Research on Accurate Inversion Techniques for Forest Cover Using Spaceborne LiDAR and Multi-Spectral Data

Abstract

Fractional Vegetation Cover (FVC) is an important parameter to reflect vegetation growth and describe plant canopy structure. This study integrates both active and passive remote sensing, capitalizing on the complementary strengths of optical and radar data, and applies various machine learning algorithms to retrieve FVC. The results demonstrate that, for FVC retrieval, the optimal combination of optical remote sensing bands includes B2 (490 nm), B5 (705 nm), B8 (833 nm), B8A (865 nm), and B12 (2190 nm) from Sentinel-2, achieving an Optimal Index Factor (OIF) of 522.50. The LiDAR data of ICESat-2 imagery is more suitable for extracting FVC than that of GEDI imagery, especially at a height of 1.5 m, and the correlation coefficient with the measured FVC is 0.763. The optimal feature variable combinations for FVC retrieval vary among different vegetation types, including synthetic aperture radar, optical remote sensing, and terrain data. Among the three models tested—multiple linear regression, random forest, and support vector machine—the random forest model outperformed the others, with fitting correlation coefficients all exceeding 0.974 and root mean square errors below 0.084. Adding LiDAR data on the basis of optical remote sensing combined with machine learning can effectively improve the accuracy of remote sensing retrieval of vegetation coverage.

1. Introduction

Vegetation plays a vital role in regulating the earth’s biogeochemical cycles, including water, energy, and carbon cycles, across both intricate biogeochemical processes on both temporal and spatial scales [1]. Vegetation covers approximately 20% of the earth’s surface, making the comprehensive monitoring of its parameters imperative [2]. Fractional Vegetation Cover (FVC) is a direct quantitative indicator that reflects vegetation growth [3]. It is defined as the percentage of the vertically projected area of the above-ground portion of vegetation within a designated statistical area, and proves to be a highly valuable feature in vegetation monitoring [4]. FVC stands as a pivotal parameter for estimating and monitoring ecosystems and their functionalities, enabling the detection of vegetation changes at regional scales, thereby assessing vegetation health and gauging ecological environmental quality [2]. The rapid, effective, and precise estimation of FVC holds great significance in providing short-term insights into the regional ecological environment status and advancing the sustainable development and management of our natural resources and environment [5].

There are two primary methods employed for estimating FVC: field measurements and remote sensing inversion [4,6]. In ground-based field surveys, FVC within designated quadrangle areas (typically 1 m or 10 m) is primarily assessed using visual estimation, dot matrix, and grid methods [7]. Due to the drawbacks associated with ground measurements, such as subjectivity, low efficiency, and high costs, their widespread application is challenging [8]. However, ground-based measurement methods remain essential for FVC calculations [9]. Firstly, they offer high accuracy and objectivity. Secondly, they provide precise ground reference measurements crucial for developing remote sensing algorithms [10]. The alternative approach involves remote sensing measurements using satellites, drones, or ground-based instruments to collect fundamental FVC data [11,12].

Specifically, there are three main methods for estimating FVC using remote sensing data: Statistical Model Method: This approach involves establishing an empirical estimation model through regression analysis [13,14]. It typically uses specific bands or combinations of bands from remote sensing data or vegetation indices derived from remote sensing data along with actual measured FVC data. The model quantifies vegetation coverage based on remote sensing data. Statistical models can be further categorized into linear regression models and nonlinear regression models [2]. Mixed Pixel Decomposition Method: This method assumes that each component within a pixel contributes to the information observed by the sensor. It establishes a mixed pixel decomposition model to estimate FVC [15,16]. There are five primary types of hybrid pixel decomposition models based on different decomposition methods: the linear spectral mixture model, geometric optical model, probability model, random geometric model, and fuzzy analysis model. Machine Learning Method: Machine learning methods encompass various techniques such as the random forest method (RFM), neural network method (NNM), and support vector machine (SVM), among others [17,18]. These algorithms simulate the learning process of humans through software algorithms. They establish a discrimination mechanism by selecting and training on specific training areas, which is then used to classify and calculate other areas in the image [19].

Remote sensing inversion provides several advantages, including efficiency, cost-effectiveness, and objectivity, rendering it a crucial tool for acquiring vegetation coverage data on extensive spatial scales over extended time periods [20,21,22]. A SAR also penetrates forest vegetation to a certain extent (with penetration depth increasing with wavelength), allowing it to capture information about tree trunks and main branches, thus providing insights into the vertical structure of forest vegetation [23,24,25]. SAR’s polarization and polarization interference properties have found widespread use in quantitative FVC inversion. Nonetheless, obtaining high-quality spaceborne SAR interference data remains a challenge, limiting SAR’s application for large-scale FVC quantitative inversion [26,27]. On the other hand, Light Detection and Ranging (LiDAR) technology emits laser energy and receives return signals, granting it a more direct and accurate ability to assess forest structure and understory terrain compared to other remote sensing methods. LiDAR provides direct and highly accurate structural measurements, particularly for canopy height and density [28,29].

