Each of the ten GFM Product output layers, which are listed below, is briefly described and illustrated in the following sub-sections. A more detailed technical description of the GFM Product output layers is provided in the GFM Product Definition Document (PDD)
¶ 1. Observed Flood Extent
The GFM Product output layer Observed Flood Extent identifies the pixels covered by floodwater, mapped using Sentinel-1 SAR backscatter intensity. An example of this output layer is shown in the figure below. Pixels that are normally under water (identified using the monthly Reference Water Mask) are not part of this output layer. Observed Flood Extent is derived using the GFM ensemble flood mapping algorithm.
To map flood extent pixels for a certain date, the algorithm uses as input the Sentinel-1 data overpass plus offline-generated Sentinel-1 SAR parameters and auxiliary thematic datasets such as Exclusion Mask and topography (e.g., DEMs and HAND index). The relative orbit path information, to select the corresponding offline-generated Sentinel-1 SAR parameters, is extracted from the Sentinel-1 Metadata. During the near real-time operation of the GFM Product, the acquisition month of the Sentinel-1 scene is retrieved from the Sentinel-1 Metadata and the corresponding monthly Reference Water Mask is cropped to the extent of the processed Sentinel-1 scene. All the processing is done at the 20 metres spatial resolution of the Sentinel-1 pre-processed data cube.
¶ 1.1 Maximum Flood Extent
This second-level product is the output a feature of the system that allows users to download, for each AOI, the automatically-generated composite showing the maximum flood extent of the available observed flood layers within a specific timeframe (the limit is set to two months).
Please note that the query can be submitted via both the GFM portal and the REST-APIs, giving the users also the chance to decide whether to retrieve the data as a vector (geojson), a raster (OCG) or in both formats.
Further guidance on how to properly use this functionality is also available in the Product User Manual in the Quick Start Guide.
¶ 2. Observed Water Extent
The GFM Product output layer Observed Water Extent identifies the pixels classified as open and calm water using Sentinel-1 SAR backscatter intensity and is derived using the ensemble flood mapping algorithm. An example of this output layer is shown in the figure below. To map water extent pixels for a certain date, the algorithm uses as input the Sentinel-1 data overpass plus offline-generated Sentinel-1 SAR parameters and auxiliary thematic datasets such as Exclusion Mask and topography (e.g., digital elevation models and HAND index). The relative orbit path information, to select the corresponding offline-generated Sentinel-1 SAR parameters, is extracted from the GFM Product output layer Sentinel-1 Footprint and Metadata. All the processing is done at the 20 metres spatial resolution of the Sentinel-1 pre-processed data cube.
In practice, the GFM Product output layer Observed Water Extent is created as a union of the GFM Product output layers Observed Flood Extent (derived by the GFM ensemble flood mapping algorithm) and Reference Water Mask (derived by the GFM ensemble flood mapping algorithm), which represents the extent of open water bodies under normal conditions. In the literature, reference masks of permanent water extent are often used for this purpose (Wieland and Martinis, 2019).
¶ 3. Reference Water Mask
The GFM Product output layer Reference Water Mask identifies pixels classified as open and calm water, both permanent and seasonal, by applying the GFM ensemble flood mapping algorithm to a five-year “data cube” (time series) of Sentinel-1 SAR backscatter intensity. An example of this output layer is shown in the figure below.
Whereas the mapping of permanent water extent uses as input the median backscatter of all Sentinel-1 data from a period of five years (2018-2022), the mapping of seasonal water extent uses as input the median backscatter of all Sentinel-1 data from a given month over a five-year period (2018-2022). As a result, twelve masks are available, one per month, which includes information on the permanent and seasonal reference water extent. This parameter database is updated once a year. For example, the NRT system running in 2022 relied on the Reference Water Mask extracted from the Sentinel-1 pre-processed data cube from 2020 and 2021.
Radar shadow and low sensitivity exclusion layers, and the HAND index, are applied to the reference water mask (and associated uncertainty layer) to correct pixels that were possibly misclassified. Finally, the Copernicus Global Surface Water Maximum Water Extent layer (Pekel et al., 2016) is used to remove possible false positive classifications, while the Copernicus Water Body Mask is used to correct false negatives (e.g. large lakes with roughened surface falsely classified as land) and to enforce a consistent land-sea border.
A truly permanent water area would mean that there was observed water coverage in every single observation of the considered time-period, i.e., the Water Occurrence (WO), which is the ratio between the number of water detections during a certain time-period and the number of valid observations of the same period, would be 100%. To consider uncertainties in the single water segmentations and the occurrence of hydrological extreme events the WO threshold is usually relaxed to a value of 85-90 % (e.g., Pekel et al., 2016).
