Spectral Discrimination of the Invasive Plant Spartina alterniflora at Multiple Phenological Stages in a Saltmarsh Wetland

Spartina alterniflora has widely invaded the saltmarshes of the Yangtze River Estuary and brought negative effects to the ecosystem. Remote sensing technique has recently been used to monitor its distribution, but the similar morphology and canopy structure among S. alterniflora and its neighbor species make it difficult even with high-resolution images. Nevertheless, these species have divergence on phenological stages throughout the year, which cause distinguishing spectral characteristics among them and provide opportunities for discrimination. The field spectra of the S. alterniflora community as well as its major victims, native Phragmites australis and Scirpus mariqueter, were measured in 2009 and 2010 at multi-phenological stages in the Yangtze River Estuary, aiming to find the most appropriate periods for mapping S. alterniflora. Collected spectral data were analyzed separately for every stage firstly by re-sampling reflectance curves into continued 5-nm-wide hyper-spectral bands and then by re-sampling into broad multi-spectral bands – the same as the band ranges of the TM sensor, as well as calculating commonly used vegetation indices. The results showed that differences among saltmarsh communities’ spectral characteristics were affected by their phenological stages. The germination and early vegetative growth stage and the flowering stage were probably the best timings to identify S. alterniflora. Vegetation indices like NDVI, ANVI, VNVI, and RVI are likely to enhance spectral separability and also make it possible to discriminate S. alterniflora at its withering stage.


Introduction
The Yangtze River Estuary is an important eco-region that covers the main part of Shanghai and a portion of Jiangsu province, China. There are more than 300 km 2 of coastal saltmarshes in Shanghai alone that cover 22.5% of its total area and 95.7% of its total wetland area and provide more than 95% of its total value of ecosystem service ($7.3610 9 US/year) [1]. However, one recent event, i.e., the Spartina alterniflora (hereafter, Spartina) invasion, has caused profound impacts on the saltmarshes [2]. The Spartina invasion to the Yangtze River Estuary will likely converse mudflats to Spartina meadows, destroy shorebirds' foraging habitats, bring negative impacts on endangered species, decrease abundance of native species, and cause ecosystem degradation and considerable economic loss and the expanding trend is still not controlled [3]. These negative impacts of Spartina invasion as well as its wide and dynamic distribution have posed an urgent task to get its spatial information accurately and timely in order to conserve native biodiversity and manage the land ecologically.
Techniques of both multi-spectral and hyper-spectral remote sensing have been used to monitor and map saltmarsh vegetation. However, to monitor invasive plant species, the ability to map vegetation at monospecific level is required and thus we have the problems such as different spectra for the same vegetation type, the same spectra for different vegetation types, and mixed pixels that should be minimized. Hyper-spectral sensors with contiguous and narrow bands are able to detect small and local variations which are likely masked within broad bands of multi-spectral data; thus, hyper-spectral data can increase the potential to discriminate vegetation species and have successfully mapped wetland monospecific vegetation covers [4,5] for a long period of time. On the other hand, though the old generation multi-spectral data such as Landsat TM and SPOT imageries were reported to be insufficient for mapping monospecific covers in detailed wetland environments due to mixed pixels resulting from coarse spatial resolution [4,6], the high spatial resolution sensors that were newly launched (e.g., QuickBird and IKONOS) and aerial photography offer opportunities to monitor monospecific covers in wetlands by achieving resolutions of meters or sub-meter [7].
In spite of all of this, there is a common problem for either hyper-spectral imagery or VHR (very high-resolution) imagery in that they are difficult and expensive to acquire [4,8]. This proposes an issue that the collected remote sensing data coincides with a period when suitable spectral differences between different vegetation types are captured by remote sensors. Field spectral discrimination serves as an important way to find the differences existing in spectral signatures between plant species [9,10]. On the other hand, spectral characteristics of vegetation change obviously with time affected by phenology, complicating the problem but also providing opportunities of spectral discrimination and remote sensing mapping (i.e., some phases must be more appropriate than others for mapping). Therefore, it is important to identify species by exploiting spectral differences at proper phenological stages [11]. Furthermore, spectral differences can be enhanced by using vegetation indices (VIs) that take advantage of vegetation reectance contrasts between wavebands/wavelengths [12]. VIs based on multi-spectral broad bands are commonly used to enhance vegetation differentiation [13,14], but as an influence of phonology, the effectiveness of different VIs may also change with time. Recently, a few studies have taken advantages of phenology for species discrimination [12,15]. Nevertheless, few studies concern multi-phonological spectra of saltmarsh vegetation, especially identifying the widespread invasive species, perhaps because of the difficult accessibility of saltmarsh environments for multi-date investigations. However, such work could provide important information guiding future field spectra sampling and remote sensing image collection.
This study aims to find out the most appropriate periods and corresponding spectral zones for mapping Spartina by remote sensing in the Yangtze River Estuary. We hope the expected information would help collect, process, and classify images at suitable periods for dynamically monitoring Spartina. To achieve this, spectral analyses were applied at multi phonological stages by sampling spectral curves into hyper2/multi-spectral bands, calculating VIs, and quantifying the spectral differences at each pure band and VI for comparison. More specifically, the objectives of this study were: 1) to characterize the canopy spectral properties of dominant saltmarsh species in different phenological stages and 2) (the more important) to identify the most suitable phenological stages to discriminate invasive exotic Spartina and simultaneously determine the corresponding best separable hyper2/multi-spectral bands and VIs.

