Quantifying expansion and removal of Spartina alterniflora on Chongming island, China, using time series Landsat images during 1995–2018

Abstract

The rampant encroachment of Spartina alterniflora into coastal wetlands of China over the past decades has adversely affected both coastal ecosystems and socio-economic systems. However, there are no annual or multi-year epoch maps of Spartina saltmarsh in China, which hinders our understanding and management of Spartina invasion. In this study, we selected Chongming island, China, where Spartina saltmarsh had expanded rapidly since its introduction in the 1990s. We investigated phenology of Spartina, Phragmites and Scirpus saltmarshes, and the time series vegetation indices derived from Landsat images showed that Spartina saltmarsh did not green-up in April-May and stayed green in December-January, which differed from the phenology of Phragmites and Scirpus saltmarshes. We developed a pixel- and phenology-based algorithm that used time series Landsat data to identify and map Spartina saltmarsh, and we applied it to quantify the temporal dynamics (expansion and removal) of Spartina saltmarsh on Chongming island during 1995–2018. The resultant maps showed that Spartina saltmarsh area on Chongming island increased from ~4 ha in 1995 to ~2,067 ha in 2012 but dropped substantially to ~729 ha in 2016 after a large-scale ecological engineering project (US$ 186 million) was started to remove Spartina during 2013–2016. Chongming island still had ~1,315 ha Spartina saltmarsh in 2018, and majority of it was distributed outside the Chongming Dongtan National Nature Reserve, which could serve as the sources for reinvasion in the near future. This study demonstrates the feasibility of using time series Landsat images, pixel- and phenology-based algorithm, and GEE platform to identify and map Spartina saltmarsh over years in the region, which is useful to the management of invasive plants in coastal wetlands.

1. Introduction

Biological invasion has been one of the most crucial ecological issues in the coastal wetlands worldwide (Mao et al., 2019; Vaz et al., 2018). Spartina alterniflora (hereafter, Spartina), a C₄ grass growing primarily in the Atlantic coastal regions of North America, was first introduced to China in December 1979 (Zuo et al., 2012). The original purpose of Spartina introduction to China was for soil amelioration, tidal reclamation and erosion mitigation (Li et al., 2009; Lu and Zhang, 2013; Xiao et al., 2010). However, with its high survivability and strong adaptability, introduced Spartina has been well established in the intertidal zones and become a dominant species in saltmarshes in China (Wan et al., 2014). As time passed, the overwhelming spread of Spartina populations threatened the coastal environments by altering the estuarine sediment dynamics, outcompeting native plant species, and reducing bird biodiversity due to losses of food resources and feeding habitats. China has the largest area of exotic Spartina (Liu et al., 2018). In 2003, Spartina was listed as one of the most harmful invasive plants by the Ministry of Environmental Protection of China (Chung, 2006). Accurate, consistent, and comprehensive records of Spartina saltmarsh in China are urgently needed for ecological management and coastal ecosystem conservation. However, to date, no datasets of Spartina saltmarsh at high spatial resolution and annual temporal resolution in China are available to the stakeholders and public.

Remote sensing is an effective tool for ecological monitoring and management of invasive plants (Bradley, 2014; Wang et al., 2017). Image data from satellite and airborne sensors, coupled with geographic information systems (GIS) and advanced cloud computing system, offer researchers viable tools to detect and monitor the distribution of invasive plants. Table 1 provides a brief summary on Spartina saltmarsh classification and mapping over the past two decades. One approach is to carry out visual interpretation and digitalization of single or multi-date images and generate Spartina saltmarsh maps (Liu et al., 2017; Wan et al., 2014; Zhang et al., 2004). This approach is labor-intensive and time-consuming, and the resultant maps often have large uncertainty when images from different acquisition dates are used or interpreted by different researchers. The second approach is to use single or multi-date hyperspectral or multi-spectral optical images, calculate the spatial statistics of spectral bands, vegetation indices and texture in these images, and apply supervised or unsupervised classification algorithms (e.g., maximum likelihood, random forest, support vector machine) to generate Spartina saltmarsh maps (Liu et al., 2016; Lu and Zhang, 2013; Mao et al., 2019). Many studies used Landsat images (Liu et al., 2018; Zhang et al., 2017) and selected one to multiple images based on image quality and other factors (e.g., tidal level). Because of large tidal dynamics and many human activities, the resultant maps of Spartina saltmarsh from the analyses of single or multi-date images in a year also have large uncertainty. To date, most of previous studies generated Spartina saltmarsh maps at either specific year(s) or small study region(s). The third approach is to use time series data of individual pixels, calculate the temporal statistics of spectral bands and vegetation indices in the pixels, and apply decision trees or rule-based algorithms to generate annual maps of Spartina saltmarsh. Phenological differences of plant species, which are documented in the time series data of spectral bands and vegetation indices, are carefully studied and the unique phenological features are identified, selected and used for classification of individual pixels. In recent years, the pixel- and phenology-based algorithms have been used in a number of studies for crops, tree plantations, coastal tidal flats and coastal vegetation (Dong et al., 2016; Helman et al., 2015; Kou et al., 2015; Wang et al., 2018).

Table 1.

A brief summary on image data sources and methods used in previous studies for classifying and mapping Spartina saltmarsh.

Methods		Multi-spectral images		Hyperspectral images	Multi-source images
		VHSR	HSR
Visual interpretation		Liu et al. (2017)	Zhang et al. (2004); Wan et al. (2014)
Images and spatial statistics of images	MLC	Liu et al. (2016);	Li et al. (2006);		Chust et al. (2008);
		Ai et al. (2017)	Huang and Zhang(2007b); Lu and Zhang (2013); Zhang et al. (2017)		Hladik et al. (2013)
	NN	Morris et al. (2005)
	SVM	Ai et al. (2017)	Wang et al. (2015); Liu et al. (2018); Mao et al. (2019)
	RB/DT/RF	Wan et al. (2014); Lin et al. (2015); Wang et al. (2015)	Zuo et al. (2012); Sun et al. (2016)		Hladik et al. (2013); van Beijma et al. (2014)
	SMA		Zhu et al. (2016)	Rosso et al. (2005)
Pixels and temporal statistics of pixels			This study

Note: VHSR: Very high spatial resolution satellite image (e.g., GF-1); HSR: high spatial resolution satellite image (e.g., Landsat); MLC: Maximum likelihood classification; NN: Artificial Neural Network; SVM: Support vector machine; RB: Rule-based; DT: Decision tree; RF: Random forest; SMA: Spectral mixture analysis.

