Information Extraction From Remote Sensing Images For Flood Monitoring and Damage Evaluation

INVITED
PAPER
Information Extraction From

Remote Sensing Images
for Flood Monitoring and
Damage Evaluation
The authors of this paper describe techniques used to monitor flood damage and
to assess its risks; they believe the main challenge is to fully exploit remote
sensing data.
By Sebastiano B. Serpico, Fellow IEEE , Silvana Dellepiane, Member IEEE , Giorgio Boni,
Gabriele Moser, Member IEEE , Elena Angiati, Student Member IEEE , and Roberto Rudari
ABSTRACT | Satellite remote sensing missions devoted to paper, we investigate the key issues involved in the exploita-
Earth observation (EO) currently offer a unique capability to tion of satellite data with special focus on the phases of the
monitor the evolution of the Earth’s surface by providing emergency and post-disaster damage assessment. To this end,
temporally repetitive views at the desired (global, regional, or the challenges and the methodological approaches involved in
local) spatial scale. This wealth of remote sensing data conveys the multidisciplinary combination of image analysis and
a huge potential for preventing, monitoring, and managing hydrometeorology are discussed with the purpose of guiding
natural or man-made disasters. Specifically focusing on flood and optimizing the process of information extraction from
risk, a successful exploitation of this potential requires not only satellite data according to the requirements of civil protection
accurate and reliable image-analysis methods to extract the from floods. Experimental examples of a few relevant case
desired thematic information, but also the ability to combine studies are also presented.
this information with physically based models of the observed
processes. Therefore, a multidisciplinary approach combining KEYWORDS | Damage evaluation; flood modeling; flood mon-
remote sensing with geophysical sciences, such as, in this case, itoring; image processing; pattern recognition; rapid mapping;
hydrometeorology, is fundamental. This combination of exper- remote sensing
tise allows, in particular, satellite data to be exploited within
the different phases of flood risk reduction: risk assessment,
prevention, mitigation, monitoring, and management. In this
I . INTRODUCTION
Satellite Earth observation (EO) missions currently offer a
unique capability to observe the Earth’s surface in a
spatially distributed and temporally repetitive fashion at
Manuscript received October 9, 2011; revised April 16, 2012; accepted April 17, 2012. global, regional, and local spatial scales. Spaceborne
Date of publication August 2, 2012; date of current version September 14, 2012. This passive sensors offer multispectral images (i.e., comprising
work was supported by the Italian Space Agency under the BOPERAVCivil protection
from floods[ project. several channels, each of which is represented as a gray-
S. B. Serpico, S. Dellepiane, G. Moser, and E. Angiati are with the Department of level image, corresponding to a different wavelength
Biophysical and Electronic Engineering (DIBE), University of Genoa, I-16145 Genoa, Italy
(e-mail: sebastiano.serpico@unige.it). range) with resolutions ranging from a few kilometers to
G. Boni and R. Rudari are with the Department of Computer Science, Control, and less than one meter [1]. Satellite active microwave imaging
Telematics (DIST), University of Genoa, I-16145 Genoa, Italy and also with CIMA
Research FoundationVInternational Center on Environmental Monitoring, I-17100 systems, implemented as synthetic aperture radar (SAR)
Savona, Italy (e-mail: giorgio.boni@cimafoundation.org). sensors, can collect data regardless of solar illumination
Digital Object Identifier: 10.1109/JPROC.2012.2198030 and almost independently of cloud cover [2]; the latest
2946 Proceedings of the IEEE | Vol. 100, No. 10, October 2012 0018-9219/$31.00 Ó 2012 IEEE
Serpico et al.: Information Extraction From Remote Sensing Images for Flood Monitoring
SAR constellations (specifically, COSMO-SkyMed) reach the Japan Aerospace Exploration Agency (JAXA; https://
up to metric resolution with revisit times as short as 12 h. sentinel.tksc.jaxa.jp/sentinel2/topControl.action). Within
To successfully exploit this wealth of data for environ- the financially supported programs, one of the most
mental disaster monitoring and risk assessment, the interesting efforts is the Global Monitoring for Environ-
information extracted from EO images must be merged ment and Security (GMES) program (http://www.gmes.
with physically based models of the observed phenomena, info) supported by the European Commission, the
a goal that requires a multidisciplinary approach involving European Space Agency, and the European Environment
both remote sensing and geophysical sciences, thus mov- Agency at different levels, from the launch of specific
ing from the task of creating scenes to one of creating missions (i.e., Sentinel constellation; http://www.esa.int/
scenarios. For clarity, we define a scene as Ba still image esaLP/SEM097EH1TF_LPgmes_0.html) to ongoing fund-
taken at a specific time,[ e.g., at the time of a disaster. A ing for service development and initial operation support
single scene is not sufficient to predict the evolution of the in direct cooperation with the European Parliament. The
observed process because it does not contain inherent above list of initiatives is far from being exhaustive, but it
information about the physics of the observed phenomena. provides an idea of the interest raised by the topic.
However, in the present context, a Bscenario[ is a Btime Satellite remote sensing plays a threefold role with re-
series of spatially distributed observational data of a process spect to the different phases of the emergency manage-
up to a given moment,[ cast into physical dynamic models ment cycle: 1) during the prevention and prediction
that allow the prediction of how the scenario will evolve in phases, it allows for increased flood-risk mitigation capa-
the case of prevention and response actions or how it has bility by improving the evaluation of scenarios associated
evolved in the case of post-disaster evaluations [3]. with possible hazardous events and the real-time predic-
In this paper, we investigate the key issues in the tion of the hydrometeorological processes leading to
exploitation of EO data in the operational chains for flood floods; 2) during the mitigation and risk assessment
monitoring and post-event damage assessment. To this phases, it permits a better definition of the elements at risk
end, effectively combining image analysis and hydrome- and their vulnerability that eventually translates flood
teorology is pivotal to guide and optimize the process of hazards into possible damage and loss of life; and 3) during
information extraction from EO images according to the the monitoring and management phases, it permits a
requirements of civil protection from floods. Here, the detailed assessment of the actual ground effects produced
methodological issues and approaches involved in this by the flood.
multidisciplinary effort will be discussed and experimental Concerning 1), most of the operational chains currently
examples of a few relevant case studies will be shown. in use for flood forecast are based on ground measure-
The usefulness of EO data in natural disaster manage- ments, which are often inadequate considering the time
ment, especially the preparedness, emergency, and recon- scales of the hydrometeorological processes that contrib-
struction phases, has been proven by the numerous ute to flood formation. The ground observations are pre-
ongoing national and international initiatives. These com- dominantly sparse, whereas fields such as precipitation
mitments, which involve providers of EO data and ser- and soil moisture have a high space–time variability,
vices, have been developed through voluntary efforts and which, because of inherent nonlinearities, strongly in-
financially supported programs. Included in the former fluences the formation of floods. The spatial variability of
category is the International Charter on Major Disasters these fields is especially difficult to catch from scattered
(www.disasterscharter.org), which involves the majority of point observations.
the space agencies with interest in EO; different value This paper focuses primarily on issues 2) and 3). With
adders belonging to the private sector or the United Na- regard to vulnerability assessment, standard cartographic
tions system, such as UNITAR_s Operational Satellite Ap- information may not have been updated, and in situ
plications Programme (UNOSAT) (http://www.unitar.org/ surveys are usually time consuming and possibly slow as
unosat/) and the United Nations Platform for Space-based related to the temporal evolution of urban and built-up
Information for Disaster Management and Emergency areas. Satellite remote sensing offers a feasible approach to
response (UN-SPIDER; www.un-spider.org); various pilot remove or mitigate these limitations, because of both its
projects, such as the Caribbean Satellite Disaster Pilot spatially distributed observation capability and the oppor-
(CSDP) or the National Aeronautics and Space Adminis- tunity the approach presents to exploit advanced image-
tration (NASA) and Canadian Space Agency’s (CSA) analysis methods. In particular, land-cover information is
Southern African Flood and Health Pilot Project, which an important input to vulnerability models. EO-based land-
provides free-of-charge data and processing in certain cover maps are feasible through the use of supervised
areas of the globe; the implementation of permanent image classification techniques, which are based on a pat-
platforms for support to disaster risk reduction (DRR), tern recognition approach [1], [4]. Given an EO image of
such as the Servir initiative from NASA; and contributions the considered area of interest, the field of pattern
to the Global Earth Observation System of Systems recognition provides a correct methodological framework
(GEOSS), such as the Sentinel Asia project supported by to formalize the problem of estimating the land cover of
Vol. 100, No. 10, October 2012 | Proceedings of the IEEE 2947
the ground area corresponding to each image pixel, within the problems of mapping flooded areas and discriminat-
a predefined and suitably characterized set of possible ing changed and unchanged ground regions, given a pair
land-cover classes [1], [4]. of satellite SAR images acquired before and after the
Moreover, when focusing on urban or built-up areas, a event.
further important input for assessing vulnerability is The paper is organized as follows. The key ideas about
represented by the related 3-D structure. A 3-D model of the characterization of vulnerability, elements at risk, and
an urban area can be accomplished by using remote sens- damage are outlined in Section II. Then, the extraction of
ing imagery, and 3-D building models are useful for eval- the relevant thematic information from remote sensing
uating the potential damage of a single building, a town imagery is discussed in Section III. Case studies and ex-
district, or a significant infrastructure (e.g., a bridge or an perimental examples are presented in Section IV and
industrial building). In general, 3-D modeling is performed conclusions are drawn in Section V.
with specialized software that processes information such
as contours, roof perimeters, shadows, and linear segments
to obtain a 3-D extrusion of the structure of interest. This I I. JOINING HYDROLOGY/HYDRAULICS
information is usually manually extracted by an expert AND REMOTE S ENSING FOR FLOOD
photo-interpreter, starting from aerial or satellite optical RESPONSE AND M ANAGEM ENT
images, which is often a time-expensive procedure. How- One of the most mature applications of satellite-derived
ever, current advances in the field of image processing information to the DRR cycle is in the response and re-
make it possible to accurately extract the features of inter- covery phases, in which remote sensing concurs in the
est while reducing costly human intervention. estimation of the damage suffered as a consequence of a
Concerning issue 3), currently, damage assessment for flooding event to people, structures, and goods. Damage
flood risk is underdeveloped, especially compared to other estimations are invaluable to deploy a more efficient and
risks. Few efforts have been made to systematically collect focused response to the disaster through appropriate civil
damage data, and few rigorous damage-assessment studies protection actions and to evaluate resources needed in the
have been conducted (see [5] for a review on the topic), recovery phases, also allowing for the activation of fi-
primarily because of difficulties in getting a proper picture nancial mitigation measures (e.g., insurance and reinsur-
of the flood event when it strikes large areas. However, ance coverage), which enhance the resilience of the social
important achievements in flood-damage evaluations have system to disasters.
recently been made by using observation techniques based Satellite-derived information helps in overcoming two
on spaceborne sensors, specifically when tangible direct critical aspects of damage estimation. The first is linked to
and indirect damages are of concern [6], [7]. These new the time scale. The damage assessment, especially when it
tools can also foster the development of long-term policies can be used to infer functionality information, is essential
and nonstructural finance-based mitigation frameworks in the first phases of response (i.e., response design and
(e.g., insurance products), especially in the case of large- organization) and therefore needs to be available imme-
scale severe disasters which are the main focus of inter- diately after the event has occurred. The time that elapses
national reinsurance players [8]. between the occurrence of the event and availability of
Specifically, after-event damage assessment relies on damage estimation is crucial in assessing the value of this
integrating the prior characterization of vulnerability and type of information. The second critical aspect, which is
an accurate detection of the ground regions that were addressed by the use of satellite information, regards the
affected by the event, i.e., flooded areas and/or further spatial scale. Extensive damage assessments are normally
ground changes [9]. The repetitive view offered by space- difficult and expensive to obtain, whereas information
borne EO allows such event-driven temporal discontinu- coming from satellites, by its nature, offers a synoptic view
ities to be captured. The goal can be either to support the of the disaster. Obviously, the level of detail that can be
authorities with first aid after the disaster, by promptly assessed by remote sensing techniques is usually lower
providing maps of the affected areas, or to contribute to than that available by standard ground inspection, but the
damage quantification by producing detailed mapping re- advantages mentioned above offer an invaluable contribu-
sults. Various spaceborne systems have been used for this tion to DRR.
purpose, including both optical (e.g., Landsat-5 TM [10] With the final aim of assessing damage in the best and
and SPOT [11]) and SAR imagery (e.g., ERS-1, ERS-2, quickest way after an event, satellite information can be
Envisat-ASAR, RADARSAT, ALOS-PALSAR [12], effectively combined with hydraulic and vulnerability
TerraSAR-X [13], [14], and COSMO-SkyMed [15]). SAR models. Some steps are needed to move from land-cover
data are especially relevant in this framework because of characterization and flooded-area maps to damage maps.
their day-and-night and cloud-penetrating acquisition Specifically, land-cover maps should, as a first step, be
capabilities. Again within the fields of pattern recognition aggregated into classes of elements at risk that show
and image processing, feature extraction, classification, similar behavior with respect to flooding. Second, proper
and segmentation techniques can be used to formalize loss functions should be assigned to each class of elements
2948 Proceedings of the IEEE | Vol. 100, No. 10, October 2012
Fig. 1. Block diagram of the overall processing chain from input EO (yellow boxes) and ancillary data (green boxes) to output thematic results
(blue boxes). Dashed lines denote optional inputs.
so that damage is known as a function of proper hydraulic modeling approaches allow elements at risk, vulnerability,
forcing (e.g., maximum water depths reached by the and damages to be mapped. Section II-A–D focus on these
flooding waters). As a third step, on the basis of either approaches, while the aspects of EO-image processing are
simplified morphologic models or physically based hy- discussed in Section III.
draulic models constrained to the delineated flooded-area,
water-depth maps and other hydraulic parameters can be
A. Vulnerability: A Central Concept in Damage
computed as best estimates of the real condition. Finally,
Evaluation
all derived information is merged to derive the estimates of
Vulnerability is a central aspect in flood-damage stu-
the actual loss.
dies. Vulnerability can be defined as the degree of loss to a
Fig. 1 presents a general block diagram of a processing
given element or set of elements at risk resulting from the
chain from EO and non-EO input data to output thematic
occurrence of a natural phenomenon of a given magnitude
maps associated with flood risks and damages. The chain
and expressed on a scale from 0 (no damage) to 1 (total
encompasses both EO-image-analysis algorithms and hy-
loss) [16]. Therefore, once the vulnerability of a specific
drometeorological modeling modules. EO-image-analysis
asset is determined for different flood magnitudes, it is
algorithms are aimed at land-cover mapping from multi-
possible to derive the economic damage by multiplying by
spectral or hyperspectral [possibly very high resolution
the element value. Vulnerability V can be specified as
(VHR)] EO data, building characterization from multispec-
follows:
tral VHR images, and highlighting and detecting ground
changes and flooded areas from multitemporal VHR SAR.
From these EO-derived processing results and possible
further non-EO ancillary data, suitable hydrometeorological V ¼ f ðE; EI; SÞ (1)
where f ðÞ is a suitable function and E, EI, and S represent B. Elements at Risk
vectors of elements at risk, exposure, and susceptibility Satellite imagery is fundamental in defining all aspects
indicators, respectively [17]. Any flood vulnerability determining vulnerability. The elements-at-risk vector is
analysis requires information regarding these three the starting point of the analysis. For this analysis, a
factors. The first factor ðEÞ represents the subject matter fundamental piece of information is represented by the
of any flood vulnerability analysis and specifies the amount land cover (see Section III-A). However, when defining
of social, economic, or ecological units or systems, which the elements-at-risk vector, it is necessary to understand
are at risk of being affected by a flood. The second factor that the final product is neither a land-cover map nor a
