1. Introduction
Although the terrestrial biosphere absorbs about 25% of anthropogenic carbon dioxide (CO
2) emissions, the rate of land carbon uptake remains highly unpredictable. This leads to uncertainties in climate projections [
1]. The rise in CO
2 concentration has drawn wide attention from governments and academics worldwide [
2]. The signing of the Paris Agreement in 2016 shows that [
3] many countries have reached a general political consensus on international cooperation on global climate change. In 2018,
Nature listed the major events of climate policy as one of the twelve things that need attention [
4]. It has also been ranked amongst the 125 most challenging scientific issues published by
Science [
5], positing how high the greenhouse effect makes the earth’s temperature rise.
The Global Carbon Budget report in 2018 stated that CO
2 emissions from worldwide fossil fuel combustion were expected to increase by 2.7% from the previous year. Fossil fuel combustion emitted 990 million tons of carbon into the atmosphere in 2017. In 2018, CO
2 emissions were likely to have hit a record high, however, the data is not yet available. To achieve the goal of controlling the temperature increase to 1.5 °C in the Paris Agreement, CO
2 emissions need to be reduced to 50% by 2030 and brought to zero by 2050 [
6].
Climate change has become a universal problem with global warming as the main feature. The circulation and distribution of carbon in each cycle has become a trending issue of global climate change research [
7]. Its primary endeavor focuses on how to improve the carbon sequestration capacity of terrestrial ecosystems and increase carbon storage. All countries are searching for methods to recollect the carbon emitted from energy consumption and fix it in soils and vegetation [
8,
9]. In the Paris Agreement of 2015, China stated that by 2030 it would cap CO
2 emissions. At present, China is positioned to make significant contributions to reducing CO
2 emissions through its ecological restoration projects [
10].
A wetland is a unique ecosystem formed by the interaction of land and water. It plays an important but complex role in the global carbon cycle, carbon emissions, and stabilizing global climate change. This contributes to the regulation of greenhouse gas levels through carbon sequestration [
9]. Although wetland areas only account for 4–6% of the total land area, its carbon storage accounts for 12–24% of the global terrestrial carbon storage. The huge carbon storage capacity of wetlands makes them integral to global soil carbon storage [
10].
Studies have shown that most wetlands have much higher CO
2 fixation than CO
2 and CH
4 emissions, and large amounts of organic matter are stored in the soil [
11]. This makes them the most efficient net CO
2 sinks and contributors to the balancing of carbon-containing greenhouse gases in the atmosphere [
12]. Although wetlands are regarded as one of the largest unknown ecosystems concerning future C dynamics, particularly carbon budgets [
13], they are an important organic carbon pool [
14]. They are equivalent to aquatic and forest ecosystems in terms of their roles in maintaining the earth’s health [
14]. Therefore, it is of great significance to increase attention over the carbon sequestration capacity as well as to quantify the temporal and spatial changes of carbon storage in wetland ecosystems.
Related research on wetland carbon storage began in the late 1990s. Many researchers had mainly focused on wetland types [
15], estimation methods [
16], and influencing factors [
17]. In recent years, the research on the carbon sequestration and sinks of wetlands has shown an obvious upward trend. This has attracted more and more attention [
18] and has prompted extensive academic interest.
As a highly important ecological factor in wetland ecosystems, soil organic carbon (SOC) provides information on several key processes, such as biomass production, necromass degradation, and organic carbon storage [
19]. It plays an important role in the global carbon cycle and has a significant impact on wetland ecosystem productivity, terrestrial ecosystem greenhouse gas emissions, atmosphere regulation, and global warming [
20]. SOC is the most important carbon pool in wetlands. Although it is also an important global carbon stock and could represent 20–30% of global soil carbon [
21], the rate of land carbon uptake remains highly unpredictable, leading to uncertainties in climate projections [
22].
The long-term use of wetlands leads to ecosystem degradation and carbon loss. Recent studies have indicated that the implementation of the ecological restoration projects has significantly increased ecosystem carbon sequestration across the region [
4]. It has also improved ecosystem services, by preventing carbon loss from vegetation and soil, subsequently enhancing carbon stocks and carbon sinks [
23,
24].
