Variations of net ecosystem CO2 exchange in a tidal inundated wetland: Coupling MODIS and tower-based fluxes


  • Yan-Er Yan,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    2. Also at Department of Geography Sciences, Chongqing Normal University, Chongqing, China.
    Search for more papers by this author
  • Hai-Qiang Guo,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    Search for more papers by this author
  • Yu Gao,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    Search for more papers by this author
  • Bin Zhao,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    Search for more papers by this author
  • Ji-Quan Chen,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    2. Also at Department of Environmental Sciences, University of Toledo, Toledo, Ohio, USA.
    Search for more papers by this author
  • Bo Li,

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    Search for more papers by this author
  • Jia-Kuan Chen

    1. Coastal Ecosystems Research Station of the Yangtze River Estuary, Ministry of Education, Key Laboratory for Biodiversity Science and Ecological Engineering, Institute of Biodiversity Science, Fudan University, Shanghai, China
    Search for more papers by this author


[1] Tidal activity is a major factor determining the distribution of plant species and ecosystem functions, including carbon fluxes. To explore the spatial variations of net ecosystem CO2 exchange (NEE) and related regulatory mechanisms along the tidal inundation gradient (i.e., middle or low tidal flat), an NEE estimation model using piecewise regression analysis was developed by coupling the Moderate Resolution Imaging Spectroradiometer (MODIS)- and tower-based measurements. The results showed that our model achieved an adequate NEE estimation (slope = 0.70, R2 = 0.78). Then the model was applied to estimate NEE variation along a transect with a tidal inundation gradient. The average NEE was −1.75 g C m−2 d−1, varying from −2.02 g C m−2 d−1 to −1.42 g C m−2 d−1 from island- to oceanside. Generally, our empirical model captured the spatiotemporal patterns of NEE and the variation of the regulatory factors along the gradient. The sensitivity analysis of various regulatory variables showed that the variations of NEE near the islandside were primarily caused by seasonal shift and annual cycle of vegetation, whereas at the oceanside, NEE was more influenced by tidal activity with no clear phenological influence. In the middle area, NEE seemed to be subjected to both phenological changes of vegetation and tidal activity. In conclusion, this study illustrates that the estimates derived from MODIS- and tower-based flux data are reliable for quantifying the spatiotemporal variations of NEE and reflecting the effect of tidal activity on NEE.

1. Introduction

[2] Tides, among all of the regulatory variables, play fundamental roles in regulating ecosystem composition, structure, and functions in coastal and estuarine wetlands. Numerous studies have shown that distributions of plant and animal species are associated with the magnitudes and frequency of tidal activities [e.g., Crain et al., 2004; Hays, 2007]. Tidal inundation at ocean-land interfaces is determined by the topography and forms a clear gradient from water to land. By comparing the tide table with the ground measurements, we found that the ecosystem carbon fluxes and storage (i.e., biomass), nutrient availability and use, and species composition in coastal Chongming Island, Shanghai, China, were highly correlated [Chen et al., 2008; Guo et al., 2009; Liao et al., 2008].

[3] The eddy covariance (EC) method, a direct measure of CO2 flux between ecosystem and the atmosphere, is generally restricted to locations with relatively flat terrain and homogeneous vegetation. Spatial and temporal variability in the net ecosystem exchange of carbon, water, and energy of a variety of ecosystems have been reported in recent years [e.g., Dunn et al., 2007; Hui et al., 2003; Saigusa et al., 2005]. For tidal inundated wetlands, the physical requirements of large and flat land surfaces can usually be met, but gradual changes of water level can play an important role in determining the net ecosystem CO2 exchange (NEE) and its dynamics over time [Yan et al., 2008; Guo et al., 2009]. Laine et al. [2007], for example, reported that NEE varied spatially with water level in a lowland blanket bog. Ideally, one could establish multiple towers to detect the changes along the tidal gradient, but the high cost and maintenance hinder this application.