Ethics Statement
Our fieldwork in this study was approved by the Shanghai Chongming Dongtan National Nature Reserve. We did not sample any soil, plants, or animals out of the ecosystem in the work. Only reflectance spectra and photographs were collected with minimum disturbances.

Study Area
The study area is a broad saltmarsh at the eastern Chongming Island (Dongtan) (Fig. 1), the largest wetland in the Yangtze River Estuary of China. The shallow open waters, mudflats, and saltmarshes in Dongtan make an important habitat for shorebirds [3] and serve as an important stopover for migratory birds on the East Asian-Australian Flyway [16].
The saltmarshes of Dongtan were dominated by native plant species Phragmites australis (hereafter, Phragmites) and Scirpus mariqueter (hereafter, Scirpus) before the invasion of Spartina, and they create a good habitat and provide abundant food for various birds. However, the invasion of Spartina is seriously threatening the native ecosystem and coastal aquaculture, and the native species are rapidly being replaced by Spartina since it is introduced in 2001 for land reclamation [2].
The phenology of the three presently dominated species is somewhat different [17]. Generally, Phragmites germinates in early April, undergoes a rapid vegetative growth stage spanning from June to mid-August, flowers in mid-October, and senesces in late-November. Spartina emerges in early-May, grows rapidly from June to early-September, flowers in late-September, and dies in late-December. Scirpus has a relatively shorter lifespan than Phragmites or Spartina. It usually begins to grow in late-April, turns to flowering stage in mid-June, and starts withering process in early-September. These differences in phenology may greatly affect the distinctive spectral signatures between the species, making some periods during a year more likely than others to discriminate them using remote sensing technology.

Spectral Data Acquisition
Spectral data of dominated plants was collected in situ using a UniSpec-DC Spectral Analysis System (PP System, Inc. Amesbury, MA, USA). This machine has a detection region of 310,1 100 nm with optional FOVs of 3u, 6u, 12u, and 20u. The field work was initially planned to be done in 2009. However, limited by weather and tidal conditions, the data in some expected periods was missed in 2009. To fill the data gaps, the missing phenological stages were re-surveyed in 2010. In total, the spectra data was collected on 12 different dates at various phenological stages under clear weather between 10:00 and 14:00 (Table 1).
Four patches (larger than 10 m610 m) with similar vegetation height and coverage were selected for each plant species to decrease multi-factors interaction and influence (Fig. 1). Three random sites were taken in each patch (sometimes, only in three patches) for spectral data collection and 20 spectrometer measurements were taken from each site at nadir points. The sensor was placed above plant canopies at approximately 100 cm height using a 20uFOV optic, which resulted in approximately 0.34 m spatial resolution (Fig. 2). The spectral data were calibrated by a standard white reference and converted into reflectance data.