The wetland ecosystems of arid regions are an indispensable part of the broader terrestrial ecosystems. However, in arid regions, wetland SOC of restoration and conservation projects have not been well-studied and estimating SOC from space remains a challenge. Previous study areas include the arid regions around Xinjiang Ebinur Lake [
25], the Baotou section of the Yellow River wetland [
24], and the Central Asia arid zone of the Aral Sea [
26]. There are relatively few studies on the construction of wetland biomass types and soil inversion models, soil carbon and sink, together with the relationship between spatial distribution and environmental factors. Therefore, a comprehensive analysis and evaluation of arid wetland soil is still required [
27]. In the context of global climate change and increasing carbon emissions, there is an urgent need for carbon sequestration and carbon sink studies for wetland restoration and conservation projects.
The Ningxia Basin of the Yellow River is part of the arid region of Northwest China. Since 2002, a series of wetland ecological restoration and conservation projects have been successively carried out in the Ningxia Basin of the Yellow River, such as the restoration of farmland from the lake and dredging operations. Additionally, various other projects have improved connectivity in smaller water systems. The wetland areas and the coverage of vegetation have continuously increased. These projects have successfully protected the regional environment and have restored degraded wetland ecosystems, countered wetland shrinkage in the region, and brought substantial increases in vegetation biomass since 2010. This has brought ecological, social, economic, and environmental benefits.
There are two national wetland nature reserves, five national wetland parks, and an international wetland city (Yinchuan City), which was accredited in the first batch of international wetland cities on 25 October 2018. Ningxia Yellow River Irrigation District was included in the World Cultural Heritage sites in October 2017. However, the total carbon sequestration benefit arising from the wetland restoration projects has not yet been systematically evaluated for the Ningxia Basin of the Yellow River. The carbon sequestration achieved in some individual projects has been investigated [
28]. The potential of wetlands to store carbon, especially in soils, often exceeds that of other terrestrial ecosystems in the arid regions of the upper Yellow River. Related studies which assessed wetland soil carbon stocks have been neglected to date. Nonetheless, as these systems are sensitive to global climate change and represent a significant proportion of the global carbon stocks, they require urgent attention [
17].
The Ningxia Basin of the Yellow River has long been affected by complex and multi-variable factors such as water and sediment changes, river channel swings and sand invasion, human activities, and climate change. It exhibits the characteristics of a diverse wetland, distinct spatial differentiation, growing wetland area, and fragile ecosystems. It forms a unique irrigation and drainage system [
20] and has extremely high scientific research value. However, most research on the Yellow River wetlands in the Ningxia arid area has focused on dynamic changes [
5], ecological assessments [
29], biodiversity surveys, and ecological restoration [
29]. There are few studies on wetland soil carbon sequestration, soil sinks, and their temporal and spatial dynamic evolution, and more emphasis on the construction of single factor inversion models. The remote sensing (RS) images and parameters selected are different in the constructed model. The inversion objects are mostly single species of the same type, but the inversion accuracy needs to be improved. It is difficult to accurately estimate wetland SOC in such complex areas using single-factor models; therefore, either or both a multi-factor and multi-characteristic variables model are required [
28]. This will improve the accuracy and usefulness of developing a SOC inversion model using a combination of spectral information, texture features, principal components, vegetation indices, and field measurements.
In our study of large-scale carbon content in ecosystems, the estimation of SOC by remote sensing (RS) shows the advantages of this method, including a high degree of accuracy. At the same time, it shows how this method overcomes the difficulties associated with obtaining ground station data. The main methods include RS information parameters for fitting biomass, and a combination of RS data and process models, artificial neural network model methods, and reference plot methods [
11]. The combination of “3S” technology and actual field data can solve the problem associated with organic carbon estimation from point to region [
19,
20,
21]. The combination of spectral information, texture features, vegetation indices, and measured data to establish a SOC estimation model can promote the precision of SOC estimation.
The status of wetland carbon storage is especially important in the global carbon cycle and carbon balance accounting. Nevertheless, the current spatial information of regional wetland carbon storage is extremely lacking and is easily overlooked when conducting large-scale assessments and establishing models. To a certain extent, it hinders the correct estimation of wetland carbon storage.
