2.1.

Test Site

The test site is located in Poyang Lake wetland, Jiangxi province, China. The latitude and longitude are 115°47′ to 116°45′E, 28°22′ to 29°15′N (Fig. 1). The climate is characterized as a subtropical, humid monsoon climate with 1620 mm mean annual precipitation and about 17°C annual average temperature. Poyang Lake exhibits large interannual variations in water level. In summer, it is the largest freshwater body in China and extends up to $3500{\mathrm{km}}^2$ by the end of the rainy, wet season (June to September). In the dry season (November to April), Poyang Lake can be less than $1000{\mathrm{km}}^2$ in extent, with only several wandering water courses remaining. In 2007, the fluctuation of the water level was up to 8.37 m (Fig. 2). In the dry season, wetland vegetation emerges above water and starts to grow rapidly from early spring, with the aboveground biomass reaching the highest level in April.¹⁹ In the wet season, wetland vegetation is flooded and hardly grows except at the lakesides with higher ground level. Therefore aboveground biomass is relatively low. In November, the water recedes and vegetation starts to grow again. In general, the biomass dynamics present positive correlations with variations of water level.

Fig. 1

Spatial location and extent of the study area of Poyang Lake, Jiangxi province, People’s Republic of China, with the sample points for April and November 2007.

Fig. 2

Water level variation of Poyang Lake in 2007.

The predominant vegetation in Poyang Lake is mostly carex and reed, which account for $>90\%$ of the vegetation coverage.¹⁹ The biophysical properties of vegetation vary in different hydrological stages from April to November. In April, plants grow rapidly with green leaves up to 50 to 70 cm in length. By July, plants are submerged under water. In November, they turn to senescent stage with low water contents.

2.2.

SAR Data

Four scenes of ENVISAT ASAR alternating polarization (AP) precision image (PRI) mode data, with range and azimuth pixel spacing of 12.5 and 12.5 m, were collected over this area in three periods of 2007: April, July, and November (Table 1). The ASAR images were radiometrically corrected with the BEST toolbox provided by the European Space Agency. All images were georeferenced with respect to a topographic map. The data in April were composed of two different scenes (April 4 and 6, 2007), each covering part of the test site, and mosaicked together. These two scenes have the same HH/VV polarization and incidence angle (39 to 43 deg), with different overpass directions.

Table 1

ENVISAT ASAR data used in the study.

2.3.

Field Measurements

Field measurements of biophysical parameters of wetland vegetation standing above water were collected during two field campaigns coinciding with satellite overpasses in April and November. Both time periods are in the growing stages of wetland vegetation. Taking the accessibility into account, 46 (in April) and 45 (in November) sampling sites with with intersite spacings $>50\mathrm{m}$ (four times the pixel size) were randomly selected over the study area. Due to the different ground status in April and November, caused by the flood from July to September (Fig. 2), there are few sample points with the exactly same locations for these two field campaigns. Ground data, including plant water content, aboveground biomass, and plant height, were collected during the field campaign. For biomass measurements at each sampling site, we clipped the total standing biomass above water or above ground within an area of $0.5\times 0.5{\mathrm{m}}^2$. The sites are covered with water or very wet soil. The clipped samples were weighed in situ and oven dried (at 120°C for 24 h) to calculate the wet and dry total biomass (WTB and DTB, respectively). The DTB is referred to as biomass in the following sections. The weight of plant water per unit area was calculated as the difference between WTB and DTB.

The plant height was measured three to five times at each sampling site using a meterstick, then the mean value was calculated as the truth data. According to ground data, the DTB was correlated with plant height, with a correlation coefficient of 0.65 in a 90% confidence interval (Fig. 3).

Fig. 3

Relationship between dry biomass and plant height from the ground measurements.

3.1.

Canopy-Scattering Model and Backscatter Simulation

The Michigan Microwave Canopy Scattering (MIMICS) model²⁰ has been widely used for the tree canopy comprising a crown layer, a trunk layer, and rough-surface ground boundary. The model assumes that total backscatter is a linear composition of the following four scattering components: direct scattering from vegetation canopy, backscatter from multiple-path scattering between surface and vegetation canopy, double-bounce trunk–ground interactions, and backscatter from the ground surface.

