In general, it is possible to distinguish between effects leading to both over- and under-estimation of detected flood extent (i.e., false alarms and missed alarms, respectively).
While false alarms draw the attention of users unnecessarily and thus could create frustration and mistrust in the product, missed alarms on the other hand would lead to situation where the flood event is not detected and leaves the users without notice, possibly losing time for reaction measures.
¶ False alarms
The here mentioned water-look-alikes would yield false alarms:
- very dry or sandy soils
- frozen ground
- wet snow
- flat impervious areas (e.g. smooth tarmac covers as airfields or roads)
These surfaces and artifacts usually feature very low backscatter signatures and appear thus as water-look-alikes in SAR imagery, rendering the water and flood mapping a difficult task.
Another common effect in SAR remote sensing is radar shadowing, which appear over strong terrain (especially at the far-range section of the SAR image) as well as in the vicinity of high objects above the ground, like high buildings and along forest borders.
¶ Missed alarms
On the other hand, floods occurring in urban areas, densely vegetated areas, or under weather conditions featuring strong winds or heavy rainfall can lead to missed alarms. In particular, wind and heavy rainfall are hard-to-spot dynamic process as they roughen water surfaces and hence undermine the initial assumption of low backscatter due to specular reflection on smooth water surfaces.
¶ How to tackle false/missed alarms with the GFM
A variety of methods can be used to address the abovementioned issues of SAR-based water mapping which potentially cannot be directly solved by the proposed flood detection algorithms using only NRT-available backscatter information and hence indication need on pixel level for potential misclassification due to reduced sensitivity.
The aforementioned challenges are the classified into Static effects and Dynamic effects with the first bound to the ground surface, land cover or topography, and the latter resulting from meteorological dynamics. This is also done with the perspective on a performant global NRT processing.
As a note: static is understood here with respect to the GFM reprocessing cycle, i.e., static layers remain unchanged during NRT processing, but they might be updated in the course of a reprocessing after evolution activities.
Static effects bound to ground surface characteristics such as land cover (e.g., flat impervious areas, urban areas, densely vegetated areas), and shadowing (radar shadowing), are addressed by the Exclusion Mask. Pixels that could not be classified by the SAR sensor into flood area, permanent/seasonal water body, and non-water area, are highlighted in this product layer as no-data pixels.
Dynamic effects triggered by weather conditions – i.e. meteorological features (strong wind, heavy rainfall), meteorological-induced state of the soil (soil dryness, frozen ground, or wet snow) - are flagged by the dynamic Advisory Flags. The Advisory Flags indicate locations where the SAR data might be disturbed by such processes during the acquisition, but leaves the flood and water extent layers unmasked.
Both layers are delivered with the other flood product layers and aim at improving the reliability, usefulness and user acceptance of the GFM product. The design of this mask/flag-system recognises also the User perspective: with our approach, we provide a simple Exclusion Mask indicating all the pixels that could not be classified by the input Sentinel-1 data, consulting statistical parameters from the data cube as well as auxiliary datasets. The pixels addressed by Exclusion Mask thus can be directly discarded as no-data, leaving the interpretation of the produced flood extent and the (adjacent) no-data-gaps to the users, who commonly know best their area-of-interest. We believe that users are have in general good skill to deal with no-data-gaps, as long as the general reliability of the product is assured.
The Advisory Flags layer aims to raise awareness that meteorological processes comprising wind or frozen conditions might impair the water body detection. As the Advisory Flags can only be retrieved at a coarser resolution, we do not forward the information of the flags to the masking or to the Exclusion Mask. As coming in the form of the additional layer, it should guide the users when interpreting the product, allowing additional insight on its local reliability at the time of Sentinel-1 acquisition.
¶ Other known issues
¶ Land-sea border
To assure a globally consistent land-sea border throughout all water extent and reference water mask products, the CopDEM water body mask is integrated into the observed water extent and all pixels are set to “water” where the CopDEM water body mask indicates the “sea” class. While this static land-sea border provides consistency, it does not take into account daily coastline dynamics caused by tides. It may, therefore, occur in rare cases that the land-sea border shows a low-tide case whereas the Sentinel-1 flood mask covers a high-tide. In such a scenario flood pixel may falsely be identified on near-shore sandbanks that are actually inundated during high-tide.
¶ Linear gaps between Sentinel acquisitions
Sentinel-1 IW CSAR products provided by ESA/Copernicus are the main input to the GFM service. After raw satellite data downlink, ESA as the original data provider, slices the data per 25 seconds sensing time (equivalent to about 170 km in track direction) without leaving any overlap/data duplication.
During the SAR geocoding step, computing the correct backscatter values along the slice edges requires the adjacent measurements. If not available, it generates no-data values, locally. As the GFM service ingest the SAR datasets separately in near-real-time (NRT), neighbouring slices are not available, and narrow stripes of no-data are generated at a dataset’s start- and end rows, yielding thin linear gaps. Waiting to have all adjacent files mutually available would decrease the timeliness of the preprocessing, and is not done within the GFM service.
As a consequence, GFM layers remain unclassified over the location of the Sentinel-1 no-data pixels, leaving thin linear gaps. The figura below exemplifies the issue for a flood event in Myanmar in June 2022.
