Empirical Analysis of a Super-SBM-Based Framework for Wetland Carbon Stock Safety Assessment

Abstract

With climate change and urbanization expansion, wetlands, which are some of the largest carbon stocks in the world, are facing threats such as shrinking areas and declining carbon sequestration capacities. Wetland carbon stocks are at risk of being transformed into carbon sources, especially those of wetlands with strong land use–natural resource conservation conflict. Moreover, there is a lack of well-established indicators for evaluating the health of wetland carbon stocks. To address this issue, we proposed a novel framework for the safety assessment of wetland carbon stocks using the Super Slack-Based Measure (Super-SBM), and we then conducted an empirical study on the Quanzhou Bay Estuary Wetland (QBEW). This framework integrates the unexpected output indicator (i.e., carbon emissions), the expected output indicators, including the GDP per capita and carbon stock estimates calculated via machine learning (ML)-based remote sensing inversion, and the input indicators, such as environmental governance investigations, climate conditions, socio-economic activities, and resource utilization. The results show that the annual average safety assessment for carbon pools in the QBEW was a meager 0.29 in 2015, signaling a very poor state, likely due to inadequate inputs or excessive unexpected outputs. However, there has been a substantial improvement since then, as evidenced by the fact that all the safety assessments have exceeded the threshold of 1 from 2018 onwards, reflecting a transition to a “weakly effective” status within a safe and acceptable range. Moreover, our investigation employing the Super-SBM model to calculate the “slack variables” yielded valuable insights into optimization strategies. This research advances the field by establishing a safety measurement framework for wetland carbon pools that leverages efficiency assessment methods, thereby offering a quantitative safeguard mechanism that supports the achievement of the “3060” dual-carbon target.

1. Introduction

As one of the world’s largest carbon stocks, wetlands store more than 30% of the world’s carbon and play an irreplaceable role in the global carbon cycle [1]. However, wetlands are also some of the most fragile ecosystems. Once damaged, substantial reductions in their biodiversities, ecological services, and carbon sink capabilities occur [2]. According to the National Wetland Survey Report published by the State Forestry Administration, China’s wetland area decreased by nearly 3.4 million ha from 2003 to 2013 [3]. Moreover, the average soil organic carbon stocks of wetlands decreased by 8.03% from 2000 to 2020 [4]. Hence, wetlands are at risk of transforming from carbon sinks into carbon sources [5]. The risk increases with urbanization, population growth, and the intensification of land use–wetland conservation conflicts.

Consistent with the need to adopt a more scientific approach to the wetland restoration proposed by Bayraktarov et al. [6] at the United Nations General Assembly, the protection of wetland carbon stocks should have a digital guarantee. That is, protecting wetland carbon reserves through digital technology and information technology. Scientific estimates of carbon stocks can provide a reliable basis for assessing the health of wetland carbon stocks. Carbon stock estimates are usually based on biomass accounting. The sample site inventory [7] and biomass-based carbon density estimation [8] models are two widely used models. The sample site inventory method can obtain accurate measurements through field surveys, but its high cost and cumbersome operations limit its applications in large areas [9]. The carbon density estimation method is cost-effective and can be applied to large areas. However, it may be inaccurate for small-scale area studies because of using a constant carbon density for different types of wetlands [4].

Due to the rapid development of remote sensing technology and machine learning algorithms, methods based on a combination of field surveys and remote sensing image inversion are more accessible and reliable and have gradually evolved into the mainstream [10,11]. For example, Lyu et al. [12] analyzed field sampling data and MODIS reflectance, built single-factor, multi-factor parameters, and multi-factor non-parametric models, and estimated the above-ground biomass of typical grassland grasslands in Mongolia from 2015 to 2019. Quang et al. [13] obtained a reasonable prediction of the mangrove biomass in the Red River estuaries of Vietnam by combining field sampling data and Sentinel-2 remote sensing data. Chen et al. [14] proposed a generative adversarial network with a constrained factor model (GAN-CF) to estimate salt marsh biomass using expanded in situ observations and Sentinel-2 remote sensing data to improve tidal flat biomass estimation accuracy. Zhang et al. [15] analyzed the correlation between 1201 ground measurement data and 15 vegetation indices generated by MODIS reflectance. They found that grassland aboveground biomass was significantly related to SAVI, NDVI, and OSAVI, summarized the differences between different grasslands, and concluded suitable variables for various biomass inversion models to support research in subdivided grasslands. These combined applications of remote sensing images and field sampling points not only overcome the difficulty of conducting surveys in some specific spaces or times but also make it possible to estimate biomass in large areas.

Biomass estimation traditionally applies regression models, such as Multiple Linear Regression (MLR) [12], but it is difficult to capture complex nonlinear relationships using these methods [14]. With the development of computer science and technology, more and more machine learning algorithms that can address this issue are being applied to biomass estimation. The Random Forest (RF) algorithm [16] is a robust machine learning algorithm that is widely used in predictive biomass estimation studies [16]. Ma et al. [17] found that the RF algorithm achieved an excellent performance in feature extraction from airborne LiDAR point cloud data to estimate the forest vegetation biomass. XGBoost [18], short for eXtreme Gradient Boosting, is another machine learning algorithm known for its high computational speed, efficiency, accuracy, and strong generalization capabilities. Li et al. [19] estimated the biomass of the mangrove forests in the Shengjin Lake Wetland based on XGBoost and obtained satisfactory results. Carbon stocks can be estimated by combining biomass estimates and the vegetation carbon conversion factor [20]. The carbon conversion factor refers to the ratio of the conversion of carbon dioxide absorbed by vegetation growth into carbon. This factor measures the vegetation’s absorption of carbon dioxide [21].

