Assessment of the Impact of Coastal Wetland Saltmarsh Vegetation Types on Aboveground Biomass Inversion

Abstract

Coastal wetlands play an important carbon sequestration role in China’s “carbon peaking” and “carbon neutrality” goals. Monitoring aboveground biomass (AGB) is crucial for wetland management. Satellite remote sensing enables efficient retrieval of AGB. However, a variety of statistical models can be used for biomass inversion, depending on factors such as the vegetation type and inversion method. In this study, Landsat 8 Operational Land Imager (OLI) images were preprocessed in the study area through radiation calibration and atmospheric correction for modeling. In terms of model selection, 13 different models, including the univariate regression model, multiple regression model, and machine learning regression model, were compared in terms of their accuracy in estimating the biomass of various wetland vegetation types under their respective optimal parameters. The findings revealed that: (1) the regression models varied across vegetation types, with the accuracy of the biomass estimates decreasing in the order of Scirpus spp. > Spartina alterniflora > Phragmites australis; (2) overall modeling, without distinguishing vegetation types, addressed the challenges of limited samples availability and sampling difficulty. Among them, the random forest regression model outperformed the others in estimating wet and dry AGB with R² values of 0.806 and 0.839, respectively. (3) Comparatively, individual modeling of vegetation types can better reflect the biomass of each wetland vegetation type, especially the dry AGB of Scirpus spp., whose R² and RMSE values increased by 0.248 and 11.470 g/m², respectively. This study evaluates the impact of coastal saltmarsh vegetation types on biomass estimation, providing insights into biomass dynamics and valuable support for wetland conservation and restoration, with potential contributions to global habitat assessment models and international policies like the 30x30 Conservation Agenda.

1. Introduction

Coastal wetland vegetation biomass carbon is a significant component of carbon storage in coastal saltmarshes [1,2,3]. Through the process of photosynthesis, wetland vegetation can convert inorganic carbon in wetland ecosystems into organic carbon, which plays a crucial role in the carbon cycle of saltmarsh wetlands in China [4,5]. However, there are variations and differences in the biomass carbon of coastal wetland vegetation due to factors such as vegetation type, geographical location, and nutrient conditions [6,7,8]. The utilization of remote sensing images can not only effectively compensate for the limitations of accessibility and difficulties in obtaining wetland vegetation biomass samples but also integrate images for analyzing the spatial and temporal variations in biomass [9,10,11].

Remote sensing can be used to estimate the distribution and growth of wetland vegetation and other resources. A statistical model can be established between the spectral data of remote sensing images and the aboveground biomass (AGB) of vegetation, and the biomass can be quickly estimated over a large area. This method solves the problems of poor accessibility and safety concerns associated with ground surveys and the sampling of wetland vegetation biomass [12,13]. Remote sensing is widely used in the fields of vegetation resource investigation, dynamic monitoring, and biomass estimation [14,15,16]. However, unlike the remote sensing inversion of aboveground biomass in forests and farmland, wetland vegetation is affected by complex environmental factors, such as differences in underlying surface conditions [17,18]. In particular, in the Chongming Island area, wetland vegetation is significantly affected by the law of natural differentiation [10]. The east is affected mainly by seawater, whereas the west is affected mainly by the Yangtze River, resulting in high heterogeneity in wetland vegetation growth and complex spectral response processes [19]. In addition, different plants differ in their spectral, absorption, and reflection characteristics. Distinguishing wetland vegetation types for biomass remote sensing inversion may reduce the impact of spectral response differences between different vegetation types to a certain extent and improve the estimation accuracy of aboveground biomass [20]. However, it is still unclear how the biomass remote sensing inversion of multiple wetland vegetation types in the same area differs and what specific impacts distinguishing vegetation type modeling and non-distinguishing vegetation type modeling will have on the wetland vegetation biomass inversion process.

In the selection of vegetation biomass inversion models, traditional statistical models such as univariate and multiple linear regressions have been used [21]. With the development of machine learning, nonparametric machine learning methods such as the neural network (NN) regression model, support vector regression (SVR) model, and random forest (RF) regression model have shown great potential for the inversion of AGB from wetland vegetation in recent years [22,23,24,25,26]. Gao et al. [27] utilized the RF model along with the hyperspectral verification method of plots to increase the accuracy of Moderate Resolution Imaging Spectroradiometer (MODIS) images for estimating the AGB and coverage of alpine grasslands on the Qinghai–Tibet Plateau. Pecina et al. [28] used a high-resolution dataset acquired by an unmanned aerial vehicle (UAV) platform combined with the RF algorithm to estimate the AGB of coastal meadows in Estonia. Lu et al. [29] proposed a solution that combined UAV and Sentinel-2 data to successfully estimate and map the AGB of Phragmites australis in the Nandagang Wetland Reserve in Hebei Province, China, effectively solving the problem of insufficient AGB samples due to the difficulty of accessing wetlands. Prakash et al. [30] introduced a robust methodological protocol, which employs machine learning models alongside field measurements and multisensor SAR data, to achieve highly accurate and low uncertainty estimates of AGB in mangrove forests (R² = 0.93). Statistical methods are the main methods for biomass estimation, but the choice of model depends on factors such as the study area and vegetation type. Previous research has noted variations in biomass among different wetland vegetation types [31,32,33]. However, it is unclear how modeling these types separately versus an overall model impacts biomass estimation. Currently, there is a gap in systematic research on coastal wetland biomass inversion, and the distribution pattern of wetland vegetation biomass on Chongming Island has not been determined.

From the perspective of image data sources, Landsat, Sentinel-2, MODIS, and other data have been widely used when optical satellite images are used to invert vegetation aboveground biomass [34,35,36]. Among them, Landsat images offer higher spatial resolution than MODIS images and are better suited for wetland vegetation communities with limited distribution areas [37,38,39]. While Sentinel-2 images provide even higher spatial resolution, their use in large-scale studies demands a greater volume of image data, involves significant computational effort, and is constrained by the lack of earlier historical records [40,41,42]. In contrast, Landsat data are widely preferred because of their easy accessibility and their broad and historical perspective on vegetation [43,44]. Statistical methods are the main methods used for biomass estimation via optical remote sensing images, but the choice of model depends on factors such as the study area and vegetation type.

In addition to optical data, synthetic aperture radar (SAR) and light detection and ranging (Lidar) data can enhance the accuracy of AGB estimation [45,46,47]. Zhu et al. [48] integrated SAR and UAV data to estimate the AGB of the largest artificially planted mangrove forest in China on Qi’ao Island, which could provide more detailed and accurate information for AGB estimation. Salum et al. [49] combined Lidar data and improved the inversion accuracy of mangrove biomass data. Lucas et al. [50] retrieved mangrove distribution, age, structure, and biomass parameters by combining optical and SAR data in a Malaysian reserve. While SAR has potential for mangrove biomass retrieval [51,52,53], coastal herbaceous wetland vegetation, such as P. australis and Spartina alterniflora, can be challenging to retrieve via SAR data because of their low height and weaker radar signal returns. On the other hand, Lidar has high accuracy in retrieving vegetation height but has limited availability [33,54,55]. Furthermore, obtaining ground lidar data in coastal wetlands is difficult, and large-scale monitoring via UAV data is currently challenging [56,57]. Considering the cost-effectiveness and workload involved in data processing, Landsat imagery is the superior choice for the inversion of aboveground biomass in coastal wetland vegetation.

In this study, various wetland vegetation AGB inversion models, such as univariate regression models, multiple linear regression models, and machine learning regression models, were compared based on Landsat images to explore the most suitable AGB retrieval method for each wetland vegetation type. Additionally, the impact of classification modeling on the AGB inversion of wetland vegetation was quantitatively analyzed to examine the utility of different models, spectral bands, species classification, and both parametric and nonparametric statistical methods for biomass prediction. By comprehensively comparing various vegetation types and inversion methods, this research quantitatively reveals the influence of different saltmarsh vegetation types on the remote sensing inversion of wetland AGB. Such an approach is essential for assessing wetland AGB over large areas, significantly enhancing our understanding of biomass dynamics in coastal ecosystems and providing more reliable data support for ecological monitoring and resource management.