In our study, we explored the potential of texture features, a vegetation index, and principal component analysis (PCA) of multi-source remote sense data for estimating SOC. We also evaluated the contribution of wetland ecology restoration and protection projects to C sinks from 2000 to 2015. The spatial distribution rules of SOC were analyzed in the Ningxia Basin of the Yellow River in early (2000 to 2005), intermediate (2010), and recent (2015) years during the restoration stages of the wetland ecosystem. The objective of this research was to estimate SOC in wetlands using a combination of texture features, vegetation index, PCA (principal component analysis), vegetation biomass, and soil factor detection data.
Five specific objectives were established to accomplish this:
- (1)
Identify the best texture features, vegetation index, and PCA to use for estimating SOC;
- (2)
Establish the regression models between texture features and/or PCA, vegetation index, and field-measured SOC;
- (3)
Assess the potential and accuracy of estimating SOC using multi-factor biomass prediction models including texture features and/or PCA;
- (4)
Estimate the net SOC density of a wetland ecosystem by comparing the changes between the years of the project from 2000 to 2005, an intermediate year (2010), to a recent year (2015);
- (5)
Estimate the wetland project-induced carbon contribution against the total ecosystem soil carbon sequestration by comparing the change in ecosystem carbon stocks between the carbon sink of the project region and a reference area (the carbon sink of the reference area was equivalent to the average soil carbon density of the wetland areas of the Ningxia Basin of the Yellow River during the same period).
The wetland SOC was then used to construct a spatial distribution map, which was analyzed to explore environmental factors. The model and approach used in this study will be a useful reference for future studies on the effect of climate change, the Yellow River, and human activities in large river-basin areas. The research results have important theoretical and practical significance for deepening the study of wetland formation. These results will help improve the monitoring methods in wetland ecosystems, conserve freshwater ecosystems, and protect rare, endangered birds along with other wildlife.
2. Materials and Methods
2.1. Study Area
The climate type is continental temperate arid, with an annual average temperature of 9.0 °C and rainfall of 180–200 mm. The potential annual evaporation is 1825 mm, the average annual humidity is 55%, and the Palmer Drought Severity Index is 7.8–8.0.
Located in the arid inland region of Northwest China and the upper Yellow River, the Ningxia Basin of the Yellow River (37°46′ N to 39°23′ N and 105°5′ E to 106°56′ E) has a fragile ecological environment that is sensitive to climate change. The Yellow River runs from south to north through the Yinchuan Plain for 193 km and crosses canals and ditches that provide water for irrigation and extensive wetlands. The soil types are swamp, saline–alkali, and silty soil.
The biodiversity is high. The dominant vegetation types are Cattails, Phragmites australis, and Suaeda glauca. Phragmites australis are widespread over the Ningxia Basin of the Yellow River, and account for 70–90% of the vegetative cover; many other types of vegetation are widely distributed. There are 9 wetland vegetation types, 30 sub-types, and 132 groups in the Yinchuan Plain. There are 202 species of 119 genera, and 52 families of vascular plants that live in the wetland plant resources. There are 29 families and 67 genera in phytoplankton, mainly distributed in the Yinchuan, Wuzhong, and Shizuishan areas on both sides of the Yellow River.
The Ningxia Basin of the Yellow River consists of river, lake, marsh, and constructed wetlands. In 2002, the Ningxia government implemented a series of wetland ecology restoration and protection projects, such as Wuzhong Yellow River National Wetland Park, Huangsha Gudu National Wetland Park, Yuehai National Wetland Park, Mingcui Lake National Wetland Park, Xinghai Lake National Wetland Park, Shahu Lake National Wetland Nature Reserve, and Qingtongxia Reservoir National Wetland Nature Reserve. In addition, various other restoration projects have addressed connectivity in smaller water systems. These projects have successfully countered wetland shrinkage in the region and brought substantial increases in vegetation biomass since 2010. Such increases have provided ecological, social, economic, and environmental benefits.
2.2. Image Data Acquisition
The RS image data used in this study were determined from thematic mapper (TM) images (resolution 30 × 30 m) acquired on 17 July 2000, 18 July 2005, and 27 July 2010. The Landsat-8 OLI (Operational Land Imager) images (resolution 30 × 30 m) were acquired on 28 July 2015. Images taken during the wet season were chosen to ensure comparability. The image data met the requirements of this study as it has shown the characteristics needed to determine vegetation indices, texture features, and principal components.