Gao and Zhang (2006) and Ouyang et al. (2013) conducted in-situ field studies to collect hyperspectral data of coastal saltmarshes on Chongming island, Shanghai, China over different seasons and they found that Spartina has unique spectral characteristics that are different from other native saltmarsh plants in spring and fall seasons. In all multi-spectral bands and vegetation indices, Normalized Difference Vegetation Index (NDVI) in the senescence stage was considered to be useful for identifying Spartina saltmarsh (Ouyang et al., 2013). The phenological information of coastal saltmarshes acquired from field studies provide valuable reference but is hardly to be used at large spatial scale. There are also some studies that used monthly NDVI data from the China GaoFen-1 satellite and China Huanjing-1 satellite to identify and map coastal saltmarshes (Ai et al., 2017; Sun et al., 2016), and NDVI in the senescence stage (November to mid-December) and the green-up stage (late April-May) was recognized as an useful indicator for identifying Spartina saltmarsh (Ai et al., 2017). Although these efforts have presented the significance of unique phenological stages for identifying Spartina saltmarsh, the developed methods faced challenges in the extension of the resulting classifiers and parameters and the high cost of commercial satellite images. Landsat imagery at 30-m spatial resolution over the period of 1984-present (Wulder et al., 2019; Wulder et al., 2016) is an undoubtedly suitable data source for mapping long-term dynamics of Spartina saltmarsh. One pilot study evaluated the potential of time series Landsat images in 1984–2015 to track Spartina saltmarsh dynamics within the three pixels (500-m spatial resolution) of the Moderate Resolution imaging Spectroradiometer (MODIS) in the Yancheng Coastal Wetland Nature Reserve, Jiangsu province, China (Wu et al., 2018). However, the potential of time series Landsat images in tracking the phenological differences between Spartina and other native saltmarsh plants has not been fully assessed. Documenting the long-term and large-scale dynamics of Spartina saltmarsh would need to process a large volume of satellite images, which requires advanced and efficient computer processing power. To develop a mapping tool that uses time series Landsat images, a pixel- and phenology-based algorithm, and the Google Earth Engine cloud computing platform to quantify and monitor the dynamics of Spartina saltmarsh over years would certainly address such needs.

In this study, we addressed the following questions: (1) what are the phenological differences between Spartina and other native saltmarsh plants as observed by time series Landsat images? (2) Do the time series Landsat images suffice to track phenology of Spartina and other native saltmarsh plants over years? (3) Is the phenology-based mapping algorithm with time series Landsat images reliable to discriminate Spartina saltmarsh from native saltmarshes? As a pilot and methodological study, we selected the Chongming island in the Yangtze River estuary, Shanghai, China, as case study area, which is the largest alluvial island in the world and has experienced both expansion and removal of Spartina in recent decades. The specific objectives of this study are: (1) to better understand the phenological characteristics of Spartina and other native saltmarsh plants on Chongming island with time series Landsat images; (2) to develop and evaluate a pixel- and phenology-based algorithm that uses time series Landsat images to effectively discriminate Spartina saltmarsh from other saltmarshes; and (3) to quantify the spatial-temporal dynamics of Spartina saltmarsh on Chongming island since the 1990s, which could illustrate the expansion dynamics of Spartina saltmarsh from the 1990s to 2000s and assess the effectiveness of the ecological restoration project to control and remove Spartina in the 2010s. The resultant maps of Spartina saltmarsh on Chongming island could be used to support conservation efforts and decision-making in Shanghai and the Yangtze River estuary. The resultant mapping tools could also be readily applied (and modified if needed) to identify and map Spartina saltmarsh in other coastal wetlands in China.

2. Materials and methods

2.1. Study area

Chongming island is part of the megacity Shanghai, China (31°27′–31°5l′N, 121°09′–121°54′E), and is located at the mouth of Yangtze River (Fig. 1). It is the largest alluvial island in the world and covers an area of about 1,267 km². It has a typical subtropical monsoon climate with an annual mean temperature of about 15.3°C and annual mean precipitation of about 1,022 mm. The eastern fringe of Chongming island is the Chongming Dongtan National Nature Reserve (hereafter, CDNNR), one of the largest nature reserves for migratory birds in Eastern Asia (Hu et al., 2015). The CDNNR was listed in the Chinese Protected Wetlands report in 1992, recognized as the Wetland of International Importance under the Ramsar Wetlands Convention in 2001, and designated as a National Nature Reserve for migratory bird conservation in 2005 (Xiao et al., 2010). The total area of the CDNNR is about 242 km², accounting for 20% of Chongming island area.

Figure 1.

The location of the Chongming island, Shanghai, China. It is within Landsat scene (path/row 118/38 and 119/38). (a) The locations of three field sites for Spartina (121.9481°, 31.5523°), Phragmites (121.9101°, 31.4531°) and Scirpus (121.9863°, 31.4932°) are shown, which were used for spectral signature analysis; (b, c, d) zoom-in view of landscapes from a very high spatial resolution image in the Google Earth (dated as 2012/11/01); and (e, f, g) field photos taken in August 2015 at the three field sites.

The commonest native plants in the study area are Phragmites australis (hereafter, Phragmites), Scripus mariqueter (hereafter, Scirpus) (Sun et al., 1992). In 1995, Spartina was first spotted in the north of the CDNNR. In an effort to reclaim more land for regional development by accelerating siltation, Spartina was intentionally planted in the CDNNR in 2001 (337 ha) and 2003 (542 ha), respectively (Wang, 2011). In 2007, the area of Spartina saltmarsh in Shanghai ranked third among all Spartina-invaded megacities and provinces in China, and the area of Spartina saltmarsh in Shanghai, Jiangsu, Zhejiang and Fujian accounted for 94% of the total area in China (Zuo et al., 2012).