2. Materials and Methods

2.1. Study Area

Chongming Island, located in the core area of the Yangtze River Delta, is the most well-developed tidal coastal wetland in the Yangtze River estuary. It is located northeast of Shanghai, between 31°27′00″ and 31°51′15″N and 121°09′30″ and 121°54′00″E. Surrounding three rivers and facing the sea to the east, it is influenced by both the Yangtze River and the East China Sea, and it is the world’s largest estuarine alluvial island [10]. The different types of coastal wetlands form discontinuous wetland zones, which are key areas for the biodiversity of China’s coastal region and provide a suitable growth environment for various animals and plants [19,58]. It is 80 km long from east to west and 18 km wide at its widest point from north to south, with a total area of 1411 km². Chongming Island has beautiful natural scenery and important wetland and forest resources, such as Dongping National Forest Park, Dongtan Bird Reserve, and Xisha Wetland Park. Its development goal is to construct a world-class ecological island with management and demonstration functions.

Chongming Island currently has nearly 100 species of plants with rich biodiversity. The coastal wetland vegetation of Chongming Island belongs to the North China Plain and to the shallow-water plant wetlands in the middle and lower reaches of the Yangtze River. The coastal wetland vegetation on Chongming Island is dominated by herbaceous vegetation, including P. australis, Scirpus mariqueter, Scirpus triqueter, S. alterniflora, Miscanthus saccharifleus, Juncellus serotinus, Suaeda glauca (Bunge) Bunge, Solidago decurrens Lour., and more [59]. These species play important roles in maintaining the ecological balance of wetland ecosystems and provide habitats for various wildlife species [60,61,62]. Among them, S. alterniflora, P. australis, and Scirpus spp. are the three most important types of wetland vegetation on Chongming Island. The distribution area of other herbaceous vegetation is small and scattered, making it difficult to dominate the communities.

This study investigated the AGB of three types of wetland vegetation on Chongming Island: S. alterniflora, P. australis, and Scirpus spp. (Figure 1). Among them, Scirpus spp. are considered pioneer species of coastal wetland vegetation due to their ability to colonize and establish in areas of bare mudflats, while P. australis is a widely distributed coastal vegetation type native to Chongming Island. In contrast, S. alterniflora was originally introduced as an invasive alien species to promote siltation and land formation but has now become a dominant and problematic species in many coastal wetlands in China [63].

2.2. Data Resources

This study employed field sampling methods for investigation and sample collection. First, sample sites were selected based on uniform vegetation types, consistent growth, and continuous distributions, with each site covering an area greater than 30 × 30 m². This approach aimed to minimize the impact of local heterogeneity and capture the spatial variability of AGB across different wetland vegetation types [19,64]. Next, biomass sampling was conducted within the selected plots. At each sample site, three 1 × 1 m² quadrats were set up at the diagonal corners and center, and the AGB of the wetland vegetation in each quadrat was collected. All the vegetation in the sample quadrat was harvested, after which the plants were numbered separately, packed into bags, and brought back to the laboratory of the Chongming Ecological Research Institute of East China Normal University. The coordinates of the sample site, vegetation coverage, surrounding environment of the sample plot, and vegetation growth status were recorded via a handheld global positioning system (GPS) and cameras (Figure 2a,b). To obtain accurate biomass data, the vegetation collected in the field was immediately weighed and packaged, and its net wet weight was recorded with an accuracy of 0.01 g. Then, the vegetation was oven-dried at 105 °C for 30 minutes for fixation and further dried at 85 °C for more than 48 h until it reached a constant weight. Finally, the dried vegetation was weighed to obtain the dry weight of the vegetation (Figure 2c–e). The average biomass of the three sample quadrats at each sample site represented the biomass per unit area.

The biomass data collection process for the sample plots in the coastal wetland vegetation communities on Chongming Island was carried out from August to September 2021 (Figure S1). The data collection included 124 vegetation community sites covering the entire coastal wetland area, comprising 50 P. australis sites, 40 S. alterniflora sites, and 34 Scirpus spp. sites. The selected season for the data collection corresponds to the vigorous growth period of herbaceous vegetation and correlates well with remote sensing images. Therefore, ground data and corresponding images taken during this season were used for biomass inversion modeling.

Considering the spatial resolution of the images and the computational demands, this study selected the surface reflectance (SR) product of Landsat 8 OLI (path: 118 and row: 38) images covering Chongming Island. The Landsat 8 surface reflectance (SR) product underwent atmospheric correction using the Land Surface Reflectance Code (LaSRC) [65]. Foga et al. [66] compared various cloud removal algorithms and reported that, compared with other methods, the CFMask method, which is based on prior knowledge of physical phenomena, operates without geographical limitations and effectively removes cloud and shadow influences. Landsat 8 SR products utilize the CFMask algorithm to eliminate noise from clouds and shadows based on pixel quality and radiative saturation attributes [67]. Given the susceptibility of the Chongming Island area to cloudy and rainy weather, effective images are limited. Moreover, the AGB of wetland vegetation shows significant seasonal fluctuations but limited annual variability. Therefore, the study selected Landsat data from August to September for the years 2020 to 2022, along with ground biomass data collected from 124 vegetation community sites across Chongming Island [19]. Ultimately, the Landsat image from August 16, 2020, was chosen as the remote sensing data source for ground biomass inversion. The spatial resolution of the Landsat image is 30 m, with spectral bands including blue, green, red, near infrared, and short-wave infrared bands at 1640 nm and 2130 nm. Six original spectral and ten spectral indices of the Landsat image were extracted for use in the model (

Table 1. Spectral indices calculation formula used in biomass inversion.

Spectral Index	Calculation Formula
NDVI	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}}}$
NDWI1640	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1640}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1640}}$
NDWI2130	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2130}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2130}}$
SAVI	${\displaystyle \frac{1.5\times (\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d})}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}+0.5}}$
RVI	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}\mathrm{e}\mathrm{d}}}$
DVI	$\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}$
PVI	${\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{a}\times \mathrm{R}\mathrm{e}\mathrm{d}-\mathrm{b}}{\sqrt{1+{\mathrm{a}}^2}}}$ $\mathrm{a}: \mathrm{S}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e} \mathrm{o}\mathrm{f} \mathrm{t}\mathrm{h}\mathrm{e} \mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l} \mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e};\mathrm{b}: \mathrm{G}\mathrm{r}\mathrm{a}\mathrm{d}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t} \mathrm{o}\mathrm{f}\mathrm{t} \mathrm{h}\mathrm{e} \mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l} \mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}$
EVI	${\displaystyle \frac{2.5\times (\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d})}{\mathrm{N}\mathrm{I}\mathrm{R}+6\times \mathrm{R}\mathrm{e}\mathrm{d}-7.5\times \mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}+1}}$
NDSVI	${\displaystyle \frac{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1640-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1640+\mathrm{R}\mathrm{e}\mathrm{d}}}$
NDRI	${\displaystyle \frac{\mathrm{R}\mathrm{e}\mathrm{d}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2130}{\mathrm{R}\mathrm{e}\mathrm{d}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2130}}$

). The normalized difference vegetation index (NDVI), widely used for monitoring vegetation health and density, is sensitive to vegetation presence [68]. The soil-adjusted vegetation index (SAVI) minimizes soil interference, enabling effective differentiation of vegetation from other land cover types [69]. The ratio vegetation index (RVI) is highly sensitive to green vegetation, providing a simple measure of vegetation cover [70]. The normalized difference water indices at 1640 nm (NDWI1640) and 2130 nm (NDWI2130) effectively detect leaf water content, reflecting moisture variations, with distinct spectral bands and varying sensitivities to vegetation types and environmental conditions [71,72]. The enhanced vegetation index (EVI) is particularly responsive to canopy structure changes, making it suitable for monitoring biomass in dense vegetation [73]. The difference vegetation index (DVI) is well correlated with biomass in low-coverage areas and is commonly used for ecological monitoring [74]. The perpendicular vegetation index (PVI) captures vegetation structural information, useful for estimating biophysical parameters [75,76]. The normalized difference senescent vegetation index (NDSVI) [77,78,79] and normalized difference residue index (NDRI) [80,81] help mitigate the effects of vegetation senescence, providing insight into late-stage growth phases. These indices collectively enable precise monitoring of vegetation dynamics, with each serving a distinct purpose to enhance the accuracy of biomass estimation.