2.3. Sample Layout and Experimental Data
Applying the OLI-8 image in 2015, seven Wetlands Restoration and Conservation projects (including the five National Wetland Parks and the two National Wetland Nature Reserves) were selected in the Ningxia Basin of the Yellow River. Sampling plots were evenly distributed throughout the wetland restoration and conservation projects areas. Field measurements were conducted in the seven project areas from August 2014 to 2015 to obtain wetland vegetation data and soil monitoring data, which were essentially the same as the RS image.
Furthermore, thirteen representative sample sites were selected within each project area using a grid selection method; therefore, a total of 91 sample areas were identified. Three sample squares (1 × 1 m) were then randomly selected in each sample area. During this time, a total of five sample strips and 273 effective sample points and sample squares were arranged, including 60 river sites, 75 lake sites, 87 marsh sites, and 51 sample sites in the constructed wetland. The sample strips were laid from the center of the lakes and marshes to the lakeshore area, and from the Yellow River channel to both sides. This was done to understand the changes through time in wetland vegetation biomass, as shown in
Figure 1.
Sample sites were located using the same GPS device. A selection of vegetation was harvested at each sample plot. The vegetation type, coverage, plant height, and the number of plants in the sample were recorded, and then the aerial parts of the plants were collected and immediately weighed to obtain the fresh weight. All sampled vegetation was put in polyethene bags, and were accordingly marked, sealed, and taken to the laboratory. They were then dried at 85 °C to reach a stable weight (accuracy of 0.01). The plant’s soil, and its salt, nutrient, and water conditions directly affect the distribution of plant species and their biomass.
Therefore, we chose three replicate sample spots for soil sampling based on gathered plant samples. Soil organic carbon (SOC), total nitrogen (TN), total phosphorus (TP), salinity, soil pH, soil bulk density, and soil water content were determined from the total of 273 soil samples taken in the field within the sampling quadrats. These were based on collected vegetation samples, which were according to the concentration of the distribution of fabric roots in the study area, as well as referring to the method of stratified sampling of wetlands in the arid area [
25].
In each sample, soil columns of 0–40 cm were collected using a cylindrical metal sampler (3 m in length and 5 cm in diameter). Each soil column was divided into four layers (0–10, 10–20, 20–30, 30–40 cm) along the diagonal of the sample and drilled into the soil. After which, the same soil layers were evenly mixed, placed in a sealed and numbered sample bag, and sent back to the laboratory. Then, the samples were air-dried, passed through a 60-mesh sieve, sealed, and placed in a small refrigerator in the laboratory.
At the same time, the
ring knife method was used to take another four-delamination cut soil column samples from each quadrat [
30]. These samples were placed into sealed bags for soil bulk density (BD) analysis. Each soil sample was analyzed for SOC and soil total nitrogen (STN), which was determined using the Elementar Vario MACRO analyzer. The total soil phosphorus (STP) content was determined using a UV-visible spectrophotometer (UV-2450, Shimadzu, Japan). Soil samples were oven-dried at 105 °C at a constant weight so that the BD and water content could be determined from the mass change. Total salts were measured using a conductivity meter (DDS-307A, Rex Shanghai, China), and soil pH was measured using an acidity meter (PHB-5, Beijing Xin Yu Tengda Instrument Equipment Co. Ltd., Beijing, China) [
30]. The detection methods were strictly carried out in accordance to the industry standards (organic matter, NY/T1121; nitrogen, LY/T1228-1999; phosphorus, HJ 632-2011; water-soluble salt, NY/T1121.16-2006; pH, NY/T 1377-2007; soil moisture content, NY/T 52-1987).
2.4. Data Processing Method
2.4.1. RS Image Preprocessing
The RS image data were geometrically corrected, cut out, mosaicked, merged, strengthened, radiated calibration, and corrected for atmosphere applying the ERDAS 8.0 software. The corrected error was 0.5 pixels. After pretreatment, Band 2, Band 3, and Band 4 of the OLI images and Band 3, Band 4, and Band 5 of the TM images were selected for use. The RS images were manually pictured for the study region using the ENVI 5.2 and ArcGIS 10.2 software packages to incorporate terrain, water, and locations of wetland projects in the Ningxia Basin of the Yellow River. Locations and ranges of wetland projects in 2000, 2005, 2010, and 2015 were determined from the longitude and latitude of GPS field investigation and from manually obtained geographic indicators, which is more precise than an automatic abstraction.