This study combines active and passive multi-source remote sensing data to conduct inversion research on vegetation coverage. Firstly, based on actual measurement data, the vegetation coverage sampling capabilities of active ICESat-2 and GEDI spaceborne LiDAR were explored, and satellite ground collaborative vegetation coverage sampling was conducted. Secondly, effective remote sensing estimation factors for different vegetation types from multi-source data such as passive optical remote sensing Sentinel-2, active microwave remote sensing Sentinel-1, and DEM were explored. The objectives of this study are as follows: (1) Determine the optimal segmentation inversion unit for ground FVC estimation and determine the parameter values set for each factor. (2) Extract feature variable datasets from various remote sensing sources for different ground FVC inversion purposes and identify the key feature variables. (3) Compare FVC estimation models employing different algorithms and select the most effective inversion model.

2. Study Area

The study was conducted in the upper reaches of the Daqing River Basin, located in the Baiyangdian region, in the central part of the Haihe River Basin in China, bounded by coordinates 38°10′~40°10′ N and 113°39′~117°34′ E [17]. This river system spans across Shanxi, Hebei, Beijing, and Tianjin, encompassing a total area of 43,060 km² [30]. The region comprises 18,659 km² of mountainous terrain and 24,401 km² of hilly plains, accounting for 43.33% and 56.67% of the total river system area, respectively. The basin experiences a temperate semi-arid continental monsoon climate, with average annual temperatures ranging from 6.58 °C to 13.26 °C and average annual precipitation ranging from 365 mm to 626 mm. Precipitation is highly variable throughout the year.

The study area features four distinct soil types: brown loam, brown soil, tidal soil, and loess. In the mountainous region, the average soil layer thickness does not exceed 50 cm, and is characterized by low organic matter content, loose texture, and susceptibility to erosion [31]. The study area exhibits a clear vertical zonal distribution of forest vegetation, with plant species changing with varying altitudes. In the middle and lower mountain areas, broadleaf forests and coniferous forests dominate, accompanied by an abundance of shrubs like vitex, among others. The hilly areas and lower elevations are primarily characterized by various types of artificially planted broadleaf forests and economically significant fruit trees [32].

3. Data and Methods

3.1. Research Data Source and Processing

The data utilized in this study primarily encompass three main categories: field survey data, remote sensing data, and additional data (

Table A1. Remote sensing supplemental material data used in the study.

Data Types	Description	Data Source
Field survey	Measurement of FVC data	Field investigation
Spaceborne LiDAR	ICESat-2	https://nsidc.org/data/icesat-2, accessed on 12 January 2024
	GEDI	https://gedi.umd.edu/data/download, accessed on 12 January 2024
Optical remote sensing	Sentinel-2	https://www.gscloud.cn/, accessed on 12 January 2024
Synthetic aperture radar	Sentinel-1
Terrain	DEM	https://www.gscloud.cn/, accessed on 12 January 2024
Land use/Land cover change	GlobeLand30	http://www.globallandcover.com/, accessed on 12 January 2024

and

Table A2. Land cover types, landscape metrics and other factors abbreviations.

Abbreviation	Description
NDVI	Normalized difference vegetation index
RVI	Ratio vegetation index
EVI	Enhanced vegetation index
TNDVI	Transformed normalized difference vegetation index
RDVI	Renormalized vegetation index
GEMI	Global environmental detection index
SAVI	Soil adjusted vegetation index
TSAVI	Adjustable vegetation index for transformed soil
CI	Red edge chlorophyll index
REP	Red edge position index
PSRI	Plant senescence reflectance index
NDCSI	Normalized vegetation canopy shadow index
NDVHVV	Normalized radar vegetation index
RAVHVV	Ratio radar vegetation index
H	Elevation
S	Slope
SinA	Sine in aspect (eastward degree)
CosA	Cosine in aspect (northward degree)

). The field survey data comprise observations from two key sources: field data collected in 2019, consisting of more than 800 sampling points within 30 m × 30 m quadrates, and forest resource survey data from 2014, encompassing more than 900 patches. These datasets include detailed vertical structure information, considering the tree, shrub, and herbaceous layers. Remote sensing data comprise observations from a variety of sources, including spaceborne LiDAR data from ICESat-2 and GEDI, optical remote sensing data from Sentinel-2, and C-band synthetic aperture radar data from Sentinel-1. Additionally, auxiliary data encompass terrain data and land use data. The collected optical and microwave remote sensing data, along with topographic data, were segmented by unit, and feature variables were extracted. These data, in combination with LiDAR and ground-measured data, were employed to construct the vegetation coverage dataset for model development. The technical workflow is illustrated in Figure 1.

3.2. Object-Oriented Segmentation

This research relies on Sentinel-2 image data and utilizes eCognition 9.0 software to perform image segmentation [33]. The best index factor method, iterative method, and evaluation tool ESP2 were applied to determine the parameters of segmentation. Additionally, the accuracy of the segmentation outcomes was assessed by comparing them with measurement data and GF-2 images, which have a spatial resolution of 0.8 m.