2.3. Methods

This study used biomass survey data and satellite images to extract spectral information for corresponding locations. The SR product of the Landsat 8 OLI was selected for this study. The preprocessing steps for this study included the following: cloud and shadow removal using the CFMask algorithm to ensure data clarity; and data normalization, where all spectral variables were scaled using min-max normalization. This step ensured comparability and prevented potential bias in the machine learning models due to differing magnitudes of the variables. Following these preprocessing steps, feature selection was performed to identify spectral bands and indices that were significantly correlated with biomass while remaining mutually independent, thus ensuring the inclusion of only relevant and non-redundant features. A test of significance was conducted on the six original spectral bands and ten spectral indices, and the spectral variables that were both significantly correlated with biomass and independent of each other were selected for biomass estimation. To comprehensively reveal the impact of vegetation type on the inversion of wetland AGB, this study constructed optimal biomass inversion models for different vegetation types and performed an overall biomass inversion without distinguishing between vegetation types. This approach allows for a thorough consideration of the differences in the radiation transfer processes among various vegetation types, enabling a quantitative analysis of their effects on AGB inversion. Thirteen regression models, including seven univariate models, stepwise multiple linear regression models, and five machine learning models, were compared for P. australis, S. alterniflora, Scirpus spp., and saltmarsh wetland vegetation. The optimal AGB model was chosen for each wetland vegetation type, considering the distinction between wetland vegetation types. The impacts of individual modeling of vegetation types and overall modeling without distinguishing vegetation types on biomass estimation were assessed (Figure 3). All analyses were conducted using MATLAB 2019b software [82,83,84].

2.3.1. Univariate Regression Model

For correlation analysis, the Pearson coefficient was used to describe the correlation between the spectral data and AGB. The Pearson coefficient reflects the correlation direction and degree between two sets of variables, and the calculation formula is as follows:

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

(1)

In Formula (1), $r$ is the Pearson correlation coefficient, which measures the linear correlation between two variables, $x$ and $y$. $x_i$ and $y_i$ represent sample values in each group of variables, and $\overline{x} \mathrm{a}\mathrm{n}\mathrm{d} \overline{y}$ represent the mean values of each group of variables. n represents the number of samples. The positive and negative values of $r$, respectively, represent whether the two sets of variables are positively or negatively correlated. The range of r is between −1 and 1, and the closer $\left|r\right|$ is to 1, the higher the correlation between the variables; a value of 0 indicates no correlation [85,86,87].

Based on the correlation analysis, spectral data that were significantly and independently correlated with biomass were selected for regression modeling. Univariate models provide a straightforward approach to analyzing the direct relationships between biomass and various variable characteristics, enabling the rapid identification of appropriate fitting models [84]. Therefore, this study initially selected seven commonly used functional model forms, including linear regression and polynomial regression, for univariate regression inversion. To analyze the relationships between AGB and the spectral information of P. australis, S. alterniflora, Scirpus spp., and saltmarsh vegetation overall, seven univariate regression models were selected for the inversion of wetland vegetation AGB, including linear models, quadratic models, cubic models, power function models, exponential models, logarithmic models, and S-curve models (

Table 2. Calculation formula of univariate regression models.

Regression Model Name	Formula
Linear model	$y=a+bx$
Quadratic model	$y=a+bx+c{x}^2$
Cubic model	$y=a+bx+c{x}^2+d{x}^3$
Power function model	${y=ax}^b$
Exponential model	${y=ae}^{bx}$
Logarithmic model	$y=a\mathit{ln}x+b$
S-curve model	${y=e}^{(a+{\displaystyle \frac{b}{x}})}$

).

2.3.2. Multiple Linear Regression Model

The stepwise regression model combines the forward selection and backward elimination methods, iteratively adding or removing variables based on their significance in explaining the variation of the dependent variable [88,89]. The stepwise regression method is particularly useful when dealing with large datasets with many predictors, as it can help to automatically select the most important variables for the model and exclude any variables that are irrelevant or redundant, reducing the risk of overfitting and improving the accuracy and efficiency of the model [90]. In this study, a total of 16 original spectral and spectral features were selected. Thus, the stepwise multiple linear regression method was used to establish a multivariate linear regression model between the spectral data and AGB.

2.3.3. Machine Learning Regression Models

Machine learning regression models are effective at performing nonparametric estimations and can efficiently handle high-dimensional feature data. Given the multitude of spectral features in this study, it is challenging to parameterize AGB estimation using a single spectral feature. Therefore, five machine learning regression models, namely, SVR, Gaussian process regression (GPR), RF regression, NN regression, and decision tree (DT) regression, were selected to estimate the AGB of wetland vegetation.

The SVR model can use kernel functions to project training data onto a new high-dimensional space (hyperplane) and adjust the prediction function through flexible boundary lines. The SVR model, which has multiple kernel functions and strong generalizability, is particularly useful for handling high-dimensional data and solving nonlinear regression problems. However, this approach is sensitive to missing data, and there is no uniform standard for kernel function selection. In this study, the initial values for key parameters were set as follows: the penalty coefficient C = 1.0, the kernel parameter γ = 0.1, and the slack variable ϵ = 0.1. Multiple kernel functions, including linear, polynomial, and Gaussian kernel functions, were evaluated to identify the most appropriate function for each wetland vegetation type. This comparison was conducted to enhance the accuracy and reliability of the SVR model.

The DT regression model is a popular and interpretable machine learning method that uses tree-based models to handle regression problems. Multiple decision rules branch outward, with leaf nodes that output predicted values. The use of the mean squared error for feature selection and tree pruning helps prevent overfitting and improve model generalizability. The DT regression model can provide a structured interpretation of the results, handling noise robustly, but it is at risk of overfitting the training data. In this study, the DT regression model was initialized with min samples split set to 2 and min samples leaf set to 1. These parameter settings allowed the model to grow fully, maximizing its capacity to capture data patterns without prematurely limiting tree depth, thereby ensuring comprehensive learning.

The RF regression model uses ensemble learning by combining multiple decision tree models for regression analysis, using the average of multiple decision trees to represent the result of RF regression. When the number of RF trees is sufficient, overfitting problems in regression analysis can be effectively avoided. The RF regression model is not sensitive to missing data and can sort the relative importance of input variables, making the results interpretable. In the RF model, the impact of varying the number of trees to specifically 100, 300, 500, and 1000 on prediction accuracy was systematically evaluated. To balance reliability and computational efficiency, this study used 500 trees for the RF regression model. However, it should be noted that the RF regression model may not fully leverage its advantages when the number of samples is small but rather when the number of features is large.

The NN regression model is inspired by biological neural networks that can learn data features automatically through instance data training without given rules, achieving regression prediction. The NN model is effective at handling nonlinear regression problems involving multiple variables and has been widely applied to estimate vegetation biomass. The complexity of the NN is determined by the number of layers, which directly affects the accuracy of the prediction. In this study, the neural network started with two layers, and additional layers were added iteratively up to a maximum of nine layers. The best-fitting configuration was then selected to construct the model.

GPR flexibly chooses suitable kernel functions for prediction based on data, and it provides probabilistic predictions, allowing for the calculation of empirical confidence intervals. GPR fits a distribution of random variable functions, assigning probability values to multiple potential fitting functions, with the mean of the probability distribution representing the most likely representation of the data. However, its applicability in high-dimensional spaces is limited. In this study, parameters such as the length scale (l = 1) of the RBF kernel and the smoothing parameter of the Matern kernel (ν = 1.5) were adjusted to optimize the model. The modeling performances of linear, square exponential, and other kernel functions were compared, and suitable kernel functions were flexibly selected for the GPR model.

2.3.4. Model Accuracy Evaluation

In this study, the 10-fold cross-validation method was employed to validate the inversion models. Specifically, 90% of the data was iteratively used for training, while the remaining 10% was reserved for validation [91]. This approach was chosen to enhance model reliability by reducing errors and minimizing the risk of overfitting, particularly in the context of a limited sample size [61]. All ground sample data were divided into 10 parts and used iteratively for model building and validation. The use of cross-validation allows for the full utilization of limited data and can help reduce overfitting of the model [91,92]. Because obtaining actual measurement data for wetland vegetation biomass is difficult and the quantity of data is limited, cross-validation was deemed suitable for this study. To evaluate the inversion accuracy of the model, performance evaluation metrics such as the mean absolute error (MAE) (Formula (2)), root mean square error (RMSE) (Formula (3)), coefficient of determination ($R^2$) (Formula (4)), and accuracy (Formula (5)) were chosen. A smaller MAE and RMSE indicate a smaller inversion error, while a higher $R^2$ and accuracy indicate a higher accuracy of the inversion model.

$$MAE={\displaystyle \frac{\sum _{i=1}^n\left|A_i-{\widehat{A}}_i\right|}{n}}$$

(2)

$$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(A_i-{\widehat{A}}_i)}^2}{n}}}$$

(3)

$$R^2={\displaystyle \frac{{\sum }_i^n{({\widehat{A}}_i-\overline{A_i})}^2}{{\sum }_i^n{(A_i-\overline{A_i})}^2}}=1-{\displaystyle \frac{{\sum }_i^n{(A_i-{\widehat{A}}_i)}^2}{{\sum }_i^n{(A_i-\overline{A_i})}^2}}$$

(4)

$$Accuracy=\left(1-{\displaystyle \frac{RMSE}{\overline{A_i}}}\right)\times 100\mathrm{\%}$$

(5)

Note: $A_i$ represents the measured AGB of the i-th sample site, ${\widehat{A}}_i$ represents the predicted biomass of the i-th sample site, n represents the total number of sample sites, and $\overline{A_i}$ represents the average value of measured biomass across all sample sites.

