Introduction: In the tropics and subtropics, mangroves constitute a critical ecosystem that is under significant pressure despite providing a host of local to global ecosystem services, the annual value of which has been estimated at US$25 trillion (zu Ermgassen et al., 2020). These include coastal protection, provision of harvestable wood, tourism, fisheries and carbon sequestration. Mangroves form a highly dynamic biome with extensive changes occurring over short temporal baselines due to anthropogenic activities (e.g., clearing for expansion of agriculture/aquaculture and urban areas) and natural events (e.g., storms) and processes (e.g., erosion and deposition), including those exacerbated by climate change (e.g., sea level fluctuation).

The mapping and monitoring of these ecosystems have been recognized as being of critical importance for their protection and sustainable management. To this end, the Global Mangrove Watch (GMW) was established in 2011 as part of the Japan Aerospace Exploration Agency (JAXA) Kyoto & Carbon (K&C) Initiative. The first global mangrove maps (version 2.0) were made available in 2018, with a baseline map generated from 2010 (Bunting et al., 2018a) based primarily on optical (Landsat) satellite data but informed by Japanese L-band Synthetic Aperture Radar (SAR). Maps were subsequently produced for 1996, 2007, 2008, 2009, 2015 and 2016 by updating the 2010 baseline from changes detected using JERS-1 SAR (1996), ALOS PALSAR (2007–2010) and ALOS-2 PALSAR-2 (2015–2016). These GMW maps constituted the most complete maps of global mangrove extent and change available.

The version 2.0 layers nonetheless had areas missing or mapped with poor quality, most often due to a lack of cloud-free optical satellite data for the 2010 baseline. Therefore, an effort was initiated to update the version 2.0 GMW maps by improving the spatial extent, quality of the mapping and extending the annual time series to 2020.

Methods: As an update, the baseline year was maintained as 2010. However, areas where the v2.0 product was of lower quality were remapped using ESA Sentinel-2 optical data. In total, 11,262 Sentinel-2 acquisitions were downloaded and used for this analysis. Classification of mangrove extent was undertaken on a scene-by-scene basis rather than through the creation of image composites (i.e., merging multiple scenes using a metric such as greenest pixel). To derive training data for the classification, the existing GMW v2.0 map was sampled and regions outside of the GMW v2.0 map were manually annotated. The XGBoost binary classification algorithm was used for the analysis given its ability to make use of large training datasets John et al., (2020). The testing accuracies of the models (using 50,000 samples per class) were estimated to 97–99 %. The individual classifications were merged in two steps to create a probability for each pixel as to whether it is mangroves or not. To derive the final binary mask of mangrove extent, the probability surface was globally thresholded with a value of 0.5. The 2010 ALOS PALSAR imagery was finally used to perform a change detection to create a 2010 map.

The accuracy assessment was undertaken using 60 sites globally distributed. From 18 of those sites, the accuracy was found to be 94.6 % using 15,100 points, with individual sites ranging from 99.8 % to 87.4 %. The F1-score for the mangrove class was 94.8 %, with a 95th confidence interval ranging from 93.4 % to 96.0 %.

Using the updated 2010 baseline, the change was calculated using the JAXA L-band SAR data time-series, where the map-to-image change approach outlined in Thomas et al. (2018) and applied to the GMW v2.0 analysis of Bunting et al. (2018b) was employed. However, rather than only applying the change from 2010 to each year (1996, 2007, 2008, 2009, 2015, 2016, 2017, 2018, 2019 and 2020), the change detection was subsequently applied from each year to every other year, resulting in ten change maps for each year. The ten maps were summarised and only pixels that had been identified as mangroves > 5 times were classified as mangroves to produce the new (v3.0) GMW maps.

Results: The global mangrove extent for 2010 in the v2.0 product was estimated to 136,839 km2, increasing to 146,400 km2 in the v3.0 product. The additional 10,000 km2 of mangroves mapped in the v3.0 product is consistent across all the years (1996 to 2016) and largely results from the identification of areas that were not mapped in v2.0. The global trend from 1996 to 2016 was similar to that of v2.0, which identified an estimated net loss of 6,078 km2 of mangroves from 1996 to 2016, while v3.0 mapped an estimated net loss of 7,350 km2 for the same period. After 2016, the long-term global pattern of mangrove forest loss appears to have changed, with a small gain of about 400 km2 of mangroves identified from 2016 to 2020. However, given the ±1 pixel misregistration between the PALSAR and PALSAR-2 datasets, we have estimated the global confidence interval to be approximately ±5000 km2 and so this change of trend has yet to be confirmed.