3.2.1. Optimal Exponential Factor Method

The optimal exponential factor method (OIF) is an effective approach for striking a balance between the information content of image bands and the information redundancy among bands [34]. The calculation formula is:

$$OIF={\displaystyle \frac{{\sum }_{i=1}^n{S}_i}{{\sum }_{i=1}^n{\sum }_{i=1}^n\left|R_{ij}\right|}}$$

(1)

where n is the number of bands, S_i is the standard deviation of band i, and R_ij is the correlation coefficient between band i and band j.

3.2.2. Optimal Segmentation Parameter Evaluation

The receiver operating characteristic (ROC) curve can be used to evaluate the quality of the segmentation effect and determine the optimal segmentation scale parameters [35]. When the ROC curve presents a local peak value, the segmentation scale corresponding to the point is the optimal segmentation scale [36]. The calculation formula is:

$$ROC=\left[{\displaystyle \frac{L{V}_{\left(L\right)}-L{V}_{L-1}}{L{V}_{(L-1)}}}\right]\times 100$$

(2)

where LV_(L) is the average standard deviation of objects in the target layer L, while LV_(L−1) is the average standard deviation of objects in the next layer L − 1 of the target layer L.

3.2.3. Accuracy Verification Method of Image Segmentation Results

With the use of ArcGIS 9.3, a total of 200 segmentation objects are randomly selected within the measured sample field. Subsequently, a comparison is conducted to verify the accuracy of the segmentation. If the degree of area overlap between the segmentation object and the measured data exceeds 80%, it is categorized as a correctly segmented unit.

3.3. Vegetation Cover Sampling Method for Spaceborne LiDAR

3.3.1. ICESat-2 Sampling Method for Vegetation Coverage

The spaceborne LiDAR ICESat-2 employs photon counting technology to capture three-dimensional point cloud data of terrestrial objects [37]. Through the correlation of two ICESat-2 data products, namely, the vegetation canopy height and surface elevation data (ATL03) and the surface elevation data product (ATL08), distinct categories of data are derived, including noise points, ground points, and canopy points [38]. After eliminating the noise points, a smooth spline curve method is employed to interpolate the seed points, yielding the ground profile curve. Subsequently, the relative elevation of each photon point concerning the ground is calculated, and the photon points captured by ICESat-2 within each sampling unit are tallied to compute vegetation coverage. This process facilitates the accuracy assessment of ICESat-2-based vegetation coverage concerning survey quadrat data.

Topographic Section Extraction Method

The smooth spline function curve is used to interpolate and fit the ground photon points by MATLAB 7.0 software [39]. The smooth spline algorithm is to find the optimal fitting function f(x) to minimize the penalty coefficient RSS, and the formula is as follows:

$$RSS(f,\lambda )=\lambda \sum _{i=1}^N{w}_i{\{y_i-f(x_i\left)\right\}}^2+(1-\lambda )\int v_t{\left\{f^{″}\right(t\left)\right\}}^{2}dt$$

(3)

where λ is the smoothing coefficient, which is a constant and takes a value in the range 0–1. When λ = 1, f(x) can be any function that bends indefinitely through all sample points. When λ = 0, f(x) is relatively smooth and is a simple least square straight line fit.

Photon Point Classification Method

Two photon point classification methods are used in this study. The photon points were classified by using the photon point classification algorithm and the height threshold method. In this study, 10 heights were set, ranging from 0.5 m to 5.0 m, with a step size of 0.5 m [40]. Using the inversion FVC data and the measured quadrat, linear regression simulation was carried out to find the best fitting effect between the photon points of different heights and the measured data.

ICESat-2 Method for Calculating Vegetation Coverage

On the basis of photon point classification, the number of vegetation photons and ground photons was calculated [41]. The calculation formula is as follows:

$$C_a={\displaystyle \frac{N_a}{N_a+N_b}}$$

(4)

where $C_a$ stands for vegetation coverage, $N_a$ stands for vegetation photon count, and $N_b$ stands for ground photon count.

Sampling Method of GEDI Vegetation Coverage

The spaceborne LiDAR GEDI assumes that the laser can only reach the ground by penetrating the canopy gap, ignoring the contribution of multiple reflections (<10%). Vegetation coverage is the ratio of reflected echo energy of vegetation canopy to the total reflected echo energy [42]. The formula is as follows:

$$C_b={\displaystyle \frac{R_v}{R_v+R_g\times \frac{{\rho }_v}{{\rho }_g}}}$$

(5)

where R_v is the reflected echo energy of the canopy, R_g is the reflected echo energy of the ground, and $C_b$ is the vegetation coverage. $\frac{{\rho }_v}{{\rho }_g}$ is the reflectance ratio between the canopy and ground, which is an empirical value.

3.4. Extraction of Feature Variables for Vegetation Coverage Modeling

3.4.1. Feature Variable Extraction Method Based on Passive Remote Sensing Data

Nearly 200 remote sensing factors related to vegetation coverage were extracted from Sentinel-1 and Sentinel-2 data. These factors encompass three variables obtained through tassel-cap transformation, three variables derived from principal component analysis, their corresponding texture characteristics (3 × 8), 10-band reflectance data, 147 spectral indices, and 2 radar vegetation indices [27,43]. The contribution degrees of these factors are analyzed, and the important factors are selected as the characteristic variables for model estimation (

Table A3. The formula for calculating the vegetation index.