3. Results

This study employed 13 models, including univariate regression, multiple linear regression, and machine learning regression models, to estimate the aboveground biomass of vegetation through remote sensing. These models were specifically applied to different vegetation types, such as P. australis, S. alterniflora, and Scirpus spp., and the overall saltmarsh vegetation, which represented all of the species combined. The optimal models for each vegetation type were then integrated to estimate the aboveground biomass of the study area. This was compared with the results from the overall modeling approach to determine whether classification-based modeling enhanced the accuracy of remote sensing biomass estimation for coastal wetland vegetation. This study also aimed to quantify the degree of accuracy improvement and its impact across various vegetation types in coastal wetlands.

3.1. Single-Species AGB Retrieval

This study analyzed the relationships between spectral variables and AGB for different types of wetland vegetation. For P. australis, most spectral variables were significantly correlated with biomass, except for those in the blue band (

Table 3. Correlations between the spectra and AGB of P. australis, S. alterniflora, and Scirpus spp.

Spectral Information	P. australis Wet Biomass	P. australis Dry Biomass	S. alterniflora Wet Biomass	S. alterniflora Dry Biomass	Scirpus spp. Wet Biomass	Scirpus spp. Dry Biomass
Blue band	−0.133	−0.026	−0.261	−0.229	−0.598 **	−0.569 **
Green band	−0.499 **	−0.485 **	0.195	0.186	−0.668 **	−0.643 **
Red band	−0.751 **	−0.696 **	−0.211	−0.196	−0.745 **	−0.719 **
Near-infrared band	0.502 **	0.453 **	0.522 **	0.487 **	−0.086	−0.071
SWIR1640	−0.496 **	−0.361 **	0.099	0.124	−0.523 **	−0.492 **
SWIR2130	−0.688 **	−0.631 **	−0.344 *	−0.28	−0.587 **	−0.548 **
NDVI	0.737 **	0.682 **	0.648 **	0.598 **	0.825 **	0.809 **
NDWI2130	0.692 **	0.634 **	0.632 **	0.561 **	0.742 **	0.705 **
NDWI1640	0.657 **	0.552 **	0.613 **	0.544 **	0.745 **	0.714 **
SAVI	0.650 **	0.596 **	0.599 **	0.556 **	0.737 **	0.723 **
RVI	0.702 **	0.648 **	0.690 **	0.643 **	0.833 **	0.830 **
DVI	0.592 **	0.538 **	0.554 **	0.517 **	0.609 **	0.599 **
PVI	0.614 **	0.559 **	0.565 **	0.527 **	0.724 **	0.709 **
EVI	0.601 **	0.578 **	0.585 **	0.545 **	0.722 **	0.712 **
NDSVI	0.703 **	0.699 **	0.536 **	0.540 **	0.648 **	0.660 **
NDRI	−0.417 **	−0.409 **	0.176	0.084	−0.123	−0.169

Note: * indicates significant correlation at the p < 0.05 level; ** indicates a significant correlation at the p < 0.01 level.

). The red band showed the highest correlation with the wet biomass of P. australis, while the NDVI and RVI showed positive correlations. The NDSVI, red band, and NDVI were strongly correlated with dry biomass. Based on these findings, 15 noncollinear spectral variables were selected for modeling P. australis biomass. For S. alterniflora, the correlations between the spectral variables and biomass were consistent for both dry and wet biomass, except for the SWIR2130 band, which was significantly correlated with only wet biomass. Therefore, 11 variables were chosen for modeling the wet biomass of S. alterniflora, and the same 10 variables were used for modeling dry biomass, excluding SWIR2130. For Scirpus spp., all spectral variables, except for the near-infrared band and NDRI, were significantly correlated with both wet and dry biomass. Therefore, 14 spectral variables were chosen for the biomass inversion of Scirpus spp., excluding the near-infrared band and NDRI. The structural differences in vegetation canopies among species lead to variations in spectral responses between species. Additionally, different species types exhibit varying optical sensitivities across different models.

3.1.1. P. australis AGB Inversion

This study evaluated different regression models for predicting the AGB of P. australis. When 105 univariate models for wet and dry AGB were compared separately (Table S1), the highest accuracy was achieved when the RVI was used as a cubic regression model. However, the accuracy of dry biomass prediction was generally lower than that of wet biomass prediction when univariate models were used for P. australis. For wet biomass estimation of P. australis, the stepwise multiple linear regression model was consistent with the univariate linear model. In contrast, for dry biomass estimation, the stepwise multiple linear regression model selected the NDSVI, with an R² value of only 0.340. The performance of the stepwise multiple linear regression model was affected by the samples, leading to inferior performance compared with that of the optimal univariate linear regression model. Five machine learning models were compared for their ability to predict the AGB of P. australis. The RF model performed the best, utilizing spectral data effectively for biomass estimation (

Table 4. Remote sensing inversion accuracy of the AGB of P. australis based on machine learning models.

Type	Model Name	Spectral Information	MAE	R2	RMSE	Accuracy (%)
Wet biomass	RF regression	15 spectral variables	296.172	0.541	382.762	84.720
	GPR (with squared exponential kernel function)		307.607	0.506	396.674	84.165
	SVR (with linear kernel function)		383.840	0.274	480.716	80.810
	NN regression (with 3 layers)		395.702	0.125	519.887	79.246
	DT regression		335.444	0.445	414.256	83.463
Dry biomass	RF regression	15 spectral variables	185.858	0.462	227.484	82.160
	GPR (with squared exponential kernel function)		195.058	0.413	240.149	81.167
	SVR (with linear kernel function)		226.540	0.287	267.102	79.053
	NN regression (with 3 layers)		237.132	0.089	291.665	77.127
	DT regression		227.022	0.206	268.170	78.969

Note: The parameters of each model in the table have been optimized. For example, SVR has chosen a linear function as its kernel function, and GPR has chosen a squared exponential kernel function.

). The GPR model had slightly lower accuracy but was still viable. The DT model worked well for wet biomass but not for dry biomass. The SVR and NN models struggled to utilize spectral data effectively. Overall, the machine learning models were slightly more accurate for wet biomass than for dry biomass, which was consistent with the univariate and multiple linear regression models. The RF regression model outperformed the univariate regression models and stepwise multiple linear regression models in predicting the AGB of P. australis. However, limited measured data pose challenges for fully harnessing the advantages of some machine learning methods in nonparametric regression.

3.1.2. S. alterniflora AGB Inversion

The RVI outperformed the other spectral variables in the estimation of AGB under S. alterniflora. Among the seven univariate regression models tested, the RVI was the optimal spectral variable for all of the models except for the S-curve model when wet biomass was estimated and for all the models when dry biomass was estimated (as shown in Table S2). The comparison revealed that the quadratic model using the RVI achieved the best performance in estimating both the wet and dry biomass of S. alterniflora, with a slightly greater accuracy in estimating wet biomass. The stepwise multiple linear regression model for estimating the AGB of S. alterniflora included the RVI and NDVI as selected spectral variables; these two spectral variables had the highest correlation with the wet and dry biomass of S. alterniflora, which was consistent with the correlation analysis results. The stepwise multiple linear regression model for dry biomass estimates could use information from both the RVI and the NDVI, resulting in greater accuracy than that of univariate linear regression. However, for wet biomass estimation, the accuracy was between the accuracy of the RVI and the accuracy of the NDVI univariate linear regression models. This is due to the susceptibility of both stepwise multiple linear regression and univariate linear regression models to the influence of sample data [93]. During cross-validation, stepwise multiple linear regression did not always select RVI and NDVI values, resulting in lower overall accuracy than that of the RVI univariate linear regression model.