In addition to carbon stock estimation, it is crucial to balance wetland resource protection and urban expansion and to scientifically assess the health of wetland carbon stocks. Various factors may affect wetland carbon stocks. For example, climatic conditions shape the environment of the basic ecological processes in wetlands. Sanders et al. [22] found that increased rainfall may increase wetland mangrove carbon stocks. Socio-economic activities, such as wetland reclamation, land reclamation, and fishery activities, have seriously disturbed wetland ecological environments and have become one of the factors that contribute to the conversion of wetlands from carbon sinks to carbon sources [23]. In recent years, with the improvement in environmental awareness, the government has actively adopted environmental management measures, such as the withdrawal of fishing from wetlands and water quality control, which have promoted improvements in wetland ecological environments and thereby increased their carbon storage capacities [24]. As an important indicator for measuring regional economic development, the per capita GDP has a special relationship with wetland environmental protection, which affects the wetland carbon stock status to a certain extent [25]. This has a potential impact on wetlands. The growth of per capita GDP indicates rapid economic development in the region, a portion of which comes from the development and utilization of wetlands. Moreover, pressure for increased carbon emissions follows when regional economies develop rapidly.

The abovementioned factors, such as climate conditions, socio-economic activities, and government investigations, among others, directly or indirectly affect the security of wetland carbon stocks [26]. However, to the best of our knowledge, no comprehensive health assessment for wetland carbon stocks has been proposed. Drawing on the wetland ecological health assessment method [27] and the green development index [28], a non-parametric statistical method, Data Envelopment Analysis (DEA) [29], is available to assess the safety statuses of wetland carbon stocks. The DEA model can obtain an efficiency value that can be used to measure the relative efficiency of each study unit. This efficiency value represents the health status of a developing carbon stock. The efficiency value in the traditional DEA model ranges from 0 to 1, which is limited and not conducive to unit comparison [30,31]. Andersen and Petersen [32] extended the traditional radial measure method and proposed a procedure for ranking efficient units to achieve mutual comparisons. Tone [33] proposed the Super Slack-Based Measure (Super-SBM) model, which introduces a slack variable that can calculate the degree of efficiency required to achieve optimal values. Then, the model was improved to support unexpected outputs [34]. The Super-SBM model has received increasing attention as a model for calculating relative efficiency because it considers multiple input–output indicators simultaneously and can guide the optimization of the indicators in each decision-making unit (region) [35,36].

Therefore, in this study, we established a framework for evaluating the safety of wetland carbon stocks based on the Super-SBM model. By integrating factors such as government investigations, climate conditions, socio-economic activities, resource utilization, and carbon emissions, we conducted an empirical analysis in the Quanzhou Bay Estuary Wetland (QBEW), where human–land conflict is prominent, to validate its effectiveness.

2. Materials and Methods

2.1. Study Area

As shown in Figure 1, our study area was the Quanzhou Bay Estuary Wetland (QBEW), Fujian, on the southeastern coast of China. It has a southern subtropical maritime monsoon climate, and its administrative division includes five counties: Hui’an County, Luogang District, Fengze District, Jinjiang City, and Shishi City, with a total area of 7065.31 ha [37]. The QBEW is a subtropical estuarine mudflat wetland with China’s largest continuous area of artificially planted mangrove forests. It has been listed as one of “Asia’s Important Wetlands”, one of “China’s Priority Ecosystems”, one of “China’s Important Wetlands”, and one of “Fujian’s Important Wetlands” [38]. The QBEW and its surrounding villages are densely populated, and land use–wetland conservation conflicts in the area are prominent. The main protection objects of the QBEW area are the wetland ecosystem, the mangroves, and the various rare wild animals that inhabit the area. Artificial mangrove planting began in 2000. The artificial planting communities are mainly distributed on the east bank of the Louyang River, including a single community of Aegiceras corniculatum, a community of Kandelia obovata, and a mixed community of Kandelia obovata and Aegiceras corniculatum [39]. Based on comprehensive considerations of safety and wetland accessibility, we conducted a field survey of the mangroves near Luoyang Bridge.

2.2. Data Sources and Preprocessing

2.2.1. Field Sampling Data

Kandelia obovata and Aegiceras corniculatum are the dominant species of mangroves in our study area, with tree heights ranging from 3 to 5 m. There are also a few scattered Avicennia marina trees in the QBEW. According to the spatial distribution characteristics and tidal conditions, a field survey was conducted on the mangrove area in the QBEW from 1 to 5 March 2023. A total of 20 plots were surveyed, each with an area of 10 × 10 m. The height and diameter of each tree at breast height were measured using the wood-checking ruler method, and the longitude and latitude coordinates of each sampling point were determined through the GPS.

2.2.2. Remote Sensing Data

The remote sensing data were Sentinel-2 products with 13 spectral bands collected from the European Space Agency (ESA) (https://dataspace.copernicus.eu/). The data observation dates were 4 March 2023; 4 March 2022; 9 March 2021; 8 April 2020; 29 January 2019; 10 March 2018; 29 April 2017; 25 January 2016; 26 November 2015. The maximum cloud coverage was 19.84% (no clouds cover the QBWE in these images). Many features, such as texture and vegetation indices, can be calculated through multiple spectral bands [40].