[4] The two EC towers at our study sites provided us with temporal variations of NEE and their relationship with water level due to tidal inundation. However, the tower-based flux data are insufficient to provide full understanding of heterogeneous coastal landscapes where an obvious tidal inundation gradient exists [Yan et al., 2008]. One solution is to take the spatial coverage of remote sensing products that provide continuous observations of land surface properties [Running et al., 1999], but at much lower temporal resolutions. By coupling tower-based measurements and remotely sensed data, one could address the variations of NEE at broader spatial and temporal scales. Using this approach, Wylie et al. [2007] estimated the NEE of grasslands in the Northern Great Plains using multiple-year data sets from EC towers and Système Probatoire d'Observation de la Terre (SPOT) satellite images. Yamaji et al. [2008] explored the spatial patterns of carbon sequestration within the deciduous broadleaf forests over Japan by coupling EC towers and the Moderate Resolution Imaging Spectroradiometer (MODIS). Xiao et al. [2008] estimated NEE for the United States by coupling MODIS and Ameriflux data covering a range of vegetation types, including forest, shrubland, savanna, grassland, and cropland. In this study, we used a similar school of thought to understand the NEE variations along a tidal inundation gradient by coupling tower-based NEE and MODIS product datasets.

[5] Our study was performed in an estuarine tidal inundated wetland in Chongming Island near Shanghai, China. We aimed to explore the effects of tidal inundations on NEE using continuous records from two EC towers and MODIS products. Data from the two EC towers, installed in 2004 with contrasting inundating frequency, were used along with MODIS products to examine the empirical regulatory mechanisms. Our objectives were to (1) seek new methods for quantifying NEE in tidal inundated wetlands by coupling observed NEE from the EC towers and MODIS data and (2) provide estimates of NEE at the landscape level so that the tidal effects could be assessed quantitatively. We hypothesized that the tidal inundations had significant effects on NEE through changing ecological processes.

2. Data Sets and Methods

2.1. Study Area

[6] Chongming Island is the largest estuarine alluvial island in the world located in the Yangtze River estuary. The climate is characterized by an annual precipitation of 1022 mm and an annual average temperature of 15.3°C. Our study area is located in Dongtan on the east side of island (121°50′ ∼ 122°05′E, 31°25′ ∼ 31°38′N) (Figure 1). The tidal movement of this area is regular and semidiurnal, with maximum and average tide heights of 4.62–5.95 m and 1.96–3.08 m, respectively [Zhao et al., 2008]. The vegetation at the time of our study was dominated by Scirpus mariqueter, Phragmites australis, and Spartina alterniflora. There formed a distinct vegetation zone from water interface in the inland: uncovered mudflats → S. mariqueter-dominated community → S. mariqueter and S. alterniflora mixture → S. alterniflora-dominated community → S. alterniflora and P. australis mixture [Zhao et al., 2009]. The average plant height was 1.5–2.5 m for P. australis and S. alterniflora and 0.3–0.7 m for S. mariqueter.

Figure 1.

Six selected Moderate Resolution Imaging Spectroradiometer (MODIS) pixels in our study area, where the two eddy covariance towers (small triangles in cells 952 and 954) were installed in 2004; the background is a Landsat TM image taken in July 19, 2004.

2.2. Data Source

[7] Two types of data were obtained in this study: NEE from two EC towers and explanatory variables derived from MODIS (

2.2.1. Tower-based NEE

[8] Two EC flux towers were installed in August 2004 in Dongtan on Chongming Island (Figure 1). The first tower was erected at the low elevation (S tower: 31°31.014′N, 121°58.300′E) with P. australis, S. alterniflora, and S. mariqueter as the dominant species, and the other was located at a high elevation (D tower: 31°31.000′N, 121°57.643′E) with P. australis and S. alterniflora dominating the community. The horizontal distance between the two towers is 1100 m. The EC sensors were mounted on a 5 m tall tower above the ground with a 3-D sonic anemometer and an infrared gas analyzer (IRGA) mounted at 3 m above the vegetation (i.e., with an empirical fetch of 500 m). The flux data were recorded on CR5000 dataloggers at 10 Hz, and 30 min averages were calculated for further analyses. Data from January 2005 to December 2006 were included in this study.