Because the MIMICS model requires so many input parameters, some assumptions must be made to use the model. For ENVISAT ASAR data, the radar’s parameters (frequency and look angle) are known. For wetland applications in our study area, there is no woody layer in herbaceous vegetation but only two layers: ground surface and grass canopy. Thus, we modified the MIMICS model input files, turning off the parameters about trunk, primary branch, and secondary branches. The simulated backscatter included ground-surface scattering, multipath scattering between the ground surface and vegetation canopy, and canopy volume scattering. Because the study area is wetland with high soil moisture (especially between April and October), the ground is mostly flooded, covered by water. By changing the settings, the parameters about ground parameters (soil and snow) are turned off, so there is no backscatter from the ground surface. Finally, the parameters about radar (frequency and look angle), leaf (gravimetric moisture, dry density, number density, diameter, thickness, and temperature), ground standing water (salinity and temperature), and crown (layer height) were used by the MIMICS model as input parameters. Some of the parameters were treated as constant or replaced by the average of survey data, e.g., standing water salt content, dry density of leaf material, leaf thickness, temperature for vegetation and standing water (Table 2). Wetland vegetation consists primarily of carex, grass, and reed and was modeled as a cluster of vertical dielectric cylinders that were uniformly distributed.²⁰

Table 2

Major input parameters and corresponding values of MIMICS model used for backscatter simulation.

In flooded wetlands, radar backscatter is mostly attributed to vegetation properties. With the modified MIMICS model, we simulated backscatter coefficients responding to increases in a set of vegetation biophysical parameters, including plant height [Fig. 4(a)] and plant water content [Fig. 4(b)]. Figure 4(c) shows that the incidence angle has little impact on backscatter at all polarizations (HH, VV, and HV) owing to the limited double-bounce in the absent trunk layer in wetland vegetation. Due to the dense coverage of vegetation, the total $C$-band backscatter is mainly from canopy volume scattering and canopy–ground multiple interaction,²¹ which are not strongly affected by the variation of incidence angle. The incidence angle has only a little effect on canopy–ground interaction, but less microwave energy can penetrate the dense canopy layer to reach the ground and be reflected by the ground to produce canopy–ground interaction. Put simply, the incidence angle has little impact on the backscatter due to the dense canopy layer and the absence of the trunk–ground and canopy–ground interactions. Figure 4(c) agrees with the simulation of the flooded forest at $C$-band.²² Figure 4(d) describes the MIMICS simulation as a function of biomass. Backscatter in all polarizations increases with the increase of biomass.

Fig. 4

Variations in backscatter as a function of height, plant water content, incidence angle, and biomass for wetland vegetation based on predictions from MIMICS model.

After simulating the backscatter coefficient (${\sigma }^0$) using the MIMICS model, we take the results from April 2007 for accuracy evaluation by comparing it with that from ENVISAT ASAR image through the root mean square error (RMSE). The RMSE is 1.86 dB for VV polarization and 1.42 dB for HH polarization, which is smaller compared to the actual dynamic range (6.84 dB for VV polarization, 4.52 dB for HH polarization). Figure 5 also shows that the coefficient of determination ($r^2$) between the backscatter coefficients from the ASAR images and simulated by the MIMICS model is 0.9271 for VV polarization and 0.94961 for HH polarization.

Fig. 5

Comparison between the backscatter coefficient (${\sigma }^0$) from ENVISAT ASAR image and simulated by the MIMICS model for April 2007.

3.2.

Generating Training Data for the Neural Network

Training data, as prior knowledge of the network, are very important and mainly determine the accuracy of the training results. To estimate aboveground biomass of wetland vegetation, the training data need to satisfy the following conditions:²³ (1) the training data should consist of a wide range of biomass from different plant growth stages; (2) the change of backscatter with biomass, height, plant moisture content, and system factors should be consistent with ground measurements and ASAR images; and (3) highly correlated data should be removed to reduce data redundancy and improve simulation accuracy and efficiency.

Training data meeting the above conditions could not be easily obtained from ground measurements owing to limitations in time, labor, and large random errors introduced in field. In this study, we apply the MIMICS model to simulate backscattering coefficients with a set of system factors (polarization, incidence angle, etc.) and biophysical parameters (plant height, water content, biomass, etc.). Training data between biophysical parameters and backscattering coefficients are thus generated. Here we first determine the range of the training data based on ground measurements to cover all biomass levels in the study area (Table 3). Then, a set of 50 training data pairs were generated using the MIMICS model, with each pair matching biomass, plant water content, and height to HH and VV backscattering coefficients, respectively (Fig. 6). The HV polarization is not considered because it is not recorded in the ASAR imagery.

Table 3

Range of the training data used in the neural network.

Fig. 6

MIMICS model for training data generation.