Table 1. Formulas and references for calculating remote sensing variables.

Features	Algorithms/Definitions	Descriptions
Spectral bands	$\mathrm{B}2,\mathrm{B}3,\mathrm{B}4,\mathrm{B}5,\mathrm{B}6,\mathrm{B}7,\mathrm{B}8,\mathrm{B}8\mathrm{a},\mathrm{B}11,\mathrm{B}12$	Original band reflectance of Sentinel-2. B2, B3, and B4 are visible light bands of blue, green, and red, respectively; B5, B6, and B7 are vegetation red-edge bands; B8 and B8a are near-infrared bands; and B11 and B12 are short-wave infrared bands. https://sentinel.esa.int/web/sentinel/missions/sentinel-2
Vegetation indices	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\frac{\left(\mathrm{B}8-\mathrm{B}4\right)}{\left(\mathrm{B}8+\mathrm{B}4\right)}$	Normalized difference vegetation index [41]
	$\mathrm{E}\mathrm{V}\mathrm{I}=2.5\times \frac{\left(\mathrm{B}8-\mathrm{B}4\right)}{\left(\mathrm{B}8+6\times \mathrm{B}4-7.5\times \mathrm{B}2+1\right)}$	Enhanced vegetation index [42]
	$\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{B}8-\mathrm{B}4$	Difference vegetation index [43]
	$\mathrm{R}\mathrm{V}\mathrm{I}=\frac{\mathrm{B}8}{\mathrm{B}4}$	Ratio vegetation index [44]
	$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=1.5\times \frac{\left(\mathrm{B}8-\mathrm{B}4\right)}{\left(\mathrm{B}8+\mathrm{B}4+0.5\right)}$	Soil-adjusted vegetation index [45]
Texture features	$\mathrm{C}\mathrm{M}\mathrm{M}={\displaystyle \sum _{i,j}}p\left(i,j\right)i$	The Haralick texture feature [46] variables were calculated using the 10 m B2, B3, B4, and B8 bands of Sentinel-2. $p\left(i,j\right)$ is the pixel of the GLCM (gray level co-occurrence matrix), and $M_{GLCM}$ and $V_{GLCM}$ are the mean and variance values of the GLCM. Then, the texture feature variables are the co-occurrence measure mean (CMM); co-occurrence measure variance (CMV); co-occurrence measure homogeneity (CMH); co-occurrence measure contrast (CMC); co-occurrence measure dissimilarity (CMD); co-occurrence measure entropy (CME); co-occurrence measure second moment (CMSM); and co-occurrence measure correlation (CMCO) [47].
	$\mathrm{C}\mathrm{M}\mathrm{V}={\displaystyle \sum _{i,j}}p\left(i,j\right){(i-M_{GLCM})}^2$
	$\mathrm{C}\mathrm{M}\mathrm{H}={\displaystyle \sum _{i,j}}p(i,j)\frac{1}{1-{\left(i-j\right)}^2}$
	$\mathrm{C}\mathrm{M}\mathrm{C}={\displaystyle \sum _{i,j}}p\left(i,j\right){\left|i-j\right|}^2$
	$\mathrm{C}\mathrm{M}\mathrm{D}={\displaystyle \sum _{i,j}}p(i,j)|i-j|$
	$\mathrm{C}\mathrm{M}\mathrm{E}=-{\displaystyle \sum _{i,j}}p(i,j)\mathrm{l}\mathrm{o}\mathrm{g}\left(p\right(i,j\left)\right)$
	$\mathrm{C}\mathrm{M}\mathrm{S}\mathrm{M}={\displaystyle \sum _{i,j}}{p(i,j)}^2$
	$\mathrm{C}\mathrm{M}\mathrm{C}\mathrm{O}={\displaystyle \sum _{i,j}}{p(i,j)}^2\frac{(i-M_{GLCM})(j-M_{GLCM})}{V_{GLCM}}$

lists the three types of critical features, including 48 variables, used for the biomass inversion. The image data were preprocessed on the SNAP (Sentinel Application Platform, https://step.esa.int/main/toolboxes/snap (accessed on 23 March 2023)) using radiometric calibration, atmospheric correction, and mosaic image cropping and resampling.

2.2.3. Input–Output Indicators for the Super-SBM

Table 2. Input–output indicators.

Inputs and Outputs	Indicators		Descriptions and Sources
	Level 1	Level 2
Inputs	Climate conditions	TEMP	Temperature (°C), NCEI (https://www.ncei.noaa.gov (accessed on 8 May 2023))
		PREC	Precipitation (mm), NCEI (https://www.ncei.noaa.gov (accessed on 8 May 2023))
	Environmental governance investigations	EP	Environmental protection investment (CNY million), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
		TI	Technology investment (CNY million), accounts of the Quanzhou Bay Estuary Wetland Nature Reserve Development Center (http://lyj.quanzhou.gov.cn (accessed on 7 May 2023))
	Socio-economic activities	PD	Population density (cap/km2), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
		URB	Urbanization level (%), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
	Resource utilization	FOV	Fishery output value (CNY million), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
		UP	Unit power consumption (kwh/CNY million), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
Expected outputs	Economic development	GDPPC	GDP per capita (CNY), Quanzhou Statistical Yearbook (https://tjj.quanzhou.gov.cn/tjzl/tjsj/ndsj (accessed on 7 May 2023))
	Environmental improvement	CS	Carbon stocks (Mg/ha)
Unexpected outputs	Environment pollution	CE	Carbon emissions (10 kilotons) [48]