[9] The computation of covariance of vertical wind speed and concentration of CO2 was performed using the EC_Processor software package (version 2.2;∼anoorme/), including two-axis rotation, sonic temperature correction due to humidity and pressure changes [Schotanus et al., 1983], and density effects due to water vapor fluctuations [Leuning, 2004; Webb, 1980]. Furthermore, faked flux due to warming of IRGA at low temperature was also corrected for 30 min fluxes [Burba et al., 2008].

[10] The measured NEE and auxiliary micrometeorological data were subjected to several steps of quality control, mainly following the work by Noormets et al. [2007, 2008] Atmospheric stability and stationary flux during well-developed turbulence conditions were used as filters, and the gaps mainly happened during periods of dew and precipitation and poorly developed turbulence. The long gaps were caused by calibration of IRGA and power failure. Following these screenings, about 56% and 54% of data were remained at sites D and S, respectively.

[11] To obtain an estimation of annual carbon sequestration, gaps were filled with a dynamic parameter model [Noormets et al., 2007, 2008]. Briefly, filtered nighttime data of NEE and air temperature were first used to fit an Arrhenius equation [Equation (1)] from Lloyd and Taylor [1994] with a reference temperature (Tref) of 283.16 K. The estimated parameters were then used to estimate daytime ecosystem respiration (RE):

equation image

where RE,ref is the ecosystem respiration rate at Tref (μmol m−2 s−1), Ea is the activation energy (J mol−1), R is the gas constant (J mol−1 K−1), and TK is the air temperature (K). Through fitting the filtered daytime NEE and photosynthetically active radiation (PAR) (μmol PPFD m−2 s−1) with the Michaelis-Menten function [Equation (2)], gaps of daytime NEE were filled:

equation image

where α is apparent quantum yield or maximum light use efficiency (μmol CO2μmol PPFD−1), PPFD is density of PAR, GPPmax is maximum gross primary productivity (GPP) (μmol m−2 s−1), and Re is mean daytime ecosystem respiration (μmol m−2 s−1). Then GPP was calculated as daytime RE minus daytime NEE. Evaluation of performance of this gap-filling model can be found in the work of Moffat et al. [2007]. In short, the gap-filling approach provided satisfactory results compared with other approaches. Considering the periodicities of NEE in this wetland and also sufficient data for modeling, the gap-filling approach was launched at one month interval.

[12] The half-hourly NEE values were used to calculate daily NEE and then averaged into 8 d NEE means to match the composited MODIS data.

2.2.2. MODIS Data

[13] We downloaded MODIS tile h28v05 (cells 952–957) for 2005 and 2006 (Zhao et al. [2009] for cell selection) (Table 1). Among the cells, 952 and 954 overlapped with the footprint of our two EC towers.

Table 1. Specifications of Moderate Resolution Imaging Spectroradiometer Cells in Tile h28v05 of Product MOD09A1 Used in This Study
CodeGridLatitude (°E)Longitude (°N)
952952, 203531.517121.960
953953, 203531.517121.965
954954, 203531.517121.970
955955, 203531.517121.975
956956, 203531.517121.980
957957, 203531.517121.985

[14] To estimate NEE, we obtained the MODIS product MOD09A1 for surface reflectance at a spatial resolution of 500 m and then calculated four vegetation indices (VIs), including the normalized difference vegetation index (NDVI), the enhanced vegetation index (EVI), and two kinds of the land surface water index (LSWI). LSWI and LSWI (2100) provided similar indications for water content, except that the former uses MODIS band 6 and the latter uses MODIS band 7. We estimated NEE for each 500 m × 500 m cell across the line transect of Dongtan from island- to oceanside, covering 500 m × 3000 m for each 8 d interval during the 2005–2006 period.

