Ionela Craciun 3 [627272]

area. The RAS Mapper tool can produce a raster through spatial interpolation of the water
level outputs from the 1D -RAS simulations. By subtracting the elevations of the terrain in raster
file format from the spatial interpolation of water level, it is possible to get the water depth raster.
This raster (Fig. 17) represents the hazard map and is one of the input data used by the damage
model.

3.3.2. Exposure and Vulnerability analysis
The exposure analysis refers to the elements at risk, which are obtai ned by overlapping the
information about assets present in the area (i.e. land -use map and roads map) with hazard
information (i.e. flood extent and water depth) (Fig. 18). In this way the elements that are affected
by floods are identified. The vulnerabil ity analysis shows how these elements will be affected,
depending on their characteristics (i.e. type of element, asset value).

Figure 18. Identification of the exposed elements for the city of Căianu Mic
(adapted after Albano et al., 2017a)

At meso -scale, the assets with similar characteristics are usually in an aggregated form (i.e.
land use maps); in this case study the CLC database were used. For a more accurate analysis of

the urban category, the buildings were digitized and introduced in the COR INE land use map (Fig.
19). The urban category is the most affected class in the damage assessment, being responsible of
around 80% of the total damage (de Moel and Aerts, 2011). Therefore, the accurate representation
of the buildings is essential in order to obtain reliable results.

Fig. 19 . The improved CORINE land use map developed for Ilișua basin

Afterwards, the asset v alue was determined for each CORINE land use class. The data
regarding the value of the exposed assets were provided by the local a uthorities of the respective
communities. Due to the fact that the event took place in 2006, and the documentation regarding
this event was made available in 2012, the value of the assets needs to be corrected with inflation.
In order to achieve this, the GDP deflator was used (GDP -Deflator, 2018 ), which is a demonstrated
tool used to quantify price inflation. For the selected case study, the inflation adjusted values
(euros/m2) for each CORINE land use class exposed to flooding, are presented in Table 7.

Table 7. Land use classes reclassification and site -specific assets value
for the selected case study (Albano et al., 2017a)
Code Description CORINE land use classes Assets value (euro/m²)
112 Discontinuous Urban Fabric 108
122 Road and rail networks and associated
land 16
211 Non-irrigated Arable Land 0.08
231 Pastures 0.06
242 Complex Cultivation Patterns 0.16
243 Land Principally Occupied by Agriculture
with significant areas of natural
vegetation 0.13
311 Broad -Leaved Forest 0.04
312 Coniferous Forest 0.04
313 Mixed Forest 0.04
321 Natural Grassland 0.06
324 Transitional Woodland -shrub 0.06
331 Beaches Dunes and Sand Plains 0.06

For each land use class affected by flooding a corresponding depth -damage function is
associated. For the selected case study, the JRC depth -damage functions (Huizinga, 2007)
developed for different European Countries were used. The JRC developed this functions with the
aim of offering a harmonised approach across EU for the estimation of direct damages pr oduced
by floods. For each of the considered countries, depth -damage functions were p rovided for 5 land
use classes: residential buildings, industry, commerce, infrastructure, and agriculture.
In order to use these functions with the CORINE land use map fo r this study, the functions
were adjusted by associating the depth -damage curves of Huizinga 2007 to the land use class that
was considered to be most comparable, as shown in Table 8. For example, for the land use class
"discontinuous urban fabric", the "r esidential building" class from JRC was attributed, along with
its corresponding depth -damage functions; and so on.
Major infrastructure elements such as roads, were inputted manually, i.e. by digitizing the
roads from the topographic map provided by the R omanian Waters National Administration of
Someș -Tisa Water Branch, and were assigned to the “Road and rail networks and associated land”
class and their corresponding asset value provided by the Romanian Waters National
Administration of Someș -Tisa Water B ranch.