Data Pre-analysis
Only the data between 400 nm and 900 nm were used for analysis due to extraneous noises at the extremes of the detection range. The sampled data of each surveyed day were analyzed separately following the same steps below.
First, the 20 measurements from each site were averaged as one reflectance curve sample, resulting in 9,12 samples for each plant species. Second, the reflectance curves were divided into 100 consecutive 5 nm-wide hyper-spectral bands by taking the average from its original five wavelengths without losing much information (the raw spectral resolution is 3.3 nm). Moreover, reectance data were averaged to represent four broad multi-spectral wavebands according to band ranges of TM (blue -B: 450,520 nm; green -G: 520,600 nm; red -R: 630,690 nm; near-infrared -NIR: 760,900 nm). The following VIs NDVI [18], RVI [19], RB [20], VNVI, ANVI, and EVI [21] were also calculated for analysis to examine whether they can enhance spectral separability:

Hyper-spectral Data Analysis
To estimate how Spartina is separable from Phragmites and Scirpus by using hyper-spectral remote sensing on different days, the statistical difference and separability of the hyper-spectral bands between plant species were calculated. First, the Mann-Whitney U-test was employed on the 5 nm-wide bands between every pair of vegetation types and the significance level was ,0.05. The null hypothesis assumes that there is no significant difference of the same 5 nm-wide band between one type and another. If this null hypothesis was rejected, the alternative hypothesis was that they are significantly different. The significantly different bands suggested the potential to discriminate plant species, but in order to compare the discrimination ability, it' d better to quantify how far they are in the statistical spaces (the spectral separability).
The separability of the statistically different bands was then quantified using Jeffries-Matusita (JM) distance. The JM distance between a pair of probability functions is the measure of the  average distance between the density functions of two classes, indicating how successful the two classes will be discriminated [22]. The JM distance has upper and lower bounds that vary between 0 and ffiffi ffi 2 p (<1.414), with 1.414 implying classification of two classes with 100% accuracy [23]. The JM distances were calculated on each individual band without considering various band combinations.
One advantage of JM distance over other measurements of statistical separability, such as divergence, is that an upper bound of error probability for classification could be estimated [22]. According to the estimated error probability, the separability of statistically different hyper-spectral bands was classified into four categories: perfect-separability (error probability ,1%), good -separability (1% # error probability ,7%), normal-separability (7% # error probability ,14%), and poor-separability (error probability $14%). The formula for computing the JM distance and the error probability (P e ) could be found in Swain and Davis [22].

Multi-spectral Data Analysis
To estimate how Spartina is separable from Phragmites and Scirpus by multi-spectral remote sensing on different days, the same process as in a hyper-spectral analysis was applied to calculate the statistical difference and JM distance between each pair of plant species based on multi-spectral broad bands (B, G, R, NIR) and the commonly used VIs of NDVI, RVI, RB, VNVI, ANVI, EVI).