In order to avoid the adverse effects of Spartina encroachment, Shanghai government had invested about ¥1.3 billion Chinese Yuan (US$ 186 million) to carry out an ecological restoration project (hereafter, DEPA) in the CDNNR since 2013 (Fig. 1a). The project had three major components: ecological eradiation of Spartina, restoration of native saltmarsh plant communities, and reconstruction of bird habitats (Tang, 2016). Its preliminary work was started in December 2010 and various technical methods were tested and evaluated during 2011–2012 in one demonstration area (~4 km²). Large-scale restoration work was launched in December 2013, covering a total area of 24.19 km². The project region was enclosed with artificial levee for eradicating Spartina through cutting the plants and flooding the marshes (Hu et al., 2015).

2.2. Data

In this study we used time series Landsat datasets, images from Google Earth, and in-situ data to identify and map Spartina saltmarsh on Chongming island, and Figure 2 shows the dataflow and workflow for image processing and classification.

Figure 2.

The workflow for the pixel- and phenology-based Spartina saltmarsh mapping in this study.

2.2.1. Landsat data and preprocessing

The study area is covered by 2 paths/rows (P118R38 and P119R38) of the Landsat Worldwide Reference System (WRS-2) (Fig. 1). We used all the available Landsat 5/7/8 surface reflectance (SR) products between January 1, 1995 and January 1, 2019, which have been archived in Google Earth Engine (GEE), a popular cloud computing platform for planetary-scale remote sensing analysis (Gorelick et al., 2017). The SR datasets have been atmospherically corrected through Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) algorithm and Landsat Surface Reflectance Code (LaSRC) algorithm, respectively (Masek et al., 2006; Vermote et al., 2016). The bad-quality observations, including clouds, cirrus, and cloud shadows were identified, according to the pixel quality assessment (pixel_qa) generated from the CFMask algorithm (Zhu and Woodcock, 2012). All the Landsat image pre-processing tasks were carried out on the GEE platform. Fig. 3 shows the annual distributions of all the available Landsat images by sensors (Fig. 3a) and by seasons during 1995–2018 (Fig. 3b). We also counted the numbers of good-quality observations of individual pixels in April-May (Fig. 3c) and in December-January (Fig. 3d) in each year during 1995–2018. The majority of pixels had at least one good observation in April-May (> 93.7% pixels) and in December-January (> 93.9% pixels) during 1995–2018.

Figure 3.

The annual distributions of number of Landsat images (a) by sensors (Landsat5 TM, Landsat7 ETM+, and Landsat8 OLI), (b) by seasons (spring, summer, autumn, and winter), and percentage of pixels with various numbers of good-quality observations in April-May (c) and December-January (d) on Chongming island during 1995–2018.

The time series Landsat SR data with good-quality observations were used to calculate three vegetation indices (VIs): Normalized Difference Vegetation Index (NDVI) (Tucker, 1979), Enhanced Vegetation Index (EVI) (Huete et al., 2002), and Land Surface Water Index (LSWI) (Xiao et al., 2005). NDVI is closely related to leaf area index (LAI), and EVI is more responsive to chlorophyll content in the canopy. Both NDVI and EVI are widely used in the studies of land surface phenology (Zhang et al., 2003). LSWI is a spectral indicator of canopy and soil moisture, and a change from positive LSWI value to negative LSWI value is used to represent a state of change from green leaf to senescent leaf (Xiao et al., 2009). These three vegetation indices are calculated using following equations:

$$NDVI=\frac{{\rho }_{NIR}-{\rho }_{Red}}{{\rho }_{NIR}+{\rho }_{Red}}$$ (1)

$$\mathit{\text{EVI}}=2.5\times \frac{{\rho }_{NIR}-{\rho }_{Red}}{{\rho }_{NIR}+6\times {\rho }_{Red}-7.5\times {\rho }_{Blue}+1}$$ (2)

$$\mathit{\text{LSWI}}=\frac{{\rho }_{NIR}-{\rho }_{SWIR}}{{\rho }_{NIR}+{\rho }_{SWIR}}$$ (3)

where ρ_Blue, ρ_Red, ρ_NIR, ρ_SWIR are the surface reflectance values of blue (450–520 nm), red (630–690 nm), near-infrared (NIR: 760–900 nm), and shortwave-infrared (SWIR: 1550–1750 nm) bands.

2.2.2. In-situ field survey data for training

We collected and organized in-situ data from different sources. First, two major field surveys on Chongming island were carried out in mid-July of 2012 and August to September of 2015, taking many GPS field photos. Second, a transect perpendicular to the levee through east to west was established in 2005, and nowadays there are totally five transects in the CDNNR. Those transects provided frequent vegetation information and geo-referenced field photos. According to field photos and very high spatial resolution (VHSR) imagery from Google Earth (GE) around 2012, we selected and digitized the training Region of Interests (ROIs) (Fig. 1). Although some field photos were taken in 2015, they also offered useful references to vegetation types through visual interpretation. We collected training ROIs across the Spartina saltmarsh (17 ROIs with a total of 610 pixels) and other saltmarshes dominated by Phragmites and Scirpus (10 ROIs with a total of 68 pixels).

2.3. Phenology of Spartina and other saltmarshes on Chongming island

We collected time series in-situ phenological observations in 2018–2019 for one Spartina site from an automated digital camera (Brinno BCC200). The camera has high performance HDR video sensor that generates a real time-lapse video at 5-second interval. We selected one photo with best quality per month to construct the growth cycle of Spartina, which serves as reference to interpret the time series Landsat data in tracking saltmarsh vegetation phenology (Fig. S1). According to the ground observation, Spartina starts to green-up in late April, grows rapidly from June to early September, and senesces in late December.

Time series VIs derived from Landsat 5/7/8 images from 2008 to 2012 were used to depict the seasonal dynamics and interannual variation of three saltmarsh types at the pixel scale. We selected three sites (Fig. 1a), which were dominated by Spartina, Phragmites and Scirpus saltmarshes during 2008–2012, respectively, according to the field photos and VHSR images in GE (Fig. 1b–g). Phragmites and Scirpus saltmarshes are the main native saltmarsh communities. The seasonal dynamics of vegetation indices of Spartina saltmarsh differed from those of Phragmites and Scirpus saltmarshes in two periods (Fig. 4): (1) April to May (spring; plant green-up) and (2) December to January (winter; plant senescence). The detailed phenological analysis for these three saltmarshes is given in Section 3.1.

Figure 4.

