Located in the northern part of Jiangxi province, Poyang Lake (28°24′–29°46′ N, E 115°49′–116°46′ E) receives inflow from five major rivers (Ganjiang, Fuhe, Xinjiang, Raohe, and Xiushui) in the south and eventually discharges into the main stream of the Yangtze River in the north (Figure 1). Poyang Lake is geometrically complex with tortuous shorelines and incised bottom morphology, which has been shaped by a combination of lacustrine and riverine morphological processes [40]. Based on historical measures of the lake surface during periods with high water levels in combination with recognition of the extensive levee system (in order to separate artificial wetlands from nature wetlands), a study area covering 3124 km² was determined using GIS (Geographic Information System) techniques.

The seasonality of precipitation (influenced by subtropical monsoons), combined with the smooth relief and other factors, results in significant variations in the Poyang Lake’s inundation area throughout a year [41]. The lake has a maximum area of more than 3000 km² during the wet season from April to September and shrinks to 1000 km² to form a narrow meandering channel from October to March [22].

An extensive wetland developed by the dramatic water level fluctuations maintains high plant richness and diversity, and plays an important role in maintaining ecosystem functions. For example, dominant plants, such as Carex spp. (including C. cinerascens, C. argyi, and C. unisexualis, etc.) and its companion species, P. arundinacea, provide an ideal habitat for migratory birds to spawn, rest, feed, and avoid predators. They also supply local residents with fuel, fertilizer, and grass. Phragmites spp., which are typically associated with Triarrhena lutarioriparia and Artemisia selengensis, make up another important community. Phragmites spp. and Triarrhena spp. provide food and habitat for wintering migratory birds, and their height also protects the migratory birds from human interference [42]. Artemisia spp. is the main forage grass with high food and medicinal value.

Wessels et al. (2004) suggested general land cover patterns (e.g., grassland, deciduous/evergreen forest, and agricultural) could be successfully mapped with MODIS data [43]. In this study, vegetation dynamics of the Poyang Lake wetland were detected with MODIS 250 m imagery (2000–2012). MODIS data from the Terra satellite are collected on a near daily basis and are processed and composited into 16 day values by NASA’s EROS Data Center using 3–5 cloud-free images for each collection interval. The enhanced vegetation index (EVI) data acquired from the Oak Ridge National Laboratory’s Distributed Active Archive Center (DAAC) website are available as georectified and atmospherically corrected Terra MODIS products. These products provide significant refinements in spectral, radiometric, and geometric properties compared to previously available data sets with similar spatial resolution [44]. For example, EVI is less sensitive to soil and atmospheric effects than the widely used normalized difference vegetation index (NDVI) because it incorporates blue spectral wavelengths [45]. As a result, EVI remains sensitive to increases in canopy density beyond where NDVI becomes saturated [46]. Non-rationing vegetation indices, such as the perpendicular vegetation index (PVI) and the green vegetation index (GVI) are generally more linear with less saturation problems, but require external and sensor noise removal in the derivation of surface reflectances that are input to VI computation [47,48]. Current emphasis in the observing system (EOS) era involves operational “external” noise removal through improved calibration, atmospheric correction, cloud and cloud shadow removal, and standardization of sun–surface–sensor geometries with bidirectional reflectance distribution function (BRDF) models. This allows for the introduction of EVI for operational monitoring of Earth’s vegetation. The MODIS EVI is calculated from the following equation: (1) EVI = 2.5 × R NIR − R Red 1 + R NIR + 6 × R Red − 7.5 × R Blue , where R_NIR (near-infrared band: 841–876 nm), R_Red (red band: 620–670 nm) and R_Blue (blue band: 459–479 nm) are atmospherically corrected or partially atmosphere-corrected (Rayleigh and ozone absorption) surface reflectances; the coefficient “1” accounts for canopy background scattering and the blue and red coefficients, and 6 and 7.5 minimize residual aerosol variations.