The results and related statistics are visualised and available for dissemination to stakeholders in the science, government, corporate, NGO and practitioner communities via the Global Mangrove Watch Platform at https://globalmangrovewatch.org.

Future Work: Going forward, future work will be focused on the generation of a global 10 m mangrove baseline for the year 2020 using Sentinel-2. The existing GMW maps are provided at a resolution of 25 m and increasing the spatial resolution to 10 m will enable the mapping of finer features (e.g., river edges, fragmented mangroves) and a further improvement in quality over the existing GMW version 3.0 products. The work is ongoing and is expected to be completed in 2022.

Wetlands are amongst the planet‘s most productive ecosystems, providing a wealth of ecosystem services, e.g., nutrition, flood control, protection or biodiversity support. However, multiple threats like climate change, agricultural activities, hydrological modifications, etc., endanger these essential ecosystems. Thus, a consistent mapping and monitoring of global wetland ecosystems is critical to track changes and trends to support wetland conservation and sustainable management. Although EO data are ideal for large-scale wetland inventorying, their tremendous diversity makes remote detection particularly challenging. This diversity and resulting challenge have been tackled by many researchers applying different sensors (optical and radar) and mapping techniques to delineate wetland from non-wetland areas. We have developed an innovative methodology to derive information about wetlands and demonstrated the approach‘s capability throughout the past years through several projects. Our presentation will primarily focus on the results from the ESA projects “GlobWetland Africa“ and “GlobWetland Africa - Extension on Wetland Inventory“ and include details about the Copernicus High-Resolution Layer production for the Water and Wetness products in Europe. We will further present three national wetland inventories for the countries of Namibia, Tunisia, and Uganda.

Combining optical and radar observations with a hybrid sensor approach provides a more robust wetland delineation than single-sensor approaches. Optical imagery is more sensitive to the vegetation cover and radar imagery to the soil moisture content. Additionally, the higher frequency of observations stemming from the combined data streams contributes to a better characterization of seasonal dynamics, which is essential. Seasonal and temporary changes do not lead to false conclusions of the overall long-term trend in the wetland extent. Within the domain of optical remote sensing, wetland identification focuses on enhancing the spectral signature using bio-physical indices sensitive to water and wet soils (wetness) and subsequent derivation of a water and wetness probability index (Ludwig et al., 2019). Similarly, the radar-based algorithm builds on the detection of open water surfaces and surface soil moisture. After the separate optical and radar imagery processing, the data is fused into a combined water and wetness product (Water and Wetness Presence Index or WWPI) using both sensor systems‘ advantages. The fused products give information about the probability that a particular location contains water or wet soils, for which a rule-based classification can be applied to derive the wetland extent finally. Figure 1 shows an example of the output products, (a) the HRL WAW classification scheme applied to Uganda, and (b) the resulting WWPI highlighting the presence of water and wet soils.

The methodology focuses on the physical properties of water and wet soils but does not detect wetlands in the ecological sense. However, when mapping wetlands areas, one general issue is that no single scientific definition exists. The methodology applied on a large scale provides a consistent and objective map of water and wet surfaces, building the foundation for a final wetland classification according to specific user definitions and needs. Incorporating additional information (e.g., information about land use) leads to the final wetland map.

Reference:
Ludwig, C., Walli, A., Schleicher, C., Weichselbaum, J., & Riffler, M. (2019). A highly automated algorithm for wetland detection using multi-temporal optical satellite data. Remote Sensing of Environment, 224, 333-351.