ðEIÞ supplies information about the various elements at land-use map. The elements need to be classified into a
risk (e.g., location, elevation, proximity to the river) and new characterization that considers their behavior with
focuses on general flood characteristics (e.g., duration, respect to vulnerability to floods. This criterion allows
velocity, inundation depth). The third factor ðSÞ measures condensing different information types, such as land-use
the absolute and relative impact of the flood on individual or land-cover classifications, into a single map from which
elements at risk and includes the social context of flood- flood vulnerability can be easily derived, as discussed in
damage formation (e.g., preparedness, coping, recovery). Section II-A.
The simplest and most used way to express (1) is by The definition of the elements at risk is organized in
Bdamage functions[ or Bloss functions,[ which relate flood practice into two fundamental steps. In the first automatic
damage/vulnerability to such flood exposure indicators step, the land-cover map derived from satellite imagery
(e.g., inundation depth) for different classes of elements at (see Section III-A) is compared to other land-use/cover
risk; susceptibility is usually neglected [18]. Damage information, when available, usually retaining the maxi-
functions can be used to calculate the unit damage mum detail between these layers. Conversion tables might
percentage for any affected object for given flood be needed between different land-use standards (e.g.,
conditions. In this case, a vulnerability mapping algorithm Anderson [19], CORINE Land Cover and Forest 2000 [20])
can automatically associate a system of loss functions, and elements-at-risk classes. The mapping is not trivial
describing structure and content vulnerability, with the because a many-to-many correspondence exists between
elements at risk. However, local morphological and the input information and the elements-at-risk classifica-
topological characteristics, as well as specific building tion. Unfortunately, as shown in Table 1, in many cases,
typology, can deeply modify the building loss function. For multiple elements-at-risk classes correspond to the same
this purpose, the loss function may also be interactively land-cover class, or vice versa: for example, elements-at-
tuned on the basis of expert judgment. risk classes A through E are mapped to Burban or built-up
Table 1 Example of a Nonbiunivocal Correspondence Between Elements-at-Risk and Anderson Classes. The Latter Set of Classes Represents
a Well-Known Standard for Satellite-Based Land-Cover Mapping [19]
land,[ whereas class I is mapped to Brangeland[ and in information left by single-satellite scenes of a flooded
Bforest land.[ Because of this difficulty, an interactive step area, allowing for the estimation, for example, of the max-
is usually necessary to complete the classification within imum flooding extension and magnitude. Furthermore,
elements-at-risk classes. other characteristics of the actual scenario are estimated,
An effective approach is to compute a quality index for such as the prevalent direction of flow during the flooding
each classified pixel, by defining the classification quality and the retreatment phase, and, if allowed by the specific
after the automatic step based on the availability of the event dynamics, possible future scenario development.
various sources and their accordance/discordance. A
classification that is concordant from all available sources D. Damage Assessment
is assigned a maximum index value, whereas a classifica- A critical problem associated with the assessment of
tion deriving from discordant sources obtains a lower the damage produced by floods (or other natural disasters)
index value. This quality index can be considered together is the logistics of quickly coordinating and implementing
with the classification and guide the interactive step, in an extensive ground-based damage survey. Damage
which misclassified areas can be identified and corrected assessment is a fundamental piece of information during
based on the operator’s experience. the emergency phase to indicate the most damaged areas
An additional step can be taken when single buildings and the dysfunction of lifelines or connection lines. Da-
are available features in the area (see Section III-B). In this mage assessment is also critical in the immediate post-
case, another automatic step is performed: based on the emergency phase when a first estimation of the total
areal characterization, each building is assigned a type that damage is necessary to quickly organize the recovery
considers the geometric characteristics of the single vol- phase. Generally, when land-cover changes are caused by
umetric element and the areal elements-at-risk class pre- the event, change-detection techniques (see Section III-C)
viously identified. As an example, a 50 20-m2 building are useful tools. However, damage caused by flooding is
with a regular height of 5 m falling into a cultivated area rarely destructive to the point that it can be directly seen
would be assigned to a greenhouse in the first iteration, by satellite, especially when urban elements are of con-
while a 20 20-m2 building with a height of 10 m would cern. In such situations, a complex and indirect workflow
be assigned to a residential building. In this situation, should be implemented to canvass the damage figures on
there is a subsequent interactive step, whenever possible, an extensively flooded area.
to verify and validate the suggestions of the automatic The final step of an effective damage-assessment pro-
algorithm. cessing chain can consist of combining the loss functions
and the exposure indicators to provide percentage damage
C. Water-Depth Mapping maps. The resulting thematic map provides a fast compu-
The most influencing flood-exposure indicator is the tation of the percent of damage suffered by the elements at
water-flood depth [17]. A reliable estimate of the inunda- risk in the flooded area independently of the value of the
tion depth is therefore needed to apply the above-defined considered element. This information is computed on
loss functions. To provide dynamically consistent hydrau- areal features and, if available, on the set of identified
lic depths, an effective approach can be the real-time use buildings in the analyzed domain (see Section III-B). If the
of a 2-D hydraulic model, based on a simplified version values of the elements at risk are available, a rough eco-
of the Bshallow-water[ equations [21]–[23], constrained nomic quantification of the damage can be obtained.
by satellite-retrieved flooded areas, as determined in
Section III-C. The main features of such a model are com-
putational speed and simple startup, with no need for III . EXTRACTING THEMAT IC
complex information but only of a subset of simplified INFORMATION FROM REMOTE
boundary and initial conditions. Those characteristics SENSI NG IMAGERY
allow the model to be fast enough to be used in real time
for the simulation of flooding events. An ensemble of A. Land-Cover Mapping
model runs is produced with a set of initial conditions (i.e., According to the discussion in Section II-A and B, a
the expert guess based on the satellite image of the flooded characterization of land cover for vulnerability and
area). The ensemble member that best matches the single elements-at-risk mapping usually requires working at med-
flood extension scenes derived from the SAR images is ium (e.g., 30 m) or (very) high spatial resolutions (ap-
then selected. proximately 0.5 to 5 m), depending on the size of the
The static information provided by EO is thus inter- considered basin and the investigated area. A land-cover
preted in a dynamic and consistent way, and useful flood mapping task can be addressed with EO data by a pattern
exposure indicators (e.g., water depth, water speed, and recognition approach and formalized as a supervised image
the evolution of flooded areas) are provided. The advan- classification problem [1]. From a methodological stand-
tages of combining satellite scenes and physical modeling point, this problem has been extensively studied for several
of the phenomenon are numerous. The model fills the gap decades, involving progressively more advanced types of
remote sensing data, from multispectral to hyperspectral, found in [25]). At the same time, desired properties of an
and from coarse to VHR. The current scientific develop- image classifier also include limited sensitivity to noise,
ment in this research area, coupled with the increasing robustness to possible curse-of-dimensionality issues, and
computing capabilities provided by hardware/software the capability to benefit from the spatial information asso-
platforms and architectures, makes it feasible to suc- ciated with the image. Such multiple goals present a chal-
cessfully apply advanced remote sensing image classifica- lenging scientific task, which has been addressed in the
tion techniques, not only in laboratory experiments but literature through many methodological approaches, in-
also when operational requirements and constraints need cluding Bayesian decision theory [4], neural networks
to be met. [26], fuzzy sets and logic [27], support vector machines
Most classification techniques can be formalized (SVMs) and kernel-based learning [25], multiple classifier
within a probabilistic framework; i.e., the input image systems [28], and spatial–contextual techniques, which are
and the output classification map are modeled as realiza- based, for example, on Markov random fields [29] or on
tions of 2-D random processes Iðx; yÞ and Cðx; yÞ (x ¼ 0; 1; region-based and object-based processing [30]. Here, we
. . . ; X 1; y ¼ 0; 1; . . . ; Y 1) defined over the same rec- briefly discuss three topical approaches (i.e., SVMs,
tangular X Y pixel lattice. In this framework, I is gene- Markov random fields, and region-based methods) that
rally a vector-valued process and associates with each pixel rely on well-defined methodological bases and that have
ðx; yÞ the multidimensional vector Iðx; yÞ of all related been experimentally found to be accurate and reliable tools
remote sensing observations, and C is a discrete process in many land-cover mapping applications.
and assigns, to each pixel ðx; yÞ, a class label Cðx; yÞ taking We first focus on SVMs, which currently represent a
values in a predefined finite set of land-cover classes [1], well-established and widespread approach to remote
[4]. In the language of pattern recognition, the compo- sensing data classification. SVMs can be used in several
nents of Iðx; yÞ are called Bfeatures[ and Iðx; yÞ is the fea- problems, encompassing not only classification but also
ture vector of pixel ðx; yÞ. For example, the observations regression [31], [32] and probability density estimation
recorded by a multispectral sensor within different wave- [33]. SVMs are a family of theoretically rigorous tech-
length ranges can be features [1]. Further features (e.g., niques based on the integration of methodological contri-
texture descriptors or image transforms) can be computed butions stemming from diverse fields, such as statistical
from the input image and used along with or in place of the learning [25], convex optimization [34], functional anal-
EO observations [1]. ysis, and Hilbert-space theory [35]. This combination has
In general terms, the goal of a supervised image classi- proven to be capable of offering convincing response to
fication method is to compute an estimate Cðx; ^ yÞ of the crucial methodological issues regarding generalization ca-
process Cðx; yÞ such that the probability of a classification pability, robustness to curse of dimensionality, and preven-
error [i.e., the probability that Cðx; ^ yÞ 6¼ Cðx; yÞ] is mini- tion of local minima, issues that were previously open to
mized. To this end, a subset of pixels, called training earlier popular families of pattern recognition techniques
pixels, the class labels of which are known a priori, is also (e.g., neural networks) [25]. The vast diffusion of SVMs in
assumed to be available (the adjective Bsupervised[ ac- learning applications has acknowledged their role as a
tually denotes the availability of such training informa- mature result of the development of pattern recognition in
tion). From an operational viewpoint, training data may the last decades. In remote sensing, SVMs are currently
derive from in situ surveys or be visually selected in the considered to be powerful tools for land-cover mapping
image by a human photo-interpreter [24]. In both cases, because of the accurate results they have obtained in the
their collection is generally time expensive. However, this classification of diverse types of satellite data [32]. Their
is usually not a difficulty in the application to vulnerability success in real-world applications, despite their intrinsi-
and elements-at-risk mapping, because land-cover classi- cally substantial computational burden, is currently fea-
fication is typically a scheduled operation (e.g., once per sible also because of the efficient numerical optimization
year or per season) and is not meant to be included within strategies that have been developed and tailored for
real-time emergency processing chains. From a pattern SVM-based learning [32], the small number of input para-
recognition standpoint, training pixels play the role of meters to be tuned [32], and most recently, the intro-
Bexamples of the classes[ and are essentially used to char- duction of automatic optimization algorithms for these
acterize prior knowledge on the statistical behaviors of the parameters [31], [36], [37], [88].
feature vectors belonging to each class [4], [24]. However, a still largely open issue in SVM-based
From this perspective, the intrinsic difficulty of a su- classification is that the related theory models the input
pervised image classification problem is due to the need to data as a collection of independent random samples. When
reliably Bgeneralize[ from a finite (and often small) col- images are being handled, this statement implies that
lection of preclassified training samples a reliable Bmodel[ samples drawn from the stochastic process Iðx; yÞ at dif-
for the unknown statistical dependence of the feature vec- ferent pixel locations are independent of one another
tor and the label stochastic processes (a thorough discus- (similar to a whiteness assumption in random noise mo-
sion of generalization concepts and properties can be deling), a condition that is readily violated by real-world
images, in which neighboring pixels normally exhibit high image classification. The key concept of these methods is
correlations. to generate, from the input EO image, a collection of seg-
Accordingly, the second approach, i.e., Markov random mentation results parameterized by some spatial scale
field (MRF) modeling, exhibits complementary properties. variable. In a given VHR image, it is expected that the
MRFs are a wide class of probabilistic models for 2-D main regions and structures in the scene will be identified
stochastic processes over discrete lattices [29]. They repre- at coarse scales, whereas smaller structures, objects, and
sent flexible and powerful models for the spatial– parts of objects will be apparent at fine scales. Multiscale
contextual information associated with images, and their methods aim at advantage of this complementary infor-
use for classification benefits from the dependence among mation to model into the classification process the spatial–
neighboring pixels to minimize the probability of classi- geometrical structure of the scene, which is an approach
fication error. MRFs intuitively generalize the notion of that bears similarities with concepts of computer imaging
Markov chain, which is a popular model for 1-D random used, for example, in robotics or video surveillance.
sequences in many applications (e.g., queueing systems,
speech and audio processing, genome sequence analysis) B. Mapping of 3-D Buildings
[38], to a 2-D setting. Among the reasons for the relevance With the advent of new remote sensing modalities,
of MRFs in remote sensing, are: 1) the rigorous mathe- various methods for 3-D building extraction have been
matical formalization, which clearly characterizes the re- proposed or are under study. In addition to optical data,
lationships between local and global properties of the light detection and ranging (LIDAR) [41] and SAR data
image statistics [29]; 2) a remarkable flexibility to incor- [42] are now available for high- to very-high-resolution
porate multiple information sources in an image-analysis imagery. However, although the novel possibilities granted
task [39]; and 3) the accurate results they have provided in by polarimetric and interferometric SAR have been the
many analysis problems, such as remote sensing data focus of many research efforts [43], the common practice
classification, restoration, denoising, feature extraction, is still based mostly on optical data. Submeter imagery is
and multitemporal analysis. MRF modeling has recently now widely available from a variety of satellites (e.g.,
attracted even greater attention as a feasible classification QuickBird, WorldView-1 and 2, GeoEye-1), at an afford-
tool also as a result of the development of computationally able price, thus narrowing the gap between satellite images
efficient approaches (e.g., graph cuts [29]) to related com- and aerial photographs, the resolution of which usually
plex minimization problems. ranges from 0.2 to 0.3 m [44].
In particular, the role of spatial information modeling As widely recognized in the literature [45], although
in remote sensing is now emphasized even further by the these types of sensors have revolutionized the role of VHR
vast popularity gained by VHR satellite sensors during the satellite imagery in various applications, the available tools
last decade. With VHR imagery, not only correlated neigh- for processing satellite imagery still require a high level of
boring pixels but also a well-defined spatial–geometrical technical expertise and are usually complex. Traditionally,
structure is often apparent in the scene, especially when the extraction of 3-D measures is based on double imaging
the image is taken, for example, over urban areas, infra- such as photogrammetric or stereoscopic pairs [46]. A
structures, or cultivations (which are important areas for comprehensive investigation of the application of 1-m
vulnerability and elements-at-risk mapping). From this IKONOS satellite imagery to 3-D building reconstruction
perspective, a third interesting approach to classification, is reported in [45], in which stereoscopic images are used.
which is especially relevant for VHR images, is repre- More recently, with the availability of precise satellite
sented by region-based and object-based methods. These sensor orientation models, which describe the relationship
methods rely on the combination of classification and seg- between 2-D image coordinates and 3-D object points, the
mentation processes. A segmentation algorithm aims at extraction of 3-D spatial information is also made possible
partitioning an image into a set of homogeneous regions by using only a single optical image.