Vegetation Index	Calculation Formula	Number
NDVI	NDVI(i,j) = (bandi − bandj)/(bandi + bandj)	45
RVI	SR(i,j) = bandi/bandj	90
EVI	EVI = 2.5 × (band8 − band4)/(band8 + 6band4 − 7.5band2 + 1)	1
TNDVI	TNDVI = sqrt[(band8 − band4)/(band8 + band4) + 0.5]	1
RDVI	RDVI = (band8 − band4)/sqrt(band8 + band4)	1
GEMI	GEMI = z × (1 − 0.25 × z) − (band4 − 0.125)/(1 − band4)	1
SAVI	SAVI = 1.5 × (band8 − band4)/(band8 + band4 + 0.5)	1
TSAVI	TSAVI = 0.5 × (band8 − 0.5 × band4 − 0.5)/(0.5 × band8 + band4 − 0.15)	1
CI	CI = band7/band5 − 1	1
REP	REP = 705 + 35 × [(band4 + band7)/2 − band5]/(band6 − band5)	1
PSRI	PSRI = (band4 − band3)/band6	1
NDCSI	NDCSI = (band8 − band4)/(band8 + band4) × (bandi − bandimin)/(bandimax − bandimin)	3
NDVHVV	NDVHVV = (H − V)/(H + V)	1
RAVHVV	RAVHVV = V/H	1

Note: band_i represents the i-th band, band_imin represents the minimum value, and band_imax represents the maximum value. H represents the backscattering intensity under horizontal polarization, while V represents the backscattering intensity under vertical polarization (V-polarization).

and

Table A4. Image texture feature calculation formula and meaning.

Feature of Texture	Formula of Calculation	Meaning
Mean	${\displaystyle \sum _{i=o}^{N-1}}{\displaystyle \sum _{j=o}^{N-1}}ip(i,j)$	The average gray value of remote sensing imagery reflects the uniform distribution of pixel values, and the texture regularity is positively correlated with the mean value
Variance	${\displaystyle \sum _{i,j=0}^{N-1}}i·P_{i,j}\left(i-{\displaystyle \sum _{i,j=0}^{N-1}}i·P_{i,j}\right)$	The variance of gray value in remote sensing imagery reflects the variation degree of gray value in remote sensing image
Entropy	${\displaystyle \sum _{i=o}^{N-1}}{\displaystyle \sum _{j=o}^{N-1}}p(i,j)\mathrm{l}\mathrm{o}\mathrm{g}\left(p\right(i,j\left)\right)$	The larger the entropy, the higher the complexity, the greater the amount of information
Contrast	${\displaystyle \sum _{\left|i-j\right|=0}^{N-1}}{\left|i-j\right|}^2\left\{{\displaystyle \sum _{i=1}^N}{\displaystyle \sum _{j=1}^N}p(i,j)\right\}$	Reflect the clarity of remote sensing imagery
Homogeneity	${\displaystyle \sum _{i=o}^{N-1}}{\displaystyle \sum _{j=o}^{N-1}}{\displaystyle \frac{p(i,j)}{{\left[1+(i-j)\right]}^2}}$	Reflects the magnitude of local homogeneity of remote sensing imagery
Dissimilarity	${\displaystyle \sum _{i=o}^{N-1}}{\displaystyle \sum _{j=o}^{N-1}}(i-j)p(i,j)$	The similarity degree and dissimilarity degree of texture information of remote sensing imagery are higher, which indicates that texture information has stronger uniqueness
Correlation	${\displaystyle \sum _{i,j=0}^{N-1}}{P}_{ij}{\displaystyle \frac{(i-u_i)(i-u_j)}{\sqrt{\left({\alpha }_i^2\right)\left({\alpha }_j^2\right)}}}$	The similarity of matrix elements in rows and columns in the gray co-occurrence matrix
Angular Second Moment	${\displaystyle \sum _{i=o}^{N-1}}{\displaystyle \sum _{j=o}^{N-1}}{p}^2(i,j)$	Image gray distribution uniformity degree and texture thickness degree; the larger the second angular distance, the clearer the image texture, the more uniform distribution

Note: N is the number of gray levels, $P_{i,j}$ is the entry (i,j) in the GLCM, $u_i$ and $u_j$ are the GLCM means, ${\alpha }_i^2$ and ${\alpha }_j^2$ are the GLCM variances.

).

3.4.2. The Feature Variable Extraction Method of DEM

Topography exerts a twofold influence, impacting both the representation of vegetation in remote sensing imagery and the actual growth and distribution of plants by shaping the vegetation’s immediate environment [44]. Consequently, considering the intricate terrain characteristics in the study area, we extracted pertinent topographic factors, namely slope, slope direction, and elevation. These factors were derived from the pre-processed 10-m resolution ALOS DEM (Digital Elevation Model).