In addition, 54 scenes of cloud-free Landsat MSS/TM/ETM+ images (30 m × 30 m²) with the same overpass date were selected to validate the lower-resolution MODIS observations. These data were acquired from the Global Land Cover Facility (GLCF). MSS/TM/ETM+ band 3 (red band) and band 4 (NIR band) were used to delineate the relatively accurate vegetation cover.

In this study, the grasslands were delineated using a thresholding segmentation algorithm, which assumes the EVI images are relatively homogeneous regions that can be separated by selecting the appropriate threshold. An optimal threshold between vegetation and non-vegetation features was determined using the generated EVI histogram. In the same way, grassland area was also obtained from the transformed NDVI, which was extracted from Landsat MSS/TM/ETM+ DN values.

Observed daily water levels at four gauging stations (Xingzi, Duchang, Tangyin, and Kangshan) available for the period of 2000–2012 were acquired from the Hydrological Bureau of Jiangxi Province and the Hydrological Bureau of the Yangtze River Water Resources Commission. Xingzi station was selected as the most downstream and the closest to the northern part of the lake, and was considered to best reflect the water level fluctuations in the long and narrow watercourse that flows into the Yangtze River mainstream. Duchang and Tangyin stations are expected to reflect the water level fluctuations in the central part of the lake, which is characterized by a broad area with high vegetation coverage (e.g., Poyang Lake National Nature Reserve (PLNNR) and Nanji Wetland National Nature Reserve (NWNNR)). Kangshan station, located at the most upstream end of the lake, was selected because it is expected to respond primarily to catchment runoff in the southern part of Poyang Lake.

A combined Sen’s slope and MK test approach was applied to analyze trends in the EVI time series. Sen’s slope involves computing the slopes for all temporally ordered pairs of data points and then calculating the median of these slopes as an estimate of the overall slope. Since Sen’s slope is not greatly affected by outliers or single data errors, and missing values are also allowed, it is more rigorous than the commonly used regression slopes and thus provides a realistic measure of the trends in time series [49,50]. But, this approach cannot test the significance of the time series. The MK test is a non-parametric test and does not require the data to be normally distributed. Meanwhile, the MK test has low sensitivity to abrupt breaks due to non-homogeneous time series [51,52]. Therefore, the MK test was used to determine whether the trend in EVI time series was significant.

First, the slopes of all data value pairs were calculated as: (2) Q i = x j − x k j − k in which x_j and x_k are pairs of the EVI series, and j and k are serial numbers, j > k. If there are N values x_j in the time series we get as many as N = n (n − 1)/2 slope estimates Q_i (n is the length of the series, n = 296). Sen’s slope is the median of these N values of Q_i. The N values of Q_i are ranked from the smallest to the largest and the Sen’s estimator is (3) Q = { Q ( N + 1 2 ) ,   N   is odd Q ( N 2 ) + Q ( N + 2 2 ) 2 ,   N   is even  

A positive value of Q indicates an increasing trend whereas a negative value indicates a decreasing trend.

Then the MK test statistic is calculated as (4) S = ∑ i = 1 n − 1 ∑ j = i + 1 n sgn ( x j − x i )   where s g n   ( x j − x i ) = { + 1 ,   x j > x i   0 ,   x j = x i − 1 ,   x j < x i   and n is the sample size. When   n ≥ 10 , the standardized statistics Z for the one-tailed test is formulated as (5) Z = { S − 1 Var ( S ) ,   S > 0 0 ,   S = 0 S + 1 Var ( S ) ,   S < 0   in which   V a r ( S ) = ( n ( n − 1 ) ( 2 n + 5 ) − ∑ i = 1 n t i ( t i − 1 ) ( 2 t i + 5 ) ) / 18 ; n is the sample size, and t_i is the number of ties of extent i. In this study, trend free pre-whitening proposed by Yue and Wang (2002) was applied with significant autocorrelation to eliminate the effect of serial correlation [53].

Due to the pre-observation of MODIS between 1 January and 25 February 2000, a total of 296 EVI time-series images (i.e., 23 images/year × 13 years − 3 images = 296 images) were analyzed in this study. In case | Z | ≤ Z 1 − ∝ 2 = 1.96 for the trend, the null hypothesis of no trend should be accepted at the 95% significance level. When   | Z | > 1.96 , a positive value of Q connotes a significant “upward trend,” while a negative value of Q indicates a significant “downward trend”.