Salt marshes provide extensive ecosystem services, including nursery habitat for fish species, recreation, coastal resilience, and carbon sequestration. The United States has the largest extent of mapped salt marshes. Therefore, it is critical to understand the ecosystem's carbon stock in the Contiguous United States (CONUS). While salt marshes and other blue carbon systems store most of their carbon within the soil, aboveground biomass is an important carbon indicator. Existing aboveground biomass models in tidal marshes have medium spatial resolution and limited geographic extent. To improve spatial resolution to 10 m, we evaluate the use of Sentinel-1 and 2 data for incorporation into the aboveground biomass prediction. To incorporate these satellite observations with temporally disparate in situ samples, we evaluate the stability of these locations using the Landsat time series finding that 71% of training data were stable from field sampling to remote sensing observation. Next, we conducted a data fusion machine learning regression combining Sentinel-1, Sentinel-2, and Landsat data to predict aboveground biomass in salt marshes. We compared model performance with in situ testing data across machine learning algorithms (Support-Vector Machines, Random Forest, and XGBoost), spatial scale (10 m, 30 m), and training data stability. The best performing model was the 10 m XGBoost using the stable data, which achieved Root Mean Square Error (RMSE) of 301.0 and 107.33 at the plot and site scale, respectively. We created an updated 2020 salt marsh extent with Sentinel-1/2, SRTM, and National Elevation Dataset and estimated 3.6 (3.1-4.1) Tg of aboveground carbon across the CONUS. We will also explore the drivers of salt marsh biomass at the HUC 6 watershed scale, using climate variables (precipitation and temperature), average regional sea-level rise, tidal amplitude, coastal chlorophyll, diffuse attenuation coefficient, and land use. Our results demonstrate the need to monitor these systems to enable management, restoration, and understanding resilience to climate change.

C-Band backscatter is mostly sensitive to changes in structure and water content of the canopy in
dense tropical rain forests. With the emerging Sentinel-1 time series data we can distinguish the
backscatter on different time scales. Therefore we can detect phenomena, which are not visible in
single time steps.
We use empirical mode decomposition (EMD) - a data-driven alternative to the Fourier transform –
to decompose Sentinel-1 time series into multiple sub-signals of different, characteristic timescales
(sub-seasonal, seasonal, and slow oscillations). We compare the original signal and the seasonal
sub-signal of Sentinel-1 with the water level of the Jurua River, a major tributary of the Amazon
river in western Brazil. We estimated the correlation between river water levels and Sentinel-1 time
series as an indicator for flood-related seasonality in the forest area.
We show, that the correlation of Sentinel-1 VH forest backscatter time series with the water level of
the Juruá river is higher in seasonally flooded forests than in non-flooded forests. The correlation
further increases when we use the seasonal sub-signal instead of the original Sentinel-1 signal. This
is partly due to general de-noising of the signal. Since the high correlation values are clustered in
the seasonally flooded forest, we can assume that the seasonality is also due to the common driver
of flood appearance. Analysing Sentinel-1 VV signals, we do not find such a diagnostic
relationship. While the overall correlation between Sentinel-1 VV and the water level is higher, the
correlation is not confined to flooded areas only. This indicates, that the Sentinel-1 VV signal has an
overall higher seasonality, but this seasonality is not driven by the forest flooding near the river.
Our results lead to two hypotheses: Firstly, during the flooded state of the forest there may be
double bounce scattering between the stem and the standing water from the part of the signal which
is not scattered in the canopy. In this case, the returning signlas are then depolarized again in the
canopy, which would increase the VH component of the signal during the flooded state compared to
the non-flooded state - as observed. Secondly, during the flooded state, the water content in the
canopy may be increased because of the standing water underneath, which would increase the
volume scattering directly in the canopy. Further research is necessary to distinguish the best
explanation.
Independent of the exact scattering mechanism, these results have implications for the calibration
and validation of Sentinel-1 data from the Amazon rainforest. Specifically, we cannot expect
Sentinel-1 VH signals from tropical rainforest to be homogeneous in space and time. The
observation that flooded rainforest can be discriminated from non-flooded forest with C-band time
series opens a wealth of new important applications for Sentinel-1 time series data usage. It also
encourages further research on the power of time series exploration with intelligent algorithms and
microwave backscatter modelling.

Near-real time surface water extent monitoring from SLC Sentinel-1 SAR imagery:
The SMART project case study.

Cristian Silva-Perez1, Javier Ruiz-Ramos2, Armando Marino1, Andrea Berardi2,
1 University of Stirling, Scotland, UK; 2The Open University, Milton Keynes, UK.