The Reflectance Spectral Data Corresponding to Different Phenological Stages
The mean reflectance spectral curves of Spartina, Phragmites, and Scirpus on different days at various phenological stages were plotted in Fig. 3. Spectral reflectance curves of the plants and the difference between them were affected by phenology. As the phenological stage changed, their spectral signatures exhibited differences in both magnitude and shape of reflectance curves. In the late-withering, dormant, and early germination stages (Fig. 3a, b, c, k, and l), the reflectance curves of all plants showed slightly rising lines with no green peak, steep red edge, or near-infrared plateaus (common characteristics of green vegetation). The low chlorophyll content at these stages alone probably can explain the lack of green peak, while the higher reflectance in red due to no   Jan. 14 - -: poor separability (error probability$14%); +: normal separability (14% $error probability $7%); ++: good separability (7% $error probability $1%); +++: perfect separability (error probability #1%); #: no significant difference. Dates with good or perfect separability bands were highlighted in bold. doi:10.1371/journal.pone.0067315.t004 chlorophyll absorption and the lower reflectance around 800 nm associated with sparse leaf/canopy structure can explain the absence of the red edge. Phragmites was the earliest to show typical characteristics of green vegetation (Fig. 3c). Phragmites began to show a steep red edge and infrared plateau from April 17 as it sprouted in early April, though no green peaks were observed. Afterwards, the significant green peak, steeper red edge, and higher near-infrared plateau began to show from May 5 when Phragmites grew up to about 100 cm (Fig. 3d-i). The reflectance of near-infrared began to decline and the typical characteristics of green vegetation of its reflectance curves gradually disappeared as Phragmites was inflorescent and withering. Scirpus also emerged in early April, but its rapid growth did not begin until late April; thus, its reflectance spectra showed a normal curve of green vegetation until May 5 (Fig. 3d) when it already grew to ca. 15 cm. In its following growth, flowering, fruit development, and even early withering stage, Scirpus showed higher near-infrared plateau (Fig. 3e-i). However, its reflectance in green and near-infrared wavelength regions declined quickly from August as Scirpus turned to mature and withering stages early. Conversely, Spartina generated and grew about one month later than Phragmites and Scirpus. Thus, in late spring (e.g., May 5 and 31), it has significantly lower reflectance than the Phragmites and Scirpus in green and nearinfrared wavelength regions (Fig. 3d-e). Spartina was the latest to show a steep red edge and near-infrared plateau (Fig. 3e), but later, it shows the highest reflectance in the near-infrared region (Fig. 3gh). On the other hand, Spartina withered later and slower than Phragmites and Scirpus. Thus, its canopy reflectance curves were more similar to normal curves of green vegetation in later times of a year than Phragmites or Scirpus (Fig. 3k-l).     It is obviously that Spartina was more separable from Scirpus than from Phragmites as a whole (Comparing to Fig. 4 to 5). No statistically different bands were found on July 2, 22 and there were no good-separability bands on August 19 between Scirpus and Spartina, when they were both in stages with high chlorophyll content. Nevertheless, with the exception of July 2, July 22, and August19, there were always many statistically different bands with either perfect-separability or good-separability. The perfectseparability bands primarily distributed in blue, red, and nearinfrared wavelength regions, while the good-separability bands were more widely distributed.