Table 8. Correspondences between JRC depth damage curves and
CORINE land use classes (Albano et al., 2017a)
CORINE land use classes JRC depth -damage curves
classification on the basis of
economic sectors
Discontinuous Urban Fabric Residential buildings
Road and rail networks and associated land Infrastructure
Non-irrigated Arable Land Agricultural
Pastures Agricultural
Complex Cultivation Patterns Agricultural
Land Principally Occupied by Agriculture
with significant areas of natural vegetation Agricultural
Broad -Leaved Forest Agricultural
Coniferous Forest Agricultural
Mixed Forest Agricultural
Natural Grassland Agricultural
Transitional Woodland -shrub Agricultural
Beaches Dunes and Sand Plains Agricultural

3.3.3. Damage analysis
The quantitative estimation of damages is usually done by using damage models. This
models commonly use depth -damage functions that combine hazard with vulnerability
information. For the flood event considered in this study, the damages were calcu lated using the
free and open source tool FloodRisk (Mancusi et al., 2015) developed in QGIS platform. The tool
calculates both direct tangible damage (damage to structure) and direct intangible damage (loss of
life) caused by floods and visualizes them in form of tables, graphs and maps. Maps give a direct
and strong impression of the spatial distribution of the flood risk, providing essential information
to stakeholders (Albano et al., 2015b). For the estimation of direct damages in this study, the
follow ing input data were used: the water depth map developed in the hazard analysis step, the
improved Corine land use map containing the asset value for each class and the JRC depth -damage
functions.

The FloodRisk plugin provides a dataset containing default f unctions of different models
collected from literature (e.g. Hazus Model, RAM, DSM, etc.). These functions provided by the
plugin are associated with the Urban Atlas land use classification. However, in this study the
damages were calculated using the JRC functions developed for various European Countries (i.e.,
Belgium, Czech Republic, Germany, Netherlands, Norway, Switzerland, UK) and the CLC
classification. For this purpose, the data provided in the JRC report (Huizinga, 2007) were
collected and processe d in order to substitute the standard input data available in the FloodRisk
plugin. Therefore, the damages were calculated using seven different functions introduced in the
plugin (Fig. 20).

Fig. 20. The JRC functions for different countries, introduced in the FloodRisk plugin

In this way an uncertainty analysis regarding the functions’ transferability in space was
possible. The aim of this analysis is to determine the effect on the results when applying different
damage functions from other counties in another context than the one for which they were
developed.
For each land use class described in Table 8, the damage was calculated using the JRC
depth -damage functions and the asset value described above. These data were subsequently
compared with the real flood event of June 2006 in Ilișua Basin. Given the fact that the surveyed
data were collected for only 3 classes (urban, roads and agriculture) the results of the simulations
have been grouped according to these classes in order to ensure a consistent comparison between
the calculated damage assessment and real damage surveyed by authorities after the June 2006
flood.

In Table 9, the damages calculated and the real surveyed damages for the Ilișua Basin,
reported in the Coverage and Risk Assessment Plan, 2006 of the County Committee for Emergency
Situations Bistr ița-Năsăud, are presented. The simulations with the JRC functions exhibit higher
urban damages than flood losses in agriculture. For infrastructure however, which represents a
significant percentage of the total damage, the JRC functions strongly underesti mate the
corresponding real, surveyed losses. Overall, the model systematically overestimates the losses for
the urban and agricultural classes, highlighting the fact the estimations of infrastructure damage is
less developed and can be a large source of u ncertainty in the modelling process.

Table 9. Results of damage calculation using different JRC depth -damage functions
and the reported damages (Albano et al., 2017a)
Damages (MEuro)

urban roads agriculture Total
Belgium JRC 8.3 0.09 0.85 9.2
Czech Republic
JRC 5 0.14 0.69 5.8
Germany JRC 3.8 0.16 0.31 4.3
Netherlands JRC 3.2 0.10 0.81 4.2
Norway JRC 18.4 0.26 0.69 19.4
Switzerland JRC 13.8 0.13 0.58 14.6
UK JRC 32.4 0.13 1 33.6
Surveyed Ilișua
Basin 0.29 0.6 0.12 1.1

Fig. 21. Graphic representation of the results of damage calculation using different JRC
depth -damage functions and the reported damages (adapted after Albano et al., 2017a)

