Academic and research departmentsDepartment of Electrical and Electronic Engineering, Surrey Space Centre.
In the context of a flooding, a clear cloud-free SAR (Synthetic
Aperture Radar) image proves mainly useful to retrieve flood
features that can provide an extensive understanding of the
disaster. Among these features, extremely important is the
water depth on which this paper will focus by looking for a
semi-automated algorithm for its estimation in the neighborhood
of a given building from a pair of SAR images.
In this study, two SAR images acquired during dry and
flooded conditions are necessary, as well as a DSM (Digital
Surface Model) to give an a priori knowledge of the height of
the building and its footprint. The whole process is divided
into two main parts: First, an extraction of the building?s
double-bounce contribution using Genetic Algorithms, then
the computation of the inundated building?s height, to eventually
evaluate the water level locally in the neighborhood of
Thanks to the semi-automation of the double-reflection
line retrieval, the execution time of the whole process was
reduced from a few minutes (time to manually delineate the
double-bounce line) to a few seconds, while keeping an error
in the estimated flood depth in the order of a few decimeters
(35cm on average).
Flood is causing devastating damages every year all over the
world. One way to improve the readiness of stakeholders (rescue
authorities, policy makers, and communities) is by providing
flood extent maps promptly after the disaster, preferably
in an automated way and with a minimum number of
satellite imagery to reduce costs. The web application developed
in this paper aims to address this problem by mapping
the flood extent automatically from SAR images.
This web application is portable since it runs on the internet
browser, and allows to perform the classification of
the flooding in an automated fashion. Another strong point
is the rapidity of the processing: the whole processing time
was around 3 to 5 minutes for a subset of 20 million pixels.
The inundation map returned by our algorithm was validated
against vector files mapped by the United Nations Institute for
Training and Research (UNITAR) for the same flood event.
Regarding the dataset needed in this study, a pair of a preflood
SAR image and an optical image of the same area were
used to build a training dataset of water and non-water classes.
The learning phase is immediately followed by the classification
of the post-flood SAR image into a binary flood map.
The web application described in this paper was built with
open-source Python libraries which are backed by large communities
(Django, Scikit-learn among others). The flood map
was eventually displayed on OpenStreetMap maps provided