Biomass Inversion of Highway Slope Based on Unmanned Aerial Vehicle Remote Sensing and Deep Learning

Abstract

Biomass can serve as an important indicator for measuring the effectiveness of slope ecological restoration, and unmanned aerial vehicle (UAV) remote sensing provides technical support for the rapid and accurate measurement of vegetation biomass on slopes. Considering a highway slope as the experimental area, in this study, we integrate UAV data and Sentinel-2A images; apply a deep learning method to integrate remote sensing data; extract slope vegetation features from vegetation probability, vegetation indices, and vegetation texture features; and construct a slope vegetation biomass inversion model. The R² of the slope vegetation biomass inversion model is 0.795, and the p-value in the F-test is less than 0.01, which indicates that the model has excellent regression performance and statistical significance. Based on laboratory biomass measurements, the regression model error is small and reasonable, with RMSE = 0.073, MAE = 0.064, and SE = 0.03. The slope vegetation biomass can be accurately estimated using remote-sensing images with a high precision and good applicability. This study will provide a methodological reference and demonstrate its application in estimating vegetation biomass and carbon stock on highway slopes, thus providing data and methodological support for the simulation of the carbon balance process in slope restoration ecosystems.

1. Introduction

Vegetation biomass is an important indicator of ecosystem plant productivity and health [1,2,3,4,5]. Slope vegetation is an important part of the slope ecosystem and is significant for maintaining soil stability and ecological balance. On slopes, plants intercept rainfall, slow runoff, prevent and treat sand, and retain soil and water. During the construction of transport infrastructure, deep trenches and high embankments are usually excavated, which inevitably destroy the vegetation along the slopes and form a large number of exposed slopes, resulting in a reduction in vegetation diversity and soil erosion, or even geological disasters such as landslides, avalanches, and mudslides [6,7,8]. At the same time, slope vegetation usually consists of small trees, herbaceous vegetation, shrubs, or vines. Highway slope vegetation offsets carbon emissions from transport, and its restoration and protection are important ways to achieve carbon peak and neutrality targets. Vegetation biomass is a crucial component of ecosystem carbon balance modeling and a significant indicator of ecosystem productivity [9]. It is an important indicator for monitoring and evaluating vegetation health because it directly reflects the level of primary productivity and ecosystem structure. Spatially explicit information on changes in biomass on slopes can quantify the ecological consequences of slope restoration. Thus, the primary task of ecological restoration is to re-vegetate slopes and monitor the spatial and temporal slope vegetation biomass [10].

The traditional direct harvesting method for obtaining biomass provides accurate data on aboveground biomass within sample plots but is time-consuming, labor-intensive, and destructive to vegetation [11,12,13]. Due to the rapid development of remote-sensing satellites, geographic information systems, and UAVs technologies, the quantitative estimation of the aboveground biomass of vegetation on slopes will be quick, accurate, non-destructive to slope vegetation, and capable of macro-monitoring [14].

Integrating optical-remote-sensing, microwave-remote-sensing, and other multi-source-remote-sensing data is a new approach for estimating biomass on the ground at a large scale. Combining multiple data and field measurements for collaborative inversion can improve estimation accuracy [15]. Yan et al. used four machine-learning methods, namely RF, Gradient Boosted Decision Tree, Categorical Regression Tree, and Minimum Distance, to construct a forest aboveground biomass model for Taiyue Mountain forests in Shanxi Province using single-source and multi-source remote-sensing data and the Google Earth Engine platform, reporting that machine-learning algorithm parameterization based on multi-source remote-sensing variables can improve the prediction accuracy of mixed forests [16]. Basuki, T.M. et al. fused images from Landsat 7 ETM + and ALOS /PALSAR using the discrete wavelet transform and Brovey transform to estimate the biomass of tropical forests [17]. Tamiminia et al. conducted a study in New York’s Adirondack Park, utilizing data from Landsat, Sentinel-2A, and PALSAR-2/PALSAR and applying a gradient-boosting machine and extreme gradient-boosting models to estimate forest aboveground biomass [18]. Xu et al. conducted a study in Chongli District, Zhangjiakou City, utilizing multispectral remote-sensing data from the Sentinel-2A satellite and terrain information to establish biomass models for shrublands, broad-leaved forests, and coniferous forests. These results indicated that using multi-source remote-sensing data led to a higher estimation accuracy [19].