3.3.

Training the Neural Network

Figure 7 presents the topology of the neural network used in this study. The neural network is a one-hidden-layer back propagation network with two input elements (HH, VV) and three output elements (biomass, plant height, plant water content). There are eight neurons in the hidden layer. The activation function of each input element in the hidden layer is a sine function, and a logistic function of the output element is defined as

Fig. 7

Neural network training model.

The NNA is trained with the data generated from the MIMICS model. The HH and VV backscatter can also be extracted from the ASAR images acquired in April, July, and November. Therefore, the neural network could be applied to estimate biomass in the study area in an inverted process.

4.1.

Biomass Maps of the Poyang Lake Wetland

Biomass distributions of Poyang Lake wetlands in April, July, and November 2007 were mapped using ENVISAT ASAR data and MIMICS-fed neural network analysis (Fig. 8). These maps clearly depict the development phases of wetland plants with the changes of water level of the lake. In April, water level starts to rise but most of the wetland is visible. The vegetation is mainly distributed in the southern areas of the lake. In July, the water level is about 17 m, close to the peak level for this year. Therefore most areas of the lake are flooded except the shorelines on higher ground in the south. In November, the lake reaches the lowest water level, which results in the smallest area of water body. Therefore most areas of the lake are covered with wetland vegetation.

Fig. 8

Abovewater dry biomass maps of Poyang Lake derived from ENVISAT ASAR data; black areas in November present no ASAR data coverage.

Table 4 summarizes the biomass level distributions in the three periods. In April, total dry biomass of the wetland was $1.09\times {10}^9\mathrm{kg}$. In the middle of the growing season, 23.66% of the biomass in April is at the level of 200 to $500\mathrm{g}{\mathrm{m}}^{-2}$ and 39.27% at the level of 500 to $800\mathrm{g}{\mathrm{m}}^{-2}$. In July, the percent of high biomass level increases because the plants grow to maturation in this period. However, with higher water level, most of the wetland is flooded, so the total biomass decreased to $1.86\times {10}^8\mathrm{kg}$. In November, as plants wither and leaves dry up, the high-level biomass starts to decrease. About 72.68% of the biomass was in the range of 200 to $800\mathrm{g}{\mathrm{m}}^{-2}$. The total biomass in this period was nearly the same as that in April. The average biomass in the three periods was nearly consistent. In April and November, the wetland areas are widespread but most plants have low biomass. In July, plants grow very well and reach high biomass, although most of the area is flooded.

Table 4

Biomass level distribution and statistics in April to November 2007.

4.2.

Accuracy Assessment

Ground data collected in April and November were used to examine the accuracy of NNA-estimated biomass. Figure 9 shows the scatterplots between ground truth and estimated biomass in April [Fig. 9(a)] and November [Fig. 9(b)]. The estimated biomass in July is not evaluated due to the lack of ground measurements in this period. When all data samples are considered, the $r^2$ between ground-measured and estimated biomass in April was 0.718, and in November, 0.5579. The intermediate level of biomass between 400 and $800\mathrm{g}{\mathrm{m}}^{-2}$ appeared to be better simulated, whereas the high and low biomass estimation results were not satisfactory. According to the MIMICS simulations, vegetation in high biomass levels ($>1800\mathrm{g}{\mathrm{m}}^{-2}$) has backscattering values similar to those of vegetation at intermediate biomass, resulting in underestimation in the high biomass level. Another reason for the underestimation is that the NNA has been trained for biomass values in the range of 125 to $1300\mathrm{g}{\mathrm{m}}^{-2}$ (as shown in Table 3), so the NNA approach cannot estimate biomass above this value. In future work, the NNA should also be trained for higher biomass. In the two scatterplots in Fig. 9, there are a few sample points in the top left that demonstrate apparent overestimation, which may result from the low plant density in April and November. The total backscatter was strongly influenced by surface backscatter from wet soil ground, which was not taken into account in the modified MIMICS model.

Fig. 9

Comparison between ground-measured and estimated biomass in April and November 2007.

The overall RMSE of biomass estimation in the study area was $141\mathrm{g}{\mathrm{m}}^{-2}$ in April and $104\mathrm{g}{\mathrm{m}}^{-2}$ in November. In the intermediate biomass level, it reached an RMSE of 117 and $91\mathrm{g}{\mathrm{m}}^{-2}$. This indicates the potential of the NNA approach combined with a MIMICS herbaceous wetland scattering model in large-area, multitemporal biomass estimation.