The seasonal dynamics of Landsat (TM/ETM+/OLI)-derived vegetation indices (a) NDVI, (b) EVI, and (c) LSWI at the three Spartina, Phragmites and Scirpus saltmarsh sites during 2008–2012. The landscapes of these three saltmarsh sites are shown in Fig. 1a.

2.4. A pixel- and phenology-based algorithm for identifying and mapping Spartina saltmarsh

The workflow diagram for identifying and mapping spatial-temporal dynamics of Spartina saltmarsh on Chongming island during 1995–2018 is given in Figure 2. For each year, we first delineated the intertidal zone, which stretches between the low-tide waterline, mudflats, wetland, and high-tide waterline. Second, we identified and mapped the coastal vegetation area. Third, we identified Spartina saltmarsh from other saltmarshes among the coastal vegetation maps. Since the ecological project was started in 2013, we used 2012 as the baseline year to first produce the intertidal zone, coastal vegetation area, and Spartina saltmarsh maps. We generated annual maps of Spartina saltmarsh during 1995–2018, and used them to analyze the expansion process (1995–2012) and removal dynamics (2013–2018) of Spartina saltmarsh on Chongming island. The detailed procedure is described below.

2.4.1. Delineating intertidal zone

Spartina encroachment into native saltmarshes is the main focus in this study, thus we are more concerned with the vegetation distributed in the intertidal zone. We delineated the coastline by visual interpretation and manual digitization of very high spatial resolution images, and then excluded the inland area by specifying intertidal buffer zone. A 6-km outward coastline buffer zone was generated to delineate the intertidal range. The boundary of buffer zone was modified in some regions where the border intersects with Jiangsu province and two southern islands (Hengsha island and Changxing island). Delineation of intertidal zone as potential distribution regions was the first step to assist for Spartina saltmarsh extraction.

2.4.2. Identifying coastal vegetation area

Effective discrimination of Spartina saltmarsh starts with the accurate coastal vegetation maps. In an earlier study, we had developed a decision tree classification algorithm for mapping coastal vegetation in China, based on Landsat TM/ETM+/OLI imagery (Wang et al., 2018; Wang et al., 2020). As a follow-up work, we produced coastal vegetation maps of Chongming island using the same algorithm. Individual good-quality observations in a year were first identified as green vegetation or not, using Eq. (4), and then vegetation frequency in a year is calculated, using Eq. (5).

$$\mathit{\text{Vegetation}}=NDVI\ge 0.2\cap EVI\ge 0.1\cap LSWI>0$$ (4)

$$VF=\frac{N_{\mathit{\text{Vegetation}}}}{N_{\mathit{\text{Good}}}}$$ (5)

where VF is green vegetation frequency scaled to 0 and 1 for a pixel in a year, N_Vegetation is the number of observations identified as green vegetation in a year, N_Good is the number of good-quality observations in a year. The frequency threshold value of 0.05 was used to classify a pixel as coastal vegetation area (VF ≥ 0.05) or non-vegetated area such as tidal flats (VF < 0.05) (Wang et al., 2018; Wang et al., 2020). The resultant annual coastal vegetation maps on Chongming island were used as the baseline map for the phenology-based identification and mapping of Spartina saltmarsh in the next step.

2.4.3. Identifying Spartina saltmarsh

As an invasive plant with high competitiveness, Spartina could rapidly dominate and replace the native saltmarsh communities, forming large patches on the ground. A pixel- and phenology-based algorithm was developed for Spartina saltmarsh mapping based on 30-m Landsat data. The two key phenology features of Spartina saltmarsh (Fig. 4) were used in the phenology-based algorithm: (1) later green-up in spring and (2) later senescence in winter. In April-May, Spartina saltmarsh had low NDVI, EVI and LSWI values, while the other saltmarshes (Phragmites and Scirpus) started to green-up with higher VIs values especially in LSWI (> 0). We calculated the mean value of LSWI during April to May (LSWI_{mean(Apr-May)}) for individual pixels, and the histogram suggested that LSWI_{mean(Apr-May)} less than 0 could discriminate Spartina from the other saltmarshes (see Section 3.1 for more details). In December-January, NDVI and EVI showed large differences between Spartina and the other two saltmarshes. The relatively high NDVI (> 0.2), EVI (> 0.1) and LSWI (> 0) values at the Spartina pixels represented green vegetation signal, which was another critical indicator to identify Spartina saltmarsh. The green vegetation here is consistent with the concept presented in Section 2.4.2. We counted the number of observations that were identified as green vegetation in December-January and calculated the green vegetation frequency (VF_Dec-Jan). Approximately 99% of Spartina saltmarsh pixels had VF_Dec-Jan > 0 (see Section 3.1 for more details). Therefore, the decision rules for identifying Spartina saltmarsh from the coastal vegetation map were LSWI_{mean(Apr-May)} ≤ 0 and VF_(Dec-Jan) > 0, as shown in Eq. (6).

$$\mathit{\text{Spartina saltmarsh}}=LSW{I}_{\mathit{\text{mean}}(\mathit{\text{Apr}}-\mathit{\text{May}})}\le 0\cap V{F}_{(\mathit{\text{Dec}}-\mathit{\text{jan}})}>0$$ (6)

where LSWI_{mean(Apr-May)} is the mean value of LSWI in April-May, and VF_(Dec-Jan) is the green vegetation frequency in December-January, calculated as the ratio of the number of observations identified as green vegetation to the number of good quality observations in December-January, similar to Eq. (5).