The rapid development of unmanned aerial vehicle (UAV) remote-sensing technology provides a more accurate and higher resolution model for biomass inversion at a fine scale. The ultrahigh hardware and software integration, flexible flight altitude, and other characteristics can quickly obtain a large amount of remote-sensing data, and a lower flight altitude greatly improves the spatial resolution of remote-sensing data [20,21,22]. An increasing amount of attention has been paid to the research of constructing accurate vegetation biomass inversion models based on UAV remote-sensing technology. Biomass inversion models can be divided into two categories: parametric and non-parametric models. For the parametric model, Batistoti et al. used visible-light images from a UAV to establish an inversion model of pasture canopy height and biomass for the Brazilian savanna biomass estimation. Traditional biomass inversion methods typically use statistical regression models, including linear, nonlinear, and generalized linear regression [23]. Zhang et al. used a canopy height model (CHM) to calculate canopy height metrics and established a grassland aboveground biomass estimation model at a plot scale by conducting synchronous experiments of UAV and field measurements [24]. Some scholars combined a variety of variables to build a more complex multivariate model. For example, Lussem et al. investigated German temperate grassland through UAV images and used a polynomial model combining the vegetation index and canopy height to estimate biomass, providing a promising method for grassland biomass monitoring. The non-parametric model has a higher estimation accuracy, but it is more complex and less explanatory compared with the parametric model [25]. For example, Li et al. used a random forest model to invert the aboveground biomass of grasslands in Texas [26]. Alvarez et al. constructed a K-nearest neighbor–grassland aboveground biomass model and evaluated the rotational grazing system and aboveground canopy characteristics of a grassland in Colombia [27]. Sharma et al. used support vector machine (SVM) models to construct grassland aboveground biomass prediction models for three artificial grasslands in South Dakota [28].

There is a lack of vegetation biomass inversion models and methods applicable to the slope scale because there are issues such as the spatial location deviation of remote-sensing data, difficulty in extracting vegetation structural parameters, and low precision in identifying tree species types. To provide methods for estimating highway slope vegetation biomass, quantitatively monitoring slope vegetation, ecological restoration, and carbon stock estimation, this study aims to integrate multi-source remote-sensing data from multiple platforms, including the Sentinel-2A satellite and UAVs. New methods such as deep learning can be used to fuse remote-sensing data and extract information about slope vegetation, construct models for slope vegetation biomass inversion, and obtain spatial distribution maps of slope vegetation biomass. This will provide data and methodological support for the simulation of the carbon balance process in slope restoration ecosystems and a methodology and demonstration of this approach in vegetation biomass estimation on road slopes, quantitative slope vegetation monitoring, ecological restoration, and carbon stock estimation.

2. Materials

2.1. Study Area

The study area is a rocky slope measuring 0.4 hectares in the deep road-cutting area on the right side of the Guanglian Expressway’s Hanguang Parking Area Square (northbound direction) from AK0 + 250.0 to AK0 + 392.9, located within the jurisdiction of Huangguan Town, Yingde City, Qingyuan, China (Figure 1). The slope has an orientation of approximately 324° and is designed as a terraced slope with a bench width of 2.6 m. The slope ratios of the first, second, and third terraces are 1:1.5, 1:2.0, and 1:2.0, respectively. The platform width is uniformly 2.0 m, and the maximum height of the slope at all levels is 8.0 m, with a total slope height of approximately 22.1 m. The reinforcement and protection measures adopted for the slope primarily include net hanging and soil spray seeding for the first-level slope, and grass spraying for the second- and third-level slopes.

2.2. Data Sources and Preprocessing

2.2.1. Remote-Sensing Data Sources

The Sentinel-2A remote-sensing images were obtained from the Copernicus Open Access Hub (copernicus.eu) of the European Space Agency. Sentinel-2A images were acquired on 3 June 2023 and 1 October 2023, which closely aligned with the timeframe of the UAV image acquisition and biomass quadrat sampling. Sentinel-2A images encompass 13 multiple bands. Atmospheric correction was conducted using SNAP version 9.0 software, and the 13 bands were resampled using bilinear interpolation with a spatial resolution of 10 m with the WGS-84 coordinate system.