3.4. Uncertainty analysis
The results are characterized by a large variability (Table 10 and Fig. 21), therefore an
uncertainty analysis regarding the damage functions’ transferability was conducted. The damage
values that were simulated using the FloodRisk tool and the reported damages of the 2006 flood
of Ilișua Catchment were compared. The reported damages were provided by the County
Committee for Emergency Situations Bistrița -Năsăud for Ilișua Basin for the 2006 flood.
Each damage value estimated for each of the JRC depth -damage functions was divided by
the reported damage value, resulting a ratio t hat was considered as the relative error. The relative
error, indicates error rate of the simulated damage respect to the surveyed damage. The results
showed that the Netherlands -JRC functions offers the lowest relative error (37%), while large
errors are induced by the Norway -JRC, Switzerland -JRC and UK -JRC functions (Table 10).
However, a general overestimation of the damage results when using the JRC functions
can be observed, these findings being in agreement to other studies from literature (e.g. Merz et
al., 2004; Jongman et al., 2012). The large variation in the results and the large relative errors are
mostly related to the shape of depth -damage functions (Jongman et al., 2012, Messner at al., 2007,
de Moel and Aerts, 2011). For example, a great var iation in the results can be observed between 0510152025303540Damages for different land use classes
Urban Roads AgricultureMEuro

UK-JRC and Netherlands -JRC functions. This can be partially explained by analysing the shape
of this functions. In the case of the UK functions for the urban land -use class the damage factor
reaches 100% for a water depth of 3 m, while in the case of Netherlands functions, the damage
factor is around 20% for a water depth of 3 m (Fig. 22).

Fig. 22. Differences in the shape of damage functions (Netherlands -JRC and UK -JRC)

Table 10. Relative error and function uncertainty determination (Albano et al., 2017a)
Relative
error Function
uncertainty
Total reported (MEuro) 1.1
Total
calculated
(MEuro) Belgium -JRC 9.2 8.37 2.21
Czech Republic –
JRC 5.8 5.28 1.39
Germany -JRC 4.3 3.92 1.03
Netherlands -JRC 4.2 3.79 –
Norway -JRC 19.4 17.53 4.63
Switzerland -JRC 14.6 13.18 3.48
UK-JRC 33.6 30.35 8.01

In order to identify the effect of the depth -damage functions and to quantify uncertainties
associated with the modelling of flood damage, the results calculated with the FloodRisk model

were analysed and compared. For this purpose, each calculated damage value was divided by the
lowest calculated value, the results being expressed in term of "function uncertainty" (Table 10).
In the present case the lowest value is the one calculated through the JRC functions for
Netherlands, therefore, all the other estim ated values were divided by this one. The values of this
factor varies between 1.03 and 4.63 for all the functions, excepting the outlier value for the UK
JRC function which is equal to 8.01, i.e. the difference between the highest and lowest damage
estima tes.
These findings are consistent with others in the research field, such as the study done by
de Moel and Aerts, 2011, a function uncertainty factor of 5.6, the study by Cammerer et al., 2013,
which reported a factor of 2.3 for the cases when the models are thoroughly validated, the study
by Jongman et al., 2012, based on two study areas reporting a factor of 3.7 and one of 10.5, and
finally, Scorzini and Frank, 2015, for which this factor varies quite extensively between 1.1 and
14.8.
The results also sh owed that the transferability of damage functions in space represents a
major problem in flood risk assessment, lea ding to great uncertainties which need to be
characterized and reduced when possible. This can be partially achieved by establishing a coherent
framework for data collection and evaluation.

3.5. Conclusions
In this case study a quantitative approach for flood damage assessment was presented. The
aim was to create a more comprehensive methodology that includes all flood risk factors in the
analysis and in this way to offer a better understanding of the risks that can occur and help in the
decision -making process.
Furthermore, the qu antitative economic damages estimation can be used in flood risk
estimation, if more flood hazard scenarios are used. This study is guided by the need for the
development of new methods and tools that can efficiently support stakeholders in their
complianc e with the Flood Directive2007/60/EC. In particular, this approach could be the initial
step to develop a rationale procedure to achieve objective results on which stakeholders can base
their decisions for prioritizing investments, and performing cost -bene fit analyses of mitigation
alternatives.