lists the two-level input–output indicators and their descriptions. The input indicators include environmental governance investigations, climate conditions, socio-economic activities, and resource utilization, and the output indicators include the expected economic development and environmental improvement indicators and the unexpected environmental pollution indicator. According to the literature review in the Introduction Section, these indicators affect the safety of wetland carbon stocks. Following the principles of systematization, operationalization, and efficiency, we collected many two-level indicators, such as the technology input (EP), temperature (TEMP), precipitation (PREC), fishery output value (FOV), and carbon emissions (CS), to feed the Super-SBM model. We set up decision-making units (DMUs) in four counties within the study area for eight years (2015–2022). All the input-output data (carbon emissions as undesired output) of 32 DMUs were imported into MaxDEA Ultra 9.0 software (http://www.maxdea.com/Index_CN.htm) (accessed on 25 January 2024) to calibrate the Super-SBM model.

2.3. Methods

2.3.1. Carbon Stock Estimation

In this study, we combined field surveys and remote sensing inversion to estimate the carbon stocks. The carbon stocks in this study refer only to the carbon in living biomass. This method is based on the biomass estimates using the anisotropic growth equations listed in

Table 3. Allometric growth equation for Aegiceras corniculatum, Kandelia obovate, and Avicennia marina.

Species	Allometric Equation	Source
Aegiceras corniculatum	${\mathrm{W}}_{\mathrm{A}\mathrm{G}\mathrm{B}}=31.34\times {\left({\mathrm{D}\mathrm{B}\mathrm{H}}^2\times \mathrm{H}\right)}^{0.465}$	[52]
	${\mathrm{W}}_{\mathrm{B}\mathrm{G}\mathrm{B}}=9.33\times {\left({\mathrm{D}\mathrm{B}\mathrm{H}}^2\times \mathrm{H}\right)}^{0.303}$
Kandelia obovata	${\mathrm{W}}_{\mathrm{A}\mathrm{G}\mathrm{B}}=0.187\times {\mathrm{D}\mathrm{B}\mathrm{H}}^{1.855}+0.267\times {\mathrm{D}\mathrm{B}\mathrm{H}}^{1.906}$	[53]
	${\mathrm{W}}_{\mathrm{B}\mathrm{G}\mathrm{B}}=4.6\times {\mathrm{D}\mathrm{B}\mathrm{H}}^{1.136}$
Avicennia marina	${\mathrm{W}}_{\mathrm{A}\mathrm{G}\mathrm{B}}=0.308\times {\mathrm{D}\mathrm{B}\mathrm{H}}^{2.11}$	[54]
	${\mathrm{W}}_{\mathrm{B}\mathrm{G}\mathrm{B}}=1.28\times {\mathrm{D}\mathrm{B}\mathrm{H}}^{1.17}$

Note: W_AGB and W_BGB denote the aboveground and belowground biomasses, respectively. DBH and H denote the diameters at breast height and tree height, respectively.

. The total carbon in living biomass stock can be estimated as follows:

$$\mathrm{C}={\mathrm{W}}_{\mathrm{T}}\times 0.5$$

(1)

where ${\mathrm{W}}_{\mathrm{T}}$ denotes the carbon conversion factor, representing the carbon stock per gram of dry matter. The international standard default of 0.5 is often used in many studies [49,50,51].

2.3.2. ML-Based Methods

Various models can be employed to estimate the total biomass. Compared with the traditional regression models, such as Multiple Linear Regression (MLR), machine learning (ML)-based models can capture nonlinear relationships and are therefore becoming increasingly popular. In this study, we applied and compared two mature ML-based models, the Random Forest (RF) and XGBoost models, to obtain better scientific estimations. The RF algorithm is an ensemble learning method that creates an ensemble of decision trees [55] and makes robust overall predictions from all the trees [16]. XGBoost is characterized by the use of a second-order Taylor formula in its optimization process and the addition of a regularization term in its loss function to reduce the variance and model complexity.

2.3.3. Super-SBM Model

The Super-SBM model is an advanced form of efficiency measurement that can provide a more detailed and discriminating approach than traditional DEA models. Due to the model’s ability to consider slacks directly and handle complex variables, it has become a valuable tool for managers seeking to understand and improve their operational efficiencies [34]. A decision-making unit (DMU) in the Super-SBM model is an entity that consumes various inputs to produce outputs. A DMU can be any operational unit that performs activities, making DMUs applicable in various contexts. The DMUs include three vectors: the inputs ($x$), expected outputs ($y^g$), and unexpected outputs ($y^b$).

$$X=\left[x_1,x_2\dots ,x_n\right]\in R^{m\times n}>0, x\in R^m$$

(2)

$$Y^g=\left[y_1^g,y_2^g,\dots ,y_n^g\right]\in R^{c_1\times n}>0, y^g\in R^{c_1}$$

(3)

$$Y^b=\left[y_1^b,y_2^b,\dots ,y_n^b\right]\in R^{c_2\times n}>0, y^b\in R^{c_2}$$

(4)

Here, $n$ denotes the total number of DMUs; in this paper, $n$ is equal to 32. $g$ and $b$ are the expected and unexpected variables, respectively; $m$, $c_1$, and $c_2$ denote the numbers of inputs, expected outputs, and unexpected outputs, respectively. Given that the intensity vector is $\lambda$, which denotes a non-negative vector in $R^n$, the production possibility set (P) can be defined as follows [34]:

$$P=\left\{\left(x,y^g,y^b\right)|x\ge X\lambda ,y^g\le Y^g\lambda ,y^b\ge Y^b\lambda ,\lambda \ge 0\right\}$$