Visible and multispectral UAV images were collected using a DJI Phantom 4 UAV on 29 May 2023 and 23 September 2023. The DJI Phantom 4 integrates one visible camera (RGB) and five multispectral cameras (blue, green, red, red edge, and NIR). The camera parameters are as follows: 1/2.3-inch CMOS, 12.4 effective megapixels, 94° lens angle of view, 20 mm lens focal length (35 mm format equivalent), and f/2.8 maximum aperture. The route planning was set using a DJI GS pro ground station with an altitude of 200 m, a flight speed of 5 m/s, a heading overlap of 80%, and a side overlap of 80%. DJI Terra software 3.4.0 was used to mosaic the acquired images, and a digital orthophoto map (DOM) of the study area was obtained based on the WGS-84 coordinate system with a resolution of 0.05 m.

2.2.2. Multi-Source Remote-Sensing Data Fusion

The Gram–Schmidt (GS) transformation is a method of converting a set of linearly independent vectors into orthogonal vectors, which is effective in the fusion of UAV and Sentinel-2A images [29,30,31]. The workflow of GS is shown in Figure 2.

The GS transformation method used in this study for fusing the UAV and Sentinel-2A images involves a sequence of detailed steps designed to optimize the combination of spectral and spatial information. The method is applied as follows:

(1)

Preprocessing and preparation: The required data are preprocessed by layer stacking, radiometric calibration, and geometric correction, and the overlapping areas between the UAV and Sentinel-2A images are identified. Specifically, Sentinel-2A images are combined into a single, simulated lower-resolution panchromatic band, while the mean value of all UAV image bands is calculated to create the UAV panchromatic band.

(2)

GS transformation: the GS transformation is applied, with the low-resolution Sentinel-2A image serving as the first component and all other bands used as subsequent components in the GS transformation.

(3)

Adjustment of panchromatic bands: The mean and standard deviation of both the UAV panchromatic band and the first component of the GS are calculated. The UVA panchromatic band is then adjusted to match the first component of the GS.

(4)

Inverse GS Transformation: Using the adjusted panchromatic band as the first component, an inverse GS transformation is performed. This step produces 13 0.05 m resolution multispectral bands, maintaining the spectral characteristics of the Sentinel-2A images while achieving the spatial resolution of the UAV images.

The fused images provide a solid data basis for extracting spectral and texture features.

2.2.3. Biomass Quadrat Sampling

The biomass quadrat was collected on the day the UAV image was acquired. The main vegetation types in the study area are herbs and shrubs. Herbaceous vegetation includes Bermuda grass and Broadleaf paspalum, and shrubby vegetation includes Silver wattle and Salix myrtillacea. Based on the previous sampling method [32,33,34] and the specific conditions of this study area, thirty-one sample plots, each measuring 5 m × 5 m, were established for this study (Figure 3). For herbaceous vegetation, 3 quadrats (0.1 m × 0.1 m) were randomly selected in each sample plot. The live plant parts in the quadrats were collected via the harvesting method. The fresh weight was obtained, and the samples were dried at a constant temperature (65 °C) for 24 h to constant weight. For shrubby vegetation, 10 representative Silver wattle and Salix myrtillacea were selected, respectively, from the study area, and the biomass of single silver wattle and Salix myrtillacea samples was obtained using the same laboratory method. In addition, a handheld GPS terminal was used to record the location information of the sample plot during the field survey, in order to investigate the type and amount of shrubby vegetation, and, thus, determine the shrub biomass in the sample plot.

3. Methods

3.1. Vegetation Probability Extraction Based on Deep Learning

Deep learning was used to extract slope vegetation based on the fused images to obtain a spatial distribution map of slope vegetation probability.