¶ 4. Exclusion Mask
The GFM Product output layer Exclusion Mask indicates those locations (pixels) where the SAR data does not contain the necessary information for a robust flood delineation, due to the combined deleterious effects of the following main “static” factors:
- No sensitivity to flood mapping, where Sentinel-1 does not receive sufficiently strong signals from the ground surface to distinguish a flooded from a non-flooded surface.
- Water look-alikes, where Sentinel-1 SAR backscatter from the non-flooded ground surface is so low as to be indistinguishable from the backscatter from smooth open water.
- Strong topography, when Sentinel-1 signals are heavily distorted by terrain effects, effectively enhancing the noise and signal disturbances to such a degree to that it becomes larger than the change in backscatter due to potential flooding.
- Radar shadows, when Sentinel-1 receives no signals from certain regions of the land surface because of mountains, high vegetation canopies or anthropogenic structures.
For generation of the Exclusion Mask, the GFM Product implements various methods that address the identified problems of SAR-based flood mapping. The parameter database stores for all locations, on a pixel basis, the areas excluded by the four groups of factors, with the radar shadow layer per local Sentinel-1 orbit configurations (up to six per location). During NRT operation, the relative orbit is determined from the S-1 metadata, and the respective Exclusion Mask layers are subset to the extent of the processed Sentinel-1 scene and form a single binary mask for exclusion areas.
As no-sensitivity is a problem leading more often to an under-estimation rather than over-estimation of flooding (e.g. in urban areas), the no-sensitivity -masking is only applied to pixels classified as non-flooded. Pixels classified as flooded are kept unmasked. Any no-data areas from the flood mapping algorithm are forwarded to this layer and added as no-data values.
¶ 5. Likelihood Values
Aggregated likelihood values are generated along with the binary map product as a simplified appraisal of trust in the ensemble flood mapping method. An example of this output layer is shown in the figure below.
First, the likelihood information from each of the individual algorithms is expressed in the same numerical range [0, 100] to ensure comparability and to facilitate further harmonization. More specifically, probabilistic values from the LIST algorithm and fuzzy membership values from the DLR algorithm are first converted into classification likelihood and next multiplied by a factor of 100. For the DLR algorithm, since fuzzy membership values are only assigned to water pixels, all unflooded pixels are assigned a value of 0. The conversion of these probabilistic (or fuzzy membership) values is carried out as:
U = Wi * (100 - Pi) + (1 - Wi) * Pi
where:
- Wi represents the binary water map classification (0 for dry and 1 for water classes).
- U is the likelihood value.
- Pi is the probabilistic (resp. fuzzy membership) value.
The classification likelihood values resulting from the TU Wien algorithm (ranging from 0 to 1000) are divided by 10 to fall with the range [0,100].
Finally, the average of the likelihoods from the three algorithms is taken. This has a value in the range of [0, 100], where values towards 100 indicate high confidence in the GFM ensemble flood mapping approach. Consequently, the likelihood information provided along with the map product communicates how much confidence is associated with the Sentinel-1 classification. End users can then use highly certain flood map products to identify resource requirements over areas of flood exposure to make timely emergency response decisions.
¶ 6. Advisory Flags
The GFM Product output layer Advisory Flags indicates pixels that potentially suffer from decreased contrast between water and non-water surfaces due to meteorological factors as wet snow, frost and dry soil or wind-roughened water. Pixels marked by the Advisory Flags are not excluded by the Exclusion Mask, but users are advised to use with caution the GFM’s flood and water extent products data over advisory-flagged areas. An example of this output layer is shown in the figure below.
For each incoming Sentinel-1 scene that is input to the flood mapping algorithm, the advisory flag information is generated during NRT runtime. After generating the Observed Water Extent layer, the pre-processed 20m Sentinel-1 backscatter coefficient array is analysed and compared with calm water signature database to find pixels with rough water surface.
After the thresholding and spatial buffering, the obtained intermediate pixel-map indicating regions potentially affected by wind is stored. The 20m-oversampled backscatter data from the temporally corresponding 25km ASCAT observation dataset is compared to the local ASCAT backscatter 20th percentile value, which is pre-computed and accessed from the data cube. All pixels with detected low regional backscatter are forwarded to the intermediate storage and the final Advisory Flag pixel map for the Sentinel-1 scene extent is computed along with a 14 km-radius buffer.
¶ 7. S-1 Footprint and Metadata
The GFM Product output layer “Sentinel-1 Footprint and Metadata” shows all of the metadata attributes provided with each Sentinel-1 GRD data product used to generate the GFM Product main output layers. Metadata of each Sentinel-1 GRD scene are provided in the distributed Sentinel “Standard Archive Format for Europe (SAFE)” format, an XML file containing the mandatory product metadata.An example of this output layer is shown in the figure below.
Attributes contained in the manifest file are classified into four categories:
- Summary.
- Product.
- Platform.
- Instrument.