(or segments), which often correspond to objects or por- For decades, the extraction of buildings from digital
tions of objects [40] and intrinsically provide a character- images has been performed manually, requiring well-
ization of the spatial–geometrical structure associated with trained operators employing slow and expensive proce-
the image.1 A basic intuitive region-based classification dures [44]. Today, semiautomated or automated extraction
strategy could be to assign a class label to each segment of building features from optical images is made possible
rather than separately to each pixel; however, more so- through specific image processing algorithms. Various
phisticated architectures that integrate segmentation, fea- techniques with different automation levels have been
ture extraction, and possibly multiscale analysis modules presented in the literature to address fundamental steps in
may be devised [30]. In this framework, multiscale region- the reconstruction of buildings, such as roof extraction.
based methods represent especially effective tools for VHR For extracting the contour of an object of interest, such
as the roof of a building, the classical strategy adopted in
1
A further category of segmentation methods (i.e., the seed-based image processing when avoiding specific prior image mod-
region segmentation techniques) will be introduced in Section III-B. els, is represented by an edge-based approach. However,
such techniques are successful in detecting building roofs ing it possible to process large images. Seed points, i.e.,
only in simple scenes [47], and were soon abandoned in examples of pixels belonging to the object of interest (e.g.,
favor of approaches that integrate other features with a building roof), can be manually inserted by the user or
edges when dealing with complex scenes. automatically detected to initialize the method. The seg-
In particular, the active-contour approach [48] is able mentation results generally depend on the definition of
to integrate contour information with shape models. With specific aggregation criteria and on the order followed in
this approach, the snake method uses a dynamic curvilin- selecting the pixels to be analyzed for aggregation.
ear contour to search the edge image space until it settles Among the semiautomated methods for seed-based re-
on the boundaries of regions; such an approach is driven by gion segmentation, so-called seeded region growing (SRG)
an energy function that is attracted to edge points. Some [50] is a simple but effective method to adaptively grow a
modified versions of the snake method [48] have been region of interest starting from a given seed pixel. Fur-
proposed for extracting buildings from satellite images, thermore, the results of SRG may depend on the pixel-
although the complexity related to initialization and selec- scanning order applied during the growing procedure.
tion of weight coefficients is often open-ended and creates Some improvements have been proposed to solve this
major problems with repeatability and robustness. problem in [51], in which an adaptive seed-based region
Model-based approaches have been proposed in an segmentation method, exploiting the analysis of color
effort to solve the problems related with image variations images to detect building roofs in optical VHR remote
attributable to differences in type, scale, and aspects of the sensing images, was presented.
objects of interest. Available prior knowledge might refer
to geometrical constraints or building models. The major C. Mapping of Changed and Flooded Areas From
drawback to these approaches is that all variability must be Multitemporal SAR
completely taken into account in the adopted model. Assuming that a flood occurred in a given area at time
In the present application, the use of segmentation t , a pair of multitemporal images, acquired by a given SAR
methods, although they have been less explored, repre- sensor over the same area at times t0 G t (before-event
sents an interesting approach. For example, in [46], this image) and t1 > t (after-event image) is assumed to be
approach is applied to stereoscopic aerial images for roof available. In an operational emergency situation, the image
mapping. The major recent advances in image segmenta- at time t1 could, for example, be the first feasible acqui-
tion are related to two main aspects: the use of color and sition after the flood, whereas the image at t0 is generally
the growing of single regions. an archived acquisition. Hence, an underlying operational
Color is perceived by humans as a combination of hypothesis is that an archive of SAR images exists over the
tristimuli, i.e., R (red), G (green), and B (blue), which are monitored area.
called primary colors. The RGB color space is the most Detecting the flooded areas and the changes that oc-
commonly used model for color picture representation curred between the two acquisitions is a challenging prob-
and expresses a generic color image Iðx; yÞ (x ¼ 0; 1; . . . ; lem, both because of the speckle phenomenon, which is
X 1; y ¼ 0; 1; . . . ; Y 1) as vector valued with three intrinsically related to the scattering of the radar signal and
scalar components: IðRÞ ðx; yÞ, IðGÞ ðx; yÞ, and IðBÞ ðx; yÞ. The behaves similarly to a multiplicative and often strong
RGB space is suitable for color display but is not optimal for noise, and because of the need to proceed without training
color image segmentation because of the high correlation data about the two input scenes. Indeed, changed and
among the R, G, and B components [49]. Moreover, the flooded-area detection is usually included in the rapid
measurement of a color in the RGB space does not repre- mapping procedures to be performed shortly after the
sent color differences in a uniform scale; hence, it is flood to assist the authorities in charge of the emergency.
impossible to evaluate the similarity of two colors from Hence, detection of these areas is intended to be
their distance in the RGB space. To address this problem, delivered approximately a few hours after the acquisition
other representations are used (e.g., the Lab color space) of the after-event image, which is a timing that is
that are more perceptually uniform than the RGB space generally incompatible with collection of training data.
from which they are derived. Obviously, if multiple EO images are available after the
The seed-based region segmentation approach is par- event, similar comments apply to the problem of mapping
ticularly suitable for cases in which it is necessary to seg- the evolution of the affected areas (e.g., flooded areas
ment a few distinct objects in an image. The main drying out).
difference from the segmentation process defined in As in Section III-A, when adopting a pattern recogni-
Section III-A is the incomplete scanning of the entire tion and probabilistic approach, the two acquired images
image and the object-oriented strategy. Therefore, this are modeled as realizations of 2-D stochastic processes
type of segmentation is useful for building extraction, I0 ðx; yÞ and I1 ðx; yÞ (x ¼ 0; 1; . . . ; X 1; y ¼ 0; 1; . . . ;
because it is able to investigate only particular parts of the Y 1) defined over the same lattice. These processes
considered image while looking for a specific structure of may be either real or complex, scalar, vector, or matrix
interest, thus minimizing the computation time and mak- valued, depending on the specific type of input SAR data
(e.g., SAR amplitude images are real nonnegative, whereas 2) Flooded-Area Color Display: For flooded areas,
polarimetric SAR data may be complex vector or matrix multitemporal SAR data can be processed to obtain map-
valued) [2]. ping results suited to diverse purposes. When the goal is a
fast and easy interpretation of an event, multitemporal
1) Change Detection: Similar to I0 and I1 , the output data takes can be integrated to create a specific image
change map is modeled as a realization of a stochastic display. However, when a flooded-area map needs to be
process Cðx; yÞ, which attaches to each pixel a binary label integrated in a damage-assessment chain, as discussed in
indicating either the Bchange[ or Bno-change[ class. Section II-D, an unsupervised classification result, in
Different types of change might also be discriminated in which a discrete label is assigned to each pixel (similar to
certain cases; we refer to the specific literature (e.g., [52]) the change-detection formalization above), needs to be
for such situations. In an unsupervised framework, an generated.
^ yÞ of the process Cðx; yÞ is searched for, based
estimate Cðx; Focusing on the former purpose, the multitemporal
solely on the multitemporal pair fI0 ðx; yÞ; I1 ðx; yÞg. This pair fI0 ðx; yÞ; I1 ðx; yÞg is used in a data-fusion procedure to
classification task is made especially challenging by a integrate and enhance the image content related to flooded
number of factors, such as: 1) the need to obtain accurate regions. Thus, in the literature, a colored display of a
models for the statistics of changed and unchanged image multitemporal dynamic event, obtained from an RGB
areas directly from the image data without prior training image composition, is usually employed. The channels of
information on the two classes; 2) the strong non- the RGB space are associated with the two available images,
Gaussianity of most SAR data distributions, some of which obtaining a false-color display that makes it possible to
(e.g., the K-distribution) are not conveniently fit by visually appreciate flooded areas. Both SAR backscattering
classical parameter-estimation methods, such as maximum intensity and further information extracted from the
likelihood; 3) the speckle and the resulting need to incor- multitemporal pair (e.g., interferometric SAR coherence
porate spatial–contextual information to minimize its im- [64]) can be used. Each RGB component IðiÞ is the output of
pact on the change-detection results; and 4) the necessity a specific functional Fi applied to the original multi-
to correctly capture the temporal variations caused by temporal pair
actual ground changes, without simultaneously critically
emphasizing the temporal variations caused by differences
in acquisition modalities (e.g., different viewing angles or IðiÞ ðx; yÞ ¼ Fi fI0 ðx; yÞ; I1 ðx; yÞg; i 2 fR; G; B; g: (9)
orbit directions) or by mismatches in the spatial alignment
of the two images.
A variety of approaches has been explored to address Based on its simplicity and good performances, such a
these tasks, involving not only SAR-specific operators (e.g., method is often chosen because it allows a fast and easy
SAR coherence [53] or polarimetric transformations [54]) interpretation of an event in real time or shortly after its
but also methodological contributions from quite diverse occurrence [65]. The result of the composition can also be
signal and image processing areas, such as: Bayesian deci- used as a preliminary data-fusion step before other image-
sion theory [55]–[57], information theory [58], copula analysis algorithms, such as classification or change detec-
functions [59], Markov random fields [60], Markov chains tion [66], are applied. More generally, RGB compositions
[61], Mellin transforms [60], [62], or wavelet transforms have proven to be valid tools in several contexts, including
[52], [63]. These methodological contributions, together multispectral [67], hyperspectral [68], and polarimetric
with the advances in hardware and software mentioned SAR imagery [69], and possibly in situations involving the
above, currently make it possible to successfully and accu- fusion of heterogeneous (e.g., multiparametric or multi-
rately address operational change-detection tasks. Accord- source) data.
ingly, we note that the capability of complying with the Although RGB compositions are widely used in flood
related requirements of accuracy and computational time monitoring, a still scarcely addressed, but relevant, issue is
(critical performance figures for emergency response) is the study of the optimal color composition criteria, robust-
currently ensured by both the scientific progress in the ness, and reproducibility. A sensitive and complex problem
methodological areas mentioned above and the current is the precise radiometric calibration of the two images,
availability of a number of time-efficient probabilistic which is often affected by measurement errors. Image
modeling and parameter-estimation algorithms recently processing methods are needed to ensure image cross
introduced in the SAR and change-detection literature. calibration. Therefore, approaches based on theoretical
Two examples of such algorithms may be the approach models or histogram matching have been proposed [70]; in
based on Mellin transforms and second-kind statistics for both cases the calibration methods are applied before
the estimation of the probability density of SAR data multitemporal data fusion. Even though these operations
[60], [62] and the recent computational advances in can achieve accurate results, they necessarily involve a
MRF modeling mentioned in Section III-A regarding complex analysis and some a priori knowledge about the
classification. SAR formation process, knowledge that is difficult to
obtain completely and precisely. In [71], a method for the Similar to 3-D-building detection (see Section III-B),
detection of flooded areas through an RGB composition is for detailed flooded-area mapping, seed-based region seg-
presented that automatically cross-normalizes multitem- mentation also bears some advantages over the segmen-
poral SAR images without an a priori model. An appro- tation of the entire image. If the user/operator identifies
priate preprocessing step is developed that preserves the seed pixels belonging to the region of interest (i.e., a
image information content despite the strong histogram flooded or permanent-water area), a set of specific class
asymmetry caused by the non-Gaussianity of SAR data. signatures can be automatically derived. Such signatures
Both acquisitions are then cross-calibrated by matching to can be exploited to focus the segmentation process on the
a common uniform distribution model. region of interest, thus taking advantage of local adaptive
Because of such current advances in the related liter- statistical measures [15].
ature, RGB compositions have become valid tools for the The integration of seed-growing and fuzzy connected-
fast and immediate generation of preliminary results for ness is particularly significant because it allows the pro-
the evaluation of flooded areas. The development in hard- cessing to be focused on specific parts of the image, such as
ware and software and the short revisit times of current those areas spatially connected to a river. Therefore, the
SAR constellations allow such preliminary maps to be computational cost is sharply reduced as compared with
generated and delivered shortly after the event. alternate methods and large images can be processed for
flooded-area mapping in short times. The intrinsic
3) Detailed Flooded-Area Detection: As discussed in adaptivity of this approach makes it possible to apply the
Section II-D, beyond the generation of fast results to same algorithms to images acquired with different moda-
qualitatively localize inundated regions, quantitative lities and spatial resolutions. The user interaction required
results precisely identifying flooded areas are also a fun- for the location of a few seed pixels is intuitive and fast and
damental piece of information for evaluation to damage- can be supported by easy-to-use graphical interfaces. How-
assessment chains. In this case, the multitemporal pair ever, the automation of the seed-selection phase for
fI0 ðx; yÞ; I1 ðx; yÞg is analyzed to generate a map Cðx; yÞ on flooded-area mapping is quite straightforward, which
which flooded and permanent-water areas are detected. allows the whole process to be performed in a user-
For this purpose, segmentation, classification, and change- independent, precisely repeatable, and faster fashion [77].
detection methods can be applied. The current availability of such fast and semi- or
Outside the stringent time requirements of emergency completely automatic approaches allows image processing
response, supervised classification algorithms could be methods to be used for the operational analysis of multi-
used to identify flooded areas, based on the extraction of temporal remote sensing images acquired before, during, and
suitable features from the input images (e.g., texture [72] after a flood, meeting the related requirements in terms of
and morphological features [73]). accuracy, precision, and execution time and successfully
Despite their complexity in parameter selection and coping with both the possibly large amount of collected data
repeatability, statistical active contours allow flood boun- to be processed and the time requirements mentioned above.
daries to be precisely delineated [73]. As an alternative,
segmentation techniques are widely used to detect flooded
or permanent-water regions [74]. For this purpose, model- I V. EXPE RIMENTAL EXAMPLES AND
based, edge-based, split-and-merge, and region-growing CASE STUDIES
segmentation techniques are the most commonly applied Two relevant case studies of recent floods in Albania and
strategies. In this framework, an interesting approach is Italy are presented. Here, we highlight the interdisciplin-
represented by fuzzy methods. When handling real-world ary synergies mentioned above between EO-image-analysis
imagery, the intrinsic uncertainty and variability in the and environmental modeling approaches. We do not de-
input data and the frequent unavailability of mathemati- scribe in detail each adopted technique and do not discuss
cally well-defined models for the imaged scene are the the quantitative accuracy of each processing result; how-
main motivations for the fuzzy approach to image pro- ever, descriptions of adopted methods and details of their
cessing. Even though the genesis of fuzzy methods is experimental validation, including comparisons with state-
related more to medical images, two of the best known of-the-art techniques, can be found in the related technical
fuzzy-segmentation methods, based on the so-called papers [15], [37], [51], [60], [71], [78], [88].
Bfuzzy-connectedness[ concept [75], [76], are also well The resulting maps will be shown superimposed onto
suited to remote sensing image analysis. Both of these geographical layers within the NAZCA web geographic
methods successfully address the long-standing problem of information system (WebGIS) 3-D application [79], which
integrating the statistical and contextual–topological is included in the Opera Web InterfaceS (OWIS) web
information associated with an image. The fuzzy segmen- platform, implemented by Acrotec S.r.l. (Savona, Italy) in
tation approach has proven robust for shape complexity, the framework of the OPerational Eo-based RAinfall-
noise, and low-contrast borders and also prevents the need runoff forecast (OPERA) project funded by the Italian
for time-consuming iterative approaches. Space Agency.
A. Shkodër, Albania, 2010 international flood relief organizations, the floodplain

