Spatiotemporal-based automated inundation mapping of Ramsar wetlands using Google Earth Engine

Wetlands are one of the most critical components of an ecosystem, supporting many ecological niches and a rich diversity of flora and fauna. The ecological significance of these sites makes it imperative to study the changes in their inundation extent and propose necessary measures for their conservation. This study analyzes all 64 Ramsar sites in China based on their inundation patterns using Landsat imagery from 1991 to 2020. Annual composites were generated using the short-wave infrared thresholding technique from June to September to create inundation maps. The analysis was carried out on each Ramsar site individually to account for its typical behavior due to regional geographical and climatic conditions. The results of the inundation analysis for each site were subjected to the Mann–Kendall test to determine their trends. The analysis showed that 8 sites exhibited a significantly decreasing trend, while 14 sites displayed a significantly increasing trend. The accuracy of the analysis ranged from a minimum of 72.0% for Hubei Wang Lake to a maximum of 98.0% for Zhangye Heihe Wetland National Nature Reserve. The average overall accuracy of the sites was found to be 90.0%. The findings emphasize the necessity for conservation strategies and policies for Ramsar sites.

Despite covering only about 2.6% of the earth's land area, wetlands are vital to the hydrological cycle and play a significant role in regulating water flow and quality 1 .Moreover, wetlands are responsible for the production of over 20% of the earth's organic carbon 2 , making them an essential source of nutrients and energy for many aquatic and terrestrial ecosystems 3 .The availability of adequate food and water makes them the best place for diverse species forms 4,5 .Wetlands provide a wide range of vital ecosystem services, such as purifying water, controlling floods, conserving biodiversity, supplying food, and sequestering carbon 6 .Unfortunately, as one of the most vulnerable ecosystems, wetlands have suffered significant losses and degradation worldwide due to climate change and human activities [7][8][9] .As wetlands have a close relationship with the climate, any changes in their behavior reflect the changing climatic conditions and vice versa 10,11 .
The Ramsar Convention, also known as the Wetlands Convention, is an international treaty signed on February 2, 1971, in the city of Ramsar, Iran 12 .Its objective is to conserve and sustainably use wetlands, which are designated as Ramsar sites (Please refer to S1 in the Supplementary Information to understand the criteria adopted for identification of wetland as a Ramsar site).The convention provides a framework for the protection and responsible use of wetlands 11 .The mission of the convention is to achieve sustainable development worldwide by conserving and wisely using wetlands through local and national actions and international cooperation 13 .Implementation of the Ramsar Strategic Plan contributes to the achievement of the Sustainable Development Goals (SDGs) 14 .The Ramsar Conventions' fourth strategic plan (2016-2024) identifies addressing the drivers of wetland loss and degradation, effectively conserving and managing the Ramsar site network, wisely using all wetlands, and enhancing implementation as four overarching goals 13 .Most of the proposed SDGs are relevant in some way or another to wetlands, but the following are of particular importance: wetlands ensure fresh water, help replenish ground aquifers, and purify and filter harmful waste from water (Goal 6 of SDG) 14 .Rice grown in wetland paddies is the staple diet of nearly three billion people (Goal 2 of SDG) 13 .They also help reduce drought and contribute to the land formation and coastal zone stability by regulating sediment transport (Goal 11 of

Study area
China became a member of the Ramsar Convention in 1992 and has since designated 64 Ramsar sites as of 2021, located in 24 provincial-level regions (Fig. 1).Heilongjiang has the most sites with 10, followed by Gansu, Guangdong, Hubei, Yunnan, Tibet, and Inner Mongolia with four sites each.These wetlands include all types defined by the Ramsar Convention and have a carbon sink capacity of more than 1.71 million metric tons per year 38 .The Ramsar sites in China cover various natural wetland types such as swamps, marshes, lakes, rivers, mangroves, tidal flats, estuaries, and shallow marine water.Of these, 17 receive more than 1500 mm of mean annual precipitation; 25 and 23 sites were found in the humid subtropical and continental climate regions,