After the five machine learning regression models were compared, the RF regression model was found to perform well in estimating both the wet and dry biomass of S. alterniflora, with higher accuracy for wet biomass than dry biomass. The GPR model with a squared exponential kernel function achieved the second highest accuracy. However, the ability of the SVR, NN, and DT models to estimate the biomass of S. alterniflora was limited (

Table 5. Remote sensing inversion accuracy of the AGB of S. alterniflora based on machine learning models.

Type	Model Name	Spectral Information	MAE	R2	RMSE	Accuracy (%)
Wet biomass	RF regression	11 spectral variables	699.687	0.491	810.830	79.773
	GPR (with squared exponential kernel function)		733.731	0.414	876.531	78.135
	SVR (with Gaussian kernel function)		786.275	0.309	938.623	76.586
	NN regression (with 3 layers)		930.903	0.164	1054.186	73.703
	DT regression		793.006	0.208	996.405	75.144
Dry biomass	RF regression	10 spectral variables	252.041	0.399	303.649	77.885
	GPR (with squared exponential kernel function)		263.273	0.305	321.373	76.594
	SVR (with Gaussian kernel function)		264.972	0.284	327.255	76.166
	NN regression (with 3 layers)		248.216	0.220	326.226	76.241
	DT regression		315.811	0.083	365.945	73.348

Note: the parameters of each model in the table have been optimized, such as the kernel functions of GPR and SVR and the number of layers in the NN.

). Although the RF regression model had a greater estimation accuracy for AGB than the stepwise regression model, it was lower than that of univariate models such as the RVI quadratic and cubic models. This is primarily because machine learning regression models struggle to learn the relationships among data when the amount of data is limited.

3.1.3. Scirpus spp. AGB Inversion

The analysis revealed that the RVI and NDVI were more effective than the other spectral variables for estimating the wet biomass of Scirpus spp. Among the 98 univariate regression models, the logarithmic model based on the RVI had the highest accuracy in estimating the wet biomass of Scirpus spp. (Table S3). When the dry biomass of Scirpus spp. was estimated, the cubic regression model based on the NDVI had greater accuracy than the other univariate models did. Additionally, the R² of the optimal model for wet biomass estimation was greater than that of the optimal model for dry biomass estimation. According to the stepwise regression model for the inversion of wet and dry biomass of the Scirpus spp., the RVI was selected from the 14 spectral variables. The inversion of dry biomass via stepwise regression was consistent with the use of the RVI for univariate linear regression. However, when inverting wet biomass, the spectral variable selected by the stepwise multiple linear regression model was the RVI, while the best spectral variable selected by the univariate linear regression model was the NDVI. The inconsistency between the stepwise multiple linear regression model and the univariate linear regression model was due to the different variables produced by tenfold cross-validation for each fold, and the RVI was selected more often than the NDVI in the stepwise multiple linear regression model.

When machine learning regression models are used for the wet and dry biomass inversion of Scirpus spp., both the RF and GPR methods can effectively invert AGB, and their inversion accuracy is better than that of the SVR, NN, and DT models. Among them, RF performed slightly better than GPR for wet biomass, while GPR performed better for dry biomass estimation (

Table 6. Remote sensing inversion accuracy of the AGB of Scirpus spp. based on machine learning models.

Type	Model Name	Spectral Information	MAE	R2	RMSE	Accuracy (%)
Wet biomass	RF regression	14 spectral variables	201.396	0.537	239.999	83.969
	GPR (with linear kernel function)		216.095	0.514	244.573	83.664
	SVR (with Gaussian kernel function)		226.082	0.393	276.066	81.560
	NN regression (with 4 layers)		243.090	0.324	288.226	80.748
	DT regression		237.267	0.301	290.76	80.579
Dry biomass	RF regression	14 spectral variables	40.822	0.512	47.722	83.517
	GPR (with linear kernel function)		37.013	0.555	43.888	84.841
	SVR (with Gaussian kernel function)		43.511	0.468	49.379	82.944
	NN regression (with 6 layers)		45.562	0.198	57.794	80.038
	DT regression		38.038	0.428	50.657	82.503

Note: the parameters of each model in the table have been optimized, such as the kernel functions of GPR and SVR and the number of layers in the NN.

). However, the overall estimation accuracy of machine learning regression models is lower than that of univariate regression models, mainly due to the limitation of the sample size of the measured data, which makes it difficult to exploit the advantages of machine learning data processing.

In summary, when distinguishing wetland vegetation types for the AGB inversion of a single species, the inversion accuracy of different wetland vegetation types is ranked as follows: Scirpus spp. > S. alterniflora > P. australis. Additionally, different wetland vegetation types were suitable for different regression models. For example, P. australis RF regression models for the inversion of aboveground wet and dry biomass perform better than other models; Scirpus spp. are suitable for univariate regression models, where the logarithmic model using the RVI is the most accurate for the inversion of wet biomass and the cubic model using the NDVI is the most accurate for the inversion of dry biomass; and S. alterniflora is suitable for an RF regression model for wet biomass and a univariate regression model for dry biomass, where the RVI quadratic model has the smallest inversion error. Thus, a single approach may not be suitable for all wetland vegetation types, and it is necessary to choose an appropriate regression model and spectral index for accurate biomass estimation.

3.2. Remote Sensing Inversion of the Total AGB of Saltmarsh Vegetation

In previous studies, appropriate regression models for AGB estimation were selected separately for each type of wetland vegetation, but machine learning regression models may not perform well when the number of site samples is low. In addition, site sampling in wetlands requires considerable manpower and resources, resulting in limited sample sizes for each type of vegetation. To address these issues, an overall modeling approach was used with all available data on saltmarsh vegetation. This method efficiently utilizes site data, avoids multiple analyses and modeling, and provides a simple way to estimate aboveground wet and dry biomass for wetland vegetation.

The analysis of the correlation between the remote sensing spectral data and the AGB of saltmarsh wetland vegetation on Chongming Island revealed that there was a significant correlation (p < 0.01) between the wet biomass and all 15 spectral variables, except for the NDRI. Among these variables, the RVI had the highest correlation with wet biomass (

Table 7. Correlation between saltmarsh vegetation spectral and AGB.

Spectral Information	Saltmarsh Vegetation Wet Biomass	Saltmarsh Vegetation Dry Biomass
Blue band	−0.386 **	−0.181 *
Green band	−0.445 **	−0.319 **
Red band	−0.675 **	−0.665 **
Near-infrared band	0.739 **	0.828 **
SWIR1640	−0.253 **	0.128
SWIR2130	−0.622 **	−0.460 **
NDVI	0.759 **	0.834 **
NDWI2130	0.807 **	0.776 **
NDWI1640	0.838 **	0.679 **
SAVI	0.784 **	0.870 **
RVI	0.871 **	0.836 **
DVI	0.783 **	0.858 **
PVI	0.791 **	0.862 **
EVI	0.766 **	0.875 **
NDSVI	0.618 **	0.844 **
NDRI	−0.093	−0.469 **

Note: “*” in the table represents significant correlation at the <0.05 level, and “**” represents significant correlation at the <0.01 level.

).The dry biomass was significantly correlated (p < 0.01) with all 14 spectral variables except for the SWIR1640 band and was significantly correlated (p < 0.05) with the blue band. The EVI had the highest correlation with dry biomass. Therefore, the 15 spectral variables that were significantly correlated with the aboveground wet and dry biomass of saltmarsh wetland vegetation were selected for analysis.

The RVI was identified as the optimal spectral variable for estimating the aboveground wet biomass in saltmarsh vegetation, and the highest accuracy was achieved with the cubic model (

Table 8. Remote sensing inversion accuracy of aboveground biomass of saltmarsh wetland vegetation based on univariate regression models.