The first case study examines the flood that occurred downstream of Shkodër Lake was inundated to an ex-
south of Shkodër, Albania, in January 2010. On January 11, tension of approximately 10 500 hectares [80]. Approxi-
the northwestern regions of Shkodër and Lezhë were mately 2500 houses were flooded, leading to the
inundated as a result of heavy rainfall and high temper- evacuation of approximately 6000 people starting on
atures in Albania. Large volumes of rainfall in the Drin January 5, 2010. The flooding of the area, as docu-
River basin, associated with high temperatures that re- mented in this paper (see below), lasted until the end
duced snow accumulation during December 2009, created of January 2010.
high flows in the Drin River. Recorded accumulated To map the vulnerability and elements at risk, an image
rainfall in northern Albania for that month (approximately acquired in September 2009 by the thematic mapper (TM)
200 mm/month) reached more than 200% of the average passive multispectral sensor onboard the Landsat-5 satel-
accumulation for the same period [80]. This situation led lite was considered [see Fig. 2(a) for a detail of this image].
to increased water levels in the three related hydroelectric The spatial resolution was 30 m (a sufficient value,
power lakes (Fierzë, Koman, and Vau i Dejës). The autho- considering the size of the studied flooded area). The
rities were forced to release water from these lakes, in- image depicted approximately a 27 24-km2 scene (906
creasing the discharge downstream of the three lakes up to 805 pixels) and was composed of seven spectral channels
approximately 2500 m3 /s, whereas the maximum river within the visible, near-infrared, and thermal infrared
capacity does not exceed 800 m3 /s. The volume of released wavelengths. An expert photo-interpreter visually identi-
water increased flooding in the Buna River, which from fied five main land-cover classes as Burban and built-up
the Drin downstream forms Shkodër Lake. According to land,[ Bagricultural,[ Bforest,[ Bwater,[ and Bbare soil,[
Fig. 2. ‘‘Shkodër’’ case study: (a) RGB composition of three channels of the Landsat-5 TM image; (b) land-cover map obtained by classifying
this image; (c) elements-at-risk map obtained from the land-cover map; (d) vulnerability map. Color legend for (b): yellow ¼ urban and built-up;
magenta ¼ agricultural; green ¼ forest; blue ¼ water; red ¼ bare soil. Color legend for (c) and (d): dark brown ¼ elements-at-risk class A
(continuous urban areas); pink ¼ B (noncontinuous urban areas); violet ¼ G (cultivations); olive ¼ H (vineyards, olive gardens, fruit gardens);
pale green ¼ M (woods); dark green ¼ N (bare soil); blue ¼ Z (nonsensitive to flood). In (d), the RGB composition of the TM channels is visible in the
areas with negligible vulnerability. All images are displayed by using the NAZCA WebGIS 3-D application.
and selected corresponding training pixels. The classifica- inundated areas, this second image was used in the expe-
tion method in [37] and [88], which is based on an ana- riments in place of the pre-event image. The considered
lytical integration of the SVM and MRF approaches, was scene is approximately 34 44 km2 (3359 4422 pixels),
applied with these training samples to generate a land- and in this case, the original spatial resolution of 5 m was
cover map. The set of the input features included the TM reduced to 10 m, a more than adequate resolution for the
multispectral observations and additional texture features subsequent hydraulic analysis.
aimed at optimizing the discrimination among the consid- Starting from the image pair, the so-called Bfast-ready
ered classes. This study used the semivariogram approach flood map,[ which is a flood picture generated automat-
to texture extraction, which is based on the computation of ically and very rapidly without any user interaction, can be
spatial second-order statistics related to the autocovariance produced for use by the authorities in emergencies and
function of the random field of the pixel intensities. This monitoring and risk-assessment situations. This colored
approach was found to be accurate in the application to map, shown in Fig. 3(c), displays the changes that occurred
multispectral image classification [81] and, compared to between the two data acquisitions and represents a useful
alternate texture-extraction methods (e.g., greylevel co- tool for generating an overall understanding of the pheno-
occurrence [1]), its tuning involves fewer input parameters menon in question. The method proposed in [71] was ap-
(details can be found in [81]). The classifier in [37] and [88] plied because it has several advantages over the other
was used because it allows the generalization capability of approaches. First, it avoids the calibration step, which is
SVMs and the contextual modeling capability of MRFs to be highly dependent on the acquisition parameters, time
combined in a unique framework. The automatic param- consuming, and can be affected by residual errors. Because
eter optimization algorithms in [82] and [83] were also it is not based on some a priori model, this recently pro-
integrated in this method, thus preventing the need for posed preprocessing chain achieves cross calibration and
manual parameter tuning. In the experimental analyses cross normalization (see Section III-C2) and can be ap-
conducted in [37], [88], and [84], this method provided plied both to unfiltered data and to images that have been
accurate classification results from diverse types of remote filtered by a despeckling method. Finally, it assures robust-
sensing data, encompassing hyperspectral, multichannel ness and reproducibility so that similar colors are repro-
SAR, and high-resolution multispectral imagery and out- duced when similar situations occur at ground, regardless
performing previous contextual classifiers based on Bayes- of the SAR sensor, the spatial resolution, and the specific
ian and non-Bayesian approaches (details can be found in despeckling filter being used. In particular, magenta pixels
[37], [88], and [84]). A detail of this land-cover map is indicate flooded areas, dark red areas indicate permanent
shown in Fig. 2(b). Visual inspection of the map, compared water, and unchanged ground regions are shown in cyan.
with the input Landsat-5 TM image, showed a remarkable To generate a detailed flood map, the multitemporal
accuracy in the mapping of the land covers. Similar seed-growing segmentation technique presented in [15]
analyses also suggested that visually less accurate results has been applied. Because this method is based on fuzzy
were obtained by alternate classifiers based on the processing, it can take uncertainty into account in the data
combination of MRFs with Bayesian parametric models representation and processing steps. This method requires
for the statistics of the pixel intensities or with the a very simple user interaction for the manual definition of
Bk nearest neighbors[ decision rule [1], [4]. a few points (i.e., seeds) that surely belong to water. The
On the basis of such classification, an elements-at-risk processing required to generate the detailed map is adap-
map was derived [Fig. 2(c)] that retains the same resolu- tive to the statistics local to the seed and is clearly adaptive
tion of the land-cover map. Eventually, a vulnerability to the specific image characteristics. The obtained flooded-
map, which contains the loss-function information, was area map is shown in Fig. 3(d). The two SAR images are
produced [Fig. 2(d)]; false color was used to display the separately processed, and then two segmented results are
degree of vulnerability, and the RGB composition of the fused in a single map so that every pixel is classified as
original TM image is visible in the areas with negligible Bwater in both images[ (blue) or Bflooded areas[ (cyan).
vulnerability to floods. This result identifies flooded areas connected with perma-
Two COSMO-SkyMed images, acquired in stripmap nent water and can be used for both a preliminary analysis
mode [which implies 5-m spatial resolution; see Fig. 3(a) of the flooded area over the investigated region and as
and (b)] were used for the generation of both RGB display input for assimilation into a hydraulic model. By visually
and detailed maps of the flooded area. Both images were comparing the two multitemporal SAR images, it is possi-
acquired after the flood event, because no pre-event image ble to say that both maps in Fig. 3(c) and (d) exhibit ac-
regarding the investigated area was available. In particular, curate results in terms of identification of flooded and
one image was acquired on January 10, 2010, immediately permanent-water areas, and allow a synthetic display of
after the flood, and shows the extension of the inundated the flood situation in a single image.
area. The second image was acquired on January 31, 2010, Based on the flooded areas retrieved from COSMO-
twenty-one days after the first image was acquired. SkyMed imagery, the 2-D hydraulic model run was set in
Because a large portion of the water retreated from the terms of boundary and initial conditions. This delicate
Fig. 3. ‘‘Shkodër’’ case study: COSMO-SkyMed images acquired on (a) January 10, 2010 and (b) January 31, 2010; (c) RGB color display and
(d) map of the flooded areas. Color legend for (c): dark red ¼ low backscattering in both images; cyan ¼ high backscattering in both images;
magenta ¼ increased backscattering from (a) to (b); green ¼ decreased backscattering from (a) to (b). Color legend for (d): blue ¼ water
in both (a) and (b); cyan ¼ water in (a) but not in (b). All images are displayed by using the NAZCA WebGIS 3-D application.
procedure requires a careful evaluation by a flood-model much more to the south as compared with what can be
expert to obtain a preliminary idea of the river segments of seen from the SAR-based results. This effect occurs be-
the flooding source, the flooded water volumes, and the cause the break line is detectable in the DTM but is
consistency of the input digital terrain model (DTM). fragmented and discontinuous as a result of sampling
Flooded areas obtained from satellite images provide cru- problems. Simply forcing the continuity of the break lines
cial help in defining all of these aspects. creates the result seen in Fig. 3(b) and almost perfectly
From this perspective, the Shkodër case study is par- matches the extensions of the flooded areas in the different
ticularly interesting because it offers a remarkable example temporal scenes. Such results were made possible by the
of the impact of an effective synergy between remote integrated use of remote sensing and hydraulic-modeling
sensing information and hydraulic modeling. For this case components and represent a significant example of a real-
study, only a low-resolution (90 m) DTM was available. world situation in which using only one of these two
The consistency of this DTM had two major gaps, which components (i.e., the latter) would lead to erroneous con-
were promptly identifiable by examining the SAR flooded clusions. The corresponding damage map is also reported
areas. Two major roads are present in the Shkodër flood in Fig. 4(c) and resembles the pattern of maximum water
plain, one running parallel to the north part of the Buna depths, although the damage percentage becomes saturat-
River developing in the northwest/southeast direction and ed because of the high vulnerability of the most diffuse
the second connecting two coastal dunes south of the river element at risk (i.e., cultivations) in the domain.
outlet. It is evident from the SAR images that these two
roads represent major break lines for the flood develop- B. Tanaro River, North Italy, 2009
ment. If the DTM was used without checking consistency The second case study is a flood of the Tanaro River (in
with the satellite-derived map of flooded areas, the water- northern Italy) that occurred on April 28, 2009. Between
depth map would be the one reported in Fig. 4(a), in which April 26 and 28, heavy and widespread rainfall occurred in
the hydraulic model of the flooded area would expand the Piedmont region (Italy), involving in particular the
Fig. 4. ‘‘Shkodër’’ case study: (a) hydraulic simulation of water depth without DTM correction; (b) hydraulic simulation of water depth,
constrained on satellite-derived flooded areas with DTM correction; (c) damage estimation. Shades of cyan and blue in (a) and (b) denote
water-depth levels. Shades of red and pink in (c) denote damage percentages. All images are displayed by using the NAZCA WebGIS
3-D application.
alpine and pre-alpine regions, the southern plains, and the were hit by the flooding, which primarily affected build-
hills south of the Po River, and produced critical ings and farmlands (Fig. 5).
conditions in the hydrographic network. This led to a A land-cover map was generated based on a high-
generalized excess of rainfall thresholds, triggering land- resolution image (spatial resolution: 4 m) acquired by the
slides on slopes and raising the water levels of rivers, multispectral passive IKONOS sensor [see a detail in
which in some cases reached critical levels, especially in Fig. 6(a)]. The image is composed of four channels
the southern Piedmont area. In the Tanaro River basin, a (essentially corresponding to blue, green, red, and near-
spatial average of 150–200 mm of accumulated rain was infrared radiation) and depicts a 5 5 km2 scene (1250
recorded between April 26 and 28. This produced a critical 1250 pixels). The imaged area includes part of Alessandria,
situation in the area around the city of Alessandria. The nearby agricultural fields and other vegetated covers, a few
river mainly flooded the riverbanks, creating water depths bare and wet soil regions, and part of the Tanaro River
of up to 2–3 m and extensions of approximately 2 km. At path. Training data were selected by an expert photo-
least 40 buildings were flooded, and 6000 people were interpreter for seven classes: Burban and built-up land,[
temporarily evacuated from some districts of Alessandria. Bagricultural,[ Brangeland,[ Bforest,[ Bwater,[ Bwet soil,[
Several local roads were damaged by the flooding. Approx- and Bbare soil.[ Some of these classes exhibited a well-
imately ten other municipalities surrounding Alessandria defined spatial–geometrical structure in the image, mostly
Fig. 5. ‘‘Tanaro’’ case study: water-depth and flooded-area map from ground recorded information (source: ARPA PiemonteVRegional
Environmental Protection Agency).
because of the rather regular geometrical shapes of build- important urbanized features, such as isolated buildings
ings and agricultural fields. To correctly capture this geo- or small connection roads, which present huge differences
metrical information, the classification method in [78], in vulnerability compared with the background, which is
which combines an MRF model with a multiscale region- represented in this area mainly by rangelands and
based approach, was applied to classify the IKONOS cultivated fields.
image. The method exploits a graph-based segmentation No SAR before-event acquisition was available, but
algorithm to generate a collection of segmentation maps multiple COSMO-SkyMed images were taken after the
associated with different spatial scales and fuses this ex- flood. In particular, two images acquired by COSMO-
tracted multiscale information through an MRF model. SkyMed in the stripmap modality on April 30 and 31 [i.e.,
Here, this multiscale region-based approach was deemed two and three days after the flood, respectively; see details
especially interesting due to the high spatial resolution of in Fig. 7(a) and (b) and Fig. 8(a) and (b)] were used to