Platform- and instrument-related attributes are considered as static for the different Sentinel-1 satellites. A total number of 29 attributes are contained in the manifest file such as information about the absolute orbit number, pass direction, polarisation, sensing start and end date and the product timeliness category. An abstract of the included attributes is given in hereafter as an example.
After successful retrieval of a Sentinel-1 GRD product at EODC the manifest file is parsed and inserted into the operated metadata database. Access to the actively maintained metadata base is given via the well-defined OGC CSW standard referred to as metadata catalogue. The database itself builds upon PostgreSQL following a relational database system.
The requested Sentinel-1 metadata are provided via PostGIS representing a spatial database extender of the PostgreSQL database. The PostGIS layer allows on-demand querying of geographic objects conform to OGC mapping standards such as WMS-T, WCS or WFS.
The footprint of a Sentinel-1 GRD scene as provided in each data product is included in the mentioned manifest file. The footprint is represented as human readable Java Topology Suite (JTS) object named “JTS footprint”.
The JTS footprint is converted into Well-known Text (WKT) and Well-known Binary (WKB) in the process of parsing and ingesting the acquired manifest files into the operated metadata database. WKT and WKB are originally defined by the OGC to describe simple features.
Hence, the Sentinel-1 footprint will be available as PostGIS layer accessible via all supported OGC mapping standards. On-the-fly conversion to standalone data formats can be supported via standard GIS tools based on user request.
¶ 8. S-1 Schedule
Sentinel-1 observations follow a strict acquisition planning often referred to as acquisition segments. Information on the planned future acquisition is provided by ESA in form of Keyhole Markup Language (KML) files. A single file usually covers an acquisition period of about 12 days, with the start and stop time of the future planned acquisitions already given in the file name. The GFM Product output layer “Sentinel-1 schedule” shows planned Sentinel-1 acquisitions for the next three days. An example of this output layer is shown in Figure 14.
KML files are published regularly by ESA, well before activation, with potential last-minute changes due to requests from the CEMS. Information provided in the KML files is organised based on the planned data takes. Parameters listed in
Table 5 are included in the KML. The KML files are regularly checked and downloaded at EODC and ingested into the described metadata database for further analysis. All parameters are exposed as PostGIS layer to extract the requested schedule information indicating the next planned Sentinel-1 GRD acquisition for a given location.
Information provided with the next planned Sentinel-1 GRD acquisition.
|
PARAMETER |
DESCRIPTION |
|
Datatake ID: |
Unique product identifier (hexadecimal). |
|
Mode: |
Instrument acquisition mode. |
|
Observation (duration): |
Duration of the planned data take (in seconds). |
|
Observation (start): |
UTC start date and time of the planned data take. |
|
Observation (end): |
UTC end date and time of the planned data take. |
|
Orbit (absolute): |
Absolute orbit number at the start time of the data take. |
|
Orbit (relative): |
Relative orbit number at the start time of the data take. |
|
Polarisation: |
Instrument polarisation for the acquired data take. |
|
Satellite ID: |
Satellite identifier. |
|
Swath: |
Instrument swath (from 1 to 6 for SM, not applicable for IW and EW). |
¶ 9. Affected Population
The GFM Product output layer Affected Population is derived based on data extracted from the Global Human Settlement (GHS) layer and, in particular, from the GHS-POP dataset. This data contains a raster representation of the population's distribution and density as the number of people living within each grid cell. The information is available at various spatial resolutions and for different epochs. An example of this output layer is shown in the figure below.
For the GFM processing, the dataset at 100m resolution (highest possible resolution) and for the latest available timestep, 2020 are used. Combining this re-projected and re-sampled raster dataset and the flood extent allows to provide the number of affected people for each specific pixel detected as flooded.
The updated dataset replaces the former Global Human Settlement population dataset at 250m resolution with the reference year 2015, which was available at the start of GFM operations and was used for production during October 2021 – March 2023.
¶ 10. Affected Land Cover
The GFM Product also provides, in addition to affected population, an output layer highlighting the affected landcover for a particular flood case. This information can provide a first assessment of affected land cover or land use types, for example how much agricultural area is affected by the flood extent. An example of this output layer is shown in teh figure below.
Based on the GFM consortium’s production heritage and experience in land cover mapping, this output layer is derived from the 100m-resolution database from the Copernicus Global Land Cover Service enriched with information from the Copernicus Pan-European High-Resolution Layers (i.e., Imperviousness, Forests, Grassland, Water and Wetness) over Europe.
The Global Land Service includes 23 classes and provides annual updates, with an overall accuracy of 80%. The Copernicus Pan-European High-Resolution Layers have an overall accuracy of 85%+ and are available at 20 m (the 2018 version at 10 m) spatial resolution. In addition to both datasets, GFM also includes the most relevant classes from OpenStreetMap (roads, railways, etc.) to allow a first assessment of affected areas and infrastructure.