The purpose of deep learning is to let computers learn from humans, with the hope that machines can acquire skills through experiential learning without human participation [35,36]. The deep learning process includes three steps: creating label grids, training deep learning models, and using the trained models for object extraction. In this experiment, we created 10,000 labels (8000 labels as the training set and 2000 labels as the validation set) through visual interpretation combined with field survey data. The model with the strongest generalization ability was selected as the vegetation extraction model by initializing the deep-learning model and using randomized parameter training tools. The probability of slope vegetation was extracted based on the selected vegetation extraction model.

Based on the UAV data, slope vegetation information was extracted using a deep learning module, and the spatial slope vegetation probability distribution (VPD) was obtained. The spatial distribution map of the slope vegetation probability obtained using deep learning is roughly consistent with the real situation, which can characterize the distribution probability of surface vegetation well. The result can be used as an influencing factor in the inverse model of slope vegetation biomass.

3.2. Vegetation Index Extraction Based on Band Math

The spectral characteristics of vegetation can be used to effectively distinguish it from other ground objects in remote-sensing images. The vegetation index, a spectral characteristic of vegetation, is based on the strong absorption of vegetation chlorophyll (0.65 μm) through a linear combination of infrared and near-infrared bands. Linear or nonlinear mathematical transformations on multiband reflectivity were performed to enhance the vegetation information and reduce non-vegetation signals, thereby realizing the expression of vegetation information status. Considering the characteristics of the slope vegetation in the study area and the vegetation indices used in the biomass remote-sensing estimation model, seven vegetation indices were selected to construct the slope vegetation biomass inversion model. Various band operations were performed on the fused data of the Sentinel-2A images and UAV images to extract the corresponding vegetation index (

Table 1. Vegetation indices and calculation methods.

Vegetation Index	Calculation Formula	References
Soil-Adjusted Vegetation Index (SAVI)	$S_{AVI}={\displaystyle \frac{B_8-B_4}{B_8+B_4+0.5}}\times (1+0.5)$	[37]
Atmospheric Impedance Vegetation Index (ARVI)	$A_{RVI}=B_8-{\displaystyle \frac{2B_4-B_2}{B_8}}+(2B_4-B_2)$	[38]
Perpendicular Vegetation Index (PVI)	$P_{VI}=\sin 45°\times B_8-\cos 45°\times B_4$	[39]
Meris Terrestrial Chlorophyll Index (MTCI)	$M_{TCI}={\displaystyle \frac{B_6-B_5}{B_5-B_4}}$	[40]
Red-Edge Inflection Point Index (REIP)	$R_{EIP}=700+40\times {\displaystyle \frac{\frac{B_7+B_4}{2}-B_5}{B_6-B_5}}$	[41]
Normalized Difference Vegetation Index (NDVI)	$N_{DVI}={\displaystyle \frac{B_8-B_4}{B_8+B_4}}$	[42]
Green-Light-Normalized Vegetation Index (GNDVI)	$G_{NDVI}={\displaystyle \frac{B_8-B_3}{B_8+B_3}}$	[43]
Leaf Chlorophyll Index (LCI)	$L_{ci}={\displaystyle \frac{B_8-B_6}{B_8+B_4}}$	[44]
Normalized Difference Red-Edge Index (NDRE)	$N_{dre}={\displaystyle \frac{B_8-B_7}{B_8+B_7}}$	[42]
Optimized Soil-Adjusted Vegetation Index (OSAVI)	$O_{SAVI}={\displaystyle \frac{B_8-B_4}{B_8+B_4+0.16}}$	[45]

Note: B in this Table represents the fused data of the Sentinel-2A images and UAV images, where $B_2$ is the blue band; $B_3$ is the green band; $B_4$ is the red band; $B_5,B_6, \mathrm{a}\mathrm{n}\mathrm{d} B_7$ are the red edge bands; and $B_8$ is the near-infrared band.

).

3.3. Texture Feature Extraction Based on a Gray-Level Co-Occurrence Matrix

A gray-level co-occurrence matrix (GLCM) is a simple and effective method for calculating image texture features [46]. By calculating the grayscale image of the remote-sensing image, its co-occurrence matrix is obtained, and then the co-occurrence matrix is calculated to obtain several of the eigenvalues of the matrix to represent certain texture features of the image. In this study, after fusing Sentinel-2A images with UAV images, a texture feature extraction tool was used to extract texture features from the fused images based on ENVI 5.6 software. The textural features included mean, homogeneity, contrast, dissimilarity, entropy, angular second moment (ASM), correlation, and variance [47,48].