the considered IKONOS image and the well-defined geom- experimentally investigate the behavior of the affected
etrical structures of some of the land covers in the scene areas. In this case, the original spatial resolution of 5 m
(e.g., Burban and built-up land[ and Bagricultural[). Simi- was preserved. The analyzed scene was approximately
lar to the classifier applied in the Shkodër case study, the 25 11 km2 wide (4963 2286 pixels).
use of the automatic parameter optimization method in Similar to the Shkodër case study, to generate a fast-ready
[82] prevents the need for manual parameter tuning. RGB color display of the flooded areas, preliminary filtering
When experimentally tested in [78] with multispectral was first used to reduce speckle, and then the method in [71]
IKONOS data of urban and vegetated areas, the method was applied [see Figs. 7(c) and 8(c)]. Furthermore, a flooded-
outperformed previous noncontextual and contextual clas- area map was obtained from the COSMO-SkyMed image pair
sifiers based on single-scale Bayesian and MRF modeling. by the multitemporal seed-growing segmentation technique
The resulting land-cover map [see Fig. 6(b)] accurately in [15] [see Figs. 7(d) and 8(d)]. The interpretation of colors
discriminates the land covers in the area, including those in the RGB display and the color legend in the flooded-area
characterized by natural textures (e.g., the Bforest[ class) map are the same as in the previous case study. Both results
and those exhibiting a spatial–geometrical structure, as compare favorably with the visual interpretation of the two
mentioned previously. As expected, visually less accurate SAR images, thus confirming the effectiveness of the adopted
results were obtained through contextual but nonregion- approach. Moreover, the change-detection method in [60]
based techniques using MRFs or texture extraction was applied to the COSMO-SkyMed multitemporal pair.
through semivariogram. This method uses MRFs to model spatial–contextual
On the basis of the high-resolution land cover, an information and combines the approach based on Mellin
elements-at-risk map was produced [Fig. 6(c)] and transforms with a specific formulation of the expectation–
eventually converted into a vulnerability map [Fig. 6(d)]. maximization (EM) parameter-estimation algorithm [85] to
The high spatial resolution allows the detection of characterize the image statistics. The method in [60] was
Fig. 6. ‘‘Tanaro’’ case study: (a) RGB composition of three channels of the IKONOS image; (b) land-cover map obtained by classifying this image;
(c) elements-at-risk map obtained from the land-cover map; (d) vulnerability map. Color legend for (b): yellow ¼ urban and built-up land;
magenta ¼ agricultural; light green ¼ rangeland; dark green ¼ forest; blue ¼ water; cyan ¼ wet soil; and red ¼ bare soil. Color legend for
(c) and (d): dark brown ¼ elements-at-risk class A (continuous urban areas); pink ¼ B (noncontinuous urban areas); violet ¼ G (cultivations);
medium green ¼ L (pasture lands and meadows); light green ¼ M (woods); dark green ¼ N (bare soil); yellow ¼ Q (humid areas); blue ¼ Z
(nonsensitive to flood). In (d), the RGB composition of the IKONOS channels is visible in the areas with negligible vulnerability.
All images are displayed by using the NAZCA WebGIS 3-D application.
chosen because of its robustness against speckle and its each considered building, given the obtained roof perim-
capability to effectively characterize the non-Gaussian eter and prior information about the building height, a 3-D
statistics of multitemporal SAR observations in a completely model was generated [see Fig. 9(b) and (d)].
unsupervised manner. The experimental validation carried Hydraulic-modeling gains in this case study enhanced
out in [60] also noted the remarkable accuracy of the method importance. In this case, the time scale of the flooding was
compared to previous approaches based on alternate MRF very short, which makes it difficult to capture through
models or on likelihood ratio tests for polarimetric SAR data, satellite a representative scene of the maximum flood
when applied to change detection with multichannel SAR. In extension. In addition, many hydraulic disconnections are
the present case study, the resulting change map, in which present because of the drying out of certain areas and the
changes caused by retreating flooded areas are well permanent ponding of others. However, the hydraulic
identified, is shown in Figs. 7(e) and 8(e). modeling makes it possible to reconstruct the water
To support a more comprehensive damage analysis, 3-D passage while maintaining the hydraulic connectivity in all
models of certain agricultural buildings near the Tanaro phases of the simulation. This advantage is even more
River were extracted from two VHR aerial images. As enhanced by the high resolution of the vulnerability
mentioned previously, the seed-based segmentation strat- information that, in this case, reached the level of single-
egy is particularly useful when looking for objects of building identification. By examining Fig. 10(a) one can
interest independently of the surrounding objects. Thus, easily see that the identified buildings were outside the
the color-based segmentation approach in [51] was used to flooded area detected from satellite data and therefore
extract roof perimeters [see Fig. 9(a) and (c)]. Then, for could not be marked with damage values. However, the
Fig. 7. ‘‘Tanaro’’ case study: details of the COSMO-SkyMed images acquired (a) two and (b) three days after the flood; (c) RGB color display and
(d) map of the flooded areas; (e) change map. Color legend for (c): dark red ¼ low backscattering in both images; cyan ¼ high backscattering
in both images; magenta ¼ increased backscattering from (a) to (b); green ¼ decreased backscattering from (a) to (b). Color legend for
(d): blue ¼ water in both (a) and (b); cyan ¼ water in (a) but not in (b). Areas classified as ‘‘change’’ in (e) are shown in red and
superimposed to the image in (a). All images are displayed by using the NAZCA WebGIS 3-D application.
hydraulic simulation [Fig. 10(b)] makes it evident that the satellite-based flooded area map prevented the hydraulic
detected flooded areas were accumulation areas, although model from being negatively affected by the available low-
their flooding also involved other areas (including those resolution DTM, in this case, conversely, the hydraulic
where the identified buildings stand) that were already dry model allowed the extension of flooded-area estimates and
at the times of the satellite acquisitions. By using the the derivation of damage estimates outside the time
hydraulic simulation, the buildings were marked with window that had satellite observations.
minor damage (less than 20%: green color code) or modes
damage (35%: yellow color code) according to how they
were involved by the water in the early stage of the V. CONCLUSION
flooding. Similar to the previous case study, the importance The main issues related to the exploitation of remote
of the integrated use of remote sensing and hydraulic- sensing data and their synergistic use with physically based
modeling components is clearly confirmed. More precisely, hydrometeorological models for flood-risk monitoring and
whereas in the Shkodër case, the integration of the assessment were discussed in this paper. In particular, the
Fig. 8. ‘‘Tanaro’’ case study: details of the COSMO-SkyMed images acquired (a) two and (b) three days after the flood; (c) RGB color display and
(d) map of the flooded areas; (e) change map. Color legend for (c): dark red ¼ low backscattering in both images; cyan ¼ high backscattering
in both images; magenta ¼ increased backscattering from (a) to (b); green ¼ decreased backscattering from (a) to (b). Color legend for
(d): blue ¼ water in both (a) and (b); cyan ¼ water in (a) but not in (b). Areas classified as ‘‘change’’ in (e) are shown in red and
superimposed to the image in (a). All images are displayed by using the NAZCA WebGIS 3-D application.
role of remote sensing imagery in mapping vulnerability to would have been erroneous for the first of the two case
floods, elements at risk, and after-event damages was studies. Therefore, as confirmed in the two considered
outlined, and the main methodological approaches aimed real-world situations, an integrated use of satellite-derived
at extracting the relevant thematic information from EO thematic maps and hydraulic models allows the flood
data were reviewed. Examples of image-analysis results scenario to be correctly understood and characterized,
were shown with regard to two case studies related to whereas using only one of these two components
recent floods with different spatial extensions and separately would lead to erroneous or more limited results.
dynamics. In addition, examples of the outputs that can Compared to traditional approaches, which are pri-
be derived from hydraulic modeling (e.g., water-depth marily based on in situ surveys, such an integrated
maps), for which the satellite images have been utilized to approach exhibits complementary properties. On the one
determine initial conditions, were also shown. Obviously, hand, when focusing on a small and specific target area
the latter results cannot be generated by making use only (e.g., a given building), in situ surveys generally ensure
of the satellite images. However, without the information higher accuracy in the characterization of vulnerability
derived from the satellite images, the output of the model and damage. On the other hand, EO data allow such
Fig. 9. ‘‘Tanaro’’ case study: (a) and (c) identification of roof perimeters for 3-D building mapping and (b) and (d) corresponding
3-D extrusions. Figures (b) and (d) are displayed by using the NAZCA WebGIS 3-D application.
information to be mapped onto large areas, in a spatially supported by the current advances in hardware and
distributed fashion, and to be updated in a temporally computing resources, presently allows the potential for
repetitive manner, a capability that is usually unfeasible classification, segmentation, and feature-extraction tech-
or extremely expensive by ground means. niques to be exploited, not only for laboratory experiments
The pattern recognition and image processing methods but also for operational applications. Finally, the capability
that have been used for experiments in this paper were to interoperate these EO image-analysis methods with
previously validated with several data sets, both qualita- processing techniques drawn from the hydrometeorology
tively (i.e., by visually analyzing the mapping results) and and environmental engineering fields allows the thematic
quantitatively (i.e., by computing accuracy parameters results extracted from EO data to be translated to
with respect to suitable ground-truth maps; see [15], [37], information that is relevant for the authorities in charge
[51], [60], [71], [78], and [88]). The input EO data fed here of flood-risk management and permits the development of
to these methods are analogous to those used for validation image-analysis techniques to be guided according to the
purposes, which suggests that comparable accuracies can requirements of the flood-risk application.
be expected as well. This unique combination of interdisciplinary effort,
The role of EO for flood risks relies on several important scientific maturity, and data availability can effectively
factors that collectively contribute to the feasibility of support the operational integration of EO data into flood-
current EO-based flood-risk monitoring and assessment risk monitoring and assessment systems that can be
architectures. First, the increasing availability of EO data deployed at regional, national, or international levels,
from a number of diverse satellite sources (both multi- with the aim of successfully handling flood events and
spectral and SAR) currently allows the support for a consequently minimizing their impacts.
comprehensive satellite coverage of a given flood event Driving this development further is an important
because of the VHRs and short revisit times granted by the technical and scientific challenge because of the impor-
latest and upcoming missions. Then, the scientific maturity tance of proposing more accurate and reliable processing
of the field of remote sensing image analysis, rooted in the methods and of keeping these methods continuously
fields of pattern recognition and image processing and updated with new sensors and satellite missions, new
Fig. 10. ‘‘Tanaro’’ case study: 3-D buildings superimposed (a) onto the flooded areas mapped from COSMO-SkyMed images and (b) onto the
water-depth map generated by hydraulic modeling. Note that in (a) buildings do not seem affected by water, whereas in (b), hydraulic continuity
of the flood is respected, and a damage percentage can be computed for each building (20%: green code; 35%: yellow code). Color legend
for (a): blue ¼ water at both observation dates; cyan ¼ water on one date but not on the other. Shades of cyan and blue in (b) denote
water-depth levels. Both images are displayed by using the NAZCA WebGIS 3-D application.
types of remote sensing data (e.g., further improved spatial The framework and image-analysis methods described
resolutions and revisit times), new input information in this paper could also be applicable to other types of
sources (e.g., wireless sensor networks and social net- natural disasters, such as forest fires, landslides, or
works), and new computing capabilities (e.g., cloud earthquakes. Similar synergistic approaches, involving
computing). In addition, it is a challenge to fully exploit remote sensing, civil and environmental engineering,
the potential of EO for flood risks also from an and Earth sciences, could be devised to exploit satellite
organizational viewpoint because it may require modifying observations within the processing chains for civil
or reconfiguring currently consolidated operational chains protection from these environmental risks. In particular,
for flood-risk management (most of which are non-EO the image-analysis methods discussed here in the applica-
based) to take advantage of the availability of additional tion to floods play relevant roles in extracting thematic
information extracted from satellite observations. information of interest for other natural disasters from
satellite data. For instance, unsupervised change detection for floods to other environmental risks is a further
and multitemporal RGB visualizations are important tools development worth pursuing. h
for detecting and visually emphasizing burnt areas after a
forest fire (see, e.g., [86]) or damaged buildings and
structures after an earthquake (in the latter case, VHR Acknowledgment
images are clearly necessary). Similarly, the 3-D modeling The authors would like to thank the Italian Space Agency
of buildings plays a primary role in the framework of risk (ASI) and Dr. L. Candela (OPERA Project Manager) for
prevention and management for earthquakes in urban providing the COSMO-SkyMed images used in the experi-
areas, and supervised classification provides relevant mental examples; C. Versace and Acrotec S.r.l. (Savona,
information for forest-fire prevention because it allows Italy) for developing the NAZCA interface used for
an accurate inventory of vegetated land covers in forested visualizing the mapping results; Dr. M. De Martino for her
areas to be generated and updated over time [87]. This help in the experiments; and Dr. G. Troglio, D. Di Gloria, and
generalization of the interdisciplinary approach discussed M. Fiorini for their contributions to implementation.