Trend analysis
The Mann-Kendall (MK) test was carried out under the presumption that a significant trend is the one with a p value less than 0.05 (also represented by an absolute Z c score greater than 1.96) 42,43 .Since each site had a varied number of maps, the analysis was done using the information that was available, and the patterns were extrapolated to cover the entire 30-year period to allow for a consistent comparison of all the locations.Observed trends from the trend analysis suggest that 35 sites (Fig. 3) follow an increasing trend with a positive MK test statistical   4).
While the MK test is a useful tool, there are some limitations and concerns associated are: (i) the MK test does not account for seasonal patterns or repeated variations in the data.(ii) the MK test assumes that datasets are collected at equal time intervals.If the data does not meet this assumption, then data interpolation or resampling need to be performed.(iii) the trend analysis was conducted using the data available at each site.Variations in  www.nature.com/scientificreports/Trari Nam Co Wetlands, Zhaling Lake, Maidika, and Tibet Selincuo Wetlands.Similarly, most of the sites with decreasing or significantly decreasing trends have elevation values less than 4000 m except Bitahai Wetland, Gansu Yellow River Shouqu Wetlands, Mapangyong Cuo, and Sichuan Changshagongma Wetlands (Fig. 5b).
An analogy was also found with respect to the precipitation that the wetlands received.Most of the sites with increasing or significantly increasing trends have average precipitation values greater than 500 mm, except Gansu Yanchiwan Wetlands, Jilin Momoge National Nature Reserve, and Tibet Trari Nam Co Wetlands.Most of the  sites with decreasing trends have average precipitation values less than 1500 mm except Fujian Zhangjiangkou National Mangrove Nature Reserve, Guangxi Beilun Estuary National Nature Reserve, Huidong Harbor Sea Turtle National Nature Reserve, Jiangxi Poyang Lake Nanji Wetlands, Shankou Mangrove Nature Reserve (Fig. 5a).
It can also be inferred that most of the sites with significantly increasing trends have maximum temperature values less than 35 °C except Anhui Shengjin Lake National Nature Reserve, Hangzhou Xixi Wetlands, Hubei Wang Lake.On the other hand, most of the sites with significantly increasing trends have maximum temperature values less than 30 °C except Shandong Yellow River Delta Wetland, Henan Minquan Yellow River Gudao Wetlands, Henan Minquan Yellow River Gudao Wetlands, Dafeng National Nature Reserve.Most of the sites with decreasing or significantly decreasing trends have maximum temperature values greater than 20 °C except Bitahai Wetland, Dashanbao, Gansu Gahai Wetlands Nature Reserve, Mapangyong Cuo, Sichuan Changshagongma Wetlands (Fig. 5c).
The analogy was also found with respect to the mean Temperature of the wetlands.Most of the sites with increasing or significantly increasing trends have mean temperature values less than 15 °C except Anhui Shengjin Lake National Nature Reserve, Chongming Dongtan Nature Reserve, Hangzhou Xixi Wetlands, Hubei Chen Lake Wetland Nature Reserve, Hubei Wang Lake, Mai Po Marshes, Dafeng National Nature Reserve, Henan Minquan Yellow River Gudao Wetlands, Guangdong Haifeng Wetlands, and Inner Deep Bay.Among the sites showing increasing trends, Eling Lake, Gansu Yanchiwan Wetlands, Inner Mongolia Grand Khingan Hanma Wetlands, Maidika, Tibet Selincuo Wetlands, Zhaling Lake have negative mean temperature values.All the sites