3.4.3. Feature Variable First Method

The correlation coefficient is used for correlation analysis, and the calculation formula is as follows [45]:

$$\mathrm{r}={\displaystyle \frac{{\displaystyle {\sum }_{\mathrm{i}=1}^n(x_i-\stackrel{-}{x})(y_i-\stackrel{-}{y})}}{\sqrt{{\displaystyle {\sum }_{i=1}^n{(x_i-\stackrel{-}{x})}^2{\displaystyle {\sum }_{i=1}^n{(y_i-\stackrel{-}{y})}^2}}}}}$$

(6)

where n is the number of samples, xi and yi are the characteristic variables and measured vegetation coverage values in object i respectively. $\overline{x}$ and $\overline{y}$ is the average value of the characteristic variables of n samples and the measured vegetation coverage.

3.5. Vegetation Coverage Based on Different Algorithms

In order to find the optimal FVC modeling method, this study used stepwise multiple linear regression (DPM), support vector machine (SVM), random forest (RDF), decision tree (DT), K-Nearest Neighbors (KNN), and the random forest regression model (RDM) to establish FVC inversion models for different vegetation types [46]. The fitting ability of different regression models was compared and analyzed.

Root mean square error (RMSE) is used to measure the deviation between the measured vegetation coverage and the estimated vegetation coverage [47,48].

3.6. The Number of Samples Collected Based on Land Use and the Number of Preferred Characteristic Variables

Due to certain differences in spectral characteristics among different vegetation types, this study first classified vegetation types according to land use products, and then screened characteristic variables. The optimal feature variables of vegetation coverage modeling under different vegetation types were selected. Using the optimization method of feature variables constructed in Section 3.4, the variables were optimized according to ICESat-2 vegetation coverage quadrat points collected under different vegetation types (

Table 1. The number of sampling points and characteristic variables in different land use types.

Types	Samples of ICESat-2	Number of Characteristic Variables
Wetland	33	15
Cropland	772	27
Shrubland	1087	33
Grassland	7724	41
Deciduous broadleaved forest	1308	31
Others	4971	40

Note: The number of samples for different land types is determined by the proportion of land types.

).

4. Results

4.1. Optimal Segmentation Data Set and Factor Setting Results

Table 2. The standard deviation of each band of Sentinel-2 data.

Types	Visible Light			Near Infrared					Short Wave Infrared
Bands	B2	B3	B4	B5	B6	B7	B8	B8A	B11	B12
Standard deviation	771.97	743.02	771.30	775.15	785.69	816.51	872.44	833.38	655.16	601.51
Mean value	762.10			816.63					628.34

provides the standard deviations for the 10 bands of Sentinel-2 images, illustrating that the near-infrared band contains the most information, followed by visible light and short-wave infrared. In

Table 3. OIF values in combinations of different Sentinel-2 bands.

Fid	Composition	Optimal Factor Index	Fid	Composition	Optimal Factor Index
A	B2, B5, B6, B8A, B11	490.99	B	B2, B5, B6, B8A, B12	501.02
C	B2, B5, B7, B8A, B11	497.12	D	B2, B5, B7, B8A, B12	509.05
E	B2, B5, B8, B8A, B11	509.33	F	B2, B5, B8, B8A, B12	522.50
G	B3, B5, B6, B8A, B11	476.50	H	B3, B5, B6, B8A, B12	486.69
I	B3, B5, B7, B8A, B11	482.43	J	B3, B5, B7, B8A, B12	494.43
K	B3, B5, B8, B8A, B11	494.22	L	B3, B5, B8, B8A, B12	507.43
M	B4, B5, B6, B8A, B11	486.59	N	B4, B5, B6, B8A, B12	495.14
O	B4, B5, B7, B8A, B11	492.65	P	B4, B5, B7, B8A, B12	503.03
Q	B4, B5, B8, B8A, B11	504.71	R	B4, B5, B8, B8A, B12	516.25
S	B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12	220.23

, we calculate the Pearson correlation coefficients between these 10 bands of Sentinel-2 imagery. Notably, the correlation coefficient between bands B5 and B8A, with core wavelengths of 703.9 nm and 864.8 nm, is found to be below 0.9 in comparison to the other bands. This suggests that there is limited information redundancy between these two bands and the remaining bands (

Table A5. Sentinel-2 data correlation coefficients between bands.

Correlation Coefficient	B2	B3	B4	B5	B6	B7	B8	B8A	B11	B12
B2	1
B3	0.989 **	1
B4	0.967 **	0.983 **	1
B5	0.939 **	0.966 **	0.963 **	1
B6	0.727 **	0.767 **	0.706 **	0.824 **	1
B7	0.641 **	0.681 **	0.613 **	0.737 **	0.983 **	1
B8	0.608 **	0.653 **	0.579 **	0.688 **	0.941 **	0.954 **	1
B8A	0.597 **	0.639 **	0.572 **	0.705 **	0.972 **	0.992 **	0.951 **	1
B11	0.659 **	0.726 **	0.750 **	0.840 **	0.788 **	0.734 **	0.699 **	0.732 **	1
B12	0.720 **	0.773 **	0.830 **	0.858 **	0.636 **	0.558 **	0.514 **	0.542 **	0.932 **	1

Note: ** represents p < 0.01.