Type	Model Name	Spectral Information	Formula	MAE	R2	RMSE	Accuracy (%)
Wet biomass	Linear	RVI	$y=397x+119.2$	456.014	0.741	593.880	78.116
	Quadratic	RVI	$y=32.33x^2-22.03x+1229$	455.662	0.748	576.578	78.753
	Cubic	RVI	$y=10.7x^3-178.4x^2+1208x-777.3$	442.556	0.760	561.985	79.291
	Power	RVI	${y=398.6x}^{1.016}$	477.304	0.728	606.588	77.647
	Exponential	RVI	${y=899.5e}^{0.155x}$	446.726	0.758	564.339	79.204
	Logarithmic	RVI	$y=2033\mathit{ln}x-875$	532.241	0.651	687.981	74.648
	S-curve	RVI	${y=e}^{(8.846-{\displaystyle \frac{5.694}{x}})}$	560.412	0.627	711.979	73.764
	Stepwise	RVI and NDSVI	$y=499.25\times RVI-3066.2\times NDSVI+958.09$	452.341	0.744	588.631	78.309
Dry biomass	Linear	EVI	$y=3082x-430.9$	200.738	0.720	266.490	74.289
	Quadratic	EVI	$y=-961.5x^2+3919x-587.4$	199.895	0.720	266.469	74.291
	Cubic	EVI	$y=-22000x^3+28710x^2-8401x+925$	182.374	0.738	257.272	75.178
	Power	EVI	${y=2988x}^{1.47}$	210.835	0.705	273.375	73.624
	Exponential	NDVI	${y=49.3e}^{4.236x}$	226.170	0.670	292.411	71.788
	Logarithmic	EVI	$y=1187\mathit{ln}x+2002$	210.605	0.701	277.910	73.187
	S-curve	EVI	${y=e}^{(8.385-{\displaystyle \frac{0.673}{x}})}$	192.405	0.729	261.603	74.760
	Stepwise	EVI and NDSVI	$y=2185.293\times EVI+1369.506\times NDSVI-682.05$	183.945	0.760	246.940	76.175

Note: this study compared the inversions of 15 spectral variables, and only the best inversion accuracy of each model is shown in the table.

). Additionally, the RVI proves to be more effective for wet biomass estimation compared with other spectral variables, while the EVI and NDVI are better suited for estimating aboveground dry biomass. Univariate regression models highlight the cubic model for the EVI as the most accurate for estimating aboveground dry biomass, with wet biomass estimation showing greater accuracy than dry biomass estimation in saltmarsh vegetation.

Compared with univariate models, the stepwise multiple linear regression model, which incorporates the RVI and NDSVI, demonstrated greater accuracy in estimating aboveground wet biomass in saltmarsh vegetation. Additionally, for aboveground dry biomass estimation, the EVI and NDSVI selected by the stepwise multiple linear regression model outperformed the two univariate linear regression models, with dry biomass accuracy surpassing that of wet biomass in saltmarsh vegetation.

The RF regression model was the most effective among the five machine learning regression models for utilizing image spectral data to estimate the aboveground wet and dry biomass of saltmarsh vegetation. This method achieves the highest accuracy and has great potential in estimating aboveground wet biomass. Despite slightly lower accuracy, the GPR model performs well in estimating the AGB of saltmarsh vegetation and outperforms the univariate and stepwise models. The accuracy order for estimating aboveground wet biomass was RF > GPR > SVR > NN > DT, while for aboveground dry biomass, it was RF > GPR > NN > DT > SVR (

Table 9. Remote sensing inversion accuracy of saltmarsh wetland vegetation AGB based on machine learning models.

Type	Model Name	Spectral Information	MAE	R2	RMSE	Accuracy (%)
Wet biomass	RF regression	15 spectral variables	399.396	0.803	551.913	79.660
	GPR (with squared exponential kernel function)		436.839	0.765	562.052	79.289
	SVR (with Gaussian kernel function)		484.313	0.710	623.693	77.017
	NN regression (with 5 layers)		483.252	0.690	614.694	77.349
	DT regression		481.776	0.689	637.917	76.493
Dry biomass	RF regression	15 spectral variables	164.490	0.829	226.416	78.155
	GPR (with squared exponential kernel function)		175.719	0.787	235.921	77.238
	SVR (with Gaussian kernel function)		242.016	0.669	293.042	71.727
	NN regression (with 5 layers)		202.900	0.719	266.506	74.287
	DT regression		215.948	0.691	284.926	72.510

Note: the parameters of each model in the table have been optimized, such as the kernel function of GPR and SVR, as well as the number of layers in the NN, etc.

). Furthermore, the overall modeling accuracy of the machine learning regression models for the overall AGB of saltmarsh vegetation was greater than that of the single vegetation type inversion methods. This is mainly due to the larger sample data available for overall modeling, which better leverages the advantages of machine learning nonparametric methods.

3.3. Comparison of AGB Retrieval Results With and Without the Distinction of Wetland Vegetation Types

This study evaluated the performance of various models, including seven univariate models, stepwise multiple linear regression models, and five machine learning regression models, in predicting AGB for specific wetland vegetation types (P. australis, S. alterniflora, and Scirpus spp.) and overall saltmarsh vegetation. Optimal inversion models were obtained for each wetland vegetation type and overall saltmarsh vegetation, but there are advantages and disadvantages when distinguishing wetland vegetation types during modeling. Distinguishing vegetation types in modeling provides consistent impacts on radiation transfer due to the uniform plant structure and morphology of wetland vegetation. However, this approach may be influenced by other factors, such as underlying surfaces and limited sample sizes. Conversely, not distinguishing vegetation types allows for a large sample size, but variations in plant structure and morphology can lead to diverse effects on the radiation transfer process. This study further investigated the influence of differences in wetland vegetation types on AGB inversion by comparing results under conditions of distinguishing and not distinguishing vegetation types.

The RF regression model outperformed the other 12 models when modeling without distinguishing wetland vegetation types in terms of the accuracy of the inversion of AGB from saltmarsh vegetation. Consequently, this model was selected to predict the aboveground wet and dry biomass of saltmarsh vegetation. In contrast, when distinguishing between wetland vegetation types, optimal aboveground wet and dry biomass inversion models were selected for each wetland vegetation type, and their inversion results are presented in

Table 10. Comparison of the accuracy of remote sensing inversion of AGB with and without the distinction of wetland vegetation types.

Types	Wetland Vegetation Type	Biomass Type	MAE	R2	RMSE	Accuracy (%)
Distinguishing wetland vegetation types	P. australis	Wet biomass	296.172	0.541	382.762	84.720
	P. australis	Dry biomass	185.858	0.462	227.484	82.160
	S. alterniflora	Wet biomass	699.687	0.491	810.830	79.773
	S. alterniflora	Dry biomass	244.510	0.461	280.106	79.585
	Scirpus spp.	Wet biomass	178.187	0.640	207.218	86.159
	Scirpus spp.	Dry biomass	36.825	0.575	42.314	85.385
	All saltmarsh vegetation	Wet biomass	390.785	0.806	545.515	79.888
	All saltmarsh vegetation	Dry biomass	162.865	0.839	219.965	78.777
Without distinguishing wetland vegetation types	P. australis	Wet biomass	298.452	0.537	386.362	84.576
	P. australis	Dry biomass	194.776	0.430	236.374	81.463
	S. alterniflora	Wet biomass	702.629	0.464	819.092	79.567
	S. alterniflora	Dry biomass	241.967	0.447	284.593	79.258
	Scirpus spp.	Wet biomass	193.651	0.602	218.630	85.397
	Scirpus spp.	Dry biomass	43.874	0.327	53.784	81.423
	All saltmarsh vegetation	Wet biomass	399.396	0.803	551.913	79.660
	All saltmarsh vegetation	Dry biomass	164.490	0.829	226.416	78.155

.

In general, the difference in accuracy between distinguishing and not distinguishing wetland vegetation types was relatively small for the inversion of saltmarsh wetland vegetation wet and dry biomass. However, the accuracy of distinguishing vegetation types was slightly better than that of not distinguishing them. When distinguishing wetland vegetation types, the R² values of the measured and fitted wet and dry biomasses could reach 0.806 and 0.839, respectively, and the inversion accuracy of type-based individual modeling was better than that of overall modeling for all three types of saltmarsh wetland vegetation.

For wet biomass, the difference in inversion accuracy between distinguishing and not distinguishing wetland vegetation types was relatively small. When modeling the wet biomass of P. australis separately, the R² value was 0.541, which was slightly better than that of the overall model (0.537), and the difference in the RMSE between the individual and overall models was only 3.600 g/m². Therefore, the overall modeling method can adequately invert the wet biomass of P. australis. Similarly, the inversion accuracy of the wet biomass of S. alterniflora was like that of P. australis, and the overall modeling method could be used to invert it as well. The differences in R² and RMSE between distinguishing and not distinguishing wetland vegetation types were 0.027 and 8.262 g/m², respectively. However, when the wet biomass of Scirpus spp. was inverted, the overall modeling method tends to overestimate low values and underestimate high values, and the difference in the RMSE between the individual and overall modeling methods is 11.412 g/m². Therefore, the individual modeling method for estimating the wet biomass of Scirpus spp. is better than the overall modeling method and provides more accurate inversion results (Figure 4).