Accuracy assessment
The accuracy assessment allowed for the thresholding method to be relied upon as an effective way to produce inundation maps.Due to several circumstances, including the nature of the site, the digitizing areas, and the difference in the spectral values of wet and dry areas, the accuracy of each site turned out to be varied.The Accuracy ranged from a minimum Overall Accuracy of 72.0% at Hubei Wang Lake to a maximum of 98.0% at Zhangye Heihe Wetland National Nature Reserve.The average Overall Accuracy of the sites was found to be 90.0%, with the average dry and wet Producer's Accuracies of 86.4 and 81.3%, respectively and the dry and wet User's Accuracy of 88.7 and 86.4% (Fig. 6).
The degree of accuracy was determined by contrasting a map made from a 4-month composite with a single image from a specific day (even though it was taken from the same period), which might not accurately depict the area that was entirely inundated.It either represented less extent due to no rainfall in the near time or more due to heavy rainfall that might have happened that day, thus creating a point of error.Secondly, digitizing wet and dry areas required a human process, which was simple in wetlands with open water and parts that were alternately permanently wet and dry.However, places in those wetlands with a lot of variation in the inundated and where every part had been inundated at some point in the series made digitization difficult since it confused www.nature.com/scientificreports/ the classifier.The disparity between the median wet and dry values was also lessened in lakes where the water was surrounded by greenery, and the dry area was also covered with vegetation.The shadows altered the pixel values, leading to an inaccurate estimate of the median wet and dry values and a classification error.Mixed pixels may also be a source of inaccuracy.Digitizing wet and dry areas was difficult due to the marshy environment because it was difficult to make judgments from the Landsat imageries, which reduced accuracy.Due to a number of factors, including some of those stated above, the average Overall Accuracy of this study was lower than the 95.9% discovered by Inman and Lyons 27 .

Conclusion
The effect of climatic change and commercial changes in the behavior of Ramsar wetlands has been given importance throughout the Ramsar report.This study focused on analyzing all 64 Ramsar sites in China based on their inundation pattern over the last three decades using pre-processed Landsat imageries for a period of 30 years (1991-2020).The technique of SWIR Thresholding was applied on these composites to generate inundation maps.The analysis was carried out on each Ramsar site individually to account for its typical behavior owing to regional climatic and geographical conditions.From the analysis, 37 sites showed an increasing trend, and 27 sites showed a decreasing trend.14 out of 37 sites were significantly increasing and 8 out of 27 showed significantly decreasing behavior.The accuracy ranged from a minimum Overall Accuracy of 72.0% at Hubei Wang Lake to a maximum of 98.0% at Zhangye Heihe Wetland National Nature Reserve (Fig. 2).The average Overall Accuracy of the sites was found to be 90.0%, the average dry and wet Producer's accuracies of 86.4 and 81.3%, respectively and the dry and wet User's Accuracy of 88.7 and 86.4%.This study helps to understand, through circumstances, the importance of wetlands and their wise management.
The limitation of study comprises various factors, including: the composites in the 1991-2004 time-period showed the most flaws, with some composites completely missing the B7 (SWIR) band due to poor Landsat 5 imageries.The coastal region adds to the complexity by forming clouds on a regular basis, making cloud masking difficult to process and resulting in regions of transparent (masked) pixels.Further, the degree of accuracy was determined by contrasting a map made from a 4-month composite with a single image from a specific day (even though it was taken from the same period), which might not accurately depict the area that was entirely inundated.The shadows altered the pixel values, leading to an inaccurate estimate of the median wet and dry values and a classification error.The mixed pixels may also be a source of inaccuracy.
For further study, the main focus would be on significantly decreasing sites, i.e., Dalai Lake National Nature Reserve, Dong Dongting Hu, Eerduosi National Nature Reserve, Hubei Dajiu Lake Wetland, Mapangyong Cuo, Nan Dongting Wetland and Waterfowl Nature Reserve, Poyanghu, and Sichuan Changshagongma Wetlands as they were found to be at the highest risk of extinction.For the conservation of these 8 sites, some machine learning models can be used to predict the future inundation of the sites and do the future analysis to find out possible reasons for the decrease in inundation area of the sites so that suitable measures can be taken to conserve the wetlands under threat.Future work could expand the management and conservation of wetlands by: (1) incorporating different satellite products at higher resolution; (2) using microwave and recently available Landsat-9 data where satellite temporal coverage is inadequate or cloud-covered; (3) mapping the spatiotemporal component of hazard with hotspots of climate impacts and risks; (4) establishing the science evidence base through recognized modelling and participatory risk assessment where modelling is not achievable; and (5) implementing actions that can lessen the vulnerability of wetlands to changing climate along with their management recommendations based on the risk assessment.