(5)

Then, the unexpected output of the Super-SBM model can be defined as follows:

$${\rho }^{*}=min\frac{1+\frac{1}{m}\sum _{i=1}^m\left(\frac{s_i^{-}}{x_{ik}}\right)}{1-\frac{1}{c_1+c_2}\left({\sum }_{r=1}^{c_1}\frac{s_r^{+}}{y_{rk}^g}+\sum _{t=1}^{c_2}\frac{s_t^{b-}}{y_{tk}^b}\right)},s.t.\left\{\begin{array}{c}\begin{array}{c}{x}_{ik}\ge {\sum }_{j=1,\ne k}^n{x}_{ij}{\lambda }_j-s_i^{-}(i=\mathrm{1,2},\dots ,m)\\ y_{rk}^g\le {\sum }_{j=1,\ne k}^n{y}_{rj}^g{\lambda }_j+s_r^{+}(r=\mathrm{1,2},\dots ,c_1)\\ y_{tk}^b\ge {\sum }_{j=1,\ne k}^n{y}_{ij}^b{\lambda }_j-s_t^{b-}(t=\mathrm{1,2},\dots ,c_2)\end{array}\\ 1-\frac{1}{c_1+c_2}\left({\sum }_{r=1}^{c_1}\frac{s_r^{+}}{y_{rk}^g}+{\sum }_{t=1}^{c_2}\frac{s_t^{b-}}{y_{tk}^b}\right)>0\\ \begin{array}{c}\begin{array}{c}{\lambda }_j,s_i^{-},s_r^{+},s_t^{b-}\ge 0\end{array}\end{array}\end{array}\right.$$

(6)

where $k$ denotes the DMU; ρ* represents the optimization efficiency value of the DMU; $i$, $r$, and $t$ denote the $i$th, $r$th, and $t$th input or output variables, respectively; s_i⁻, s_r⁺, and s_t^b− denote the slacks in the inputs, expected outputs, and unexpected outputs, respectively. The model can evaluate the efficiency status of the DMUs. If ${\rho }^{\mathrm{*}}<1$, and $s_i^{-}$, $s_r^{+}$, and $s_t^{b-}$ are not equal to 0, the DMU is supposed to be inefficient. If ${\rho }^{\mathrm{*}}\ge 1$, the DMU is considered relatively efficient. Further, if $s_i^{-}$, $s_r^{+}$, and $s_t^{b-}$ are 0, then DMU is strong efficiency. The higher ${\rho }^{\mathrm{*}}$, the more effective the DMU is.

2.3.4. Safety Assessment Framework

We propose a novel framework to scientifically quantify the safety of wetland carbon stocks (Figure 2). The framework mainly consists of two parts. The first part is the carbon stock estimated based on field surveys and DL-based remote sensing data inversion. The second part is a safety assessment based on the Super-SBM model. This framework can incorporate the input–output indicators that affect the safety of wetland carbon stocks and produce an efficiency value and slack variables that support the creation of optimized strategies.

3. Results

For the experiments, we used an NVIDIA GeForce RTX 2050 graphics card with 16 GB memory. Using the sklearn and XGBoost Python packages, we trained the data with 47 variables to build the RF and XGBoost models. The number of decision trees in the fitted RF model was 99, with a maximum tree depth of six trees. As for the fitted XGBoost, the model’s learning rate was 0.3, the gamma was 0, and the L1 and L2 regular terms were 1. The R-squared (R2), root-mean-square error (RMSE) and mean absolute error (MAE) matrices were used to evaluate the two ML-based models’ performances.

3.1. Exploratory Data Analysis

Table 4. Statistical characterization of measured carbon stocks.

Tree Species	Number of Plants	Mean DBH (cm)	Mean Biomass (kg)	Mean Carbon Stock (kg)
Kandelia obovata	2981	4.15	30.15	15.57
Aegiceras corniculatum	1908	2.21	70.36	35.18
Avicennia marina	76	2.28	5.10	2.55
Total	4965	8.64	105.61	53.30

shows statistical information for the 4965 mangrove trees measured in our field survey. The diameters at breast height (DBH) range from 0.32 to 21.34 cm, and the heights range from 1.5 to 6.9 m. The mean carbon stock is about 49.30 Mg/ha, calculated based on the abovementioned allometric growth equation (Table 3) and the default carbon conversion factor (0.5).

3.2. Wetland Carbon Stock Estimates

Table 5. Comparison of three biomass estimate inversion models.

Model	R2	MAE (kg)	RMSE (kg)
MLR	0.57	14.96	30.46
RF	0.77	16.41	17.77
XGBoost	0.91	7.73	13.72

shows the comparison results of the three biomass estimate inversion models. The two ML-based inversion models, the XGBoost and RF, obtained better results than the traditional MLR model. The XGBoost model outperformed the other models with the highest R-squared (R²) value (0.91) and the lowest MAE (7.73 kg) and RMSE (13.72 kg).

As shown in Figure 3, the two ML-based models had more accurate carbon stock prediction results than the traditional MLR model. Compared with the RF model, the cumulative feature importance of the XGBoost reached the 90% threshold faster, with a smaller number of features. The B4 (blue) and B3 (green) bands are the top two features of the XGBoost model, contributing about 60% of the importance. The texture features B4_CMM, B2_CMV, B8_CMCO, and B2_CMH are highly important, ranking from 3 to 6. The importance rankings of the DVI and SAVI are 10 and 13, respectively.

