2.1.

Study Area

Poyang Lake, the largest freshwater lake in China, is located at 115°47 to 116°45’E and 28°22’ to 29°45′N on the southern bank of the Yangtze River (Fig. 1). The lake is fed primarily by five tributaries (Ganjiang, Fuhe, Xinjiang, Raohe, and Xiushui Rivers) and is connected to the Yangtze River at Hukou. Poyang Lake has a subtropical monsoon climate with an average annual temperature of 17.6°C and mean annual precipitation level from 1450 to 1550 mm, with the rainy season generally occurring in summer. Interactions among the hydrology, soil, and plants of Poyang Lake have formed a unique wetland ecosystem, which provides essential functions, such as water supply, floodwater storage, and biodiversity maintenance. Poyang Lake wetlands are home to 102 species of aquatic plants from 38 families and to 122 species of fish from 23 families. More than 280 bird species are also available, representing 12 genera and 51 families, including 50 rare species.

Fig. 1

Location of the PLNNR and the distribution of field sampling points in 2016. (a) Ganjiang River delta, (b) Sidu Island, (c) Dachahu sublake, and (d) Dahuchi sublake.

The PLNNR is located northwest of Poyang Lake, at the intersection of Ganjiang and Xiushui Rivers (Fig. 1). The PLNNR, with an area of $224{\mathrm{km}}^2$, was established in 1988 to preserve wintering birds.⁴⁷ Twenty-three threatened species in the International Union for Conservation of Nature and Natural Resources red list⁴⁸ were found in the PLNNR, and approximately 95% of the entire population of critically endangered Siberian cranes (Grus leucogeranus Pallas), nearly 80% of endangered Oriental storks (Ciconia boyciana), and over 70% of vulnerable white-naped cranes (Grus vipio) wintered in the PLNNR.^49,50 For these reasons, Poyang Lake was one of the first wetlands to be included in the Ramsar Convention’s List of Wetlands of International Importance.⁵¹

Complex inflow, outflow, and backflow patterns lead to large seasonal water level fluctuations.⁵² The plant distribution of the PLNNR wetland, accompanied by the fluctuating water level, is characterized by a typical concentric pattern along the elevation gradient from the lake to the shoreland.⁵³ Four main types of plants are abundant in the PLNNR wetlands, namely, Carex cinerascen, Phalaris arundinacea, Miscanthus sacchariflorus, and Artemisia selengensis. They form three belts, namely, bulrushes (Miscanthus sacchariflorus or Phragmites australis communities), sedges (Carex cinerascen or Artemisia selengensis communities), and sparse emergent vegetation (Phalaris arundinacea communities), which occur naturally along a moisture gradient from the higher lands to the lake shoreline. Wetland vegetation in the PLNNR is distributed in different types of bottomlands, which are often inundated during flood season. These bottomlands include the littoral land of the main lake, inflow river delta, sublakes detached from the main lake during autumn and winter, and small islands seasonally submerged during summer. From October to December, water levels are low, thereby exposing the areas of these vegetation communities in Poyang Lake. At this time of the year, emergent vegetation (e.g., Miscanthus sacchariflorus) experiences a heading stage and subsequently withers and dies; sedges (e.g., Carex cinerascen), which have a long growing period, continue blooming; and sparse emergent vegetation (e.g., Phalaris arundinacea and Artemisia selengensis) begin to wither and die. During this period, the spectral characteristics and AGB of these vegetation communities do not change considerably. Thus, this type of phenology phenomenon necessitates the implementation of a field work that covers all four dominant vegetation communities in the PLNNR.

2.2.