Methods
Numerous classification methods, including unsupervised, supervised band thresholding, band ratios, indices, various regression trees, and combinations of these methods, have been proposed in the past [44][45][46] .Out of these methods, band thresholding has distinguished itself as an effective and accurate method.Using MODIS imagery, Murray-Hudson et al. proposed a technique for thresholding the Short-Wave Infrared (SWIR) band and generated highly accurate findings 46 .The SWIR band is susceptible to moisture content on the Earth's surface.It can accurately differentiate inundated areas covered with dense vegetation from dryland 46 .This method is based on a simple formula for thresholding the SWIR band.Thus, its simplicity makes it time efficient, computationally straightforward, and readily applicable using concise algorithms; therefore, getting an edge over other complex methods of vegetation is another advantage of this method.The above-mentioned advantages of the band thresholding method are incorporated into the current study.This innovative study tries to assess all of China's Ramsar sites simultaneously and compare their various trends and features in order to draw some conclusions.The schematic flowchart of the process of generating inundation maps using GEE is shown in Fig. 7.

Cloud masking
Clouds and shadows can be seen in the Landsat sceneries; these elements must be hidden to produce correct composites and improve classification accuracy 47 .The pixels on the Landsat cloud mask band categorized as clouds or cloud shadows, were concealed for each scene [48][49][50] .The median value for the pixel from a year before or after the scenes' date was used to fill in these pixels as part of a gap-filling procedure 51,52 .

Landsat composites
The SWIR band (B7) is chosen for each scene, and a gap-filling method was then applied to cloud-masked images.All the scenes available from June to September of each year are used to build the composites 27,32 .This period corresponds to the growing season in many regions when wetlands experience peak vegetation growth and dynamic changes.Monitoring wetland condition during this period provides valuable insights into vegetation health, water availability, and the overall functionality of wetland ecosystems.Most of the wetlands in China endure yearly floods that coincide with the southwest monsoon, and they are most heavily inundated from June to September 53 .The value of the relevant pixel in the composites to be formed is determined by evaluating the median of the corresponding pixel values from all the scenes of that year for each pixel in the study area.

Filtering bad composites
The SWIR (B7) band wasn't present in all the composites that were produced.By deleting those composites from the image collection, they were manually screened.The majority of cloud masking is handled by the algorithm.But in practically every coastal site, there were some regions where the pixels were often categorized as clouds.The masking method turned those pixels transparent in these circumstances.A filtering method was used to filter these pixels.This resulted in a set of composites devoid of masking whatsoever 54 .From site to site, a different number of final flawless composites were produced 55 .For example, an average of 27 composites were obtained for each site, with a minimum of 6 composites on Tibet Selincuo wetland and a maximum of 30 composites on Dongfanghong Wetland National Nature Reserve, Donzhaigang, Eerduosi National Nature Reserve, Gansu Yanchiwan Wetlands, Guangdong Haifeng Wetlands, Guangxi Beilun Estuary National Nature Reserve, Heilongjiang Hadong Yanjiang Wetlands, Heilongjiang Qixing River National Nature Reserve, Heilongjiang Zhenbaodao Wetland National Nature Reserve, Huidong Harbor Sea Turtle National Nature Reserve, Inner Mongolia Bila River Wetlands, Shankou Mangrove Nature Reserve, Tianjin Beidagang Wetlands, Xi Dongting Lake Nature Reserve, Xingkai Lake National Nature Reserve, Zhaling Lake, and Zhangye Heihe Wetland National Nature Reserve.Out of 64 wetlands, 60 have more than 20 composites.