2.5. Accuracy assessment of the Spartina saltmarsh maps

For the validation datasets we first used the stratified random sample function in GEE to define and collect samples. When the sample size in each category is in proportion to the surface area of that category, the accuracy assessment is reliable (Olofsson et al., 2014; Olofsson et al., 2013). As Spartina saltmarsh and the other saltmarshes (Non-Spartina) had comparatively similar-sized areas, we set the same number of random points (100 points) for both categories. Next, the validation ROIs were delineated with a 30-m circle buffer of the random sampling points in ArcGIS, exported as Keyhole Markup Language (KML) file, and loaded into Google Earth. The VHSR imagery in GE is an effective data source to validate the land classification results (Huang et al., 2010; Kennedy et al., 2010). We inspected each ROI through visual interpretation of VHSR images and the geo-referenced photos. The ROIs without clear land cover information due to unavailable reference or cloud interference were excluded. In the VHSR images, the Spartina patches could be clearly identified from their clumpy configuration (Fig. 1b). Owing to the high invasion rate, the spatial distribution of Spartina saltmarsh was subject to interannual variation, making it necessary to take an annually intensive accuracy assessment. However, because of the limited numbers of VHSR images and field photos, we only assessed the accuracies of mapping results in four years (i.e., 2012–2014, and 2016). The total number of validation samples for Spartina and Non-Spartina were 71 and 61 in 2012, 75 and 80 in 2013, 82 and 86 in 2014, 61 and 91 in 2016, respectively (Fig. S2). The confusion matrix of Spartina saltmarsh maps was calculated to estimate the accuracy of the maps. Following the previous study (Olofsson et al., 2013), we adjusted accuracies of the resultant maps by considering the area of each category, and the area estimates were also corrected based on the area-adjusted accuracies.

2.6. Comparison with other available Spartina saltmarsh datasets

Besides the validation by in-situ data and VHSR images, a comparison between our Landsat- and phenology-based Spartina saltmarsh maps and other Spartina saltmarsh datasets would be informative. Unfortunately, no other Spartina saltmarsh maps are available to the public for spatial comparison. A couple of publications reported only the area estimates of Spartina saltmarsh, thus we did the areal comparison in this study. The Resource Monitoring Reports from Shanghai Chongming Dongtan National Nature Reserve were published annually during 2006–2015, and provided data of plants, macrobenthos, zooplankton, fishes, and waterfowl (http://www.dongtan.cn/). In most cases, they analyzed either satellite images (e.g., HJ-1a/b and ZY-02C/3) or airborne images with maximum likelihood supervised classification method for monitoring and reporting the dominant plant communities. The classification results were corrected with field survey data and validated by various sources (e.g., aerial photos) with an accuracy over 85%.

3. Results

3.1. Phenology of Spartina saltmarsh as observed by Landsat time series images

We combined all the available good-quality observations during 2008–2012 at three sites (Fig. 1a) to show the seasonal dynamics of NDVI, EVI and LSWI of Spartina, Phragmites and Scirpus saltmarshes (Fig. 4). In winter, NDVI and EVI values at these three sites were very low and in similar range during February-March. In spring, NDVI, EVI and LSWI values at the Phragmites site rose substantially during late April-early May, which were driven by new leaf flush and rapid growth of leaves. LSWI values at the Scirpus site also rose substantially in late April-early May. In comparison, NDVI, EVI and LSWI values at the Spartina site remained to be very low in late April-early May, including LSWI remained negative value (< 0). The seasonal dynamics of NDVI, EVI and LSWI in spring season suggested that Phragmites and Scirpus started to green-up in late April and early May, and Spartina started to green-up in late May. The Phragmites and Scirpus saltmarsh sites reached their highest NDVI, EVI and LSWI values in the early summer, but NDVI, EVI and LSWI values at the Spartina saltmarsh site peaked at the late summer. NDVI and EVI values at the Phragmites and Scirpus sites dropped substantially by November-December, reaching NDVI < 0.2 and EVI < 0.1, which were driven by leaf senescence. In comparison, the Spartina site still remained green in November-December, having relative higher NDVI (> 0.2) and EVI (> 0.1) values. The seasonal dynamics of NDVI, EVI, and LSWI in fall and winter seasons suggested that Pragmites and Scipus started to experience leaf senescence in November and December, and Spartina did not enter senescence until January of the next year. These phenological characteristics as observed by three vegetation indices suggested that April-May and December-January are two temporal periods that we can use Landsat images to separate Spartina saltmarsh from Phragmites and Scirpus saltmarshes on Chongming island.

Fig. 5 shows the relationships between greenness-related indices (NDVI, EVI) and water-related index (LSWI) in spring (April-May) and winter (December-January) during 2010–2012 based on the training dataset (see 2.2.2 section). In spring, most observations of the Spartina saltmarsh pixels had negative LSWI values, while most observations of the other saltmarshes (Pragmites and Scirpus) pixels had positive LSWI values (Fig. 5a–b). In winter, most observations of the Spartina saltmarsh pixels had NDVI values greater than 0.2, EVI values greater than 0.1, and LSWI values greater than 0, while most observations of the other saltmarshes pixels had NDVI values smaller then 0.2 and EVI values smaller than 0.1 (Fig. 5c–d). These scatterplots also illustrated the potential to separate Spartina saltmarsh from the other two native saltmarshes by vegetation indices in the spring and winter seasons.

Figure 5.

The two-dimensional scatterplots between greenness-related indices (NDVI, EVI) and water-related index (LSWI) of Spartina and Non-Spartina saltmarshes based on the training dataset in April-May (spring season) (a, b) and December-January (winter season) (c, d) during 2010–2012.

We also used the training data and the histograms to explore and visualize the separability between Spartina and the other two saltmarshes in spring (April-May) and winter (December-January) during 2010–2012 (Fig. 6). In April-May, LSWI was a more effective indicator than NDVI and EVI for identifying Spartina saltmarsh (Fig. 6a–c). The mean value of LSWI in April-May (LSWI_{mean(Apr-May)}) indicated that a threshold of 0 could be used for identifying Spartina saltmarsh (Fig. 6d). In December-January, NDVI and EVI were more effective indicators than LSWI for identifying Spartina saltmarsh (Fig. 6e–g). The histogram of green vegetation frequency in December-January (VF_(Dec-Jan)) showed that the threshold of 0 can separate 99.7% of Spartina saltmarsh pixels (VF_(Dec-Jan) > 0) from 98.5% of Non-Spartina saltmarsh pixels (VF_(Dec-Jan) < 0) (Fig. 6h). The results from the training datasets clearly illustrated phenological differences between Spartina saltmarsh and the other saltmarshes (Phragmites and Scirpus), which were the foundation for our pixel- and phenology-based algorithm to identify and map Spartina saltmarsh on Chongming island.

Figure 6.

The histograms of vegetation indices (NDVI, EVI, and LSWI), LSWI_{mean(Apr-May)} and VF_(Dec-Jan) of Spartina and Non-Spartina saltmarshes based on the training dataset in April-May (a-d) and December-January (e-h) during 2010–2012.

