Automatic near real-time flood detection in high resolution X-band synthetic aperture radar satellite data using context-based classification on irregular graphs

This thesis is an outcome of the project “Flood and damage
assessment using very high resolution SAR data” (SAR-HQ), which is
embedded in the interdisciplinary oriented RIMAX (Risk Management
of Extreme Flood Events) programme, funded by the Federal Ministry
of Education and Research (BMBF). It comprises the results of three
scientific papers on automatic near real-time flood detection in
high resolution X-band synthetic aperture radar (SAR) satellite
data for operational rapid mapping activities in terms of disaster
and crisis-management support. Flood situations seem to become more
frequent and destructive in many regions of the world. A rising
awareness of the availability of satellite based cartographic
information has led to an increase in requests to corresponding
mapping services to support civil-protection and relief
organizations with disaster-related mapping and analysis
activities. Due to the rising number of satellite systems with high
revisit frequencies, a strengthened pool of SAR data is available
during operational flood mapping activities. This offers the
possibility to observe the whole extent of even large-scale flood
events and their spatio-temporal evolution, but also calls for
computationally efficient and automatic flood detection methods,
which should drastically reduce the user input required by an
active image interpreter. This thesis provides solutions for the
near real-time derivation of detailed flood parameters such as
flood extent, flood-related backscatter changes as well as flood
classification probabilities from the new generation of high
resolution X-band SAR satellite imagery in a completely
unsupervised way. These data are, in comparison to images from
conventional medium-resolution SAR sensors, characterized by an
increased intra-class and decreased inter-class variability due to
the reduced mixed pixel phenomenon. This problem is addressed by
utilizing multi-contextual models on irregular hierarchical graphs,
which consider that semantic image information is less represented
in single pixels but in homogeneous image objects and their mutual
relation. A hybrid Markov random field (MRF) model is developed,
which integrates scale-dependent as well as spatio-temporal
contextual information into the classification process by combining
hierarchical causal Markov image modeling on automatically
generated irregular hierarchical graphs with noncausal Markov
modeling related to planar MRFs. This model is initialized in an
unsupervised manner by an automatic tile-based thresholding
approach, which solves the flood detection problem in large-size
SAR data with small a priori class probabilities by statistical
parameterization of local bi-modal class-conditional density
functions in a time efficient manner. Experiments performed on
TerraSAR-X StripMap data of Southwest England and ScanSAR data of
north-eastern Namibia during large-scale flooding show the
effectiveness of the proposed methods in terms of classification
accuracy, computational performance, and transferability. It is
further demonstrated that hierarchical causal Markov models such as
hierarchical maximum a posteriori (HMAP) and hierarchical marginal
posterior mode (HMPM) estimation can be effectively used for
modeling the inter-spatial context of X-band SAR data in terms of
flood and change detection purposes. Although the HMPM estimator is
computationally more demanding than the HMAP estimator, it is found
to be more suitable in terms of classification accuracy. Further,
it offers the possibility to compute marginal posterior
entropy-based confidence maps, which are used for the generation of
flood possibility maps that express that the uncertainty in
labeling of each image element. The supplementary integration of
intra-spatial and, optionally, temporal contextual information into
the Markov model results in a reduction of classification errors.
It is observed that the application of the hybrid multi-contextual
Markov model on irregular graphs is able to enhance classification
results in comparison to modeling on regular structures of
quadtrees, which is the hierarchical representation of images
usually used in MRF-based image analysis. X-band SAR systems are
generally not suited for detecting flooding under dense vegetation
canopies such as forests due to the low capability of the X-band
signal to penetrate into media. Within this thesis a method is
proposed for the automatic derivation of flood areas beneath shrubs
and grasses from TerraSAR-X data. Furthermore, an approach is
developed, which combines high resolution topographic information
with multi-scale image segmentation to enhance the mapping accuracy
in areas consisting of flooded vegetation and anthropogenic objects
as well as to remove non-water look-alike areas.