2.3. Methods

2.3.1. Piecewise Regression Models

[15] The piecewise regression is similar to decision trees, except that it uses a multivariable linear regression function at each leaf instead of discrete constant output [Chen and Mynett, 2004]. The regression tree is the core of piecewise regression, and it splits the parameters' domain into subdomains and produces a linear regression function at each subdomain. Therefore, piecewise regression in the whole domain is nonlinear. This method has been proven to be not only more effective than simple techniques such as linear regression but also easier to understand than neural networks [Huang and Townshend, 2003], leading to its wide use in estimation and factor exploration of ecosystem carbon flux. Following Yamaji et al. [2008] and Xiao et al. [2008], we adopted the piecewise regression model for NEE estimation in this study, with an emphasis on the seasonal changes of NEE as well as its regulatory mechanisms.

2.3.2. Explanatory Variables

[16] NEE of an ecosystem is the difference between GPP and RE. GPP is primarily influenced by incoming solar radiation, air temperature, vapor pressure deficit, soil moisture, nitrogen availability at leaf level, and leaf area index (LAI), as well as canopy phenology at ecosystem level [Chapin et al., 2002; Chen et al., 1997]. RE is the sum of autotrophic respiration (Ra) and heterotrophic respiration (Rh). Empirically, Ra can be regarded as the function of temperature and tissue carbon (foliage, stem, and roots), and Rh is often modeled as the function of soil temperature and moisture as well as substrate availability [Ryan and Law, 2005]. Remote sensing-based surface reflectance provides the vegetation information and can be used to estimate NEE [Wylie et al., 2003]. For example, a series of remote sensing-derived vegetation indices have been used to calculate a number of parameters such as LAI, land surface temperature (LST), and the fraction of photosynthetically active radiation absorbed by vegetation canopy (fPAR). In this study, the standard product of Terra-MODIS LST was not used, because the surface temperature of coastal estuarine wetlands is strongly influenced by tide water. Additionally, the 1 km spatial resolution of LST does not match other MODIS products at 500 m resolution. Similarly, MODIS LAI and fPAR were also excluded because of their coarse resolutions (i.e., 1 km). Instead, we used VIs to represent LAI and fPAR. Our previous work indicated that the lateral material flux driven by tide activities in estuarine wetlands cannot be neglected and can be well explained by tidal height and evapotranspiration [Yan et al., 2008]. Consequently, LSWI and LSWI (2100) were included in this study to reflect the effect of water content on NEE. Finally, we selected all single bands of land surface reflectance (bands 1–7) to derive NDVI, EVI, LSWI, and LSWI (2100).

[17] A total of 230 data pairs of observed NEE and MODIS data were used to develop our NEE estimation model. These data were split into a training set (70%, 163 samples) and a test set (30%, 67 samples) through systematic random sampling. Mean squared error (MSE), the root mean squared error (RMSE), and the correlation coefficient of determination (R2) were used to assess model quality.

2.3.3. Sensitivity Analysis

[18] To examine the uncertainties of our predictions, we performed sensitivity analysis of all important variables after model development. Specifically, we performed the analysis by adjusting the main explanatory variables by 50%. The impact of these adjustments on our estimates of NEE and the associated coefficients of sensitivity (CS; percentage change in NEE value per unit change in an explanatory variable) are presented.

3. Results

3.1. Model Development and Evaluation

[19] MSE decreased exponentially with the number of regression trees (Figure 2). There appeared a threshold (n = 10) of optimal trees at which the estimation accuracy remained constant. Among all explanatory variables, EVI, LSWI (2100), and LSWI were identified as the statistically important factors in estimating NEE, with R2 of 1.0, 0.95, and 0.92, respectively. Therefore, EVI was selected as the first splitting point to discriminate vegetation growth status, and then band 5 and LSWI (as water index) were selected to quantify the water table (Figure 3). We used the top regression tree of 7 and found that the estimated NEE matched the observed NEE fairly well (slope = 0.70, R2 = 0.78 for the training set; slope = 0.74, R2 = 0.75 for the testing data set) (Figure 4). The RMSEs of the model were 0.18 g C m−2 d−1 (the training data set) and 0.29 g C m−2 d−1 (the testing data set). Overall, it appeared that the model slightly underestimated NEE, especially in summer months.