Moreover, the attempts in the past years to develop pan -Europeans damage models requires
more knowledge and more attention to the uncertainties regarding these models. Damage estimates
are influenced by many factors , first in the process being the input data. Therefore, a homogeneous
database across Europe is needed. Another important factor is the shape of depth -damage functions
which can lead to great uncertainties.
Another objective of this case study was to perfo rm an uncertainty analysis using the
damage functions of Huizinga, 2007. It was found that more attention should be given to the shape
of the depth -damage functions used at EU level; these ones having a large effect on the damage
results. The calculated da mages for the Ilișua catchment have a relative error which ranges from
37% to 300%, all the values being larger than the reported ones. The function uncertainty factor
obtained varies between 1.03 and 4.63. This variation on the results can be attributed t o the shape
of the functions, the data used for the construction of these functions like synthetic or empirical
data, and weather or not the emergency service or stock costs are included.
The results show that overall applicability and transferability of depth -damage curves to
other geographical regions is still a major gap in current flood damage modelling, but the
quantification of the uncertainties and its communication to stakeholders is the first step for the
maximization of quantitative risk approach effectiveness, towards flood risk management
objectives of the Flood Directive, ensuring that risk information is robust, credible and transparent.

4. Case study 2. GIS tools for large -scale analy sis of direct economic flood damage. The case
study of Romania
4.1. Introduction
This increased growth and development of tools and models that can be used for flood
hazard and flood consequence analysis proved to be particularly useful for large -scale analysis that
has gained more attention in the past years. The flood risk assessment at large scale offers support
for national and global policies and are used in the prioritisation of investment at national level,
the principal stakeholders being the governments and the reinsurance companies. The governme nts
is using the hazard information in order to implement flood damage reduction measures, for
emergency operations and for land use policies. On the other hand this information is useful for
insurance companies in order to establish the price for flood ri sk insurance.
The macro -scale analysis refers to large areas such as national or international scale.
Usually there are required less detail data with a low resolution. The land use data are aggregated
and one general asset value is used for each class. C LC is a good example of database that can be
used at macro -scale analysis (Messner et al., 2007; de Moel et al., 2015).
However, the approaches at large scales present gaps given by the lack of consistent data
over large areas and the lack of high resolut ion data. The inconsistency of data is given by the use
of different information, models and methods in order to assess the flood risk for different areas.
Therefore, in order to have consistency in the results one meth od must be applied for the entire
area. This can limit the choice of the methods that can be used, simplified approached being
needed. Furthermore the level of detail of this analysis is different from one area to another,
depending on the availability of data . (de Moel et al., 2015; Alfieri et al.,2014; De Roo et al.,
2006). In many areas there are ungauged basins that provide little information and therefore the
hydraulic modelling is rather challenging and time -consuming. In this context more simplistic
approaches were developed, using avai lable data, such as the DEM, in order to extract the
necessary informations for the analysis of flood hazard. Moreover, the land use data available for
large areas have a course resolution, which can induce uncertainties in the results. However, in the
past years, many studies focused on using high -resolution EO datasets in order to extract land use
information.

The variability in space and time of the risk components (hazard, exposure and
vulnerability) present s another problem that must be considered wh en speaking of large scale
analysis.
In this work, a quantitative flood damage assessment methodology for data -scarce
environmnets is proposed, using data that are easily available . The flood damage analysis was done
for the entire Romanian territory, for a return period of 100 years, combining EO high -resolution
data (30 m resolution) and free and open -source GIS tools . Around 98% of the Romanian territory
is situated in the Lower Danube River Basin, therefore the study was focused on the Lower Danube
Rive r and its tributaries. To the authors’ knowledge it represents the first large scale flood damage
analysis with a resolution of 30 m.
The main goal of this chapter is to offer a preliminary simplified approach and to provide
solutions for the current principals gaps in the flood risk assessment filed: limited availability of
data and quality/resolution of the data. The analysis is done at the grid resolution of 30 m using
innovative tools and methods.
4.2. Study area
The Danube river (Fig. 23), the second longest in Europe (approximately 2.860 km),
crosses on its way 10 countries (Germany, Austria, Slovakia, Hungary, Croatia, Romania,
Bulgaria, Republic of Moldova and Ukraine), from its headwaters in the Black Forest mountains
(Germany) and flowing into the Black Sea. The Danube river is divaded into three sections (FRMP,
2015) :
– The Upper Danube, from its headwaters to Bratislava city (Devin gate);
– The Middle D anube, from Bratislava to Bazias;
– The Lower Danube, from the entrance in Romania, at Bazias, to the Black Sea.
This study focuses on the Lower Danube section, which is situated on the Romanian
territory and has a length of 1.075 km. The Lower Danube is the most important regarding the
discharge (flow) and the navigation in this area and the hydrographic network includes most of
Romania’s rivers.
Adopting an integrated flood risk management is important particularly when speaking
about large rivers such a s the Danube, the potential of damage occurrence being higher due to the
large number of population and economic assets situated in the flood prone areas. The extreme
events which took place in the Danube basin highlighted the importance of Flood Directive

(2007/60/CE) implementation as well as the importance of adopting a unitary and harmonized
methodology which includes the entire Danube basin.