Field Surveying and Data Collection

The field campaign was conducted on November 23 to 30, 2016. We selected four typical bottomlands, which are representative of PLNNR wetland for sampling, and a total of 94 sampling points, which covered the four main vegetation communities in Poyang Lake wetland (Fig. 1 and Table 1). Then, in view of the concentric pattern of vegetation communities along the elevation gradient, a predetermined fixed number of $1\mathrm{m}\times 1\mathrm{m}$ sample plots at each bottomland were created from the shoreline to the relatively higher land, where flood cannot overflow. The interval between plots is 50 to 120 m (in accordance with the distribution of slop and vegetation belts in the sites) to cover all the main vegetation communities at different elevations in various types of bottomlands (Fig. 1). Once the sample plot was located, we recorded its geographic coordinates and elevation through GPS (Trimble) with an accuracy of 0 to 0.20 m for position and 0.10 m for elevation. Then, all plant types in the plot were identified, recorded, and excavated. All dead materials were removed from clipped plants and fresh biomass was measured immediately using a digital scale. Then, the average fresh AGB per plot was calculated from these measurements ($n=3$). The Landsat 8 image, which was acquired from USGS,⁵⁴ with 11 bands (bands 1 to 7 and 9 to 11 with a spatial resolution of 30 m and band 8 with a spatial resolution of 15 m) from December 16, 2016, covering the PLNNR, was used to complete the study.

Table 1

Number of sample points in every sample field in 2016.

2.3.

Data Preprocessing and Preparation

Image preprocessing, executed via ENVI 5.2, included geometric, radiometric, atmospheric corrections, and spatial subsets. On the basis of the georeferenced images of Poyang Lake, the root-mean-square errors (RMSEs) in the image registration were ensured at $<0.3$ pixel for the seven images. The FLAASH atmospheric correction module, a feature of ENVI 5.2, was used to finish the atmospheric correction. NDVI, SAVI, EVI, and the second band from the Kauth–Thomas transformation⁵⁵ were added to the Landsat 8 OLI image with 7 multispectral bands by layer stacking to create an 11-band layer-stacked Landsat image. Then, the layer-stacked image from December 16, 2016, was used to extract 94 spectral sampling points, based on geographical coordinates recorded by GPS in preparation for image classification. The proposed methods are briefly explained in the flowchart (Fig. 2).

Fig. 2

Flowchart used to map AGB in Poyang Lake using Landsat images.

2.4.

AGB Model Methods

We selected eight variables, namely, NDVI, SAVI, EVI, B3 (red band), B4 (NIR band), greenness (the second band from the Kauth–Thomas transformation), B6 (SWIR1, short-wave infrared 1 band), and B7 (SWIR2, short-wave infrared 2 band), as inputs. The descriptions and computational formulas of four VIs in this study are shown in Table 2. The variable values of 94 sampling points were extracted in accordance with their geographic coordinates. The effectiveness of LR, BPNN, SVR, and RF models in estimating AGB in the PLNNR was evaluated. Then, we utilized the trained models with the highest testing accuracy to map AGB in the PLNNR. We used spectral features as predictors to improve the accuracy of the models.

2.4.3.

BPNN model

BPNN has a good generalization capability,^57,58 and it consists of input, hidden, and output layers, including their nodes and activation functions. The main mathematical expression of BPNN is as follows:

Eq. (3)

$$y_j=f({\sum }_i^n{w}_{ji}{x}_i+b_j),$$

$x_i$ is the $i^{\prime }$th node value of the previous layer, $y_j$ denotes the $j^{\prime }$th node value of the present layer, $w_{ji}$ represents the weighted value connecting $x_i$ and $y_j$, $n$ refers to the number of nodes in the previous layer, and $f$ indicates the activation function. The BPNN model is discussed in further detail by Buscema.⁵⁹ Levenberg–Marquardt algorithm was used to determine the weighting and bias matrices for each iteration. We selected a bagging method ($n$_estimators:400, max_samples:0.2) to ensure the stability and robustness of the trained model.

Figure 3 shows the structures of the three models for estimating the AGB of Poyang Lake.

Fig. 3

Structures of: (a) RF, (b) SVR, and (c) BPNN models for estimating the AGB of the PLNNR.

Table 2

Descriptions and computational formulas of four VIs in this study.