Figure 2.

Changes of average squared error variation with the number of regression trees.

Figure 3.

The criteria of regression trees for estimating subpixel net ecosystem CO2 exchange (NEE). NEE is the dependent variable; land surface reflectance (bands 1–7) and their derived normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), land surface water index (LSWI), and LSWI (2100) are independent variables.

Figure 4.

Scatterplot of observed NEE versus estimated NEE derived from the piecewise model. (a) The training data set: y = 0.704x − 0.48, R2 = 0.78; (b) the testing data set: y = 0.737x − 0.55, R2 = 0.75.

[20] The analysis of NEE residuals, combined with all data at sites D and S, indicated that the residuals were not randomly distributed around zero (Figure 5). The RMSE in ∼January–April was low (∼1.00 g C m−2 d−1), whereas that in ∼July–August was high and often exceeded 1.50 g C m−2 d−1, and the RMSE in 2006 was greater than that in 2005. In comparison, the RMSE of site D did not show a significant difference between 2005 and 2006, whereas the RMSE of site S in 2006 was greater than that in 2005, especially in summer months.

Figure 5.

(a) Scatterplot of predicted NEE versus residuals (observed–predicted), and (b) the temporal changes of residuals over the study period (2005–2006).

[21] We found some differences between estimated and observed NEE at the two sites over the study period. At site D, the average observed and estimated NEE were −2.00 g C m−2 d−1 and −2.40 g C m−2 d−1, respectively, and the average RMSE was −0.03 g C m−2 d−1. At site S, the average observed NEE, estimated NEE, and RMSE were −1.80 g C m−2 d−1, −1.58 g C m−2 d−1, and 0.21 g C m−2 d−1, respectively. However, the estimation accuracy of site S was lower than that of site D (Figure 6). The estimation accuracy of 2005 was higher than that of 2006. On the basis of the simple linear regression between the observed and estimated NEE, R2 values at sites D and S were 0.83 and 0.71, respectively.

Figure 6.

The cumulative NEE obtained from eddy covariance towers and MODIS in 2005 and 2006. (a) Site D; (b) Site S.

3.2. Sensitivity Analysis

[22] We achieved an adequate NEE estimation from both training data at sites D and S. However, it was uncertain when the model was applied to estimate NEE variation along the transect, likely due to the different environmental variables. Cells 952–955 were covered by vegetation, and cells 956–957 had no vegetation. Consequently, to determine the robustness of NEE estimates, sensitivity analysis was done by adjusting the values of explanatory variables (bands 1–7, NDVI, EVI, LSWI, and LSWI (2100)).

[23] The absolute values of CS ranged from 0.006 to 0.383 for the total NEE value (Table 2). Among all the regulatory variables, LSWI (2100), LSWI, EVI, NDVI, and band 2 appeared as important variables for NEE variation. However, NEE values at cells 956–957 were generally less sensitive to NDVI and EVI, which were more subject to the influence of moisture difference.

Table 2. Percent Changes in Estimated Net Ecosystem CO2 Exchange Value and Coefficients of Sensitivity Resulting from the Adjustment of Explanatory Variablesa
Change in Explanatory VariablesTotalCells 952–955Cells 956–957
  • a

    CS, coefficients of sensitivity; EVI, the enhanced vegetation index; LSWI, land surface water index; LSWI (2100), land surface water index (2100); NDVI, normalized difference vegetation index.