REFERENCES [14] S. Martinis, A. Twele, and S. Voigt, BTowards [28] L. I. Kuncheva, Combining Pattern Classifiers:
operational near real-time flood detection Methods and Algorithms. New York: Wiley,
[1] J. Richards and X. Jia, Remote Sensing Digital using a split-based automatic thresholding 2004.
Image Analysis. New York: Springer-Verlag, procedure on high resolution TerraSAR-X
2005. [29] S. Z. Li, Markov Random Field Modeling in
data,[ Natural Hazards Earth Syst. Sci., vol. 9, Image Analysis. New York: Springer-Verlag,
[2] C. Oliver and S. Quegan, Understanding no. 2, pp. 303–314, 2009. 2010.
Synthetic Aperture Radar Images. Raleigh, [15] S. Dellepiane, E. Angiati, and G. Vernazza,
NC: SciTech, 2004. [30] T. Blaschke, S. Lang, and G. Hay, Eds.,
BProcessing and segmentation of Object-Based Image Analysis. New York:
[3] G. Boni and F. Siccardi, BScenes and COSMO-SkyMed images for flood Springer-Verlag, 2008.
scenarios,[ Public Service Rev., Eur. Sci. monitoring,[ in Proc. IEEE Geosci.
Technol., vol. 10, pp. 128–129, 2011. [31] G. Moser and S. B. Serpico, BAutomatic
Remote Sens. Symp., Honolulu, HI,
parameter optimization for support vector
[4] K. Fukunaga, Introduction to Statistical Pattern 2010, pp. 4807–4810.
regression for land and sea surface
Recognition, 2nd ed. New York: Academic, [16] Disaster Management Manual, temperature estimation from remote-sensing
1990. UNDP/UNDRO, 1991. data,[ IEEE Trans. Geosci. Remote Sens.,
[5] B. Merz, H. Kreibich, R. Schwarze, and [17] V. Meyer and F. Messner, BNational flood vol. 47, no. 3, pp. 909–921, Mar. 2009.
A. Thieken, BReview article: Assessment of damage evaluation methods: A review of [32] G. Camps-Valls and L. Bruzzone, Eds.,
economic flood damage,[ Natural Hazards applied methods in England the Netherlands, Kernel Methods for Remote Sensing Data
Earth Syst. Sci., vol. 10, pp. 1697–1724, 2010. the Czech Republic, and Germany,[ Analysis. New York: Wiley, 2009.
[6] C. H. Green, D. J. Parker, and P. J. Emery, UFZVDiscussion Papers, 2005, pp. 1–47.
[33] P. Mantero, G. Moser, and S. B. Serpico,
BThe real cost of flooding to households: [18] D. I. Smith, BFlood damage estimationVA BPartially supervised classification of
Intangible cost,[ Flood Hazard Research review of urban stage damage curves and remote-sensing images through SVM-based
Centre, Enfield: Middlesex Polytechnic, loss functions,[ Water SA, vol. 20, no. 3, probability density estimation,[ IEEE Trans.
Geography and Planning Paper, 1983, no. 12. pp. 231–238, 1994. Geosci. Remote Sens., vol. 43, no. 3,
[7] D. Dutta, S. Herath, and K. Musiake, [19] J. R. Anderson, E. E. Hardy, J. T. Roach, and pp. 559–570, Mar. 2005.
BA mathematical model for flood loss R. E. Witmer, BA land use and land cover [34] S. Boyd and L. Vandenberghe, Convex
estimation,[ J. Hydrol., vol. 277, no. 1–2, classification system for use with remote Optimization. Cambridge, U.K.: Cambridge
pp. 24–49, 2003. sensor data,[ Geological Survey Professional Univ. Press, 2004.
[8] B. Merz, H. Kreibich, R. Schwarze, and Paper, no. 964, 1976.
[35] L. Debnath and P. Mikusinski, Introduction to
A. Thieken, BReview article: Assessment of [20] G. Büttner, J. Feranec, G. Jaffrain, L. Mari, Hilbert Spaces With Applications. New York:
economic flood damage,[ Natural Hazards G. Maucha, and T. Soukop, BThe CORINE Academic, 1990.
Earth Syst. Sci., vol. 10, pp. 1697–1724, 2010. land cover 2000 project,[ EARSeL eProc.,
[36] O. Chapelle, V. Vapnik, O. Bousquet, and
[9] P. Bates, BRemote sensing and flood vol. 3, pp. 331–346, 2004.
S. Mukherjee, BChoosing multiple parameters
inundation modelling,[ Hydrol. Process., [21] P. Prestininzi and A. Fiori, for support vector machines,[ Mach. Learn.,
vol. 18, no. 13, pp. 2593–2597, 2004. BA two-dimensional parabolic model vol. 46, no. 1–3, pp. 131–159, 2002.
[10] Y. Wang, J. D. Colby, and K. A. Mulcahy, for flood assessment,[ Italian J. Eng.
[37] G. Moser and S. B. Serpico, BContextual
BAn efficient method for mapping flood Geol. Environ., vol. 1, pp. 5–18, 2006.
remote-sensing image classification by
extent in a coastal flood plain using Landsat [22] B. Cappelaere, BAccurate diffusive wave support vector machines and Markov random
TM and DEM data,[ Int. J. Remote Sens., routing,[ J. Hydraul. Eng., vol. 123, no. 3, fields,[ in Proc. IEEE Geosci. Remote Sens.
vol. 23, no. 18, pp. 3681–3696, 2002. pp. 174–181, 1997. Symp., Honolulu, HI, 2010, pp. 3728–3731.
[11] K. Sado and M. M. Islam, BSatellite remote [23] A. Leopardi, E. Oliveri, and M. Greco, [38] E. Pardoux, Markov Processes and Applications:
sensing data analysis for flooded area and BTwo-dimensional modeling of floods to Algorithms, Networks, Genome and Finance.
weather study: Case study of Dhaka city, map risk prone areas,[ J. Water Res. Planning New York: Wiley, 2009.
Bangladesh,[ J. Hydraul. Eng., vol. 41, Manage., vol. 128, pp. 168–178, 2002.
pp. 945–950, 1997. [39] A. H. S. Solberg, T. Taxt, and A. K. Jain,
[24] V. Madhok and D. A. Landgrebe, BA process BA Markov random field model for
[12] T. Hahmann, S. Martinis, A. Twele, and model for remote sensing data analysis,[ IEEE classification of multisource satellite
A. Roth, BExtraction of water and flood areas Trans. Geosci. Remote Sens., vol. 40, no. 3, imagery,[ IEEE Trans. Geosci. Remote Sens.,
from SAR data,[ in Proc. 7th Eur. Conf. pp. 680–686, Mar. 2002. vol. 34, no. 1, pp. 100–113, Jan. 1996.
Synthetic Aperture Radar, Friedrichshafen, [25] V. N. Vapnik, The Nature of Statistical Learning
Germany, 2008. [40] R. Haralick and L. Shapiro, Computer and
Theory. New York: Springer-Verlag, 1995. Robot Vision. Reading, MA:
[13] D. C. Mason, R. Speck, B. Devereux, [26] C. M. Bishop, Neural Networks for Pattern Addison-Wesley, 1992–1993, vol. 1–2.
B. G. J.-P. Schumann, J. C. Neal, and Recognition. Oxford, U.K.: Oxford Univ.
P. D. Bates, BFlood detection in urban areas [41] F. Rottensteiner and C. Briese, BA new
Press, 1995. method for building extraction in urban areas
using TerraSAR-X,[ IEEE Trans. Geosci.
Remote Sens., vol. 48, no. 2, pp. 882–894, [27] J. C. Bezdek, Pattern Recognition With Fuzzy from high-resolution Lidar data,[ Int. Arch.
Feb. 2010. Objective Function Algorithms. Norwell, MA: Photogrammetry Remote Sens. Spatial Inf. Sci.,
Kluwer, 1981. vol. 34, no. 3/A, pp. 295–301, 2001.
[42] H. Sportouche, F. Tupin, and L. Denise, [58] G. Mercier and J. Inglada, BA new statistical [73] M. S. Horritt, D. C. Mason, and A. J. Luckman,
BBuilding extraction and 3D reconstruction similarity measure for change detection in BFlood boundary delineation from synthetic
in urban areas from high-resolution optical multitemporal SAR images and its extension aperture radar imagery using a statistical
and SAR imagery,[ in Proc. Joint to multiscale change analysis,[ IEEE Trans. active contour model,[ Int. J. Remote Sens.,
(IEEE-ISPRS-URS) Urban Remote Sens. Event, Geosci. Remote Sens., vol. 45, no. 2, vol. 22, no. 13, pp. 2489–2507, 2001.
Shanghai, China, 2009, DOI: 10.1109/URS. pp. 1432–1445, Feb. 2007. [74] N. M. Robertson and C. Tak, BAerial image
2009.5137746. [59] G. Mercier, G. Moser, and S. B. Serpico, segmentation for flood risk analysis,[ in Proc.
[43] P. Gamba and B. Houshmand, BDigital surface BConditional copulas for change detection IEEE Int. Conf. Image Process., Cairo, Egypt,
models and building extraction: A comparison in heterogeneous remote sensing images,[ 2009, pp. 597–600.
of IFSAR and LIDAR data,[ IEEE Trans. IEEE Trans. Geosci. Remote Sens., vol. 46, [75] S. Dellepiane and F. Fontana, BExtraction
Geosci. Remote Sens., vol. 38, no. 4, pt. 2, no. 5, pp. 1428–1441, May 2008. of intensity connectedness for image
pp. 1959–1968, Jul. 2000. [60] G. Moser and S. B. Serpico, BUnsupervised processing,[ Pattern Recognit. Lett., vol. 16,
[44] A. Huertas and R. Nevatia, BDetecting change detection from multichannel SAR pp. 313–324, 1995.
buildings in aerial images,[ Comput. Vis. data by Markovian data fusion,[ IEEE [76] J. K. Udupa and S. Samarasekera, BFuzzy
Graph. Image Process., vol. 41, no. 2, Trans. Geosci. Remote Sens., vol. 47, no. 7, connectedness and object definition: Theory,
pp. 131–152, 1988. pp. 2114–2128, Jul. 2009. algorithms, and applications in image
[45] C. S. Fraser, E. Baltsavias, and A. Gruen, [61] C. Carincotte, S. Derrode, and S. Bourennane, segmentation,[ Graph. Models Image Process.,
BProcessing of IKONOS imagery for BUnsupervised change detection on SAR vol. 58, pp. 246–261, 1996.
sub-metre 3D positioning and building images using fuzzy hidden Markov chains,[ [77] E. Angiati, BStatistic, frequency, and context
extraction,[ ISPRS J. Photogrammetry Remote IEEE Trans. Geosci. Remote Sens., vol. 44, based methods for satellite image analysis
Sens., vol. 56, pp. 177–194, 2002. no. 2, pp. 432–441, Feb. 2006. and data-fusion,[ Ph.D. dissertation, Univ.
[46] H. Oriot and A. Michel, BBuilding extraction [62] F. Bujor, E. Trouve, L. Valet, J.-M. Nicolas, Genoa, Genoa, Italy, Apr. 2011.
from stereoscopic aerial images,[ Appl. Opt., and J.-P. Rudant, BApplication of [78] G. Moser and S. B. Serpico, BClassification
vol. 43, no. 2, pp. 218–226, 2004. log-cumulants to the detection of of high-resolution images based on MRF
[47] M. Kass, A. Witkin, and D. Terzopolous, spatio-temporal discontinuities in fusion and multiscale segmentation,[ in
BSnakes: Active contour models,[ Int. J. multitemporal SAR images,[ IEEE Proc. IEEE Geosci. Remote Sens. Symp.,
Comput. Vis., vol. 1, no. 3, pp. 321–331, 1988. Trans. Geosci. Remote Sens., vol. 42, Boston, MA, 2008, vol. II, pp. 277–280.
no. 10, pp. 2073–2084, Oct. 2004.
[48] S. Ahmadi, M. J. V. Zoej, H. Ebadi, [79] NAZCA WebGIS. [Online]. Available: http://
H. A. Moghaddamb, and A. Mohammadzadeh, [63] F. Bovolo and L. Bruzzone, www.operaproject.it/index.php?option=
BAutomatic urban building boundary BA detail-preserving scale-driven approach com_content&view=article&id=71&Itemid=
extraction from high resolution aerial images to change detection in multitemporal SAR 97&lang=en
using an innovative model of active contours,[ images,[ IEEE Trans. Geosci. Remote Sens.,
[80] European Commission, Joint Research
Int. J. Appl. Earth Observat. Geoinf., vol. 12, vol. 43, no. 12, pp. 2963–2972, Dec. 2005.
Centre, Eur. Flood Alert Syst. Bull. [Online].
pp. 150–157, 2010. [64] S. Dellepiane, S. Monni, G. Bo, and G. Buck, 2009(6). Available: http://floods.jrc.ec.
[49] H. D. Cheng, X. H. Jiang, Y. Sun, and J. Wang, BSAR images and interferometric coherence europa.eu/files/efas/bulletins/EFAS_
BColor image segmentation: Advances and for flood monitoring,[ in Proc. IEEE Geosci. bulletin2009_6_Final.pdf
prospects,[ Pattern Recognit., vol. 34, no. 12, Remote Sens. Symp., Honolulu, HI, 2000,
[81] Q. Chen and P. Gong, BAutomatic variogram
pp. 2259–2281, 2001. vol. 6, pp. 2608–2610.
parameter extraction for textural classification
[50] R. Adams and L. Bischof, BSeeded region [65] R. Cossu, E. Schoepfer, P. Bally, and L. Fusco, of the panchromatic IKONOS imagery,[
growing,[ IEEE Trans. Pattern Anal. Mach. BNear real-time SAR-based processing to IEEE Trans. Geosci. Remote Sens., vol. 42,
Intell., vol. 16, no. 6, pp. 641–647, Jun. 1994. support flood monitoring,[ J. Real-Time Image no. 5, pp. 1106–1115, May 2004.
Process., vol. 4, no. 3, pp. 205–218, 2009.
[51] E. Angiati and S. Dellepiane, BIdentification [82] S. B. Serpico and G. Moser, BWeight
of roofs perimeter from aerial and satellite [66] M. Badji and S. Dautrebande, parameter optimization by the Ho-Kashyap
images,[ in Proc. 17th Int. Conf. Digital Signal BCharacterization of flood inundated areas algorithm in MRF models for supervised
Process., Corfu, Greece, 2011, DOI: 10.1109/ and delineation of poor drainage soil using image classification,[ IEEE Trans. Geosci.
ICDSP.2011.6005016. ERS-1 SAR imagery,[ Hydrol. Processes, Remote Sens., vol. 44, no. 12, pp. 3695–3705,
vol. 11, no. 10, pp. 1441–1450, 1997. Dec. 2006.
[52] G. Moser and S. B. Serpico, BUnsupervised
change detection with very-high resolution [67] L. Alparone, L. Facheris, S. Baronti, [83] G. Moser and S. B. Serpico, BAutomatic
SAR images by multiscale analysis and A. Garzelli, and F. Nencini, BFusion of parameter optimization for support vector
Markov random fields,[ in Proc. IEEE Geosci. multispectral and SAR images by intensity regression for land and sea surface
Remote Sens. Symp., Honolulu, HI, 2010, modulation,[ in Proc. 7th Int. Conf. Inf. Fusion, temperature estimation from remote-sensing
pp. 3802–3805. Stockholm, Sweden, 2004, pp. 637–643. data,[ IEEE Trans. Geosci. Remote Sens.,
[53] E. J. M. Rignot and J. J. van Zyl, BChange [68] K. Kotwal and S. Chaudhuri, BVisualization of vol. 47, no. 3, pp. 909–921, Mar. 2009.
detection techniques for ERS-1 SAR data,[ hyperspectral images using bilateral [84] G. Moser and S. Serpico, BA Markovian
IEEE Trans. Geosci. Remote Sens., vol. 31, no. 4, filtering,[ IEEE Trans. Geosci. Remote Sens., generalization of support vector machines
pp. 896–906, Jul. 1993. vol. 48, no. 5, pp. 2308–2316, May 2010. for contextual supervised classification of
[54] M. Molinier and Y. Rauste, BComparison and [69] S. R. Cloude and E. Pottier, BA review of hyperspectral images,[ in Proc. IEEE
evaluation of polarimetric change detection target decomposition theorems in radar Workshop Hyperspectral Image and Signal
techniques in aerial SAR data,[ in Proc. IEEE polarimetry,[ IEEE Trans. Geosci. Remote Processing, Reykjavı́k, Iceland, 2010, DOI:
Geosci. Remote Sens. Symp., Seoul, Korea, Sens., vol. 34, no. 2, pp. 498–518, Mar. 1996. 10.1109/WHISPERS.2010.5594967.
2005, pp. 2386–2389. [70] N. Chrysoulakis, I. Herlin, P. Prastacos, [85] R. A. Redner and H. F. Walker, BMixture
[55] G. Moser and S. B. Serpico, BGeneralized H. Yahia, J. Grazzini, and C. Cartalis, BAn densities, maximum likelihood, and the EM
minimum-error thresholding for improved algorithm for the detection of algorithm,[ SIAM Rev., vol. 26, pp. 195–239,
unsupervised change detection from SAR plumes caused by natural or technological 1984.
amplitude imagery,[ IEEE Trans. Geosci. hazards using AVHRR imagery,[ Remote Sens. [86] F. Melgani, G. Moser, and S. B. Serpico,
Remote Sens., vol. 44, no. 10, pp. 2972–2982, Environ., vol. 108, no. 4, pp. 393–406, 2007. BUnsupervised change detection methods
Oct. 2006. [71] S. Dellepiane and E. Angiati, BA new method for remote sensing images,[ Opt. Eng.,
[56] Y. Bazi, L. Bruzzone, and F. Melgani, for cross-normalization and multitemporal vol. 41, no. 12, pp. 3288–3297, 2002.
BAn unsupervised approach based on the visualization of SAR images for the detection [87] M. De Martino, S. B. Serpico, and C. Camurri,
generalized Gaussian model to automatic of flooded areas,[ IEEE Trans. Geosci. Remote BForest mapping by partially supervised
change detection in multitemporal SAR Sens., vol. 50, no. 7, pt. 2, pp. 2765–2779, classification applied to vegetation indexes,[
images,[ IEEE Trans. Geosci. Remote Sens., Jul. 2012. in Proc. IEEE Geosci. Remote Sens. Symp.,
vol. 43, no. 4, pp. 874–887, Apr. 2005. [72] G. Schumann, G. Di Baldassarre, and Seoul, Korea, 2005, vol. 5, pp. 3128–3131.
[57] K. Conradsen, A. A. Nielsen, J. Schou, and P. D. Bates, BThe utility of spaceborne [88] G. Moser and S. B. Serpico, BContextual
H. Skriver, BA test statistic in the complex radar to render flood inundation maps remote-sensing image classification by
Wishart distribution and its application to based on multialgorithm ensembles,[ support vector machines and Markov
change detection in polarimetric SAR data,[ IEEE Trans. Geosci. Remote Sens., vol. 47, random fields,[ IEEE Trans. Geosci. Remote
IEEE Trans. Geosci. Remote Sens., vol. 41, no. 1, no. 8, pp. 2801–2807, Aug. 2009. Sens., accepted for publication.
pp. 4–19, Jan. 2003.
ABOUT THE AUTHORS