Fig. 23. Study area; a. Danube River – Europe location;
b. The Lower Danube – Romania

Several extreme events that took place in the Danube basin are described in detail by
Mikhailova et al., 2012 in a study regarding the impact of climate change in the Danube basin. The
floods from August 2002, caused by extreme precipitation, have had majo r consequences in the
Upper and Middle Danube, the consequences in the Lower Danube being reduced due to the
regulation of water capacity at Iron Gate I dam. The floods from 2006 have been caused by extreme
precipitation and the melt of the large snow accu mulation, having major consequences on the entire

Danube basin and particularly on the Lower Danube, Romania being one of the most affected
countries (Mikhailova et al., 2012).
These events highlighted the fact that the existing protection measures have li mitations,
their improvement being necessary as well as the improvement of population awareness policies
regarding the flood risk (European Union, 2018).
A first attempt to realize a harmonized methodology on the entire Danube basin was done
by the Danube FLOODRISK cross -border project, initiated by the Ministry of Environment and
Waters Management in Romania (Danube -Floodrisk, 2018) . The main objectives of this project
were: the awareness of the population situated in the Danube basin regarding the flood e xposure,
rising the population and assets protection level, priori tization of protection measures and decision
support. The main results are represented by the hazard and risk maps as well as the potential flood
damage calculation. The maps are represented on a scale of 1: 100.000. The hazard maps contain
the flood area and the water depth (divided in 4 classes) and the risk maps contain the flood risk
areas (high, medium and low risk), for each risk class the potential damage being presented.
Furthermore, these damages are differentiated on land use types, the database that was used being
CLC . For the damage calculation, the damage functions were applied for each land use class
(European Union, 2018 ; FRMP, 2015).

4.3. Methodology
The proposed approach for the flood damage assessment is following three steps: hazard
analysis, exposure analysis and damage analysis (Fig. 24).
Regarding the hazard analysis, a simplified methodology for water depth estimation based
on the basin’s geomorphology was proposed. Th is method is using the DEM as the main input in
order to extract the necessary geomorphic features for the determination of the flood area and water
depth. This approach offer rapid and accurate results and can be applied in data -scarce areas. A
GIS-base m odel was applied in order to obtain the water depth using the Geomorphic Flood Index
(GFI) (Samela et al., 2017).
For the exposu re analysis, the land use map for the study area was developed using the
Landsat 8 satellite images with a resolution of 30 m. A machine -learning algorithm was used for
the identification of 7 land use classes from the multi -spectral satellite images. In order to train the
algorithm, the Urban Atlas land use database was used as training data. The aim is to obtain high

resolution l and use data over large areas, which can improve the accuracy of the flood damage
assessment.
For the damage analysis the FloodRisk model and the JRC depth -damage functions were
used. The damages were calculated for four scenarios using the existing hazard map and land use
data (JRC water depth map and CLC ) and the hazard and land use maps developed in this study.

Fig. 24. Schematic representation of the applied methodology

Giving the fact that the hazard analysis method must be applied to a certain hydrographic
basin, the study area was divided into 5 sub -basins in order to be analysed:
– The Tisa sub -basin
– The Danube sector between upstream of the confluence with Timis river and
upstream of the confluence with Siret river, including the tributary sub-basins
– The Siret sub -basin
– The Danube sector between upstream of the confluence with Prut river and the
Black See, including the tributary sub -basins
– The Danube Delta