VI	Description	Calculation
NDVI	A remote sensing index that reflects the state of land cover vegetation. It is defined as the quotient of the difference and sum between the reflectance of the NIR and visible light channels	$\mathrm{NDVI}=\frac{{\rho }_{\mathrm{NIR}}-{\rho }_{\mathrm{RED}}}{{\rho }_{\mathrm{NIR}}+{\rho }_{\mathrm{RED}}}$
SAVI	An SAVI based on NDVI and a large number of observations to reduce soil background effects	$\mathrm{SAVI}=\frac{({\rho }_{\mathrm{NIR}}-{\rho }_{\mathrm{RED}})*(1+L)}{{\rho }_{\mathrm{NIR}}+{\rho }_{\mathrm{RED}}+L}$
EVI	An optimized VI that increases sensitivity to high-biomass areas and improves vegetation monitoring by decoupling canopy background signals and reducing atmospheric effects	$\mathrm{EVI}=\frac{2.5*({\rho }_{\mathrm{NIR}}-{\rho }_{\mathrm{RED}})}{{\rho }_{\mathrm{NIR}}+6*{\rho }_{\mathrm{RED}}-7.5*{\rho }_{\mathrm{BLUE}}}$
Greenness	The second component from the Kauth–Thomas transformation

2.4.4.

Accuracy assessment

In this study, we implemented all four modeling methods through packages in Python: scikit-learn.^60,61 Considering that we do not have enough sample points for every year, we divided the 94 sample points from the field survey in 2016 into two parts: training (80%) and testing (20%) datasets. We used three criteria, namely, RMSE, coefficient of determination ($R^2$), and mean absolute error (MAE), to evaluate the performance of these models in predicting AGB. RMSE [Eq. (4)] is a standard metric for measuring the discrepancies between the simulated and actual AGB values; however, it is easily influenced by outliers.⁶² Therefore, MAE [Eq. (5)] is suggested to be used with RMSE for determining the variations of errors in the model.⁶³ $R^2$ [Eq. (6)] is utilized to determine the collinearity between the predicted and observed AGB values. RMSE and MAE values close to 0 and an $R^2$ value close to 1 indicate that the model is an accurate predictor:

Eq. (4)

$$\mathrm{RMSE}=\sqrt{\sum _{i=1}^n\frac{{(y_{is}-y_{it})}^2}{n}},$$

Eq. (5)

$$\mathrm{MAE}=\frac{{\sum }_{i=1}^n|y_{is-}{y}_{it}|}{n},$$

Eq. (6)

$$R^2=1-\frac{{\sum }_{i=1}^n{(y_{is-}{y}_{it})}^2}{{\sum }_{1=1}^n{(y_{is-}{\overline{y}}_{is})}^2}.$$

Fig. 4

Performance of the four models in estimating AGB in the testing dataset: (a) LR, (b) SVR, (c) RF, and (d) BPNN.

$y_{is}$ is the $i^{\prime }$th simulated AGB value, $y_{it}$ denotes the $i^{\prime }$th real AGB value among the tested sample points, ${\overline{y}}_{is}$ represents the mean simulated AGB for all tested sample points, and $n$ indicates the size of tested samples. The RMSE, $R^2$, and MAE in Tables 3 and 5 have an average value after fivefold cross validation.

Table 3

RMSE (kg m−2), R2, and MAE (kg m−2) values of the four models for estimating AGB in the training and testing datasets.

Model	Training dataset			Testing dataset
	RMSE	R2	MAE	RMSE	R2	MAE
LR	0.49	0.39	0.36	0.31	0.53	0.26
SVR	0.25	0.84	0.2	0.27	0.64	0.22
RF	0.3	0.71	0.31	0.25	0.70	0.21
BPNN	0.47	0.39	0.34	0.29	0.59	0.23

3.1.