Sebastiano B. Serpico (Fellow, IEEE) received the University of Genoa and the Scientific Director of the CIMA Research
Laurea degree in electronic engineering and the Foundation formed by the Italian Department of the Prime Minister Office
Doctorate from the University of Genoa, Genoa, Italy, for Civil Protection and the University of Genoa. He is an expert in
in 1982 and 1989, respectively. hydrology, with particular emphasis in flash-flood forecasting and
He is a Full Professor of Telecommunications at statistical analysis of the extremes, in hydrological and hydrometeoro-
the Faculty of Engineering, University of Genoa, logical applications of remote sensing techniques and data assimilation.
and the Head of the Signal Processing and Since 2000, he has been a Scientific Director and Principal Investigator of
Telecommunications laboratory of the Depart- research projects in the field of flood risk prediction and management
ment of Biophysical and Electronic Engineering, and civil protection. Since 2009, he has been a Scientific Secretary of the
University of Genoa. His current research interests section NH1 VHydro-Meteorological hazardsVof the European Geos-
include pattern recognition for remote sensing images and for biomed- ciences Union (EGU). He is an author and coauthor of 17 papers published
ical images. He has been the project manager of numerous research in international refereed journals, 35 short papers published in
contracts and an evaluator of project proposals for various programmes proceedings of international conferences, and 49 abstracts presented
of the European Union. He is author (or coauthor) of more than 200 at international conferences.
scientific articles published in journals and conference proceedings. Dr. Boni has been a referee for a number of international journals:
Prof. Serpico is the Chairman of the Institute of Advanced Studies in Journal of Hydrometeorology, Journal of Hydrology, Natural Hazards and
Information and Communication Technologies (ISICT). He is an Associate Earth System Sciences, International Journal of Remote Sensing, Vadose
Editor of the IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING. He was Zone Journal, since 2008.
a Guest Editor of two Special Issues of the IEEE TRANSACTIONS ON
GEOSCIENCE AND REMOTE SENSING on the subject of the BAnalysis of
Gabriele Moser (Member, IEEE) received the
Hyperspectral Image Data[ (July 2001) and on the subject BAdvances in
Laurea degree (M.Sc. equivalent; summa cum
Techniques for the Analysis of Remote Sensing Data[ (March 2005). From
laude) in telecommunications engineering and
1998 to 2002, he was the Chairman of a SPIE/EUROPTO series of
the Ph.D. degree in space sciences and engineer-
conferences on Signal and Image Processing for Remote Sensing.
ing from the University of Genoa, Genoa, Italy, in
2001 and 2005, respectively.
Silvana Dellepiane (Member, IEEE) graduated in Since 2010, he has been an Assistant Professor
1986 with honors from the University of Genoa, of Telecommunications at the University of Genoa.
Genoa, Italy. She received the Ph.D. degree in Since 2001, he has cooperated with the Signal
electronic engineering and computer science from Processing and Telecommunications research
the University of Genoa, in 1990. group of the Department of Biophysical and Electronic Engineering,
In 1992, she became a Researcher (Assistant University of Genoa. From January to March 2004, he was a visiting
Professor) in the Department of Biophysical and student at the Institut National de Recherche en Informatique et en
Electronic Engineering (DIBE), University of Genoa, Automatique (INRIA), Sophia Antipolis, France. His research activity is
Genoa, Italy, where she is an Associate Professor in focused on the development of image processing and pattern recognition
the ING-INF/03 area. She has taught signal theory methodologies for remote sensing data interpretation. His current
and pattern recognition. Currently, she is Professor of Electrical Commu- research interests include contextual classification, multitemporal image
nications, Statistical Methods, Signal and Image Processing and Recognition, classification and change detection, synthetic aperture radar (SAR) data
in the courses of telecommunications engineering and bioengineering. She analysis, hyperspectral image classification, and geo/biophysical param-
is responsible for the NUMIP research area of the Signal Processing and eter estimation.
Telecommunications (SP&T) laboratory at DIBE. She has gained wide Dr. Moser has been a reviewer for several international journals. He has
scientific and technical experience in multidimensional data processing. been an Associate Editor of the IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
Her main research interests include the use of context and fuzzy systems for and Pattern Recognition Letters since 2008 and 2011, respectively.
multidimensional data processing, segmentation, supervised methods for
the processing of remote sensing synthetic aperture radar (SAR) images, and
Elena Angiati (Student Member, IEEE) was born in
nonlinear adaptive processing of digital signals. Her application domains
Novi Ligure, Italy, in 1981. She received the Laurea
are, mainly, telemedicine and remote sensing.
degree (B.Sc.) and the Laurea Magistrale degree
Prof. Dellepiane is a reviewer for various journals. She was invited at
(M.Sc.) in telecommunications engineering from the
some international conferences and schools for tutorials and lessons.
University of Genoa, Genoa, Italy, in 2004 and 2006,
She has participated, at the organization and scientific levels, in the
respectively. She developed her M.Sc. thesis on
research activities concerning various National Research Council (CNR),
unsupervised change detection in remote sensing
Italian Space Agency (ASI), Italian Minister of Education, University, and
images by multiscale Markov random fields. She
Research (MIUR), and European Union (EU) projects. She is a member of
received the Ph.D. degree in space sciences and
the Technical Committee BBio Imaging and Signal Processing[ of the IEEE
engineering from the University of Genoa in 2011.
Signal Processing Society.
Since January 2007, she has been cooperating with the NUmerical
IMage Processing (NUMIP) group at the Department of Biophysical and
Giorgio Boni received the Laurea degree in Electronic Engineering (DIBE) in research activities concerning image
hydraulic engineering from the University of Genoa, processing and pattern recognition algorithms. Her research topics are
Genoa, Italy, and the Ph.D. degree in hydrodynamics pattern recognition and image processing in remote sensing for
from the University of Padua, Padua, Italy. monitoring and security. Her works are related with damage assessment
He spent 1998 at the Massachusetts Institute of for applications such as flood, tsunami, and earthquake monitoring. The
technology (MIT), Cambridge, as a Postdoctoral principal remote sensing image processing and analysis investigated are
Researcher under a CNR-MIT grant for research segmentation of radar and optical data, coastline detection, despeckling,
activity on satellite data assimilation for hydro- and cross normalization of multitemporal images. Polarimetric and
logical applications. Currently, he is an Assistant interferometric SAR data are used for data processing. She is coauthor of
Professor of Environmental Engineering at the some papers presented at international conferences and journals.
Roberto Rudari received the Laurea degree

