Difference between revisions of "ALGORITHMS"

Latest revision as of 10:45, 25 March 2021

[Home]

GFM 's flood products are based on an ensemble approach integrating three robust, cutting edge algorithms developed independently by three leading research teams.
The motivation for choosing such a methodology is to substantially improve accuracy of the derived Sentinel-1 flood and water extent maps and to build a high degree of redundancy into the production service.

As stated elsewhere, the data processing architecture underlying the different scientific algorithms is based on the data cube concept, whereby SAR images are geocoded, gridded and stored as analysis ready data (ARD) in an existing spatio-temporal SAR data cube.
By using a data cube, where the temporal and spatial dimensions are treated alike, each Sentinel-1 image can be compared with the entire backscatter history, allowing to implement different sorts of change detection algorithms in a rather straightforward manner. Importantly, the entire backscatter time series can be analysed for each pixel. Therefore, model training and calibration may be carried out systematically for each pixel.
The advantages of working with data cubes are:

• (a) algorithms are better able to handle land surface heterogeneity;
• (b) uncertainties can be better specified;
• (c) regions where open water cannot be detected for physical reasons (e.g. dense vegetation, urban areas, deserts), can be determined a priori,
• (d) historic water extent maps can be derived, essentially as a by-product of the model calibration, which may serve as a reference for distinguishing between floods and the normal seasonal water extent.

The (internal) availability of three separate flood and water extent maps tackles, by readily identifying them, the shortcomings a single algorithm, by itself, might be suffering of in specific circumstances and/or part of the world due to many well-known factors like topography or environmental conditions.

For these very reasons, Users have access to consensus flood maps where a pixel is marked as flooded when at least two algorithms classify it as water.
Accordingly, the implemented quality assurance procedures (see INSERT REFERENCE) allow for differentiating between classification errors that can be attributed to shortcomings of individual algorithms and errors that are inherent to the SAR sensing instruments and their difficulty to capture the appearance or disappearance of surface water in particular situations.

Algorithms description

A detailed description of the three algorithms and examples of applications in an operational context is provided in the PDD, this sections and the following links provide their keypoints

All the three algorithms make use of historical time series of SAR intensity data and use topography-derived indices to refine the initial classification of water bodies. However, differences appear in the ways historical time series of intensity data are finally used to parameterize the retrieval algorithms and the way ancillary data such as topography data are used in the production system.
Other differences relate to the inclusion of a region growing step or not, the scale at which the thresholds are determined and applied to each pixel’s backscatter value and other nuances in the way the retrieval algorithms are setup.

The most relevant features of the algorithms are summarized in the table below.

	Hasard	Algorithm2	Algorithm3
Application domain	Water and flood extent mapping (pixel-based)	NRT Water and flood extent mapping	Pixel-based water and flood extent mapping
Input remote sensing data	Pair of SAR intensity images acquired from same orbit (any sensor) and model parameters derived from historical time series	Single-temporal SAR intensity data	Single SAR acquisition and model parameters derived from historical time series
Auxiliary data	HAND index map, exclusion layer, reference water layer, water and flood extent map computed at previous time step	HAND index exclusion map, reference water extent, DEM, optional: low backscatter exclusion mask based on S-1 time-series data	HAND index, exclusion mask, reference water map for generating the fresh flooded areas
Characteristic features	Scene-specific statistical modelling of backscatter distributions, systematic updating of water bodies maps using combination of change detection and region growing	Hierarchical automatic tile-based thresholding, fuzzy logic-based post classification and region growing	Classification based on backscatter probability distribution by exploiting the historical time series with consideration of backscatter seasonality.
Exploitation of time series of SAR observations	Yes	No, the integration of a low backscatter exclusion mask based on S-1 time-series data (produced offline) can be integrated optionally	Yes, parametrisation through multi-year time series
Exploitation of textual information through region growing	Yes	Yes	No
Automation	High	High	High
Initialization	Statistical modelling of backscatter distributions attributed to water / no water and change / no change classes (per tile)	Hierarchical automatic tile-based thresholding using statistical modelling of class distributions	Generation of backscatter probability distribution from historical time series measurements
Post-classification steps	Masking of exclusion areas, distinction between water and flood extent using reference water layer	Masking of exclusion areas, distinction between water and flood extent using reference water layer	Noise reduction, Mask the exclusion areas, extraction fresh flood area compared with reference water map
Water probability mask generated	Yes (based on Bayesian inference)	Yes (based on fuzzy logic)	Yes (based on the Bayesian posterior probability)
Outstanding/differentiating features	Hierarchical split-based approach enabling re-calibration of parameters in NRT based on most recent pair of S-1 images	Fuzzy logic-based approach enabling a post classification and region growing taking advantage of topography-derived indices in addition to SAR backscatter	Exploiting per-pixel full Sentinel-1 signal history in data cube; enabling a very fast and scalable production of flood and water extent maps through pre-computed global parameters at high quality
Additional information	further details on HASARD	further details on Algorithm2	further details on Algorithm3