3.2. Annual map of coastal vegetation and Spartina saltmarsh on Chongming island in 2012

According to the annual map of coastal vegetation and Spartina saltmarsh in 2012 (Fig. 7), there were 2,067, 1,312 ha, and 1,253 ha Spartina saltmarsh in 2012 on Chongming island, CDNNR, and DEPA, respectively. Spartina encroachment mainly occurred in the northern and northeastern part of Chongming island, which can be seen as two encroachment clusters (see box 1 and 2 in Fig. 7a). The northern cluster was located where tidal flats had been expanding outward over years, and Spartina saltmarsh gradually colonized newly emerging tidal flats. The northeastern cluster was spatially in line with the artificial Spartina planting history, accounting for ~56% of the total Spartina saltmarsh area on Chongming island.

Figure 7.

The coastal vegetation and Spartina saltmarsh maps in 2012 on Chongming island, China. (a) annual map in 2012, (b, c) zoom-in view of two regions labeled as 1, 2 in (a), respectively.

3.3. Accuracy assessment of the Spartina saltmarsh maps

We carried out the accuracy assessment for annual maps of Spartina saltmarsh in 2012, 2013, 2014, and 2016, based on the availability of the validation datasets (Table 2). The overall accuracies (OA) were 0.93, 0.94, 0.91, and 0.94, and the Kappa coefficients were 0.86, 0.87, 0.82, and 0.88 in 2012, 2013, 2014, and 2016, respectively. The Spartina saltmarsh had producer accuracies (PA) of 0.90, 0.91, 0.89, and 0.92, and user accuracies (UA) of 0.99, 0.96, 0.93, and 0.93 in these four maps. We also calculated the error-adjusted producer’s (APA), user’s (AUA) and overall (AOA) accuracies (Olofsson et al., 2013). The adjusted OAs (AOA) ranged between 0.90–0.94, which were slightly lower than the unadjusted OAs (0.91–0.94). The adjusted PAs (APA) dropped substantially to 0.70, 0.66, 0.59, and 0.46, respectively. This discrepancy between PA and APA could be explained by the variation in estimation weights associated with the areas of each category. We calculated the Spartina saltmarsh mapping areas (Area) and also error-adjusted estimated areas (EArea) with 95% confidence interval from these four annual maps. The adjusted areas for Spartina saltmarsh increased substantially, for example, increasing from 2,119 ha to 3,082 ha in 2013 on Chongming island.

Table 2.

A summary for the accuracy assessment for annual maps of Spartina saltmarsh and the other saltmarshes in 2012, 2013, 2014, and 2016 on Chongming island. Producer’s (PA), user’s (UA) and overall (OA) accuracies and error-adjusted producer’s (APA), user’s (AUA) and overall (AOA) accuracies. This table also shows the mapped areas (Area) and estimated areas (EArea) with 95% confidence interval.

Year	Class	PA	UA	OA	Area(ha)	APA	AUA	AOA	EArea(ha)
2012	Spartina	0.90	0.99	0.93	2067.26	0.70	0.99	0.90	3173.57±756.83
	Non-Spartina	0.98	0.87			1.00	0.87
2013	Spartina	0.91	0.96	0.94	2119.09	0.66	0.96	0.92	3082.16±767.51
	Non-Spartina	0.96	0.91			0.99	0.91
2014	Spartina	0.89	0.93	0.91	1809.79	0.59	0.93	0.90	2856.55±755.51
	Non-Spartina	0.93	0.90			0.99	0.90
2016	Spartina	0.92	0.93	0.94	728.78	0.46	0.93	0.94	1469.86±691.30
	Non-Spartina	0.96	0.95			1	0.95

3.4. Expansion and removal dynamics of Spartina saltmarsh during 1995–2018

To reduce uncertainties caused by image quality or other factors in individual years, we aggregated annual maps of Spartina saltmarsh within multi-year (three to five years) interval and then reported its maximum extent at multi-year epochs (1995–1999, 2000–2004, 2005–2009, 2010–2012) during 1995–2012, which is considered as Spartina expansion period before the Spartina removal project was started in December 2013. Fig. 8a shows the Spartina saltmarsh expansion dynamics at 30-m spatial resolution in four multi-year epochs. The spatial-temporal analysis showed that Chongming island experienced a significant Spartina encroachment and expansion in the north and northeast part. The expansion process of Spartina saltmarsh in the north and northeast of Chongming island was depicted in spatial detail (Fig. 8b–c). Spartina saltmarsh had clearly extended into the sea side over the past two decades, which was in accordance with the trajectory of tidal flats (Wang et al., 2018). It also showed the opposite direction of expansion toward high mudflat in some regions, replacing the native saltmarshes such as Phragmites saltmarsh. We calculated the annual, multi-year, and cumulative areas of Spartina saltmarsh during 1995–2012 in the Chongming island (Fig. 8d), CDNNR (Fig. 8e), and DEPA (Fig. 8f), respectively. The Spartina encroachment includes three stages: colonization, rapid expansion, and successful invasion. In the 1990s and early 2000s, Spartina was intentionally planted and naturally propagated with very small patches that hardly detected from the Landsat images. Spartina encroachment had entered successful invasion phase since the mid-2000s and then expanded linearly and continuously from ~729 ha in 2004 to ~2,067 ha in 2012 on Chongming island (Fig. 8d).

Figure 8.

The expansion dynamics of Spartina saltmarsh during 1995–2012 on Chongming island. (a) spatial distribution of Spartina saltmarsh at multi-year epochs, (b, c) zoom-in view for the region highlighted by the red box in (a). The annual, multi-year, and cumulative areas of Spartina saltmarsh in the Chongming island (d), CDNNR (e), and DEPA (f) during 1995–2012.