Introduction
Timely information on surface water extent constitutes a key decision support tool suitable for informed decision making. In the context of floods, it allows assessment of flood impact, effective risk and action management and prioritised investments in flood defences. With regards to natural ecosystems, such as wetlands, it allows monitoring critical hydrological dynamics that provide the appropriate conditions for rich biodiversity to thrive. The Landscape Sensor-based Monitoring Assessment using Remote Technologies (SMART) project provides a novel Synthetic Aperture Radar (SAR) based tool to establish a new global service aimed at providing surface water extent monitoring. The project will initially demonstrate potential in three sites as follows. In the Firth of Forth (Scotland), the aim is to monitor the extent of floods, heavy rainfall and fast snow melting events that cause river levels to rise resulting in surface water flooding affecting private and commercial infrastructure and disturbing transportation. The two additional test sites include Colombo urban wetlands (Sri Lanka) and North Rupununi wetlands (Guyana). The natural ecosystems in these two locations are characterised by seasonal floods and support important terrestrial and freshwater biodiversity, supplying local communities with a range of livelihood activities, including subsistence fishing and ecotourism. In these locations, a service to monitor the hydrological and ecological condition of the wetlands provide crucial information for sustainable development, particularly in the context of flooding and droughts emphasised by climate change.
Methods
This presentation will provide a comparison of the algorithms for near-real-time surface water extent monitoring implemented within SMART, using Single Look Complex (SLC) Dual-polarimetric (Dual-PolSAR) Sentinel-1 imagery. As a benchmark, we include the following two methodologies:
• Our previous work developed in [1] for natural wetlands monitoring based on the Cumulative Sums algorithm applied to SAR time series (SAR-CUSUM). It consists of a robust statistical and multitemporal approach to map open water and flooded vegetation areas. The algorithm extracts a dry condition reference from historical imagery and cumulates the difference between new acquisitions and this reference. This procedure highlights consecutive deviations from the dry conditions, thus enhancing any possible recurrent variation which takes place over time. Using a threshold based on the histograms of regions, open water and flooded vegetation areas can be masked out.
• A current state-of-the-art algorithm for surface soil moisture estimation [2] which utilises a stack of multitemporal VV backscatter intensity images. It infers reference dry and wet conditions for a test site and compares them via a normalised difference with every new SAR image acquisition. Since the results of the comparisons are normalised, it presents the estimated surface soil moisture in an intuitive 0 to 1 scale.
In SMART we also derive a series of novel algorithms that are included in this comparison as follow:
a) An expanded version of the multitemporal CuSum algorithm presented in [1] that includes Dual-PolSAR features such as alpha angle and entropy, derived from the dual polarimetric pixel covariance matrices.
b) A set of change detection approaches based on optimisations of covariance matrices as presented in [3]. The PolSAR detectors have been designed to identify not only the intensity of change between images, but also the type of change expressed by the change in scattering mechanisms. For the SMART project, we adapted two of these detectors for the Dual-PolSAR case: A change detector based on the difference of covariance matrices and a detector based on the ratio of these covariance matrices. Given the difference in the interpretation for each of them, complementary information may be obtained.
c) A deep learning-based approach that uses the well-known U-NET image segmentation model [4] for supervised surface water extent monitoring. For this method, a training and testing dataset was created from the semi-automatic flood maps produced by the Copernicus Emergency Management Service (human-adjusted maps produced by a computer algorithm). We test different combinations of input features, including backscatter intensities and dual-PolSAR features derived from SLC Sentinel-1 imagery.
Results
Preliminary results presented in [1], show a rigorous statistical flood monitoring tool that was able to demonstrate a 90% accuracy in detecting the extent of open water flooding in the Guyana demonstration site. In addition, preliminary results show that including the phase information within the SLC images improves the accuracy, especially for floods under vegetation. In the presentation, we will show the exact figures for accuracy on the algorithm comparison once the SLC data of Sentinel-1 is included in the retrieval algorithms. This considers analysis on the benefits/limitations of employing Dual-PolSAR SLC data and unsupervised/supervised learning approaches. It will also highlight the bespoke elements required to map surface water on natural ecosystems (wetlands) and other terrains (flood mapping).
An additional point of this presentation will be to introduce the easy-to-use visualisation tools developed in SMART (web and mobile mapping apps), tailored for non-specialist user communities to use the results obtained by the algorithms. This is due to the fact that the visualisation and mapping platforms are developed in close collaboration with communities affected by floods through the use of capacity-building programs. The visualisation tools are designed to include environmental and social information in order to support decision-making in relation to flooding (e.g., regarding mosquito-borne diseases, flood monitoring and planning, biodiversity conservation, infrastructure development and agriculture).