Mean: The mean is the average of all elements in the gray-level co-occurrence matrix, which reflects the regularity of the texture. The messier the texture, the smaller the mean value. Conversely, the stronger the regularity and the easier it is to describe, the larger the mean value.

$$Mean={\Sigma }_i{\Sigma }_{j}p(i,j)\times i$$

(1)

Homogeneity: This refers to the ratio of the number of occurrences of the same element to the number of occurrences of different elements in the same row or column of the gray-level co-occurrence matrix, thereby reflecting the texture similarity of the image. The higher the homogeneity value, the more similar the local areas with smaller grayscale differences.

$$Homogeneity={\Sigma }_i{\Sigma }_{j}p(i,j)\times {\displaystyle \frac{1}{1+{(i-j)}^2}}$$

(2)

Contrast: This refers to the degree of grayscale difference between adjacent areas in the image. It is usually used to describe the roughness, detail, and depth of an image and can be obtained by calculating the grayscale difference between each pixel and its surrounding pixels.

$$Contrast={\Sigma }_i{\Sigma }_{j}p(i,j)\times {(i-j)}^2$$

(3)

Dissimilarity: This refers to the degree of grayscale difference between adjacent pixels in an image, which can be obtained by calculating the grayscale difference between each pixel and its surrounding pixels.

$$Dissimilarity={\Sigma }_i{\Sigma }_{j}p(i,j)\times |i-j|$$

(4)

Entropy: This refers to the uncertainty in the gray value of each pixel in the image. It is a random measure of the amount of information in an image. It can be used to measure the richness and complexity of the textural information in an image. When all values in the gray-level co-occurrence matrix are equal, or the pixel values show the greatest randomness, the entropy is the greatest, and the image becomes more complex.

$$Entropy={\Sigma }_i{\Sigma }_{j}p(i,j)\times lnp(i,j)$$

(5)

ASM: This reflects the uniformity of the grayscale image distribution and texture thickness. It is the sum of the squares of the elements of the gray-level co-occurrence matrix known as energy.

$$ASM={\Sigma }_i{\Sigma }_j{p(i,j)}^2$$

(6)

Correlation: This refers to the degree of correlation between two texture features.

$$Correlation={\Sigma }_i{\Sigma }_j{\displaystyle \frac{(i-Mean)\times (j-Mean)\times {p(i,j)}^2}{p(i,j)\times {(1-Mean)}^2}}$$

(7)

Variance: This measures the deviation between the image pixel values and the mean; a large variance value indicates a significant change in the grayscale of the image.

$$Variance={\sum }_i{\sum }_j{\left(i-Mean\right)}^{2}p(i,j)$$

(8)

3.4. Correlation Analysis

Pearson’s correlation analysis was applied to quantitatively evaluate the correlation between slope vegetation biomass and slope vegetation characteristic indicators (including vegetation probability, vegetation indices, and texture features), and the t-test and the degrees of freedom were used to find the p-value in order to test whether the obtained correlation coefficient results had a significant correlation [49,50,51,52]. The principle of Pearson’s correlation analysis is shown in the following formula:

$$r={\displaystyle \frac{\sum _{i=0}^{n-1}(x_i-\overline{x})(y_i-\overline{y})}{\sqrt{\sum _{i=0}^{n-1}{(x_i-\overline{x})}^2\sum _{i=0}^{n-1}{(y_i-\overline{y})}^2}}}$$

(9)

$$t={\displaystyle \frac{r\sqrt{n-2}}{\sqrt{1-r^2}}}$$

(10)

where r is the Pearson correlation coefficient, x_i is the i-th vegetation characteristic indicator vector, y_i is the i-th biomass vector, n is the sample size, and $\overline{x} \mathrm{a}\mathrm{n}\mathrm{d} \overline{y}$ are the mean vegetation characteristic indicators and mean biomass, respectively.