The ecological engineering project of controlling Spartina was implemented in 2013–2016. Based on the experimental results for removing Spartina plants in 2012 at the experimental fields, staff at the CDNNR used a procedure that first cut Spartina plants at the flowering stage and then inundate (flood) the fields over 6 months, which would ensure full death of Spartina plants. The annual Spartina saltmarsh maps during 2013–2018 showed the spatial-temporal changes of Spartina saltmarsh (Fig. 9a–f). Spartina saltmarsh area decreased substantially in the project area (DEPA) during 2013–2016 at the rate of ~440 ha per year (P < 0.01) (Fig. 9i), which is in line with the project’s cutting plan. According to the reports from the engineering project, Spartina saltmarsh area was reduced by more than 95% in 2016. The Landsat-based Spartina saltmarsh maps also showed that the Spartina saltmarsh area was reduced by ~99% during 2013–2016. The agreement between the project’s report and our Landsat-based analysis demonstrates the potential and capacity of Spartina saltmarsh monitoring by using Landsat images and the algorithm.

Figure 9.

The removal dynamics of Spartina saltmarsh during 2013–2018 on Chongming island. (a-f) annual maps of Spartina saltmarsh, (g-i) annual area of Spartina saltmarsh in the Chongming island, CDNNR and DEPA, respectively. The red and blue polygons represent the CDNNR and DEPA boundary. The linear regression model was applied for the Spartina saltmarsh area data in 2013–2016.

4. Discussion

4.1. Area estimates and spatial-temporal dynamics of Spartina saltmarsh

According to our Landsat-based annual maps of Spartina saltmarsh on Chongming island during 1995–2012, Spartina saltmarsh area had a slowly increasing phase in the 1990s and a rapidly increasing phase in the 2000s (Fig. 8d–f). In our literature review, we found only two studies that reported Spartina saltmarsh area on Chongming island, 1,239 ha in 2003 (Li et al., 2006) and 3,132 ha in 2007 (Lu and Zhang, 2013), which were much larger than our estimates. A few studies reported Spartina saltmarsh area in the CDNNR (Huang, 2009; Lin et al., 2018; Luo, 2019; Wang, 2007), and the area estimates of Spartina saltmarsh varied among these publications. According to the study by Wang (2007), Spartina saltmarsh area in the CDNNR was 466 ha in 2000, 579 ha in 2002, 1,112 ha in 2003 and 1,441 ha in 2004. Another study by Huang (2009) showed that Spartina saltmarsh area in the CDNNR was 187 ha in 2000, 932 ha in 2003, 1,377 ha in 2008 Spartina saltmarsh. Spartina saltmarsh area estimates from Lin et al. (2018) was 345 ha in 2000, 1,114 ha in 2003, 1,184 ha in 2007, 463 ha in 2009 and 837 ha in 2013. In a study of mapping Spartina saltmarsh along the coastal zones of mainland China, Liu et al. (2018) analyzed 43 Landsat 8 OLI images from 2014 to 2016 and reported 965 ha of Spartina saltmarsh in the CDNNR in 2015, which is higher than our mapping result (668 ha) using images from 2015. Ge et al. (2013) developed a process-based grid model to simulate Spartina expansion in the CDNNR and reported that Spartina saltmarsh area increased from 107 ha in 2001 to 922 ha in 2006 and then slowed down in 2007–2008. Our results (Fig. 8e) show that the Spartina saltmarsh area continued to increase linearly up to 2012. The dramatic increase of Spartina saltmarsh in the 2000s was the result of competitive replacement of native Phragmites by invasive Spartina, driven by high salinity (Li et al., 2009). The CDNNR department reported that in 2012 the reserve had a total area of 4,265 ha saltmarshes, including 1,958 ha Spartina saltmarsh (accounting for ~45.9% of the total saltmarsh area), which are approximately 30% larger than the area estimates from our Landsat-based annual map in 2012 (1,312 ha). However, when we overlaid all annual maps of Spartina saltmarsh in 2011–2013, the resultant maximum area of Spartina saltmarsh is 1,650 ha. It should be pointed out that the CDNNR monitoring report classified saltmarsh vegetation by visual interpretation of very high spatial resolution images (e.g., ZY1–02c), it included small patches of Spartina saltmarsh, which are likely to be missed by the pixel-based algorithms and Landsat images.

During those years the ecological engineering project was implemented (2013–2016), our annual maps show that Spartina saltmarsh area in Dongtan ecological project area (DEPA) dropped linearly and was 19 ha in 2016 (Fig. 9g). Note that the CDNNR did not report Spartina saltmarsh area during 2016–2018. To date, our Landsat-based annual maps during 1995–2018 is the only dataset that documented annual dynamics of Spartina saltmarsh in the reserve. Our maps also show that Spartina saltmarsh area outside the CDNNR had large increase in 2015 but large drop in 2016 in the 2015/2016 El Nino year (Fig. 9). The large drop of Spartina saltmarsh in 2016 was attributed to the die-back and removal by the CDNNR staff (Ma Qiang, personal communication from the CDNNR). By comparing the Spartina saltmarsh maps in 2013, 2016, and 2018 (Fig. 9 and S3), we also found a small number of newly emerging Spartina saltmarsh patches in 2018. These newly emerging Spartina invasions in the DEPA and the large Spartina saltmarsh area outside the CDNNR clearly suggest that the CDNNR and researchers need to pay close attention to the remaining Spartina populations and develop a continuous monitoring and management plan for the CDNNR.

4.2. The potential of the pixel- and phenology-based algorithm for Spartina saltmarsh mapping

Many studies have used single or multi-date satellite images to identify and map Spartina saltmarsh in China (Table 1). Most of these studies calculated the spatial statistics of spectral bands and vegetation indices and used supervised or unsupervised classification algorithms to classify Spartina saltmarsh and generate maps of Spartina saltmarsh (Huang and Zhang, 2007a; Liu et al., 2016; Zhang et al., 2017). The metrics and parameters in these algorithms are affected by selected images used in the calculation of the spatial statistics of individual images, as saltmarsh plants have strong and diverse phenology in a year. Numerous studies have shown that vegetation phenology can be well depicted by time series vegetation indices from the MODIS sensors that acquire images at daily interval (Wu et al., 2017; Zhang et al., 2006; Zhang et al., 2003). In this study, our results show that there are reasonable numbers of time series Landsat images, which were acquired at 16-day interval, to describe phenology of saltmarshes on Chongming island, China (Fig. 4). This can also refine the studies that use single or multi-date images, as researchers could now carry out phenological analysis of all Landsat images in a year and then select appropriate images for data analysis.