2.4.2. Characteristic Variables from Landsat-8
The extracted RS information consisted of band reflectivity, textural features, and vegetation indices. The Landsat-8 OLI RS image has 11 bands, of which Band 2, Band 3, and Band 4 were sensitive to wetland vegetation. In the Landsat TM image, Band 3, Band 4, and Band 5 were sensitive to vegetation. Landsat Band 4-3-2 combination and OLI-8 Band 5-4-3 combination images were close to the natural color, had rich information, good layering, less interference, and clear texture. Therefore, the contrast between vegetation and non-vegetation on the image was increased, which improves the accuracy of vegetation identification. Therefore, Band 2, Band 3, and Band 4 of the Landsat-8 OLI, and Band 3, Band 4, and Band 5 of Landsat TM were selected for the reflectance extraction of each sample.
According to the characteristics of the study area, seven vegetation indices of NDVI (normalized difference vegetation index), RVI (ratio vegetation index), DVI (difference vegetation index), MSVI (modified soil-adjusted vegetation index), OSAVI (optimized soil-adjusted vegetation index), SAVI (soil adjusted vegetation index), and RDVI (renormalized difference vegetation index), three principle component analyses of PC1, PC2, and PC3, and eight texture features of CON (contrast), COR (correlation), DI (dissimilarity), EN (entropy), HOM (homogeneity), ME(mean), ASM (angular second moment), and VAR (variance), were selected as indicators for constructing the inversion model.
2.4.3. Meteorological Data Collection and Calculation
Using data from 10 meteorological stations in the Ningxia Basin of the Yellow River, the Kriging space interpolation method was used to calculate the annual mean temperature and annual precipitation of each sampling point in ArcGIS 10.2 software. Differences in significance were analyzed using one-way ANOVA, and the correlation between environmental factors (climate, vegetation, and soil physical and chemical properties) and SOC density was studied using correlation analysis. The effects of environmental factors and human activities (tourism in tourist wetlands such as Shahu Lake, Mingcui Lake, and Yuehai Lake, which implement wetland management measures) on SOC density were studied using stepwise regression analysis. Tourism data were obtained from the wetland administration and tourism bureau.
Firstly, RS factors were extracted from RS images. Then, the correlation between each of the RS factors and environmental factors and SOC was analyzed. Finally, factors with high correlations were used as independent variables, and the SOC of the sample was used as the dependent variable. A stepwise regression method was used for constructing a remote sensing multivariate linear regression model (RS-MLRM) for soil organic carbon in the wetlands. The optimal model was selected based on the change of decision coefficient and the decision coefficient of adjustment. In order to analyze the impact of human activities on the soil carbon content of wetlands in the Yellow River and Ningxia Basin, this study introduces a tourism intensity index (the number of tourists per unit area), and uses a regression analysis method for a fitting analysis of tourism intensity index, and soil physical and chemical properties index [
25,
29]. The number of tourists data were obtained by collecting the number of tourists received in each wetland administration. RS images were processed and mapped in ArcGIS 10.2 and ENVI 5.2.
2.5. SOC Estimation RS-MLRM
2.5.1. Correlation Analysis between SOC and Environmental Factors
Pearson correlation coefficients were calculated between wetland SOC and extracted RS information, as well as between the SOC and environmental factors. The most relevant independent variables were then selected to provide a basis for further modeling. This process was designed to improve upon conventional methods, which continue to have large prediction errors. According to previous research [
13,
29], 18 independent variables were chosen to test for use in the model.
2.5.2. RS-MLRM Mathematical Model
MLRM (multiple linear regression model) was first proposed to solve economic problems. MLRM was to analyze an independent variable and the relationship between multiple dependent variables, and then the mathematical models would be figured out.
Let y be the dependent variable, and mean SOC;
x1,
x2, …,
xn as independent variables, representing the remote sensing factor. When the independent variable is linear with the dependent variable, the MLRM mathematical model is [
31]:
Equation (1) for the regression equation, where b
0 was a constant term, b
1, b
2 …, b
n for the regression coefficient, ε is a random error. b
0 + b
1×1 + b
2×2 + … + b
m×n + ε was provided with m group of samples, where
xij was the observed value of
xi in the first
i. The mathematical model was expressed as:
The matrix form of the biomass MLRM was expressed as:
y was the SOC matrix, x was the remote sensing factor matrix for each sample point, b was the coefficient matrix, and ε is the random error.