4.3.1. Hazard analysis
In the hazard analysis process the flood prone area s are identified and the water depth for
this areas is calculated. For this purpose, a GIS -based model that is using the geomorphology of
the basin was applied (Samela et al., 2017). This model allows the flood -prone area delineation
and water depth calcul ation over large areas by using limited data. This is particularly useful for
the analysis in areas where the necessary data for hydraulic simulations are not available or this
simulation would be time consuming and expensive. The geomorphological features of the area of
interest are extracted using a DEM and based on this information the flood -prone areas are
delineated using a linear binary classifier. The DEM is the most important input data, the entire
analysis being based on the information extracted f rom it, therefore its quality and resolution will
have a great influence on the results.
The linear binary classifiers are basically separating the map in two classes: flooded areas
(assigning the code 1) and not flooded areas (assigning the code 0). For this purpose, a threshold
value for each of the morphological features is applied in order to separate the two classes. The
classifier must be trained in order to calibrate the threshold value. This step is done by using an
existing standard detailed haza rd map that can be developed for small areas using hydraulic
simulations.
Some of the morphologic features that can characterize the flood hazard exposure are
presented below (Manfreda et al., 2015; Samela et al., 2017):
“Single features
– Upslope contribu ting area, A(m2); upslope portion of the watershed that contributes to
water runoff to the point under exam.
– Surface curvature ∇2H [-]; Laplacian of the elevation.
– Local slope, S [ -]; maximum slope among the eight possible flow directions that
connect the cell under exam to the adjacent cells.
– Flow distance to the nearest stream, D [m]; hydrologic distance from the location under
exam to the nearest element of the reference drainage network.
– Elevation difference to the nearest stream, H [m]; difference between the elevation of
the cell under exam and the elevation of the final point of the above -identified path.
Composite indices

– Ln (h l/H); this index compares in each point of the basin a variable water depth
hl with the elevation difference H. h l is calculated for each basin locationas a function
of the local contributing area, A l.
– Geomorphic Flood Index (GFI), ln(h r/H); this index compares in each point of the basin
a variable water depth h r with the elevation difference H. h r is computed as a function
of the contributing area Ar in the nearest point of the drainage network hydrologically
connected to the point under exam”.
Based on the literature findings, the Geomorphic Flood Index (GFI) provides high accuracy
results, being the most suitable for flood -prone area delineation. In the study by Samela et al.,
2017, the GFI showed a good performance in the correct identification of flooded areas (the rate
of true positive RTP results being up to 89%). It also has a limited variabi lity, not being influenced
by changes in the input data (e.g. resolution, topography of the area), therefore it offers a good
transferability in other basins. In order to offer a simple way of implementation of this method a
new plugin, named Geomorphic Fl ood Area tool (GFA tool), was developed in QGIS (Samela et
al., 2018).
The GFA tool which is based on the GFI methodology was used in this study for water –
depth mapping for a return period of 100 years, over the entire territory of Romania. Samela et al.,
2018 performed a flood extend analysis for the territory of Romania using the GFI methodology,
therefore in this study this information and methodology will be used in order to further perform
a water depth analysis for this area.

4.3.1.1. The GFI metho d
The GFI method combines the geomorphological information extracted from a digital
description of the earth’s surface – a Digital Elevation Model (DEM) – with existing flood hazard
information from smaller areas in order to extend this information over l arge areas. The
Geomorphic Flood Index (GFI) is defined as the natural logarithm of the ratio between a variable
water depth hr and the elevation difference H (Samela et al., 2018).
𝐺𝐹𝐼 =𝑙𝑛(ℎ𝑟
𝐻) (1)
 hr represents the river stage in the point of the river network closest to the one under exam,
which is considered as the most probable source of flood hazard;

 the parameter H represents the difference in elevation between the two above -mentioned
points.
The estimation of the river stage hr is derived as a function of the upslope contributing area
𝐴𝑟 using a hydraulic scaling relationship (see eq. 2) (Samela et al., 2018):

ℎ𝑟≈𝑎𝐴𝑟𝑛 (2)