Creating inundation maps from the composites
By thresholding the SWIR band's pixel values from composite images, inundation maps are generated.First, we manually evaluated and digitalized each site's permanent wet (such as the lake's central region and permanent channels) and dry portions (like barren land or hill region near the wetland).To account for the dynamic seasonal and annual nature of the inundation patterns in the Wetlands, a composite specific SWIR threshold value is calcu- lated using Eq.(1).Using these digitized areas, the median SWIR values for wet ( SWIR wet ) and dry ( SWIR dry ) inundated areas were calculated for each composite.
The classifier evaluates each pixel's SWIR value against its SWIR threshold for each composite.Inundated pixels are those with SWIR values below SWIR threshold , and dry pixels are those with SWIR values above SWIR threshold .
To construct an inundation map, each pixel with a certain SWIR value is categorized and translated into one of the two values, namely 0 for dry pixels and 1 for inundated pixels.

Image-based accuracy assessment
To determine the thresholding method's applicability, particularly for China, remote sensing applications, validation was required.The simplest method might be to use Google Earth Pro's (GEP's) historical imagery to compare inundation maps with GEP.Based on the historical imageries that are now available in GEP, a random set of five years has been chosen for each Ramsar site.The random points function in GEE was used to create a set of 50 random points for each of the five years 56 .Using GEE's Sample Region function, the pixel values at each location were extracted from the inundation maps and exported as a CSV file with just two values: 1 for inundated pixels and 0 for dry pixels.After importing the KML file into GEP, each location was evaluated visually interpretation and categorized as either dry (i.e., 0) or inundated (i.e., 1).Due to the lack of in-situ data, we had to use this as our reference dataset.This procedure was carried out at each random point in the five-year period for each site for www.nature.com/scientificreports/ each year.Thus, for each location, a collection of 250 random points was obtained, together with reference data from the imagery and pixel values taken from the maps.This process confirmed 16,000 points (250 × 64 = 16,000).Each site's unique error matrix was produced.Several different types of accuracy were calculated, including Overall Accuracy (defined as the sum of the diagonal elements in the error matrix that were correctly classified and divided by the total sampled points), Producers Accuracy (defined as the diagonal entry in the error matrix for each column divided by its respective column total), and User Accuracy (defined as the diagonal entry in the error matrix for each row divided by its respective row total).

Mann-Kendall (MK) test
The Mann-Kendall (MK) methodology is a statistical test used to find trends in time series data.The MK test was run on each site separately to look for trends in the variation of inundation extent.This test was initially developed by Mann et al. in 57 and was further investigated by Kendall et al. in 58 and by Hirsch et al. in 59 .Blain et al. provided an assessment 42 .The MK test generates test statistics, which follows a known probability distribution, allowing for the calculation of p values.The MK test was carried out under the presumption that a significant trend is the one with a p value less than 0.05 (also represented by an absolute Z c score greater than 1.96) 42 .The p value is used to assess the significance of any observed trend and if less than a chosen significance level (typically 0.05) indicates a statistically significant trend, suggesting that the observed pattern is unlikely to be due to random chance 42,43 .

Data availability
To get the entire time series dataset from 1991 to 2020, we employed all three Landsat sensor imageries (Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI).Landsat 5 (availability: 1984 to 2012) and Landsat 7 (availability: 1999 to present) images contain four Visible and Near-Infrared (VNIR) bands and two short-wave infrared (SWIR or B7) bands processed to orthorectified surface reflectance, and one thermal infrared (TIR) band processed to orthorectified brightness temperature, while Landsat 8 (2013 to present) contains 11 bands.The SWIR bands in all the Landsat scenes have a 30 m/pixel resolution.The GEE has all the above datasets available in its data catalog and ready for usage (https:// devel opers.google.com/ earth-engine/ datas ets/ catal og/ LANDS AT_ LT05_ C01_ T1_ SR; https:// devel opers.google.com/ earth-engine/ datas ets/ catal og/ LANDS AT_ LE07_ C01_ T1_ SR; https:// devel opers.google.com/ earth-engine/ datas ets/ catal og/ LANDS AT_ LC08_ C01_ T1_ SR).For the analysis, most of the shapefiles of the wetlands were obtained from the official Ramsar website (https:// rsis.ramsar.org/).However, the shapefiles of Bitahai Wetland, Eerduosi National Nature Reserve, Jilin Momoge National Nature Reserve, Niaodao, Sichuan Ruoergai Wetland National Nature Reserve, Xi Dongting Lake Nature Reserve, and Zhanjiang Mangrove National Nature Reserve were manually created in QGIS using the coordinate information available on the official website at a scale of 10,000.Climate data such as average precipitation, maximum temperature, mean temperature, and average precipitation for each wetland, has been extracted from ECMWF reanalysis data (ERA5 monthly) 60 using GEE (https:// devel opers.google.com/ earth-engine/ datas ets/ catal og/ ECMWF_ ERA5_ MONTH LY).