Comparison of the AGB Simulation Accuracy of Various Models

Table 3 presents the specific value for the three criteria of four machine learning algorithms. Among the 76 sample points from the training dataset, the SVR had the lowest RMSE ($0.25\mathrm{kg}{\mathrm{m}}^{-2}$) and the highest $R^2$ (0.84), and it also performed best in MAE ($0.20\mathrm{kg}{\mathrm{m}}^{-2}$), which was considerably lower than that of the other three models. RF had the second best performance in training dataset: RMSE ($0.30\mathrm{kg}{\mathrm{m}}^{-2}$), $R^2$ (0.71), and MAE ($0.31\mathrm{kg}{\mathrm{m}}^{-2}$). The BPNN and LR were similar in magnitudes of RMSE, $R^2$, and MAE. For the 18 sample points from the testing dataset, the RF model showed a better generalization capability than the SVR (RMSE, 0.25 versus $0.27\mathrm{kg}{\mathrm{m}}^{-2}$; $R^2$, 0.70 versus 0.64; MAE, 0.21 versus $0.22\mathrm{kg}{\mathrm{m}}^{-2}$). Finding an overfitting problem that occurs in the SVR is easy. Although the predicting capability of BPNN and LR in the training dataset almost had no difference, the $R^2$ (0.59) of BPNN is considerably higher than that of LR (0.53) in the testing dataset, indicating that BPNN had a relatively better generalization capability in this study. Figure 4 shows that the deviation between the simulated and actual values of RF had a relatively even distribution compared with the other models. From the scatter plot, the prediction values of the BPNN and LR had a relatively more dispersed distribution around the fitting line than those of the RF and SVR, indicating that these two models had worse stability of predictions. The least accurate model was the LR, with the highest RMSE of $0.31\mathrm{kg}{\mathrm{m}}^{-2}$ and the lowest $R^2$ of 0.53 in the testing dataset. No discernible difference among the generalization capabilities of the three models except LR existed. Thus images with high spatial and spectral resolution might be essential to improve modeling accuracy. We concluded that RF was a slightly better model for predicting wetland AGB in the PLNNR that the other models.

Fig. 5

AGB density distribution in the PLNNR.

3.2.

Predicting AGB in the PLNNR

We utilized the most accurate model, namely, RF, for exploring the AGB distribution in the PLNNR (Fig. 5). The maps show that the AGB density is between 0 and $1.973\mathrm{kg}{\mathrm{m}}^{-2}$. On the whole, a higher than average AGB value occurred in the north part, including the Ganjiang River delta and Banghu sublake, whereas the south part, including bottomlands at the Dahuchi sublake experienced relatively low AGB values.

4.1.

Accuracy and Uncertainty of the Study

This study presents for the first time that a landscape-scale remote sensing model of the AGB for seasonal lake wetland in floodplain areas has been developed on the basis of machine learning algorithms and Landsat images. We used RMSE to compare the predictive performance of the RF model to that of other models (Table 4). During our remote sensing analysis of wetland biomass, we concluded that simple LR had the worst performance regarding simulation effects, whereas the further advanced machine learning algorithm performed better with regard to RMSE. In our study, the RF model had a $0.21\mathrm{kg}{\mathrm{m}}^{-2}$ RMSE value in the testing dataset, which is considerably lower than the mean level of $\sim 0.3$ to $0.5\mathrm{kg}{\mathrm{m}}^{-2}$. Researchers might select different input variables; thus, the final simulated results would be affected by the randomness of sample points and the species types in wetlands. However, we cannot ignore that the RF model is useful and effective for predicting AGB in wetlands.

Table 4

Comparison of the accuracy of various models for simulating AGB in wetlands.

The time inconsistency of remote sensing imaging and sampling may cause the error on AGB inversion. For example, the December 16 scene is the only appropriate Landsat image closest to our field survey time in 2016. The atmospheric conditions of imaging time can affect the gray value of each pixel, thereby resulting in a heterogeneous gray value of the same vegetation type on the ground in different times. Thus, the training model generalization capability will be reduced.

4.2.