Hyper-spectral Statistical Difference and JM Distance Between Plants
There were few good-or perfect-separability bands between the two native species on most days (Fig. 6). Among all sampled days, May 5 and 31 were the best for discrimination between Scirpus and Phragmites. The best bands for discrimination them were first concentrated in blue and red wavelength regions on May 5 and then shifted to red and near-infrared regions on May 31. Though December 22 showed perfect-separability and goodseparability on a many bands, late December is not ideal for discrimination because the distribution of Scirpus was narrowed by tidal water and was unable to get the real extent of Scirpus according to our field observation. Tables 2, 3, and 4 show the statistical differences and JM distances on multi-spectral bands (B, G, R, and NIR) and the commonly used VIs at the multi-phenological stages between Spartina and Phragmites, Spartina and Scirpus, and Phragmites and Scirpus, respectively. Excluding VIs, spectral difference and separability were limited on broad multi-spectral bands. Spartina and Phragmites had perfect-separability on May 31 (NIR) and good separability on May 5 (G, R, and NIR) and September 20 (B, G, and R) ( Table 2). The broadband spectral separability between Spartina and Scirpus was much better than that between Spartina and Phragmites. Perfect-or good-separability bands were found on all surveyed days except July 2, July 22, and August 19 when Spartina and Scirpus were both near their growth summit ( Table 3). The three tables also show that the VIs could enhance the spectral separability between the plants at many stages. The four broad bands between Spartina and Phragmites on November 25 and December 22 had either poor-separability or were not statistically different, but the VIs like NDVI, RVI, VNVI, and ANVI were ranked with good-or perfect-separability ( Table 2). The VIs significantly enhanced the spectral separability between Spartina and Scirpus on January 14, April 17, and July 22 (Table 3). The VIs also made it possible to discriminate Phragmites and Scirpus in early winter and mid spring, as perfect-separability was obtained on the VIs, like NDVI and ANVI, on January 14 and April 17 (Table 4). On the whole, NDVI, VNVI, ANVI, and RVI more frequently performed enhanced separability than their combined original broad bands. Tables 5 and 6 summarized the comprehensive information to determine the best periods for discriminating Spartina from Phragmites and Scirpus. Among all our sampled days, May 5, 31 and September 20 (which represent two periods, e.g. later spring and early autumn) are clearly the best to discriminate Spartina, as they have good-or perfect-separability between Spartina and either of the other two natives on either hyper-or multi-spectral bands. However, hyper-spectral bands clearly show priority to detect more minor differences by providing more narrow bands that could discriminate each plant from the others at more periods, and the commonly used VIs based on multi-spectral broad bands could enhance their ability for plant discrimination, in terms of expanding time periods and increasing separability on original bands (Tables 5 and 6).

Spectral Characteristics and Separability Affected by their Phenological Stages
The spectral characteristics of each species were not unique between phenological stages. They were varying significantly according to their phenological stages. The phenology of plants generally determines the spectral characteristics of vegetation by determining the canopy structure, growth stage, water content, and chlorophyll content of vegetation [24]. Spectral differences near the green waveband peak (around 550 nm), red edge (around 720 nm), and in the near-infrared plateau are considered important for discriminating vegetation types at growing stages [4]. The differences near red edge and in the near-infrared plateau are usually characterized by differences in canopy structure and the differences near the green waveband peak are usually characterized by differences in chlorophyll content [11]. In this case, the canopy structure between Spartina and Phragmites was more similar than that between Spartina and Scirpus in most phonological stages (see Figure S1), explaining why Spartina is more spectrally separable from Phragmites than Scirpus. Spartina generates later than Phragmites and Scirpus, causing it to have notably less green vegetative amounts and low plant height in late spring, thus making late spring the best period to monitor Spartina by increasing the spectral separability at wavelengths near the green waveband peak, red edge, and in the near-infrared plateau (see Tables 5 and  6). In summer, the three plants reached their peak heights with high chlorophyll content and similar canopy structures (see Figure  S1). Thus, they were difficult to separate with any hyper2/multispectral bands. A previous study, in a single summer, demonstrated that the vegetation amount in terms of the height and cover was important in determining the levels of reectance, particularly at the near-infrared band of saltmarsh communities [25]. In this study, Spartina showed average high levels of near-infrared reflectance in summer ( Fig. 3g and 3 h), as it had denser covers than Phragmites and Scirpus, as well as clearly higher canopy altitudes than Scirpus (see Figure S1), but this average high nearinfrared reflectance was not statistically significant. Early autumn, i.e. the flowering stage of Spartina, is another good period to discriminate Spartina. The bloom of Spartina not only changed its pigment content and proportion but also changed its canopy structure suddenly, while Phragmites was at a vegetative growth or inflorescence stage and Scirpus was at a withering stage (see Figure  S1), making perfect separability between Spartina and the other two on bands located at the visible wavelength range. Another phenological characteristic of Spartina is that it senesces and withers more slowly than Phragmites and Scirpus; thus, in early winter, Spartina can preserve more green vegetative parts ( Figure  S1). This phenomenon can result in statistical spectral differences in the visible wavelength region between Spartina and the other plants ( Fig. 4 and 5). Though these differences were small between Spartina and Phragmites probably because of their similar canopy structure and tended to be marked out in broad multi-spectral bands, they can be enhanced by specific VIs to have enough separability to distinguish them (Table 2).