Acknowledgement:
This research was supported by the SMART project, funded by the UK Space Agency’s Partnership in Innovation Development (Pin2D). The SMART consortium comprises The Open University, University of Stirling, and the Cobra Collective CIC. Sentinel-1 data were provided courtesy of ESA. Validation optical imagery were provided courtesy of Planet.
References
[1] Ruiz-Ramos, J., Marino, A., Berardi, A., Hardy, A., & Simpson, M. (2021, July). Characterization of Natural Wetlands with Cumulative Sums of Polarimetric SAR Timeseries. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS (pp. 5899-5902). IEEE.
[2] Bauer-Marschallinger, B., Freeman, V., Cao, S., Paulik, C., Schaufler, S., Stachl, T., ... & Wagner, W. (2018). Toward global soil moisture monitoring with Sentinel-1: Harnessing assets and overcoming obstacles. IEEE Transactions on Geoscience and Remote Sensing, 57(1), 520-539.
[3] Marino, A., & Nannini, M. (2021). Signal Models for Changes in Polarimetric SAR Data. IEEE Transactions on Geoscience and Remote Sensing.
[4] Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234-241). Springer, Cham.

Wetlands provide many essential ecosystem services, but at the same time they are threatened by processes like e.g., urbanization, expansion of farmland and extraction and pollution of freshwater. They play a key role in global carbon, water, and nutrient cycles, have a strong impact on local and regional climate and are essential for global food and water security. Due to their heterogeneous characteristics, spatiotemporal dynamics and topographic features, wetlands are often hard to identify, map and monitor, which hinders their efficient conservation and recognition in policy frameworks and management practices. Ongoing developments in the field of Earth Observation (EO) provide an opportunity to improve this situation. This has been empowered especially by freely available satellite programs like the USGS Landsat imagery and the Copernicus program.

GEO-Wetlands is a collaborative framework formed as an initiative within the Group on Earth Observations (GEO) with the ambition of growing into a GEO flagship in the coming years. The Mission of GEO-Wetlands is to develop sustained global approaches for EO based wetland inventory, mapping, monitoring & assessment to support global policy frameworks like the Ramsar Convention on Wetlands, the UN Sustainable Development Goals, the UNFCCC Paris Agreement, and the Sendai Framework for Disaster Risk Reduction.

Since 2016, GEO Wetlands has been initiated and supported through several research and innovation projects funded e.g., through the European Commission, European Space Agency, Japan Aerospace Exploration Agency, German Space Agency, and others. These project activities led to the formation of a GEO-Wetlands community and the development of a collection of tools, methods, datasets, guidelines and pilots of a geospatial wetland portal and a web-based knowledge collection. The ambition for the coming years is to develop this collection into a GEO-Wetlands toolkit that is freely available, easily accessible, and continuously developed and updated to support global wetland practitioners and decision makers.

A strategy is currently being developed to grow GEO-Wetlands into a GEO Flagship with its activities being embedded in the Ramsar strategic plan, SDG 6.6, and the UNFCCC global stocktake. This entails the establishment of a permanent GEO-Wetlands secretariat responsible for management of the flagship and its community, fundraising, communication and maintenance of the GEO-Wetlands toolkit and website.

GEO-Wetlands is an open and inclusive partnership welcoming contributions and participation from organizations, initiatives, companies, and individuals from all parts of the world and from all sectors related to the use, management, protection, restoration and monitoring of wetlands. New projects over the coming years will allow this partnership to grow and strengthen and to develop new approaches and components for the GEO-Wetlands toolkit with the final goal of developing global wetland maps, products, and statistics to support the wise use and sustainable management of all wetland types and their ecosystem services worldwide.