The inversion accuracy between modeling individually while distinguishing wetland vegetation types and overall modeling without distinguishing vegetation types is greater for dry biomass than for wet biomass. Additionally, by distinguishing wetland vegetation types, the inversion accuracy can be improved by reducing the underestimation of high values. For the dry biomass of P. australis, there was little difference between the individual and overall modeling methods, with only small variations in R² and RMSE of only 0.032 and 9.090 g/m², respectively. Similarly, for the dry biomass of S. alterniflora, the differences between the individual modeling and overall modeling methods were relatively small, with differences in R² and RMSE of only 0.014 and 4.487 g/m², respectively. Although the overall modeling method tends to underestimate high values more than the individual modeling methods do, both methods have comparable inversion effects on the dry biomass of S. alterniflora. However, when the dry biomass of Scirpus spp. is inverted, high values are occasionally overestimated when the overall modeling method is used compared with the individual modeling methods.

In summary, univariate regression models, multiple linear regression models, and machine learning regression models were compared to estimate the aboveground wet and dry biomass of three main saltmarsh vegetation types on Chongming Island. The optimal inversion models were selected for P. australis, S. alterniflora, Scirpus spp., and all saltmarsh vegetation. The results indicated that using RF regression models for overall modeling could better fit the aboveground wet and dry biomass of wetland vegetation, with inversion results for wet biomass being superior to those for dry biomass. The inversion results for wet biomass were consistently greater than those for dry biomass, primarily because wet biomass originates from fresh vegetation with higher water content, which results in more pronounced absorption in the red wavelength band. This enhanced absorption correlates with greater inversion accuracy, with the difference in accuracy between wet and dry biomass ranging from 1.111% to 3.974%. While individual modeling with wetland vegetation types provides a better reflection of the aboveground biomass of each type, it requires multiple models and has high computational complexity. On the other hand, using the RF machine learning regression model for overall modeling without distinguishing wetland vegetation types can quickly reveal the distribution of the AGB of saltmarsh vegetation on Chongming Island with results similar to individual modeling.

3.4. Estimation of the AGB of Wetland Vegetation Under Different Models

To evaluate the impacts of different regression models on wetland vegetation AGB estimation, various models, including univariate regression, multiple linear regression, machine learning regression, and optimal model combinations for each vegetation type, were compared. The results generally revealed small differences in the inversion performances of several regression models for wet biomass. Among these models, the RF regression model estimates were similar for the combinations of various vegetation types, while the univariate regression model underestimated the S. alterniflora AGB in Beiliuyao and the stepwise regression model underestimated the Scirpus spp. AGB in Dongtan compared with the other models. A comparison of the biomass inversion of saltmarsh wetland vegetation by multiple models revealed that the RF regression model outperformed the univariate and multiple linear regression models in fitting saltmarsh vegetation AGB and reducing underestimation in some regions. Furthermore, the RF regression model, similar to the model distinguishing vegetation types, could better fit the saltmarsh wetland vegetation AGB without distinguishing vegetation types. In combination with the classification of saltmarsh wetland vegetation, species differences in AGB per unit area were also observed, with S. alterniflora having the highest biomass, followed by P. australis and Scirpus spp. The distribution of AGB per unit area of saltmarsh vegetation was consistent with the spatial classifications (as shown in Figure 5).

Compared with the other methods, the dry biomass inversion results revealed that the univariate regression model underestimated the S. alterniflora AGB in Beiliuyao compared to the other methods, whereas the multiple linear regression model overestimated the Scirpus spp. AGB in Dongtan compared with the other regression models (Figure 6). This is attributed to the fact that the univariate regression model has a limited ability to fit high biomass values, resulting in underestimation, while the multiple linear regression model tends to overestimate low biomass values, leading to lower fitting accuracy for low biomass values [20]. In contrast, the RF regression model effectively reduced both the overestimation of low biomass values and the underestimation of high biomass values, providing a better fit for the AGB of saltmarsh wetland vegetation. When the inversions of the RF regression model and the combined inversions for distinguishing vegetation types were compared, both methods demonstrated improved fitting of the saltmarsh vegetation AGB. However, when distinguishing saltmarsh vegetation types, the low values of S. alterniflora and P. australis and the high value of Scirpus spp. exceeded the range of sample training data and were affected by classification accuracy. Therefore, overall modeling via the RF regression model can better estimate the aboveground dry biomass of saltmarsh vegetation.

3.5. Feature Importance and Stability Analysis of the RF Regression Model

The RF regression model is effective at estimating the AGB of wetland vegetation using spectral data and can determine the relative importance of each spectral variable (Figure 7). The most crucial variable for wet biomass estimation was the RVI, followed by the NDWI2130, NDVI, and NDWI1640. NDWI2130 and NDWI1640 are ranked highly in importance because they reflect variations in vegetation water content, which is correlated with the AGB [72]. Additionally, the RVI, NDVI, and SAVI also rank high in importance as they are sensitive to differences in vegetation greenness [68,94]. Spectral indices such as the NDVI are particularly sensitive to chlorophyll content, which serves as an indicator of biomass [95,96,97]. The variables strongly correlated with wet biomass, such as the RVI, NDWI1640, and NDWI2130, are given high importance in the RF regression model. The RF regression model automatically selected suitable variables based on the data features for wetland vegetation biomass estimation.

For dry biomass estimation, the NDSVI emerged as the most important spectral variable, followed by the RVI, EVI, PVI, SAVI, NDVI, and DVI in descending order of importance. This ranking aligns with their correlation with dry biomass, with the EVI, SAVI, PVI, DVI, NDSVI, RVI, and NDVI being the top seven variables. The NDSVI is sensitive to vegetation water loss, the EVI, RVI, and SAVI reflect differences in vegetation growth, and the PVI indicates vegetation structure; all of these factors are strongly correlated with dry biomass. The RF regression model automatically selected the appropriate variables based on their correlation with biomass, ensuring reliable estimation of wetland vegetation.

The tenfold cross-validation errors between the model fittings and the measured AGB at the sample sites are shown in Figure 8. The RF regression model used in this study is stable and reliable for estimating the AGB of saltmarsh vegetation. The small fluctuation in overall sample errors indicates the model’s consistency and robustness, making it suitable for accurate AGB estimation, which is supported by the similarity in the spectral range between the sample and fitted data.

4. Discussion

4.1. Comparison of Parametric and Nonparametric Methods for Biomass Inversion

A comparison of parametric (traditional statistical) and nonparametric (machine learning) methods for estimating AGB in saltmarsh vegetation (Figure 9) revealed that the nonparametric RF regression model outperformed the parametric RVI cubic regression model and stepwise regression model for accurately estimating wet AGB (Figure 9a–c). The stepwise regression model yielded scattered predictions and a high RMSE, with overestimation in the 2500–4000 g/m² range and underestimation in the range above 5000 g/m², highlighting the limitations of traditional regression approaches in capturing the variability of wet biomass across different ranges. The optimal univariate regression model (the RVI cubic model) improved the fitting accuracy in the high-value area but still exhibited high dispersion between the predicted and measured results. In contrast, the nonparametric machine learning regression model effectively utilized image spectral data and achieved greater accuracy with an R² value of 0.803, indicating its superior ability to model complex relationships between spectral indices and biomass while accounting for inherent data noise and variability. However, for dry biomass estimation, the correlation with dry weight biomass is greater than that with wet weight biomass, primarily because the association between dry matter and spectral indices is less affected by moisture content. This observation underscores the critical role of moisture as a confounding factor in wet biomass estimation, which is more effectively mitigated through machine learning approaches. The univariate regression model (EVI cubic model) showed notable variance and underestimation for high biomasses above 1800 g/m², along with overestimation in the 800–1400 g/m² range, resulting in limited accuracy in predictions (Figure 9d–f). Compared with the univariate model, the stepwise regression model slightly reduced the variance but still significantly underestimated high biomass values above 1800 g/m² and performed poorly for low biomass values below 500 g/m², even yielding some negative values. These results indicate that stepwise regression models are limited by their dependence on predefined linear relationships, making them insufficient for accounting for the nonlinear and heteroscedastic nature of biomass data in marsh ecosystems [90]. In contrast, the RF regression model demonstrated the capability to handle these nonlinearities, effectively reducing the overestimation of low dry biomass values and the underestimation of high dry biomass values, achieving an R² value of 0.829. This performance highlights its potential for accurately estimating the dry biomass of marsh vegetation.

