Abstract |
Countries around the world experience climate-related changing conditions which reflect
severe risks to the normal and sustainable operations of modern societies. Extreme
weather events, such as floods, rising sea-levels and storms, stand as characteristic
examples that impair the core services of the global ecosystem. Specifically, floods depict
severe impact on human societies. Climate change acts in benefit of such events, which
are increasing in terms of frequency and magnitude. Early-stage and accurate delineation
of the disaster is of top-priority since it provides environmental, economic, and societal
benefits and eases relief efforts. Satellite imaging systems provide significant assistance
towards the monitoring of natural disasters. The combination of satellite imagery with
Artificial Intelligence (AI) technologies provide a strong insight, and they are able to
successfully tackle the problem of flood detection and mapping. In this thesis, we
introduce OmbriaNet, a deep neural network architecture, based on Convolutional Neural
Networks (CNNs), that detects changes between permanent and flooded water exploiting
the temporal differences among flood events extracted by different sensors. To illustrate
the superior performance of our system, we construct OMBRIA, a bitemporal and
multimodal satellite imagery data-set for image segmentation through supervised binary
classification. It consists of a total number of 2776 images, Synthetic Aperture Radar
imagery (SAR) from Sentinel-1 and multispectral imagery from Sentinel-2, accompanied
with ground truth binary images produced from data derived from experts and provided
from the Emergency Management Service of the European Space Agency (ESA)
Copernicus Program. The dataset covers 20 flood events around the globe starting from
2017 to 2020. We collect data, co-registrate and pre-process them in Google Earth Engine.
To validate the performance of our algorithm, we perform benchmarking experiments on
OMBRIA dataset, with other competitive state-of-the-art techniques: the global adaptive
binarization threshold algorithm (Otsu's method), traditional machine learning algorithms
(i.e Support Vector Machines), and the widely recognized image segmentation U-Net
deep learning architecture, that act as baseline. The performed experimental analysis that
the proposed formulation is able to produce high-quality flood maps, achieving superior
performance compared with the state-of-the-art.
|