Spatial distribution patterns of vegetation dynamics were analyzed by the Gaussian regression model. The Gaussian regression model is available for evaluating the normal distribution patterns [54]. The model can be expressed as: (6) y = c exp [ − 1 2 ( x − u ) 2 / t 2 ] where y is the relative abundance (number of pixels) of each classification of vegetation dynamics within each elevation gradient; x is the DEM (digital elevation model) data in this paper, derived from the National Geometrics Centre of China; c is the maximum of abundance; u is the peak of elevation, which appears when y equals to c; and the full width of half maximum (FWHM) is 4t.

Sen’s slope estimator and the MK test were used to investigate the spatial distribution trend of EVI for each pixel (Figure 2a). Further, the statistically significant trend of the change rate of EVI was estimated by regressing the time series for each pixel; the resulting slope estimate was used as the EVI change rate (Figure 2b). A decrease in EVI time series was observed in about 11% of the wetlands in the study area with an averaged changing rate exceeding −0.0745 every 10 years (p < 0.05). Analyzed by a Gaussian regression model, significant downward trends in EVI were mainly found at 12.5–15.3 m elevation in the upper wetlands of the PLNNR and NWNNR (Table 1). In addition, more than 34% of the wetlands in Poyang Lake showed a significant increasing trend in EVI, particularly in the bottomlands (elevation between 9.9 m to 12.9 m). Magnitudes of the upward trends ranged between −0.0248 and 0.2716 every 10 years (0.0661 ± 0.0350, p < 0.05).

It has been noted that under a persistent decline in lake water level, the area of the Phragmites community has mainly distributed along levees and on the higher plains of the dish-shaped pit groups, and has been significantly decreased compared to historical evidence. In contrast, the area of the Carex community in the bottomlands has significantly increased [37].

In this section, we first presented the spatial and temporal variability of trends in EVI and their change rates across the whole Poyang Lake wetland. These results strongly support the former studies: the biomass, density and/or greenness of upland vegetation (e.g., Phragmites spp., Triarrhena spp., and Artemisia spp.) declined under drought stress, resulting in a significant downward trend in EVI across these regions. In contrast, positive succession occurred in the bottomlands due to the invasion of hygrophilous plants (e.g., Carex spp. and Phalaris spp.) resulting in a significant upward trend in EVI.

Our assessment indicates that areas with increased EVI were larger than areas with decreased EVI in the Poyang Lake wetland during 2000–2012 (i.e., 1135.80 km² − 361.44 km² = 774.36 km²). Instead, the magnitude of the trend in “upward” regions was smaller than the absolute value in “downward” regions. Degradation and succession of vegetation may lead to changes of grassland area, which are significant to the evolution of wetland ecosystems. Differential responses of grassland area within Poyang Lake wetland to the water level fluctuations during typical hydrological periods still remain unclear.

Figure 3 shows the accuracy assessment results of MODIS-derived grassland area, revealing a high correlation with the Landsat-derived results. The absolute and relative differences between the two datasets are 28.9 km² and 1.7%, respectively. For all of the MODIS-derived grassland area, the comparison with the Landsat values resulted in an average absolute error (AAE) of 193.67 km², a root mean square error (RMSE) of 233.01 km², and an R² of 0.93, indicating the relationships between fine resolution (e.g., Landsat MSS/TM/ETM+) and coarse resolution (e.g., MODIS) are linear and constant over time.

Several pioneering studies have shown that the polynomial regression is qualified to characterize the response of grassland area to water level fluctuations because the quadratic function produced the best fit with the largest correlation coefficient between them [55,56]. The coefficients of x² in quadratic function are related to the slope of the curves: a positive value indicates a parabola going upward (meaning greater change in grassland area at relatively low water levels), and a negative value indicates a parabola going downward (meaning smaller change in grassland at relatively low water levels, but greater at higher water levels).

