll content.

(4)

Integrating multiple spectral transformation features significantly improved the estimation accuracy of chlorophyll content, demonstrating the effectiveness of this approach for reclaimed vegetation.

(5)

When comparing different regression models, the RF model outperformed both the PLSR and SVR models. It better captured the relationship between spectra and chlorophyll content, providing more accurate estimates.