For the study case of Romania, since paired values ( ℎ𝑟,𝐴𝑟) were not available to calibrate
the hydraulic scaling relationship, Samela et al.,2018 estimated the the exponent n using literature
values (Engelund and Hansen, 1967; Ibbitt, 1997; Ibbitt et al., 1998; Knighton, 2014; Leopold et
al., 1965; Leopold and Maddock, 195 3; Li, 1974; McKerchar et al., 1998; Park, 1977; Rodriguez –
Iturbe and Rinaldo, 1997; Smith, 1974; Whiting et al., 1999) . The mean value estimated is
n=0.3544. Thus, the hr was calculated by ‘neglecting’ a and assuming that the 𝑙𝑛 (𝑎) will be
included in the calibrated threshold.
In order to calibrate the classifier, the pan‐European flood hazard maps ( Annex 1 ) derived
by Alfieri et al. 2014 for a return period of 100 years was used. The linear boundary between the
two classes of the binary classifica tion (flood -prone areas and areas not prone to floods), must be
first calibrated for a training area that, according to Samela et al., 2017, must be equal or larger
than 2% of the basin of interest. In this way the optimat threshold τ is calibrated.
This parameter has to be calibrated iteratively by generating binary maps of potential flood –
prone areas that have been compared with the standard flood map assumed as “gold standard
truth”. From the comparison, the following performance measures can be assess ed: the rates of
true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The value
of τ that minimized the sum of the overestimation (R FP, false positive rate) and the underestimation
error (R FN, false negative rate), assi gning equal weights to the two errors, can be considered as
optimal one and applied to the whole basin.
The results of study done by Samela et al., 2018 are presented in Table 11, showing that
the GFI classifier has a good performance.

Table 11 . Values of the optimal thresholds and relative performance measures obtained in
calibrating the classifier over the investigated basins (Samela et al., 2018)
Basin
number % training
area τ RTP RFN RTN RFP Obj.
function
= R FN+R FP AUC
1 14% -0.23 81.9% 18.1% 64.6% 35.4% 53.4% 0.81
2 12% -0.24 85.0% 15.0% 83.3% 16.7% 31.7% 0.92
3 14% -0.12 90.6% 9.4% 36.9% 63.1% 72.5% 0.63
4 6% -0.26 82.6% 17.4% 84.0% 16.0% 33.4% 0.92
5 5% -0.31 91.1% 8.9% 91.7% 8.3% 17.2% 0.97

4.3.1.2. Estimate of the water inundation depth using the Geomorphic Flood Index.
Methodology.
In this large -scale analysis, it was tried to exploit at the most the information derived from
the analysis of the basin morphology, in order to derive an instantaneous estimate of the maximum
inundation depth within the flood -prone areas previously identified in the Romanian territory by
Samela et al., 2018.
As mentioned in the previous section, the hr was calculated by ‘neglecting’ a and assuming
that the 𝑙𝑛 (𝑎) will be included in the calibrated threshold. In this sense, the GFI may be expressed
as:
𝐺𝐹𝐼 =𝑙𝑛(ℎ𝑟
𝑎𝐻) (2)
Nevertheless, after the boundary of decision between flood -prone areas and areas not prone
to floods that characterizes each specific basin was calibrated, along that bounda ry of decision the
below information were obtained:
𝑙𝑛(ℎ𝑟
𝑎𝐻)=𝜏 and ℎ𝑟
𝐻=1 (3)
(1
𝑎)=exp (𝜏) → 𝑎= (1
exp (𝜏)) (4)
In this way, the values of the initially estimated river stage ℎ𝑟 ′=𝐴𝑟𝑛 were corrected:
ℎ𝑟=ℎ𝑟′∗𝑎=ℎ𝑟′
exp (𝜏) (5)
Therefore, for each river network element resulted an estimate hr. From the geomorphic
analysis previously performed, there were obtained information about all the upslope basin
locations that drain into those elements and the difference in elevation that separates the locations

from the closest river. Therefore, within the delineated flood -prone areas, finally there were used
the hr values to estimate the water depth (WD) as follows:
WD = h r – H (6)
Based on this methodology and using the GFA tool the water depth map for the entire
territory of Romania was developed. The input data required by the GFA tool are: a. DEM,
depressionless DEM, flow direction and flow accumulation (for the GFI index determination); b.
a standard flood map. The framework of th e methododlogy is presented in Fig. 25.