). To prevent data redundancy, we systematically select one band from each set of three data sets and include the B5 and B8A band data. This process results in 18 different combinations of split data (labeled as A to R). Among these combinations, the combination F (comprising bands B2, B5, B8, B8A, and B12) exhibits the highest OIF value (522.50), signifying the most optimal segmentation effect. Consequently, this combination is employed to construct the dataset for image segmentation.

In the three variables of shape, compactness, and segmentation scale, the two variables were controlled to be unchanged, and the step increment experiment was performed on one variable to determine the optimal value, the initial value of shape factor and compactness was set to 0.1, and the test was carried out with step 0.1 increments to determine the shape factor and compactness of 0.1 and 0.5, respectively. The starting value of the segmentation scale parameter was set to 10, and the step size was increased to 200 for experimentation, and the ROC-LV curve was obtained, and it was found that there were obvious “oversegmentation” and “undersegmentation” phenomena when the step size was less than 50 and greater than 150. The red line in the figure corresponds to the mutation points of local variance between 50 and 150, which can be regarded as potential optimal scale parameters. The segmentation scale is set in steps of 25, divided into four levels: 50–75, 75–100, 100–125, and 125–150, and a mutation point (59, 82, 114, 147) is selected as the optimal segmentation parameter in each level. The accuracy was verified with the field sample data, and the scale, shape, and compactness were finally selected as the best segmentation values of 82, 0.1, and 0.5, respectively, with an accuracy of 87.67% (Figure 2 and

Table 4. Bands description of the Sentinel-2 image.

Segmentation Coefficient			Number of Pixels	Accuracy (%)
Segmentation Scale	Shape Factor	Compactness
59	0.1	0.5	2,241,826	72.6
82	0.1	0.5	1,192,036	87.67
114	0.1	0.5	633,804	70.78
147	0.1	0.5	385,285	50.68

).

4.2. Extraction and Evaluation of Vegetation Coverage with Spaceborne LiDAR

4.2.1. Ground Profile Extraction Results

Figure 3 shows the ground interpolation fitting curve and the distribution of the point cloud. Compared with the weak beam, more understory ground photon points can be obtained under the strong beam, and the ground terrain accuracy of interpolation fitting is higher. Therefore, in this study, only ground interpolation fitting of strong beam orbital data was performed to obtain continuous ground surface, and the ground interpolation fitting accuracy was high, and the coefficient of determination R² was above 0.95, which met the requirements.

4.2.2. ICESat-2 Sampling Accuracy of Vegetation Coverage

The accuracy analysis and verification of vegetation coverage extracted by ICESat-2 were randomly used to analyze and verify the vegetation coverage extracted by ICESat-2, and the appropriate method was chosen. It can be seen that under the separation of 0.5–2 m height threshold, the extracted vegetation coverage will be overestimated; under the separation of 2.5–3.5 m height, the extracted vegetation coverage will be overestimated at low coverage, and there is an underestimation phenomenon in high vegetation coverage, and the accuracy of vegetation coverage at different heights from largest to smallest is Cover 1.5 m > Cover 1 m > Cover 0.5 m > Cover 2 m > Cover 2.5 m > Cover ATL08 > Cover 3 m > Cover 3.5 m > Cover 4.0 m > Cover 4.5 m > Cover 5 m. In general, the accuracy of extracted vegetation coverage showed a trend of first increasing and then decreasing with the increase in separation height. The result based on the separation at a height of 1.5 m is the best. The correlation coefficient with the measured vegetation coverage is 0.853, and the R² is 0.763 (

Table 5. Scatter plot of ICESat-2 and measured FVC based on different photon point classification methods.

Vegetation Coverage	Fitting Equation	R
0.5	Y = 0.789x + 0.282	0.756
1.0	Y = 0.831x + 0.232	0.762
1.5	Y = 0.853x + 0.182	0.763
2.0	Y = 0.854x + 0.141	0.748
2.5	Y = 0.856x + 0.103	0.741
3.0	Y = 0.847x + 0.076	0.731
3.5	Y = 0.816x + 0.056	0.708
4.0	Y = 0.795x + 0.036	0.692
4.5	Y = 0.755x + 0.023	0.661
5.0	Y = 0.727x + 0.010	0.637
Total	Y = 0.793x + 0.333	0.74

).

4.2.3. Analysis of Sampling Accuracy of GEDI Vegetation Coverage

The accuracy analysis and verification of vegetation coverage extracted by GEDI were randomly used to analyze and verify the vegetation coverage extracted by GEDI, and the results showed that there was no correlation between the two, and the GEDI vegetation coverage in most of the samples was underestimated. The GEDI vegetation coverage was analyzed by using the sample square points collected by ICESat-2, which had a high correlation with the measured vegetation coverage. There are 684 identical sampling units between ICESat-2 and GEDI in the study area, and it can be seen the vegetation coverage extracted by GEDI has a positive correlation with that of ICESat-2, and the correlation is low; the size is 0.228, and the vegetation coverage of GEDI is significantly lower than that extracted from ICESat-2 (Figure 4).