3.3. Carbon Stock Variation in QBWE

3.3.1. Trends

Figure 4 shows the time series of the mangrove carbon stocks in the QBWE. The overall trend is highly consistent with that of Hui’an, and both trends have two peaks and two valleys. Carbon stocks surged in 2022, with an increment of about 3729 Mg, which was twice the first peak (2017). The incremental fluctuations in total carbon stocks are consistent with those in Hui’an County, indicating that Hui’an is the major contributor. However, there has been almost no increase in carbon storage in Jinjiang and Fengze districts. There has been a significant increase in Luoyang District in 2022, where carbon storage changes have been small for a long time.

3.3.2. Spatial Distributions of and Changes in Wetland Carbon Stocks

Since 2015, the total mangrove area has increased from 147 ha to 285 ha, an increase of about 93.88%. Figure 5 shows the spatial distribution of the mangrove carbon stocks. The areas with the most significant changes in carbon stocks are Luoyang Bridge (A), Yutou Bay (B), the west of Zeng’an Village (C), Fengze Xunmei (D), and the Jinjiang River Estuary (E). The mangrove area around Luoyang Bridge (A), located at the mouth of the Luoyang River, increased first, and it has had the largest carbon stock increase. Natural mangrove forests first grew in Fengze Xunmei (D) and the region between Luoyang Bridge (A) and the west of Zeng’an Village (C). By 2022, the mangrove unit carbon stocks ranged from 5.85 to 85.71 Mg/ha.

3.4. Carbon Stock Security and Optimization Strategies

3.4.1. Carbon Stock Safety Assessments

Table 6. Results of carbon stock safety assessment in QBEW.

Year	Mean	Increment	S.D.	Fengze District	Hui’an County	Jinjiang City	Luojiang District
2015	0.2885	-	0.1925	0.3644	0.4002	0.0006	0.3887
2016	0.3785	0.0901	0.1601	0.4544	0.4346	0.1403	0.4848
2017	0.6288	0.2503	0.2628	1.0227	0.4855	0.5044	0.5026
2018	1.0071	0.3783	0.0133	1.0014	1.0000	1.0000	1.0271
2019	1.0093	0.0021	0.0129	1.0283	1.0049	1.0000	1.0038
2020	1.0025	−0.0068	0.0024	1.0054	1.0012	1.0000	1.0035
2021	1.0089	0.0064	0.0059	1.0163	1.0045	1.0040	1.0109
2022	1.1507	0.1418	0.0855	1.0791	1.2219	1.0743	1.2274
Mean	-	-	-	0.8715	0.8191	0.7154	0.8311

Note: The bold numbers in the left half of the table indicate the highest and lowest mean, increment, and standard deviation. The bold numbers in the right half indicate the first time the efficiency value is greater than or equal to 1 in the region. The bold numbers in the last row indicate the best and worst average efficiency performance.

shows the results of the carbon stock safety assessments in the QBEW. The average safety evaluation value of the wetland carbon stocks increased from 0.2885 to 1.1507, with a growth rate of 298.86%. There are significant differences in safety between the regions, with the highest value of 0.8715 in Fengze District and the lowest value of 0.7154 in Jinjiang. The order from high to low is as follows: Fengze District > Luojiang District > Hui’an County > Jinjiang City. Fengze was the first to reach the efficient stage (a safety evaluation value greater than 1 can be regarded as efficient) in 2017, followed by Luojiang, Hui’an, and Jinjiang in 2018. Since then, the regional differences have gradually narrowed and stabilized, reaching the minimum standard deviation (S.D.) of 0.0024 in 2020.

3.4.2. Analysis of Optimization Strategies Based on Slack Variables

Table 7. Slack variables in QBEW.

Input–Output Variables		Unit	Fengze District		Hui’an County		Jinjiang City		Luojiang District
			2015	2022	2015	2022	2015	2022	2015	2022
Inputs	TEMP	℃	0.36	1.08	0.40	1.22	0	1.07	0.39	1.23
	PREC	mm	−7.54	0	−10.56	0.59	−5.65	0	−11.27	1.01
	EP	CNY million	−570.18	0	−901.02	0	−407.83	0	−625.67	0
	TI	CNY million	0	109.48	−6728.79	0	0	0	−1753.78	50.75
	PD	cap/km2	−0.24	0	−64.86	0	0	0	−8.51	0
	URB	%	−1302.97	0	−621.34	30.43	0	0	−237.16	1.65
	FOV	CNY million	−35.27	0	−26.08	1.18	−3.50	0	−22.58	0
	UP	kwh/CNY million	−2440.54	0	−110,939.12	0	−136,493.32	0	−605.21	0
Expected outputs	GDPPC	CNY	−0.03	0	−0.04	0	−0.04	0	−0.03	0
	CS	Mg	0	−244.53	0	−74.12	0	−320.92	0	−54.17
Unexpected output	CE	10 kilotons	19,329.19	0	96,287.92	−402,960.10	83,148.16	0	22,888.37	−61,776.94

Note: The bold numbers indicate representative slack value results for each district and county.

shows the slack variables in the QBEW. Compared with 2015, the input–output statuses in most areas were good in 2022. Areas with values of 0 only need to be maintained to achieve a good carbon stock security value. Regarding the input indicators, the temperature (TEMP) has increased, while the precipitation (PREC) has declined. Jinjiang City had a tolerance of a 1.07 °C variation in the annual average temperature within which the safety assessment of carbon stocks was not affected in 2022. Many optimization strategies can be drawn from this table. For example, an increase in the technology investment (TI) of CNY 109.48 million helped improve the security of the Fengze wetland carbon stocks in 2022. If the population density (PD) of Hui’an County decreases by 64.86 cap/km², the carbon stocks will be safer. Luojiang District has excess carbon emissions and needs to reduce its original carbon emissions by approximately 617.7694 million tons to reach a high carbon stock safety level.