The inversion results of the machine learning and parametric regression models compared in this study are consistent with the biomass inversion results for forests on Chongming Island reported by Zhang et al., demonstrating the significant application potential of machine learning methods in biomass estimation [98]. Specifically, Zhang et al.’s findings underscore the transferability of machine learning models across diverse ecosystems, supporting their robustness and scalability for biomass estimation. Zhu et al. compared the application results of distinguishing vegetation categories and not distinguishing them in mangrove biomass inversion, finding that the inversion accuracy was greater when the mangrove categories were differentiated, with a difference of 19.17% [99]. The results of this study also indicate that distinguishing wetland vegetation types yields better inversion results, although the difference is relatively small, primarily due to the less pronounced growth variations among herbaceous plants in saltmarshes than among the more diverse types of mangroves. This aligns with the hypothesis that structural and compositional diversity among vegetation types significantly influence spectral reflectance characteristics, necessitating differentiation in highly heterogeneous systems like mangroves while yielding marginal benefits in more homogenous environments like saltmarshes. Furthermore, these findings reinforce the broader applicability of machine learning models in ecological studies, emphasizing their adaptability to various scales and contexts. The ability of RF models to integrate spectral data with ancillary environmental variables provides a robust framework for enhancing biomass inversion accuracy, particularly in dynamic and spatially complex ecosystems.

In this study, nonparametric machine learning methods, especially the RF regression model, outperformed traditional parametric methods in estimating the AGB of saltmarsh vegetation. These methods are able to fully utilize image spectral data, allowing them to capture complex, nonlinear relationships between the spectral features and biomass, which traditional methods often struggle with. The high R² values obtained using RF regression confirm its effectiveness in achieving greater accuracy in biomass estimation. However, it is important to note that the retrieval results varied across different vegetation and biomass types, highlighting the need for a careful selection of methods based on the specific characteristics of the vegetation and environmental conditions. This variation emphasizes the importance of considering local ecological factors, such as species composition and vegetation structure, which can influence the spectral response and, in turn, biomass estimation. Zhuo et al. reported that the RF regression model demonstrated superior performance compared with the multiple regression model and BP neural network in estimating the AGB of P. australis and S. alterniflora based on UAV imagery. Their results showed an R² value of 0.82 and an RMSE of 116.14 g/m², which is consistent with the results obtained from Landsat imagery in this study, further confirming the robustness of the RF model in different remote sensing contexts [20]. This consistency across different types of imagery highlights the generalizability of machine learning methods like RF regression, which can effectively process diverse remote sensing data to estimate biomass in various ecosystems. Additionally, Zhou et al. estimated the aboveground biomass of wetland vegetation via UAV RGB imagery and reported a significant correlation between visible light vegetation indices and wetland vegetation AGB, which aligns with the findings of this study [57]. Based on the three wetland vegetation biomass retrievals in Section 3.1, the RF regression model, a nonparametric method, provided the highest accuracy in retrieving P. australis AGB, followed by the univariate regression model. In contrast, the stepwise regression model had the lowest accuracy. For S. alterniflora, the machine learning model showed greater accuracy in wet biomass retrieval, and the RVI quadratic model of the univariate model produced the best effect for dry biomass retrieval. This highlights that while machine learning models may perform better overall, univariate regression models can still have strengths, particularly when applied to simpler or more homogenous datasets. The univariate regression model performed better in retrieving the dry and wet biomass of Scirpus spp. than the stepwise regression model and the machine learning method did, possibly because the limited number of site samples hinders the leveraging of nonparametric algorithm advantages.

Overall, the choice of retrieval method for AGB in saltmarsh vegetation depends on the specific vegetation and biomass type, as well as the availability of data and sample size. The univariate regression model is advantageous because of its simplicity and direct reflection of the relationship between spectral data and biomass, but effective information screening is crucial for optimal performance. The stepwise multiple linear regression model is useful for screening spectral variables but is limited in its ability to model only linear relationships, and its performance can be heavily influenced by sample size and the number of variables included in the model [90]. Machine learning models, such as RF regression, can fully utilize image information for nonparametric estimation and ranking variable importance, resulting in reliable retrieval results and providing valuable insight into the spectral contributions to biomass. However, the effects of each variable may not be clear, and the performance can be influenced by the size of the sample data.

4.2. Limitations and Prospects

The study mentioned above has limitations. First, from the perspective of data sources, the study relied solely on Landsat remote sensing imagery for biomass inversion. While Landsat data are widely accessible and cost-effective, they have inherent limitations in terms of spatial and spectral resolution [100]. In contrast, Sentinel-2 imagery offers higher spatial and spectral resolution and provides more detailed classifications, making it suitable for capturing finer-scale variations in vegetation characteristics [101]. Future research should investigate the impact of using Sentinel-2 data on model performance, particularly for different saltmarsh vegetation types. Moreover, the current study used only spectral and AGB data from saltmarsh wetland vegetation and did not incorporate height information, such as that derived from Lidar or UAV photogrammetry, which is often strongly correlated with biomass. Integrating these height-based metrics could significantly enhance the accuracy of biomass inversion models. Second, the study focused on the community level of wetland vegetation AGB inversion, with relatively limited ground sample data. This limitation prevents some machine learning models from fully demonstrating their advantages in biomass inversion. This research primarily compares univariate regression models and nonparametric machine learning models across different vegetation types to quantitatively assess the impact of classification on wetland vegetation AGB inversion without integrating deep learning inversion models that are adept at handling large datasets. Deep learning approaches, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have shown promise in extracting high-level features from multispectral and temporal remote sensing data, and their inclusion could enable more precise and scalable biomass inversion [102,103,104]. Finally, while UAVs have been used for biomass estimation of mangroves and crop yields, their application for precise quantitative inversion research on small-area saltmarsh wetland vegetation biomass should be further explored [29]. UAV platforms equipped with multispectral, hyperspectral, or Lidar sensors offer unparalleled spatial resolution and flexibility for acquiring data tailored to specific research objectives [20]. Addressing these limitations will contribute to more accurate estimations of aboveground biomass in wetland vegetation.

In summary, this study provides suggestions for future research directions. Future studies could integrate additional data sources, such as Lidar point cloud data, for collaborative inversion and comprehensive analysis. Research could also incorporate more ground sampling data and explore platforms equipped with Lidar and multispectral sensors to enable more detailed inversions at smaller scales. This approach would enable the development of models that can account for variations across vegetation types and environmental conditions. Additionally, as the data volume increases, employing deep learning methods may yield improved inversion results. These methodologies have the potential to play a significant role in supporting international marine policies, such as the 30x30 Conservation Agenda, by improving habitat assessment models on a global scale without distinguishing vegetation types. The findings and proposed future directions underscore the importance of combining advanced technologies and diverse datasets to enhance the accuracy, scalability, and applicability of biomass inversion models in saltmarsh wetlands.

5. Conclusions

In this study, the AGB of three wetland vegetation types (P. australis, S. alterniflora, and Scirpus spp.), as well as the overall saltmarsh vegetation on Chongming Island, were estimated using various regression models (univariate, multiple linear, and machine learning). And the effects of modeling with and without the distinction of vegetation types were compared. The study revealed that: (1) The impact of vegetation type on model performance varies. Scirpus spp. was more precise for biomass retrieval than S. alterniflora or P. australis when distinguishing vegetation types. Different vegetation types require specific regression models. For example, P. australis performs well with RF, Scirpus spp. with univariate regression models, and S. alterniflora with a mix of RF and univariate models; (2) Overall modeling without distinguishing vegetation types is beneficial, addressing challenges in wetland sampling and limited training samples. In terms of model selection, nonparametric regression models such as RF demonstrate superior performance in fitting the biomass of saltmarsh vegetation compared with univariate and multiple linear regression models; and (3) Although distinguishing between vegetation types and modeling them separately slightly improves the inversion results, with an increase in R² by 0.003 and 0.01 and a reduction in RMSE by 6.398 and 6.451, it requires considerable computational effort. In contrast, overall modeling without distinguishing vegetation types can still accurately estimate wetland vegetation biomass, especially when nonparametric methods such as random forests are applied, offering strong practical potential. The assessment in this study of the impact of coastal wetland saltmarsh vegetation types on aboveground biomass inversion not only contributes to a better understanding of the biomass dynamics of different vegetation types but also provides valuable scientific support for the protection and restoration of wetland ecosystems.