3.5. Multiple Linear Regression

In order to establish a biomass inversion model, in this study, we comprehensively considered slope vegetation characteristic indicators (including vegetation probability, vegetation indices, and textural features) by conducting multiple regression analyses on biomass and high-correlation indicators [53,54,55,56]. The formula for multiple regression analysis was as follows:

$$Y={\beta }_1{X}_1+{\beta }_2{X}_2+\cdots +{\beta }_n{X}_k+\epsilon$$

(11)

Among them,$X_1$, $X_2$,…, $X_k$ are the independent variables corresponding to the slope vegetation characteristic indicators (including vegetation probability, vegetation index, and texture features); $Y$ is the dependent variable, corresponding to the biomass; ${\beta }_n(n=1\dots k)$ is the standard regression coefficient, corresponding to the least squares operation results of each band information and biomass; and $\epsilon$ is the error term.

3.6. Accuracy Evaluation

After obtaining the biomass inversion models of different typical slope vegetation types, it is necessary to evaluate their accuracy to understand the accuracy of the model construction. This study used indicators that are commonly used to measure the deviation between predicted and true values to evaluate the accuracy of the model inversion results. The evaluation indicators include the root mean square error (RMSE), mean absolute error (MAE), and system error (SE) [57,58,59,60]. The specific evaluation principle is as follows:

$$RMSE=\sqrt{{\displaystyle \frac{\sum _{i=1}^N{\left(\widehat{X_i}-X_i\right)}^2}{N}}}$$

(12)

$$MAE={\displaystyle \frac{\sum _{i=1}^N\left|\widehat{X_i}-X_i\right|}{N}}$$

(13)

$$SE={\displaystyle \frac{\sum _{i=1}^N\left(\widehat{X_i}-X_i\right)}{N}}$$

(14)

where $\widehat{X_i}$ is the estimated value of the ith sample (model inversion result); $X_i$ is the true value of the i-th sample (biomass quadrat verification sample result); and N is the sample capacity. Among them, the smaller the values of RMSE and MAE, the smaller the error of the biomass remote-sensing inversion model results, and vice versa. The closer the value of SE is to 0, the smaller the error of the biomass remote sensing inversion model results, which means that the overall bias of the biomass remote sensing inversion model results is overestimated when SE > 0, the overall bias of the biomass remote sensing inversion model results is underestimated, and the overall estimation of the biomass remote sensing inversion model results is not biased when SE < 0.

4. Results

4.1. Correlation Analysis between Various Influencing Factors and Slope Vegetation Biomass

Correlation analyses were carried out between the measured vegetation biomass and vegetation indices, texture characteristics, and vegetation probability distribution. Variables with no significant correlation were excluded; therefore, the filtered factors were obtained, as shown in

Table 2. Correlation analysis between influence factors and vegetation biomass on slopes.

Impact Factor	ASM	Variance	Mean	Homogeneity	Entropy	Dissimilarity	Correlation
Correlation coefficient	−0.752 **	0.535 **	0.862 **	−0.764 **	0.766 **	0.680 **	0.553 **
Impact factor	Contrast	VPD	ARVI	PVI	REIP	NDVI	GNDVI
Correlation coefficient	0.513 **	0.280 *	0.809 **	−0.557 **	−0.874 **	−0.715 **	−0.874 **
Impact factor	LCI	NDRE	OSAVI
Correlation coefficient	−0.873 **	−0.874 **	−0.754 **

** Correlations are significant at the 0.01 level (two-tailed). * Statistical significance was set at <0.05 level (two-tailed).

.

All factors were significantly correlated with biomass at the p < 0.01 level, except for VPD, which was significantly correlated with biomass at the p < 0.05 level. These variables were preliminarily selected as explanatory variables for vegetation biomass on the slopes.

4.2. Construction and Accuracy Testing of Slope Vegetation Biomass Inversion Model

After removing indicators with high correlation, eight were selected for multiple linear regression analyses. Slope vegetation biomass was used as the dependent variable. Spectral features, texture features, and spectral + texture features were used as the independent variables. The spectral features include VPD, ARVI, NDVI, and NDRE; the texture features include ASM, variance, correlation, and contrast. The final regression results are shown in

Table 3. Multiple linear regression analyses of each influence factor with slope vegetation biomass [58,59].