The SOC density (g/m
2) formula used is as follows:
where SOC is soil organic carbon content (g/kg), H is soil thickness (cm) of 40, SBD is soil bulk density (g/cm
3).
2.5.3. Accuracy Evaluation and Inspection of the Model
Based on the results of previous studies [
32,
33], we randomly selected 75% of the data from 271 field-observed soil plots to construct the SOC estimation model, and the remaining 25% of the data were used to evaluate the prediction accuracies of the RS-MLRM. For testing of the fitting accuracy and prediction ability of the RS-MLRM, we used the coefficient of determination (R
2), the adjusted R
2 (adj. R
2), relative root mean square errors (RMSEs), precision system errors (SEs), relative errors (RE%), and actual SOC (SOMS) to evaluate the accuracy of each SOC RS-MLRM [
31,
33], using Equation (4) through (8), respectively:
where SOC
i is the measured SOC value of the i sample, and SOC
i’ is the predicted SOC value of the i sample,
is the average SOC per unit area of the different types of wetland soil, and S
i is the area of wetland soil.
2.5.4. SOC Estimation Mapping
We estimated the wetland SOC in 2000, 2005, 2010, and 2015 using the optimal RS-MLRM estimation model, i.e., the one with the highest R2 and lowest RMSE. Then, a SOC level distribution map was made in the Ningxia Basin of the Yellow River.
2.5.5. Integrated Process for Model Construction
The results of the SOC remote sensing estimation model of four types of wetland soil in rivers, lakes, swamps, and artificial wetlands are presented in
Figure 2.
2.6. Soil Carbon Sink, Carbon Source, and Carbon Sink Control Area Delineation
According to the high-precision carbon sink measurement (known as the pool-difference approach), which is preferred by the Intergovernmental Panel on Climate Change National Greenhouse Gas Inventory, we used a two-year average annual change to represent the change of carbon sink (IPCC) (2014) [
34]:
where ΔSCS is the annual soil carbon sink change ((g/m
2) a
−1), SCSt
2 is the SOC density (g/m
2) at time t
2, and SCSt
1 is the SOC density (g/m
2) at time t
1.
When ΔSCS is positive, it indicates that the surface SOC density was growing in the past 15 years and is classified as a carbon sink area. When ΔSCS is negative, it indicates that the surface SOC density has been in a reduced state in the past 15 years and is classified as a carbon source area. When the ΔSCS value is zero, it indicates that the surface SOC density has been in a constant state for the past 15 years and is classified as a carbon balance area. The geochemical map of the SOC density at each spatial coordinate point can be used to visually express the spatial distribution of organic carbon in the study area. The statistical results are shown in
Figure 3.
Using the reclassification tool in ArcGIS 10.2 Spatial Analysis Module, the spatial distribution of the four types of wetlands’ SOC density were plotted from 2000 to 2015. Using the average soil carbon stock density (2669 g/m
2) of the Ningxia Basin of the Yellow River wetland in the four years, and the average carbon density of the Yellow River irrigation area in Ningxia (2144 g/m
2) [
15], we referred to previous research results [
11,
13,
19] and the spatial distribution of SOC density in the wetlands of the Ningxia Basin of the Yellow River from 2000 to 2015.
The soil carbon stock density was divided into five levels, namely the lower level (SOC density ≤ 1000.0 g/m
2), low level (1000.0 g/m
2 < SOC density ≤ 2000.0 g/m
2), intermediate level (2000.0 g/m
2 < SOC density ≤ 3000.0 g/m
2), high level (3000.0 g/m
2 < SOC density ≤ 4000.0 g/m
2), and higher level (SOC density > 4000.0 g/m
2). The SOC density distribution range, number of pixels, area, and percentage for each level were calculated. The carbon sink and carbon source area distribution maps were drawn. They were used to visually express and analyze the trend of soil carbon sinks in the past 15 years, and the spatial distribution characteristics of carbon sinks and carbon source areas and carbon balance areas. Referring to previous research [
35], the lower level area was considered to be a low carbon sink, which was used as the carbon sink control area. The higher and high-level areas were high carbon sinks, which represented carbon sink conservation areas. The other level areas were the general carbon sinks.