4. Discussion

Drawing on the theory of green sustainable development and the concept of ecological security assessment, in this study, we proposed a framework for assessing the ecological security of wetland carbon stocks based on the DEA efficiency model, and we empirically analyzed the wetlands in the estuary of Quanzhou Bay, which achieved results that match the real-world situation.

This study combined remote sensing and field sampling data to compare the performance of MLR, RF, and XGBoost models in estimating wetland biomass. The results showed that the XGBoost algorithm was best suited for biomass inversion in QBEW (Table 5). These results were consistent with the comparisons reported by [56] and [57], which indicated that the XGBoost algorithm can extract richer feature information (Figure 3b). This XGBoost algorithm can use randomness in model fitting that may help find the relationships between independent and dependent variables. Considering the differences between samples and independent variables, especially when dealing with variables such as biomass that change with complex changes and the overall trend is not apparent, better results may be produced. The RF algorithm also outputs good results for biomass inversion but is not as good as XGBoost. As shown in the importance ranking of indicators (Figure 3c), the RF algorithm extracts limited information on indicators, primarily concentrated in B4, B3, and B4_CMM (co-occurrence measure mean). The RF algorithm may perform well on large data sets and be less suitable for handling this study’s smaller training samples. The MLR model has the worst biomass inversion results, which is expected. This is because the MLR model is a linear model with a limited ability to capture nonlinear relationships. The independent variables (band reflectance, vegetation index, and texture features) may not be independent of each other, resulting in inaccurate model parameter estimates.

In addition to the traditional red and green bands (B3 and B4), the Haralick features and CMM based on B4 also showed their importance (Figure 3b,c) in biomass estimation. This may indicate that texture features can provide information about the spatial distribution and changes between image pixels [47], which is beneficial to improving biomass estimation based on optical data [58]. This finding was consistent with some studies. Tian et al. [56] considered texture features and obtained better results in mangrove biomass estimates in southern subtropical China. In addition, Zhang et al. [59] found that the correlation between forest biomass and texture features was significantly related to the CMM (co-occurrence measure mean) features of SWIR1. This may indicate that CMM, which can provide overall image brightness or color features, may also benefit the biomass inversion.

The improvement of carbon stock is closely related to implementing conservation policies. The two peaks occurred in 2017 and 2019, respectively (Figure 4), which may be closely associated with the local government’s active restoration of wetlands. In 2017, the Quanzhou Municipal Government actively carried out ecological environment quality assessment and construction, promoted the city’s “dual restoration” (environmental and urban restoration) plan, renovated water basins, repaired mountains, and established ecological protection in red line control areas [60]. In 2019, the city’s “dual restoration” pilot work achieved initial results: the ecological environment was restored, and the environmental quality was significantly improved. The surge in carbon stocks in 2022 (an increase of approximately 3729 Mg) may be essentially the result of implementing the national “Blue Bay” policy. Quanzhou City’s “Blue Bay” comprehensive improvement action project was predicted to be completed by the end of 2022. In the project, 5026 acres of Spartina alterniflora have been cleared, 2722 acres of mangroves have been planted, 190 acres of original mangroves have been restored, 374 acres of bird habitats have been restored, and two monitoring stations and 6.2 km of coastal ecological transformation have been built. Within the scope of mangrove planting in the Jinjiang Sea area in the southern part of Quanzhou Bay, the environmental restoration effect has gradually emerged, improving the ecological environment of the coastal wetland at the Jinjiang estuary of Quanzhou Bay, and promoting the steady growth of wetland carbon stocks.

The two declines in carbon stocks may be related to biological invasion and engineering construction. The trough values occurred in 2018 and 2021, respectively (Figure 4). It may be affected by the expansion of Spartina alterniflora and the construction of cross-sea bridge projects. The expansion of the invasive species Spartina alterniflora will occupy the growth space of mangroves and affect the growth of mangroves. Li, H. et al. [61] research shows that the Spartina alterniflora area in Fujian Province accounted for 13.3%, 14.52%, and 15.1% of the total Spartina alterniflora area in the country in 2015, 2018, and 2020, respectively. In 2020, the area of Spartina alterniflora in Fujian Province has expanded, and the Quanzhou Bay area has also been affected by the same impact, with the original natural mangrove resources gradually decreasing [62]. In 2018, the implementation of the Quanzhou Bay Cross-Sea Bridge was launched. The bridge crossed Quanzhou Bay into Shishi City and Jinjiang City, which impacted the environment of the Quanzhou Bay Estuary Wetland and Tianzhu Mountain Forest Park Water Source Reserve.