(M.Sc. equivalent; summa cum laude) in hydraulic
civil engineering from the University of Genoa,
Genoa, Italy, in 1998 and the Ph.D. degree in
hydraulic engineering from the University of
Padua, Padua, Italy, in 2002.
He was a visiting Ph.D. at Parsons Lab,
Massachusetts Institute of Technology (MIT), Cam-
bridge, in 2000–2001. Currently, he is a Research-
er at CIMA Foundation. He is an expert in basin
geomorphology, climate and weather predictability, continuous and
distributed hydrologic modeling assisted by satellite and remotely
sensed information. His other areas of interest are the characterization
of joint probability distributions of land effects and vulnerability
estimation to flood events, and the improvement of combined hydro-
meteorological forecast systems based on probabilistic concepts. He is
the Programmer of the MIKEDriFt rainfall-runoff model commercialized
by DHI. He is an author and coauthor of more than 80 publications in
international/national refereed journals and conference proceedings. He
was a consultant for the WMO Associated Program on Flood Manage-
ment and Consultant for the UNISDR Global Assessment Report for the
Global Flood Model development.
Dr. Rudari was a Guest Editor of Hydrology and Earth System Sciences
(HESS).

Information Extraction From Remote Sensing Images For Flood Monitoring and Damage Evaluation

Uploaded by

Copyright:

Available Formats

Information Extraction From Remote Sensing Images For Flood Monitoring and Damage Evaluation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Information Extraction From Remote Sensing Images For Flood Monitoring and Damage Evaluation

Uploaded by

Copyright:

Available Formats

INVITED

Information Extraction From

A. Shkodër, Albania, 2010 international flood relief organizations, the floodplain

ABOUT THE AUTHORS

Roberto Rudari received the Laurea degree

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.