	Independent Variable		Unstandardized Coefficients B	t-Test		R2
	Type	Name
				F	P
Model 1	Spectral features	Constant term	20.589	45.059	<0.01	0.763
		VPD	0.063
		ARVI	0.063
		NDVI	0.083
		NDRE	−0.217
Model 2	Texture features	Constant term	25.725	27.945	<0.01	0.687
		ASM	−107.340
		Variance	−20.209
		Correlation	113.469
		Contrast	10.818
Model 3	Spectral + Texture features	Constant term	17.588	24.793	<0.01	0.795
		ASM	66.360
		Variance	−10.375
		Correlation	9.739
		Contrast	5.605
		VPD	0.047
		ARVI	0.187
		NDVI	0.058
		NDRE	−0.298

.

In Table 3, the adjusted R² = 0.795 of model 3 is the highest, indicating that, in this equation, each parameter can explain 79.5% of the slope vegetation biomass, and the regression effect is excellent; p < 0.01 in the F-test of the regression equation, and the significance test statistic F = 24.793, which is statistically significant. ASM, variance, correlation, contrast, VPD, ARVI, NDVI, and NDRE were selected as the final independent variables for the simulation of slope vegetation biomass. The regression model equation for slope vegetation biomass was obtained as follows:

$$y=66.3605x_1-10.3750x_2+9.7392x_3+5.6054x_4+0.0470x_5+0.1869x_6+0.0579x_7-0.2985x_7+17.5878$$

(15)

where ASM is $x_1$, variance is $x_2$, correlation is $x_3$, contrast is $x_4$, VPD is $x_5$, ARVI is $x_6$, NDVI is $x_7$, NDRE is $x_8$, and the constant term is 17.5878.

The biomass of the eight validation samples was validated against the biomass inversion results obtained from the model calculations for model accuracy.

Table 4. Model accuracy validation (unit: kg/m²).

RMSE	MAE	SE
0.073	0.064	0.03

shows the validation results.

As shown in Table 4, the RMSE was 0.073 kg/m², and the MAE was 0.064 kg/m², indicating that the error of the regression model was small and within a reasonable range. The SE was 0.03 kg/m², indicating that the results of the biomass remote-sensing inversion model were generally biased towards overestimation.

4.3. Biomass Inversion Mapping

Using the obtained slope biomass regression model, the layers of each influencing factor were superimposed using the raster calculator tool in ArcMap, and the remote-sensing inversion results of the slope vegetation biomass were obtained, as shown in Figure 4.

As shown in Figure 3, the values of the remote-sensing inversion results of the slope vegetation biomass are all between 0 and 5 kg per square meter, which is consistent with the spatial distribution of the real biomass. This proves that the modeling effect is good, and it can invert the vegetation biomass of slopes more accurately.

5. Discussion

Vegetation biomass on highway slopes can reflect the ecological restoration situation of the slopes, so accurate predictions of the vegetation biomass are beneficial for slope ecological monitoring work. Based on using spectral features to calculate vegetation indices as input variables for biomass inversion models, this study introduces texture features and vegetation probability variables and constructs a biomass inversion model for herbaceous plants on slopes. The research results can provide a reference for unmanned aerial vehicle remote-sensing studies on the biomass inversion of herbaceous plants on slope restoration sites.