The submersion—emersion processes of wetlands in different zones of Poyang Lake during the rising and retreating water periods were divided according to the lake’s hydrological and topographic features. Indeed, the timing and duration of inundation directly affect the productivity, biomass, species richness and diversity of wetland plants. We infer that the variations of grassland area have differential responses to the water level fluctuations in different lake regions during typical hydrological periods (Table 2).

During the water rising period (submersion of grassland), water levels gradually increased from north to south, a pattern that was especially dominant in the lake’s main body and nearby estuary delta zones (e.g., PLNNR and NWNNR), resulting in a high correlation between grassland area and water levels at Duchang and Tangyin stations (Figure 4). Because a larger proportion of wetland vegetation is distributed in the very flat lowlands of Poyang Lake, the change slope of grassland area was great at first, and decreased with the rising water levels.

There are a lot of shallow sub-lakes in the middle and northern wetlands of Poyang Lake (relatively far away from the Kangshan gauging station), such as Dachahu (covers an area of 85 km²), Banghu (covers an area of 73 km²), and Dahuchi (covers an area of 30 km²), which account for a large portion of Poyang Lake grassland. These sub-lakes connect to Poyang Lake during periods with higher water levels, and separate from the Poyang Lake during periods with lower water levels. In Figure 5, the R² values are relatively low for Xingzi (R² = 0.55), Duchang (R² = 0.56), and Tangyin (R² = 0.55), indicating that grassland area in the sub-lakes no longer correlated to the middle and northern water level fluctuations of Poyang Lake at some water elevations. This case may also happen in severe drought years. To minimize the spatial variability of water level, the water level variations at Kangshan station were small during this period, suggesting a distinct change in grassland area during the low water period. After that, most grasslands were inundated by floods during the high water period.

Through all hydrological periods, we further found the grassland area was linearly related to water level fluctuations when the water level of Poyang Lake at Xingzi station was higher than 14 m (p < 0.01), and this relationship disappeared when the water level of Poyang Lake at Xingzi station was lower than 14 m (Figure 6). This result was consistent with former studies [31,57].

During the retreating period, the water level decreases slowly from south to north, with water levels dropping first in the southern parts of Poyang Lake and a water level gradient reappearing in the main lake body of the lake [58]. Different from the rising period, the slope of grassland area change decreased with the declining water levels due to the distribution pattern of wetland vegetation in Poyang Lake (Figure 7). The slope increased from north to south as follows: Xingzi (coefficient of x² equals to −16.21), Duchang (coefficient of x² equals to −19.08), Tangyin (coefficient of x² equals to −31.99), and Kangshan (coefficient of x² equals to −96.98).

The growing status of wetland plants is mainly controlled by ground water and weather (temperature) conditions [42]. Species with two growing cycles grow slowly and eventually stop growing in winter due to cold temperatures, and fresh biomass and canopy water content start to drop. The senescent stage of other species living at higher altitudes with one growing cycle (e.g., Artemisia spp.) starts in December due to cold weather and lack of water. In contrast, with the exception of the lake depressions and several artificial lakes, only the main channel was inundated during the dry period; the relatively stable lake water levels with low variation amplitudes did not likely explain the changes in grassland area (Figure 8). But the relatively shallow water and bare ground covered by mud and grass provide ideal habitats and food for migratory birds.

Liu et al. (2011) found that when the water level of Poyang Lake was higher than 14.18 m (Xingzi Station) in the dry season, fewer habitats for birds remained because the area with low water depth disappeared [59]. The proposed PLP would play a role from September to March in maintaining the lake water level at 15 m or higher. For species that are found only at Poyang Lake, if critical habitat was lost during infrequent hydrological events, populations would decline or be eliminated because the species lifecycle would be broken [60,61]. Moreover, sedge communities can tolerate flooding during part of the year, usually during the summer months. If these communities remained inundated permanently, it would greatly affect the fitness of adult plants, and cause failure in seed regeneration [62]. Lacking appropriate habitat, many wintering bird populations could be lost or severely reduced. Realistic projections of the impact of the project are required.