Our study investigated phenology of Spartina saltmarsh and other saltmarshes (Phragmites and Scirpus) on Chongming island, and the unique phenological characteristics of these saltmarshes in spring green-up and winter senescence stages were identified (Fig. 5). This is similar to a previous study (Ai et al., 2017) that tracked the phenological variations of Spartina and native species by using GF-1 WFV imagery. We also found that both greenness-related vegetation indices (NDVI, EVI) and water-related vegetation indices (LSWI) are useful to separate invasive Spartina saltmarsh from native Phragmites and Scirpus saltmarshes on Chongming island, which is consistent with our previous studies that aimed to identify and map inland natural and agricultural wetlands (Dong et al., 2015; Dong et al., 2016; Xiao et al., 2006; Xiao et al., 2005). Both NDVI and LSWI were also used in recent studies that identified and mapped Spartina saltmarsh in China (Liu et al., 2018).

As described in Table 1, we developed a mapping tool that uses time series Landsat images and a pixel- and phenology-based algorithm to identify and map Spartina saltmarsh over years. The simple algorithms are built upon the phenological differences between Spartina saltmarsh and other saltmarshes (Phragmites and Scirpus) in spring (April-May) and winter (December-January) (Fig. 5). Several studies that used single or multi-date Landsat images also considered phenological stages and tidal level and selected images in spring and winter (Liu et al., 2018; Mao et al., 2019; Wang et al., 2015). In our study, we used all the available time series Landsat images, in this way, we have more good-quality observations and could better capture phenological differences between Spartina and other saltmarshes in a year. The pixel- and phenology-based mapping algorithm and time series Landsat data have been used to identify water-related land cover types, including open surface water (Zou et al., 2017; Zou et al., 2018), paddy rice (Dong et al., 2015; Dong et al., 2016; Zhang et al., 2015) and coastal wetlands (Wang et al., 2018; Wang et al., 2020). The spatial-temporal dynamics of Spartina saltmarsh on Chongming island reported from this study showcases that the human management of invasive species (e.g., ecological engineering project) could be accurately assessed by the time series Landsat images and pixel- and phenology-based mapping algorithm.

4.3. Sources of errors and limitations of the resultant annual maps of Spartina saltmarsh

In general, the quality and accuracy of land cover map is affected by a number of factors: (1) good-quality image data, (2) availability and quality of in-situ training sample data, (3) algorithms, and (4) definition of land cover type (Foody, 2002; Gong et al., 2012). The annual maps of Spartina saltmarsh from this study also have a number of sources of errors and limitations. First, the implementation of pixel- and phenology-based algorithm in coastal wetlands was largely restricted by the amount of good-quality observations in specified phenological phases, which depends upon observation frequency (16-days revisit cycle) and data quality (e.g., cloud and cloud shadow). For example, more than 80% of the pixels had zero good-quality observation in April-May of 2003 and December-January of 2007 (Fig. 3), which would result in high omission errors in the annual maps of 2003 and 2007. For this reason, we also reported multi-year area estimates (e.g., 1995–1999, 2000–2004, 2005–2009, 2010–2012) and combined annual Spartina saltmarsh maps to generate cumulative Spartina saltmarsh extent for documenting expansion dynamics of Spartina saltmarsh (Fig. 8). Second, there are some small patches of Spartina saltmarsh, thus mixed-pixel phenomenon is unavoidable in Landsat 30-m spatial resolution imagery. A pixel may be a pure pixel or a mixed pixel with plants, soil, and sea water. The process of Spartina plants encroaching into a new pixel involve generally four phases: introduction, establishment, expansion, and dominance (Vaz et al., 2018). During the stage of introduction and establishment, Spartina often has small canopy and occupies small fraction area within 30-m pixels. It is a challenge to classify pixels with mixed vegetation species in the field of land-cover mapping (Gong et al., 2012; Herold et al., 2008; Wang et al., 2017). In our pixel- and phenology-based study, we aimed to identify those pixels dominated by Spartina, which means that the Spartina will only be detectable once it reaches high coverage (relative to pixel size) to influence the phenological signal. That could be one reason that our Spartina saltmarsh estimate area is lower than the area estimates from the annual reports of the Chongming Dongtan National Nature Reserve. Sentinel-2A/2B data together constitute time series data at higher temporal frequency (5-day) and spatial resolution (10-m, 20-m), thus, future effort in mapping and monitoring Spartina saltmarsh should combine both Landsat and Sentinel-2 images together. Third, LSWI, calculated as a normalized ratio between NIR and SWIR bands, is useful for measuring both leaf water contents and soil moisture (Chandrasekar et al., 2010). Sea water or moist soils in coastal wetland, such as the spatial-temporal variability of flooding, ponding and soil wetness could also affect NIR and SWIR bands, which might introduce some uncertainties (i.e., omission error) in identifying Spartina saltmarsh when using only LSWI-based criterion, especially for those saltmarshes close to seashore. In this study, we used all the available Landsat images, which help reduce the classification errors in the resultant annual and multi-year Spartina saltmarsh maps (Table 2).

5. Conclusions

In this study, we used a pixel- and phenology-based algorithm and time series Landsat 5/7/8 images to identify and map Spartina saltmarsh. Our case study on Chongming island shows that annual maps of Spartina saltmarsh track well the expansion dynamics of Spartina saltmarsh during 1995–2012 and removal dynamics during 2013–2016. The resultant annual and multi-year maps of Spartina saltmarsh can be used to support various studies that aim to understand the driving factors of Spartina saltmarsh dynamics and assess the impacts of Spartina saltmarsh expansion on biodiversity, carbon cycle and ecosystem services. The limited number of Landsat images in a year due to the 16-day revisit cycle, tidal dynamics, and mixed coastal vegetation communities could still cause some uncertainties in Spartina saltmarsh maps. As time series images from other optical sensors with higher temporal and spatial resolution, for example, Sentinel-2 A/B (10-day revisit cycle, 10-m spatial resolution), become freely available, the proposed algorithm in this study could be by combining both Landsat and Sentinel-2 time series images to identify and map saltmarshes over years in the coastal zones, where billions of people depend upon its goods and services.

Highlights.

1. Time series Landsat images tracked phenology of Spartina saltmarsh in a year.

2. A pixel- and phenology-based algorithm was developed to map Spartina saltmarsh.

3. Expansion and removal of Spartina saltmarsh during 1995–2018 were documented.