Look into the Timing of Saltmarsh Vegetation Classification
Limited by the weather, tides, and expensive cost of filed work at this saltmarsh, we were not be able to collect spectra daily, weekly, or even bi-weekly. However, with samples collected on those 12 single days (table 1), we can still infer the best periods for classification based on the fact that these vegetation change phonological stage gradually. This study clearly demonstrates that May 5, May 31, and September 20 achieved the best separability to classify or mapping distribution of Spartina in the Yangtze River Estuary. The corresponding phonological stages are the germination, early vegetative growth, and flowering stages of Spartina, which are about in late spring and early autumn of a year ( Table 7). The withering stage of Spartina (late autumn and early winter of a year, represented by samples collected on November 25 and December 22) may also be a good period for mapping its spatial extent; however, it may need to take advantage of proper VIs. Yet, this late-autumn or early winter period should never be good timing for Scirpus monitoring because aboveground biomass has been washed away by tidal water and its spectra mixes with background water and mud ( Figure S1). In an ecology view, Scirpus is as important as Spartina to be monitored because it is more preferred by waterfowls but has been forced to populate a narrow zone close to the sea due to competitive disadvantage [3]. As a result, it is even more difficult to map Scirpus by remote sensing as the tidal waters regularly influence its relatively short and sparse canopy. We collected Scirpus spectra samples in some depression plots where water remains after the tide and found that its spectral characteristic was much different than that without a water background at its early growing and withering stages, but not its peak growing stage (Fig. 7). Moreover, our analysis has suggested that the peak growing stage is the worst timing for vegetation discrimination. Thus, it is very important to consider timings with minimum tidal water influence for Scirpus monitoring. The information provided here would be helpful to collect, process, and classify images by guiding investigators toward suitable periods for dynamically monitoring Spartina as well as the other two saltmarsh species. Some of our previous studies that validated the proposed periods are qualified for Spartina monitoring in the Yangtze River Estuary [26,27]. However, as the spectra were collected from roughly 0.5 m square canopy areas, it would be difficult to directly extrapolate these reflectance signatures to real canopy cover reflectance captured by a remote sensor.
While phenology is critical in determining vegetation spectral characteristics, the timings of phenological stages are adjusted to climate, like temperature and precipitation patterns. Therefore, optimum periods for saltmarsh vegetation discrimination might shift between years and the right way to determine the periods is to investigate the suggested emergence of phonological divergence between plants but to not adhere to any predetermined periods. A strategy that integrates multi-phenological images may produce a better result. However, improper usage of multi-date images might also waste money, produce redundant information, and increase image processing effort. Still, based on the abundant information provided, users can have their choice based on effectiveness and cost.

Conclusion
The field spectra of saltmarsh vegetation including Spartina, Phragmites, and Scirpus were collected and analyzed at multiphenological stages in the largest estuarial wetland of the Yangtze River Estuary to find the best periods for mapping Spartina. The three plants showed changing spectral characteristics affected by their phenological stages. The results suggest that the germination and early growth stages of Spartina (late spring) as well as its flowering stage (early-autumn) may be the best periods to discriminate Spartina from Phragmites and Scirpus (Table 7). The corresponding best-separable hyper2/multi-spectral bands were also determined for each period (Table 7). In late-spring, the most separable hyper-spectral bands are located near the green waveband peak, red edge, and in the near-infrared plateau and the multi-spectral bands are NIR and G. In early-autumn, the most separable hyper-spectral bands are located at the whole visible region and the multi-spectral bands are B, G, and R. By utilizing VIs like NDVI, ANVI, VNVI, and RVI, which take advantage of the reflectance contrast between NIR and other bands, it is also possible to discriminate Spartina at its withering stage (late autumn and early winter). Proper usage of specific VIs might also enhance the spectral separability between those saltmarsh species in other periods.