The vegetation probability and vegetation index are both indicators obtained based on spectral features. In terms of spectral resolution, the UAV images include only three bands (RGB), which limits their ability to accurately produce biomass inversions. In contrast, Sentinel-2A data provide 10 spectral bands, including near-infrared and red-edge bands, with resolutions of 10 m and 20 m. This variety enables the differentiation of similar features. Data fusion of these two sources results in images with the same spatial resolution as the UAV images but with enhanced multispectral bands. This fusion is highly valuable for ecological monitoring and agricultural applications [26,28,61]. Although the fused images exhibit high resolution, detailed texture, and multispectral information, further quantitative verification may still be required. In this study, when the biomass inversion model was constructed with spectral features as the independent variables, the R² of the model reached 0.763, indicating that the image spectrum, especially the red, red-edge, and near-infrared bands, played a unique role in slope vegetation identification and biomass monitoring. This result was consistent with previous studies [62,63,64]. For example, CLARK et al.‘s study showed that the red-edge band was significantly correlated with vegetation chlorophyll, nitrogen, leaf area index, and other structural characteristics [31]. When the texture feature is used as the independent variable to construct the inversion model, the R² is only 0.687, indicating that the inversion of biomass by texture features is much lower than that of spectral features. The reason for this may be that texture information is easily affected by scaling and viewpoint changes, resulting in the excessive redundancy of extracted texture information and difficulty in distinguishing effective information from noise [26]. However, when the spectral features and texture features are used together as independent variables to build the model, the R² reaches 0.795, which is improved compared with model 1. It can be seen that the addition of the texture features provides the model with more information about the spatial distribution and structure of the slope vegetation, improves the influence of vegetation index saturation, and improves the performance and robustness of the model estimation [65,66,67]. However, the small study area and the medium spatial resolution of the Sentinel-2A images used in this research may limit the ability to capture detailed features of slope vegetation within the study area. This limitation could affect sample variability and fail to reflect the true diversity of slope vegetation. As a result, the biomass estimation model might not accurately capture ground features, potentially impacting the accuracy of the biomass estimates. To address this issue, future research could consider using high-resolution hyperspectral remote-sensing images combined with field survey data to enhance the detailed understanding of vegetation characteristics and improve the accuracy of biomass estimation.

Biomass is an indicator that reflects vegetation growth. There are various other growth indicators, such as leaf chlorophyll content and vegetation water content, that can also reflect vegetation growth on highway slopes. In the future, it will be possible to integrate multiple indicators to achieve vegetation monitoring on highway slopes. In addition, in current biomass inversion research, many scholars have adopted machine-learning methods [68,69,70,71]. For example, Fan et al. used four machine-learning algorithms, including the Cubist model, GBRT model, RF model, and XGBoost model, to estimate aboveground biomass in the grasslands of the Qinghai–Tibet Plateau [70]; Ge et al. used random forest, support vector machine, an artificial neural network, and extreme machine-learning algorithms to build aboveground biomass inversion models for grasslands in northern China [71]. Although machine-learning algorithms are efficient, they require a large amount of high-quality biomass data, and overfitting may occur when training machine-learning models. There has been no research on the use of machine-learning algorithms in the biomass inversion of vegetation on highway slopes. With more actual biomass data available, it may be feasible to apply machine-learning algorithms to construct biomass inversion models for vegetation on highway slopes.

6. Conclusions

In this study, we used multi-source remote-sensing images to extract slope vegetation features from three aspects—vegetation probability, vegetation index, and vegetation texture features—using deep-learning methods. A slope vegetation biomass inversion model was constructed, with an adjusted R² = 0.795, indicating an excellent regression effect, with p < 0.01 in the F test of the regression equation, indicating that the model is statistically significant. Based on the grid superposition calculation using the regression model, the obtained remote-sensing inversion results for the slope vegetation biomass were generally consistent with the spatial distribution of real biomass, proving that the modeling effect was good and could accurately predict slope vegetation biomass. The test results show that the biomass inversion model has high accuracy and strong applicability.

This study combined multi-source remote-sensing data from multiple platforms, such as satellites and UAV, using the Gram–Schmidt method to fuse Sentinel-2A and UVA images. We extracted spectral features from the fused images, such as using deep learning to extract the vegetation probability distribution and using band operations to calculate vegetation indices. Meanwhile, a gray-level co-occurrence matrix was used to extract the corresponding texture features. Based on biomass survey data, a slope vegetation biomass inversion model was constructed using the multiple regression method, which provided data and methodological support for the simulation of the carbon budget process of the slope restoration ecosystem.

In the future, by integrating remote-sensing data from a broader range of sources and acquiring higher-precision multispectral data, we will be able to conduct comparative analyses using various regression models and machine-learning techniques in more experimental cases. This approach is expected to yield a more accurate biomass inversion model and enhance our understanding of slope vegetation biomass. Improved estimation accuracy will highlight the advantages of our reference methodology and demonstrate its applicability for assessing highway slope vegetation biomass, monitoring vegetation quantitatively, and supporting ecological restoration efforts.