b1 + 50%−2.42−0.048−2.76−0.055−1.77−0.035
b2 + 50%7.360.1478.950.1794.250.085
b3 + 50%0.310.0060.430.0090.070.001
b4 + 50%0.390.0080.510.0100.170.003
b5 + 50%−0.51−0.010−0.26−0.005−0.99−0.020
b6 + 50%−2.03−0.041−1.98−0.040−2.12−0.042
b7 + 50%−1.47−0.029−1.50−0.030−1.42−0.028
NDVI + 50%6.290.1268.220.1642.530.051
EVI + 50%9.640.19312.760.2553.580.072
LSWI + 50%7.670.1938.570.2555.930.072
LSWI (2100) + 50%19.150.38319.730.39518.020.360
EVI − 50%−13.91−0.278−19.71−0.394−2.61−0.052
LSWI − 50%−17.33−0.348−18.22−0.364−15.76−0.315
LSWI (2100) − 50%−38.17−0.763−32.49−0.650−49.24−0.985

3.3. Spatial and Temporal Changes of NEE

[24] Our models captured the expected spatiotemporal variations of NEE (Figure 7). NEE of cells 952–955 showed clear seasonal and annual changes, with minimum values in winter and spring when air temperature was <5°C and vegetation was inactive. As the air temperature increased and vegetation began to grow, the magnitude of NEE (note that as NEE values are negative in growing season, the default NEE hereafter refers to its absolute value) increased to a peak value in summer and then declined into winter. NEE of cells 956–957 showed no obvious trends within the season and remained relatively constant, with a maximum value of −2.80 g C m−2 d−1 (note that this is a similar variation to the 8 day interval tidal flooding). However, significant differences were detected at cell 956. NEE of cell 956 in 2006 had a very small seasonal variation, probably due to growth and reproduction of sparse vegetation in that year. Finally, the differences among the six MODIS cells showed that our estimated NEE fluctuated significantly less near the ocean (site S) than that at the inland (site D), signaling strong correlation with the tidal height at the two sites.

Figure 7.

Daily fluctuation of NEE predicted from the piecewise-regression model.

[25] Because the daily NEE may be affected by other unforeseeable variables (e.g., weather, water level), the monthly average NEE was calculated to explore fluctuation and spatial variation (Figure 8). The monthly NEE of cells 952–955 showed strong seasonal changes and varied substantially across the landscapes. In spring (∼March–May), the growing season started early in mid-March, leading to significant carbon uptake at these cells. In summer (June–August), cells 952–955 showed relatively high NEE values. The NEE values of cells 952–953 were much higher than those of cells 954–955, with the maximum differences in July and August. Between the two study years, the range of NEE variation in 2006 was less than that in 2005. The spatial pattern and magnitudes of NEE in fall were similar to those in spring. In ∼September–November, the NEE value decreased, probably because vegetation began to senesce. In winter, when air temperature was <5°C and vegetation was in inactive, NEE was near zero.

Figure 8.

Monthly variation of NEE predicted from the piecewise-regression model.

[26] The spatial and temporal variations of NEE appeared to have changes similar to regulatory factors. For example, the NEE values of cells 952–953 in December were greater than those of cells 955–954, whereas in January and February, cell 955 had a higher value than cells 952–957.

[27] The seasonal changes of NEE showed that NEE of cell 957 had no clear pattern with the range from −1.00 to −2.50 g C m−2 d−1, with a higher fluctuation of NEE in 2006 than that in 2005. The temporal changes of NEE at cell 956 in 2005 were similar to those of cell 957, but the difference occurred at cells 956 and 957 in 2006. NEE of cell 956 in 2006 increased rapidly from March to July, with a maximum of −4.30 g C m−2 d−1, and then decreased gradually to a minimum value in November until a rebound in December.

[28] We calculated the annual NEE of cells 952–957 (Table 3). The estimated NEE in 2006 was higher than that in 2005, with an annual average of −1.95 g C m−2 d−1 in 2006 and −1.55 g C m−2 d−1 in 2005. Each cell had a similar annual change of NEE, with the minimum value arising at the central cells.

