The three GFM flood mapping algorithms - i.e. flood mapping algorithm 1 (developed by LIST), flood mapping algorithm 2 (developed by DLR), and flood mapping algorithm 3 (developed by TU Wien) - operate in parallel, and work on the same pre-processed Sentinel-1 input data to generate their own “Observed flood extent” output layers.
An ensemble-based approach is then used to combine the three “Observed flood extent” maps generated by the flood mapping algorithms, into a single “consensus map”. In the ensemble method, each image pixel is first assigned a ratio (0-1) indicating the number of algorithms that classified it as “flooded” (where a ratio of 1 indicates that all three algorithms classified it as “flooded”). The final classification is straightforward and based on a majority decision: a pixel is accepted as “flooded” when at least two of the three algorithms classify it as such. The ensemble flood algorithm applies the reference water and the exclusion mask to harmonize the results of all individual algorithms.
In the unlikely scenario that one of the GFM flood mapping algorithms fail to produce a result, a pixel is classified as waterlogged only if both the remaining algorithms detect it as flooded. Exceptionally, should only one algorithm's results be available for a pixel, the latter will be automatically classified as no data, regardless it is detected as inundated or not.
Similarly, the GFM ensemble water mapping algorithm is used to generate the GFM Product output layer Reference water mask. The LIST and DLR flood mapping algorithms, unlike the TUW algorithm, both produce a water extent map and associated uncertainties, as a by-product of the flood detection. In the GFM ensemble water mapping algorithm, the LIST and DLR algorithms are applied to a “data cube“ of Sentinel 1 images available over a two-year time period (for permanent water) or the mean backscattering value of S1 images available over a two-year time period per month (for seasonal water). The resulting single water extent maps are combined into twelve masks, one per month, showing permanent and seasonal reference water extent.
The GFM ensemble water mapping algorithm follows the same logic as the GFM ensemble flood mapping algorithm, using a consensus decision of the LIST and DLR algorithms to decide if a pixel is water or no water. Where they disagree, the pixel is considered as not water.
In a nutshell:
- The GFM ensemble flood mapping algorithm integrates the results of the three leading-edge GFM flood mapping algorithms into a single product.
- The GFM ensemble water mapping algorithm integrates the results of two of the GFM flood mapping algorithms to produce a reference water mask, showing permanent and seasonal water bodies.
- The “consensus maps” that are generated by the ensemble approach substantially improve the robustness and accuracy of the GFM product output layers “S-1 Observed Flood Extent” and “S-1 Reference Water Mask”, and add a high degree of redundancy to the data processing workflow.
- Further details are provided in the dedicated section of the Product Description Document (PDD): https://extwiki.eodc.eu/GFM/PDD/GFMoutputLayers