Figure 1 .
Figure 1.Location map of Ramsar wetland sites in mainland China.The table represents names of the wetlands (NNR: National Nature Reserve and NR: Nature Reserve).This figure is created using QGIS 3.30.1 (Quantum Geographic Information System; https:// downl oad.qgis.org/ downl oads/) and the background shows the basemap of Stamen terrain from HCMGIS Plugin.

Figure 2 .
Figure 2. Inundation maps represent the number of years each pixel was inundated during the period 1991-2020 (June to September).This figure represents the first 15 out of 64 Ramsar sites.This figure is created using QGIS 3.30.1 (Quantum Geographic Information System; https:// downl oad.qgis.org/ downl oads/).

Figure 3 .
Figure 3. Boxplot representing the minimum, 1st quartile, mean, 3rd quartile, and maximum inundated areas for each Ramsar site.The arrows on each site represent the trend of the wetland during the period 1991-2020 using the Mann-Kendall test.Also, the year of occurrence of minimum and maximum inundation is mentioned for each site.This figure is created using MATLAB R2022a (https:// www.mathw orks.com/ produ cts/ new_ produ cts/ relea se202 2a.html).

Figure 4 .
Figure 4. Change in inundation area for significantly decreasing Ramsar sites.The figure represents the area of each wetland for each year from 1990 to 2020.This figure is created using Microsoft Excel 365, Version 2308 (https:// www.micro soft.com/ en-in/ micro soft-365/ excel).

Figure 6 .
Figure 6.Graph showing the number of composites, overall accuracy, user accuracy, and producer accuracy for all the sites in the study.This figure is created using MATLAB R2022a (https:// www.mathw orks.com/ produ cts/ new_ produ cts/ relea se202 2a.html).

There are 14 out of these 35 sites (Dafeng National Nature Reserve, Eling Lake, Guangdong Haifeng
This figure is created using QGIS 3.30.1 (Quantum Geographic Information System; https:// downl oad.qgis.org/ downl oads/).Wetlands, Heilongjiang Hadong Yanjiang Wetlands, Heilongjiang Youhao Wetlands, Heilongjiang Zhenbaodao Wetland National Nature Reserve, Henan Minquan Yellow River Gudao Wetlands, Lashihai Wetland, Niaodao, Shandong Yellow River Delta Wetland, Tibet Trari Nam Co Wetlands, Xingkai Lake National Nature Reserve, Zhaling Lake, and Zhanjiang Mangrove National Nature Reserve) have a significantly increasing trend, with MK test statistical value (Z c ) greater than + 1.96.Twenty-one out of a total of 64 sites were found to have decreasing trend with a negative MK test statistical value (Z c ).Of these, eight sites (Dalai Lake National Nature Reserve, Dong Dongting Hu, Eerduosi National Nature Reserve, Hubei Dajiu Lake Wetland, Mapangyong Cuo, Nan Dongting Wetland, and Waterfowl Nature Reserve, Poyanghu, and Sichuan Changshagongma Wetlands) were found to have a significantly decreasing trend with MK test statistical value (Z c ) less than − 1.96 (Fig. value (Z c ).