Table 3. Estimated Annual Average Net Ecosystem CO2 Exchange from the Piecewise Regression Model for 2005 and 2006
YearCells (g C m−2 d−1)

4. Discussion

4.1. The Factors Responsible for Spatial Variation of NEE

[29] We have demonstrated that MODIS products have great potentials for estimating NEE and can be used to capture the spatiotemporal variation of NEE driven by hydrological factors in tidal inundated wetlands. The interaction between vegetation and tidal activities seemed to be responsible for the spatial and temporal variation of NEE in our systems. Our results are in agreement with those of Laine et al. [2007], but with tidal activities complicating the changes of NEE in time and space. The vegetation structure in estuarine wetlands is controlled by hydrology [Hughes et al., 1998], including the composition and abundance of plants. The plant communities along the tidal inundation gradient have different levels of photosynthetic capacity [Jiang et al., 2009], which in turn determine the wetlands' NEE. For example, the vegetation in 2006 in cell 954 was dominated by S. alterniflora, P. australis, and S. mariqueter, with 70% of the vegetation cover, whereas cell 952 in 2006 was predominated by S. alterniflora and P. australis, with the vegetation cover of 97%. The differences between cells 954 and 952 correlated well with the NEE values by year and location (Figure 7 and Table 3). The difference in plant phenology can be the important factor leading to the variation of NEE. In our study area, P. australis begins to germinate in early April, grows rapidly in mid-August, and becomes senescent in late November; S. alterniflora emerges in May, flowers from late August to September, and dies away in late December; and S. mariqueter begins to grow in late April and becomes senescent in September [He et al., 2010]. These differences lead to the amplitude variations of NEE (Figures 6 and 7), especially at cells 952–955. Previous studies have shown the effects of the phenology on NEE as a measure of vegetation function [e.g., Falge et al., 2002a, 2002b].

[30] Guo et al. [2009] and James et al. [2008] found that tidal activities have substantial effects on NEE at temporal scales. This study has explored the effects of tidal inundations on NEE at spatial scales (Figure 7 and Tables 2 and 3). The sensitivity analysis showed that NEE was more subjected to moisture difference, especially to moisture loss. For example, the percent change in estimated NEE value increased from 18.02% to 19.70% when the value of LSWI (2100) increased, and it decreased from 49.24% to 32.49% with reduction of LSWI (2100) (Table 2). Compared with cells 952–953 in the island side, the NEE values of cells 956–957 near the ocean seemed mainly to be determined by tidal inundation, whereas cells 955–954 were subjected to both phenological conditions of vegetation and tidal activity. Clearly, tidal activity needs to be considered to estimate the true NEE of these wetland ecosystems.

[31] A land surface without vegetation cover was hardly a CO2 sink. However, the estimated NEE on the basis of the empirical model showed that the average NEE of cell 957 was −2.08 g C m−2 d−1 (i.e., a major CO2 sink in 2006). To determine effects of vegetation indices on NEE estimates at cells 956–957, we set EVI at 0 when EVI was <0.1 according to the long-term vegetation observation by Zhao et al. [2009]. Consequently, we found that the average NEE was −1.78 g C m−2 d−1 for all and −1.96 g C m−2 d−1 for cell 957. This finding suggests that the vegetation cover is not the key factor for cells 956–957, which is in agreement with the result of the sensitivity analysis. Through a chamber-based experiment in the same study area, Yang et al. [2006] found that the low intertidal flat was a carbon sink with an average NEE of −13.23 mg m−2 h−1, although our value was even larger perhaps because of the model error. Yang et al. [2006] referred to this as the balance between sediment and carbonate of pore water. James et al. [2008] reported that the effects of tidal inundation on marsh plants caused a mean reduction of 46% ± 26% in atmospheric carbon fluxes when compared to nonflooded conditions. One possible explanation of our results would be from planktons' photosynthesis brought inland during tidal activities.