Implications of the Input Variables for Modeling AGB

The rank of a feature used as a decision node in a tree can be utilized to assess the relative importance of that feature with respect to the predictability of the target variable. Features at the top of the tree greatly influence the final prediction decision of a large fraction of the input samples. Thus, the expected percentage of the samples that they contribute to can be used to estimate the relative importance of the features. By averaging the expected activity rates over several randomized trees, one can reduce the variance of such an estimate and use it for feature selection. Fig. 6 shows the results of applying RF with least squares loss and 500 base learners to the AGB in Poyang Lake wetland. Plot (a) shows the training and testing errors at each iteration. Plot (b) shows the feature importance, which can be obtained via the feature importance property. Details on how the RF determined the importance of variables are discussed by Genuer et al.⁶⁸ When the number of trees is close to 450 or more, the test dataset error is $\sim 0.2$, which is the minimum value. Thus, we select 450 as the best $N$ of the RF model in this study. Furthermore, the top four most important features were NIR, SWIR1, EVI, and red band. NIR performed best among all the variables, followed by SWIR1. NDVI was not the best predictive factor for estimating AGB in the PLNNR, possibly because it is considerably more affected by Poyang Lake wetland’s complex environmental background conditions than the other modified VIs.

Fig. 6

Relative importance of all input variables identified using the RF model in this study.

There is no doubt that EVI, SAVI, NDVI, NIR, and red band are highly correlated. However, in general, the models would be further effective for estimation if their input variables were independent of each other. Thus, we conducted an experiment to determine whether using the NIR alone as the input variable of these four models would produce further accurate results for estimating AGB in the PLNNR. Table 5 shows the specific performance. The RMSE, $R^2$, and MAE values changed remarkably in the training and testing datasets. The RMSE values for the training dataset increased by $0.11\mathrm{kg}{\mathrm{m}}^{-2}$ on average (RF, 0.15; SVR, 0.12; BPNN, 0.07; and LR, 0.03), whereas the $R^2$ values decreased. In the testing dataset, a slight improvement was observed in the LR. The RMSE and MAE decreased by $0.01\mathrm{kg}{\mathrm{m}}^{-2}$ and $R^2$ increased by 0.02, indicating that LR could not cope with the collinearity of variables and could not efficiently extract other variables’ useful information. The RMSE and $R^2$ values of the other three models were increased by more than $0.05\mathrm{kg}{\mathrm{m}}^{-2}$ and reduced by 0.18 on average (RF, 0.22; SVR, 0.19; and BPNN, 0.13). Therefore, we insisted that the RF and SVR had better capability for processing further complicated information than the other methods. We concluded that placing these VIs into models together, which may decrease the influence of environmental background to some extent, improves the capability of the models for measuring AGB in the PLNNR. This improvement surpassed the influence of the collinearity of variables to some extent.

Table 5

RMSE (kg m−2), R2, and MAE values (kg m−2) of the four models for predicting AGB in training and testing datasets using NIR as input variable.

4.3.

Limitations of Predicting AGB Using Optical Remote Sensing Data such as Landsat

The complexity of species composition and the density of vegetation in wetland areas present a huge challenge for remote sensing.¹ In fact, VIs calculated from broadband sensors will rapidly approach a saturation level when the AGB estimation is limited by the asymmetrical nature of the relationship between the AGB and VIs calculated from medium-spatial-resolution (10 to 100 m) multispectral sensors using NIR and red bands. Therefore, the RF model is likely to overestimate biomass at low observed values and underestimate biomass at high observed values, which may explain why errors are associated with high biomass values. Despite these limitations, our findings showed that the Landsat NIR band was sensitive to the AGB in the PLNNR. Recent efforts have been geared toward using narrow band VIs from hyperspectral data or WorldView-2 (eight bands including red edge band and 2-m spatial resolution) to estimate high canopy density biomass.^15,23,69,70 Results from these studies have shown that modified VIs calculated from the red edge and NIR shoulder domains can more accurately estimate biomass at a full canopy cover than the standard red/NIR indices.^2,23 A reasonable explanation for this finding is that the indices calculated from the red edge are more sensitive to vegetation properties, such as canopy biomass and chlorophyll content, than that from other regions of the electromagnetic spectrum. A slight change in vegetation properties could result in a shift in the red edge curve, and NIR can minimize the influence of the atmospheric and water absorption as well as the soil background. However, the use of fine spatial and spectral resolution sensors ($<5\mathrm{m}$ and $>100$ bands) for estimating AGB is limited by the cost, availability, and complexity of processing high-dimensional data.^70,71 This technology may be widely used in the future when costs decrease.