The spatial distribution of mangrove carbon stocks is closely related to the growth status of mangroves. The areas with higher mangrove carbon storage in the study area are mainly Luoyang Bridge (A), Yutou Bay (B), and the west of Zeng’an Village (C) (as shown in Figure 5). The reasons may be as follows: (1) Natural forests will sequester more carbon than plantations. There are a few natural mangroves in Luoyang Bridge (A), the west of Zeng’an Village (C), and Fengze Xunmei (D), while most other areas are planted mangroves. This is consistent with the results of [63] and [64]. (2) The government has started replanting mangroves in Luoyang Bridge (A), Yutou Bay (B), the west of Zeng’an Village (C), and other areas since 2000, and has replanted mangroves in Luoyang Bridge (A) and Fengze Xunmei (D). The river channel is dredged to ensure a smooth flow of ebb and flow tides, promote the growth of mangroves in the estuary area, and improve the regional carbon sequestration capacity.

The security status of wetland carbon stocks may be significantly affected by socio-economic activities and resource use conditions. As shown in Table 7, the security status of Hui’an and Luojiang wetlands has been dramatically improved; in contrast, Jinjiang has more room for improvement in the safety protection and maintenance of wetland carbon pools. The reason may be that most of the mangroves in the study area are distributed in Hui’an and Luoyang, which are less developed. In contrast, the mangrove area in Jinjiang is smaller (almost no mangroves in 2015–2016), but human production activities have not decreased. Its fishery output value accounts for about 40% of the total output value of the study area. Compared with other places, the wetlands under Jinjiang’s jurisdiction are surrounded by residential buildings and are close to the city center, so the ecological security of the wetlands has been slowly improved.

This proposed framework opens a new window for assessing the health statuses of wetland carbon stocks. The safety value of the QBEW significantly increased by about 298.86% from 2015 to 2022, consistent with Quanzhou’s ecological civilization movement. Beyond this, optimization strategies can be drawn out from the generated slack variables. For example, Fengze improved its wetland carbon stock security by increasing its technology investment by approximately CNY 109.48 million in 2022. This result has important reference significance for many other areas that need to ensure the security of their wetland carbon stocks. However, due to the uncertainty of the input–output indicator selection, the efficiency of the safety assessment framework for wetland carbon stocks needs to be further verified.

Biomass estimation based on machine learning algorithms has been widely used and has been proven to improve estimation accuracy, especially for forests [65] and grasslands [66]. However, its application to wetlands needs to be strengthened. Hybrid machine learning models [67] or ensemble stacking approaches [68] should be able to improve the estimation accuracy. In addition, with the development of unmanned aircraft remote sensing technology, combined with field data samples, it is expected that higher biomass estimation accuracies will be achieved.

Differences in the timing of acquisition of remotely sensed data may affect biomass estimates, but this effect is relatively limited. There are specific seasonal differences in the biomass of mangroves, with spring biomass usually greater than winter biomass. The size of this difference is related to environment, climate, etc. [69,70]. The field survey time of this study was March 2023. A baseline for choosing remote sensing images for the inversion is close to the survey month. Some other months are selected for the following reasons: (1) The data in November 2015 were chosen because there was no Sentinel-2 data before June 2015, and the image’s cloud coverage in the study area was the smallest in November. (2) In 2016 and 2019, we used the data in January because of the high cloud coverage in March and April of those years. Using remote sensing images from November or January may result in lower biomass estimates, but this is not expected to be significant. The reason is that there are specific seasonal differences in the biomass of mangroves. The study area has abundant rainfall, is warm and humid, and has a mild climate. The difference in light and temperature between winter and spring is insignificant (the average annual temperature range is about 13.0–20.0 °C), which will not significantly impact biomass estimation.

Because of the limitations in the data availability, the input and output indicators for the Super-SBM-based framework are not comprehensive enough, and the DMUs are based on the county level. The obtainment of highly refined wetland carbon stock assessment results is difficult. Further studies could consider expanding the indicators in this framework and conducting more precise analyses based on a finer scale, such as grid level. For example, wetland carbon emissions include the emissions produced by human activities and those released through wetland biological respiration and organic matter decomposition. Therefore, a more precise efficiency assessment is expected if future studies include the latter carbon emissions as a new output indicator.

The optimization strategies based on slack variables do not consider influencing factors other than inputs and outputs, and the reliability of the results needs to be verified. One possible solution is to employ the Geographically Weighted Regression (GWR) [71,72] or Spatiotemporal Weighted Regression (STWR) [73,74] models to further explore the spatiotemporal heterogeneity affected by different driving factors after the safety assessments.

5. Conclusions

In this study, we combined remote sensing images and field survey data and used ML-based inversion models to estimate the carbon stocks of the QBEW. Taking the carbon stock estimates as one of the expected outputs, combined with other input–output indicators, a Super-SBM-based framework was established to assess the safety of wetland carbon stocks. This work fills in some of the gaps in carbon stock security assessment analysis. The following conclusions were drawn.

The proposed framework can effectively assess the security value of the wetland carbon stocks in each region (DMU), and the assessment results can reflect the actual changes. Our QBEW empirical study found that the assessment value of the carbon stock security increased from 0.29 in 2015 to 1.15 in 2022, and the assessed value of each DMU was greater than 1, having reached a weakly effective state since 2018. This matches the implementation of enhanced conservation measures in the area.

The proposed Super-SBM-based assessment framework fully considers the conflict between wetland protection and resource utilization in choosing the input–output indicators. This consideration follows the green and sustainable development concept and aligns more with the management and decision-making of local governments.

This assessment framework can also be used to explore and identify any potential room for improvement, providing guidance for further improving the safety and efficiencies of wetland carbon stocks. In particular, in the current context of global warming, the framework acts as a digital guarantee for protecting regional wetland carbon stocks and accelerating the achievement of the “3060” dual-carbon target [75].