4.2. The Performance of NEE Models

[32] The performance of our model for estimating NEE is encouraging for predicting NEE in this tidal-active ecosystem. Compared to the results obtained by Xiao et al. [2008] (R = 0.73), our model predicted NEE well (R2 = 0.78). We found that low NEE values were generally associated with low prediction errors (Figure 5), whereas high NEE values were associated with high prediction errors. This suggests that there exists a saturation to estimate NEE, probably due to the sensitivity of surface reflectance in the peak growing season. Moreover, the complex coverage would decrease the accuracy of NEE estimate (Figure 6).

[33] NEE accounts for a small percentage of GPP and is much more difficult to be estimated because the transient carbon pools and associated respiration are difficult to be estimated [Mahadevan et al., 2008; Running et al., 2004]. The complex biophysical environments in estuarine wetlands would add further difficulties in estimating NEE, including (1) the tidal activities bringing about lateral carbon flux loss in the tidal environments [Yan et al., 2008] and (2) wetlands consisting of permanent and seasonal swamps, leading to extremely high costs in time and labor for acquiring ground measurements.

[34] The estimated NEE, based on the empirical model, showed that cells 956–957 were CO2 sinks. In addition to the tidal effect, the soil characteristics at these cells may also have affected the land surface properties. The addition of water would result in lower reflectance of the near-infrared wavelength, and the frequent changes in water level would amplify the bias in surface reflectance measured by MODIS, which would lead to an overestimation of NEE in the areas. Moreover, our model was developed using observed NEE at only two points in the landscape. These two measurement sites do not represent other landscape features such as bare mudflats, a possible reason for the overestimated NEE at cells 956–957.

[35] We used the explanatory variables of surface reflectance (bands 1–7), NDVI, EVI, LSWI, and LSWI (2100) and did not include other regulatory factors such as vegetation and phenology. Yet the carbon assimilation by photosynthesis is influenced not only by canopy function and energy absorption capacity but also by vegetation age. It is thought that canopy can be characterized by LAI and fPAR in the near future [Myneni et al., 2002], but the vegetation age cannot be measured using remote sensing products in our system. Therefore, LAI and fPAR were simply replaced by VIs due to spatial resolution in our study (i.e., neglected direct canopy characteristics).

[36] Respiration is the second-largest flux (carbon loss) influenced by substrate availability, soil temperature, and soil moisture. The lack of LST in our model would reduce the temperature effects on ecosystem respiration, but the LST of our site is influenced more by tidal activities. Moreover, the lateral carbon flux associated with tidal activities and methane generation under the anaerobic conditions of wetland soils are also important components in quantifying carbon fluxes in estuarine wetlands. These variables, though needed, are difficult to be quantified with remotely sensed data.

[37] Finally, the MODIS 8 day composites can only capture NEE variability at scales of 8 days, suggesting that the variations of NEE at finer temporal scales cannot be acquired in our modeling effort. The exclusion of days with high and low values could cause the underestimation or overestimation of NEE values.

5. Conclusions

[38] Tidal activity would change ecological processes and consequently affect ecosystem functions such as NEE. We captured the spatial and seasonal variations of NEE in tidal inundated wetlands and revealed the regulatory factors of NEE along the tidal inundation gradient. The average NEE values were calculated between −2.02 and −1.42 g C m−2 d−1. The variations of NEE near the island side seemed to be related mainly to the seasonal shift and annual cycle of vegetation, whereas NEE near the ocean was influenced more by tidal activities, and NEE of the middle section along the transect was subjected to both phenological conditions of vegetation and tidal cycle. Our results illustrate that the differences in the spatiotemporal variability of NEE exist in response to tidal inundation gradients and that tidal activity should be included to reevaluate the role of estuarine and coastal wetlands when carbon fluxes are considered.


[39] This work was supported by the National Basic Research Program of China grant 2006CB403305, the Science and Technology Commission of Shanghai grant 07DZ12038-2, the National Natural Science Foundation of China grants 30870409 and 40471087), and the Program for New Century Excellent Talents in University grant NCET-06-0364 funded by the Ministry of Education of China.