4.3. Extraction and Analysis of Feature Variables of Vegetation Coverage Modeling Based on Multi-Source Data

Analyzing the relationship analysis results between 201 characteristic index variables and vegetation coverage, there was no significant correlation with vegetation coverage. The factor of 0.01 was removed, and the factors that have the strongest correlation with FVC were retained. SR65 in the vegetation index contributed the most to wetland vegetation cover modeling in different vegetation cover areas. In cropland vegetation cover modeling, band 7, SR124 and PCA1, VVEntroy and slope contributed more. In shrubland vegetation cover modeling, the contribution of multiple vegetation indices is relatively large. In the modeling of grassland vegetation cover, multiple vegetation indices contributed greatly. In broadleaf forests and others vegetation cover modeling, the vegetation index contributes the most (Figure 5).

4.4. Study on Remote Sensing Estimation of Vegetation Coverage by Comparing Multiple Algorithms

4.4.1. Analysis of Fitting Ability of Different Regression Models

Under different vegetation types, the estimation accuracy of the model established by the DPM was as follows: broadleaved forest > others > wetland > cropland > shrubland > grassland. The accuracy of the SVM for different vegetation types is: broadleaved forest > other > cropland > shrubland > grassland > wetland, all of which are less than the accuracy in the DPM model (Figure 6 and Figure 7). The accuracy of the DT for different vegetation types is: wetland > cropland > shrubland > grassland > others > broadleaved forest. The accuracy of the KNN for different vegetation types is: shrubland > grassland > others > broadleaved forest > cropland > wetland. The accuracy of the RDF is ranked as follows: broadleaved forest > others > shrubland > grassland > cropland > wetland, which is better than the evaluation accuracy of others (

Table 6. Accuracy of FVC estimation model based on DPM, SVM, DT, KNN, and RDM under different vegetation types.

Methods	Types	Cropland	Grassland	Shrubland	Wetland	Broad-Leaved Forest	Others
DPM	R	0.604	0.592	0.625	0.800	0.345	0.506
	RMSE	0.136	0.143	0.142	0.126	0.101	0.120
SVM	R	0.602	0.589	0.62	0.776	0.365	0.506
	RMSE	0.186	0.193	0.192	0.198	0.138	0.163
RDF	R	0.974	0.977	0.977	0.975	0.978	0.977
	RMSE	0.072	0.069	0.069	0.084	0.050	0.059
DT	R	0.702	0.602	0.615	0.801	0.450	0.601
	RMSE	0.102	0.123	0.185	0.156	0.118	0.152
KNN	R	0.856	0.896	0.901	0.852	0.862	0.895
	RMSE	0.095	0.085	0.091	0.101	0.078	0.084

Note: The DPM represents for stepwise multiple linear regression. The SVM represents for support vector machine. The RDF represents for random forest. The DT represents for decision tree. The KNN represents for K-Nearest Neighbors. The R represents for Correlation coefficient. The RMSE represents for root mean squared error.

).

4.4.2. The Accuracy of the Simulation Results of Different Models

In order to fully verify the accuracy of FVC inversion of RDF constructed in this study, the accuracy of more than 1700 survey data was evaluated, and five sample set verification methods were constructed. Overall, field survey data indicate that RDF performs best among the four machine learning models. Overall, field survey data show that the RMSE of RDF is 0.026 higher than DPM. Except for Coniferous mixed forest and Relatively coniferous forest, with a tolerance of 0.05, and Shrubland and Grassland, with a tolerance of 0.15, the accuracy of other RDMs is higher than that of pixel dichotomy (

Table 7. The accuracy of different models in simulating FVC.

Type	Samples	Types	Method	RMSE	Precision (e < 0.05)	Precision (e < 0.1)	Precision (e < 0.15)
Forestland	866	Pure broadleaved forest	RDF	0.08	59%	71%	94%
			DT	0.10	53%	61%	86%
			KNN	0.09	51%	67%	81%
			DPM	0.11	56%	66%	85%
			SVM	0.11	48%	56%	79%
	120		RDF	0.11	59%	64%	83%
			DT	0.12	42%	51%	69%
			KNN	0.11	48%	61%	72%
			DPM	0.15	40%	53%	71%
			SVM	0.19	39%	47%	62%
	120	Broadleaved relatively pure forest	RDF	0.11	64%	76%	89%
			DT	0.10	52%	63%	85%
			KNN	0.12	51%	65%	80%
			DPM	0.13	50%	64%	81%
			SVM	0.20	46%	58%	72%
	76	Mixed broadleaved forest	RDF	0.10	51%	62%	88%
			DT	0.12	49%	59%	71%
			KNN	0.14	49%	61%	75%
			DPM	0.14	46%	49%	63%
			SVM	0.2	39%	44%	58%
	28	Mixed coniferous broadleaved forest	RDF	0.08	79%	86%	89%
			DT	0.09	68%	59%	71%
			KNN	0.12	71%	75%	78%
			DPM	0.10	61%	64%	86%
			SVM	0.15	57%	63%	79%
	271	Pure coniferous forest	RDF	0.12	41%	50%	75%
			DT	0.13	49%	58%	71%
			KNN	0.13	47%	55%	72%
			DPM	0.15	43%	49%	66%
			SVM	0.16	39%	46%	65%
	106	Relatively coniferous forest	RDF	0.13	56%	64%	77%
			DT	0.10	44%	52%	71%
			KNN	0.13	48%	58%	73%
			DPM	0.13	48%	60%	78%
			SVM	0.14	46%	59%	75%
	52	Coniferous mixed forest	RDF	0.13	42%	48%	74%
			DT	0.12	46%	53%	72%
			KNN	0.11	49%	54%	75%
			DPM	0.14	65%	77%	81%
			SVM	0.19	39%	38%	68%
Shrubland and Grassland	133		RDF	0.59	65%	78%	96%
			DT	0.22	59%	76%	91%
			KNN	0.33	61%	77%	94%
			DPM	0.70	62%	74%	99%
			SVM	0.70	60%	69%	91%

and Figure 8).

5. Discussion

5.1. Relationship Between Active and Passive Remote Sensing Feature Index and Vegetation Coverage

Due to certain differences in the spectral characteristics of different vegetation types, this study adopts the classification of vegetation types based on land use products, which is also common in other studies [49]. Subsequently, it filters characteristic variables by category and selects the optimal characteristic variables for modeling vegetation coverage across different vegetation types [50]. Specifically, considering the short growth cycle of cropland, significant variations in vegetation coverage occur from June to September. Therefore, in this study, exclusively ICESat-2 data from September was utilized for cropland sampling, ensuring synchronization with other remote sensing data. It is worth noting that similar situations have been observed in other studies; for instance, Song et al. observed relatively high time uncertainty in cropland vegetation coverage due to complex phenological variations among different crops [2]. Moreover, wetland vegetation, being relatively scarce, yielded only 33 samples, which may introduce some minor deviations into the analysis [51]. In the study area, the number of ICESat-2 laser sampling points in the taiga region is limited. Consequently, the taiga category was not separately analyzed in this study; instead, it was incorporated into other forest types for vegetation cover modeling.

5.2. Selection of Machine Learning Methods

In areas with high vegetation cover, the density of evergreen trees tends to be higher compared to the understory background vegetation, leading to more accurate estimation [52]. It is important to note that this study selected a subset of commonly used models and did not delve into deep learning methods such as CNN, FCN, and Unet [11]. The choice between these approaches depends on various factors, including the research context, the size and characteristics of the study area, and considerations regarding model operational costs, such as hardware requirements and time constraints [53]. Additionally, machine learning models can be susceptible to overfitting issues during training, and the presence of noisy data can adversely impact model training, resulting in reduced generalization performance [54]. Future research efforts should prioritize enhancing the Random Forest (RDF) model, minimizing or mitigating the covariance among different spectral indicators, and improving both accuracy and stability [55]. This will, in turn, enhance the model’s applicability for the inversion of sparse vegetation cover in the study area.

5.3. Enlightenment for Vegetation Coverage Inversion

In future studies, phenological characteristics should be fully considered to estimate vegetation cover continuously over different time changes [56]. Future studies, as well as the accuracy of FVC predictions, can also be further solved by classifying vegetation types in greater detail. The problem of spatial and temporal uncertainties may be solved by further consideration of soil, groundwater, and phenological characteristics [57]. In addition, the increase in sample size and attempts by other machine learning models are also expected to improve the inversion accuracy of vegetation coverage [58]. At present, in areas with relatively large terrain fluctuations, it may lead to a decrease in the accuracy of estimating vegetation coverage; for example, the USGS’s Landsat surface reflectance product does not consider the correction of this effect, resulting in the uncertainty of estimating vegetation cover on rugged terrain, so, in areas with complex terrain, it is recommended to use more suitable high-precision vegetation cover products [32].

6. Conclusions

The improvement in the accuracy of remote sensing extraction affecting FVC is mainly limited by the data source. Although optical remote sensing has great advantages in the extraction of vegetation coverage, due to its characteristics of orthographic projection shooting and poor penetration, signal loss from understory vegetation in dense forests will occur in dense forests, and there will be the problem that long-term series of effective data cannot be provided due to the influence of cloudy and rainy weather. By integrating data sources, this study moved beyond traditional 2D optical and radar imagery to incorporate 3D LiDAR-based vegetation analysis. The research results show that, compared with the GEDI sampling results, the threshold classification method based on ICESat-2 photon point cloud data can significantly improve the inversion accuracy of vegetation coverage. By using the RDF and combining active remote sensing data (Sentinel’s 2, 5, 8, 8A and 12 bands) and passive remote sensing data (ICESat-2 data), in the cloud coverage area, when there are serious deviations in the extraction results of the NDVI pixel binary method, the FVC can be estimated more accurately. The results of this study showed that with a single data source, it is difficult to meet the application requirements of FVC inversion evaluation. Therefore, it is necessary to combine multi-source data including radar-spectrum data to integrate and give full play to the complementary advantages of LiDAR-spectrum data.