Fig. 25. Schematic representation of the hazard analysis methodology

The main input was represented by the the SRTM 1 Arc -Second Global elevation data
obtained from USGS. These DEMs provides worldwide coverage of void filled data at a resolution
of 1 arc -second (30 meters). The resolution of the DEM is an important factor, a low resolution
inducing error that may affect the analysis. Therefore, it is recommended the use of a filled DEM,
which was obtained using GIS tools. With this procedure, the sinks from the original DEM are
filled in order to obtain a proper river network topology.
Using the obtained filled DEM the flow direction and flow accumulation were calculated
(Annex 2 ). The flow direction shows the di rection in which the water of one cell will flow and
subsequently the slop is calculated. Further, the flow accumulation is obtained, showing the flow
that is accumulated in each cell and in this way the stream can be determined.
The obtained hazard map ( Fig. 26) contains the extend of the flood and the water depth for
the entire territory including the second and minor rivers which usually are not considered in large
scale analysis. As is highlighted in Fig. 27, the hazard map obtained using the GFA tool can
identify and analyse all the streams of the basin. Therefore, it is offering a more accurate hazard
map, comparing with the existing JRC pan -European map which is considering just the main
streams. The main advantages of this approach are represented b y the fact that it is a simple method
that does not requires detail data and costly resources to be applied, it can be used for areas with
limited availability of data and it can be applied for large scale analysis. Furthermore, the analysis
is done using a resolution of 30 m which improve the accuracy of the results.
It must be mentioned that this kinds of procedures cannot replace a comprehensive
hydrological -hydraulic study, since they have several limitations, among them, it cannot take into
account an thropogenic modifications (e.g. soil sealing and presence of floods defence structures)
and neither can describe the reduction of the inundation depth once the water overflows the banks
and spreads over the adjacent land. Nevertheless, in case of unavailab ility of data to perform a
detailed analysis, geomorphic procedures characterized by simple data requirement can be very
useful for local authorities and planners to realize management strategy.

Fig. 26 . Water depth map developed in this study (GFI hazard map)

Fig. 27 . Comparison between the JRC hazard map (Alfieri et al., 2014)
and the GFI hazard map obtained in this study

4.3.2. Exposure analysis. Data and methods
The exposure analysis refers to the risk susceptible elements, which are obtained
overlapping the information regarding the economic assets from an area (land use map,
infrastructure map) and the information regarding the hazard (flooded area map, water depth map).
The current land use datasets available for large areas (such as Corine land cover) present
a course resolution inducing uncertainties in the flood damag e analysis. Furthermore, due to the
rapid urbanisation that took place in the past years, up -to-data information become unavailable for
many regions. Even though detailed maps (such as the Urban Atlas datasets) do exist, they cover
certain areas or cities and therefore they cannot be used for large scale applications. In this context
a. JRC hazard map
(Alfieri et al., 2014)
b. GFI hazard map

the latest studies in the field (Biro et al., 2013; Liu et al., 2014; Wieland et Pitore, 2016) highlighted
the potential of satellite remote sensing data to provide information on the land use distribution.
For this study, a new land used map was developed from multi -spectral satellite images.
For this purpose the Landsat 8 data with a resolution of 30 m were used and machine -learning
classification algorithms were applied in o rder to identify the land use classes for the study area.
The method is using an object -based approach, which means that the analysis is done for
segments of an image and not for each pixel. These segments are characterised by different features
that can be analysed using a machine learning classifier in order to detect a certain object, in this
case a land use class. This method is following three steps: satellite image processing, image
segmentation and feature -based description using training data and c lassification.

4.3.2.1. Satellite images processing
For this study 26 satellite images from the period June 2016 – September 2016 have been
acquired ( USGS, 2018a ), covering the entire territory of Romania. For data homogeneity, it is
important that all the images to be acquired from the same period.
Each image is composed of 11 bands, containing as well a Quality band (QA, Quality
Assessment). This band can be utilized to indicate the pixels quality and to identify the pixels
covered by clouds. Each pixe l has a value indicating the presence or the lack of the parameters
taken into consideration (clouds, snow/ice, water, etc.), as well as the level of confidence of this
algorithm ( Table 12 .) (USGS, 2018).
The satellite images acquired for this study were s elected to have a quality as good as
possible, containing a small area covered by clouds. This was done through image visualization,
for example in Fig. 28 the lighter pixels indicated the fact that those surfaces are covered by clouds.
Thus, a number of i mages have been selected, which if they are overlapped provide a image with
a smaller cloud coverage, by selecting the pixels with better quality. Further, each image (286
raster files – 26 image x 11 bands) has been processed using the QA band in order to exclude the
pixels covered by clouds from the images.

Pixel
value Cloud Cirrus
Conficence Snow/Ice
Confidence Water
Confidence Terrain
Occlusion Dropped
Frame Pixel description
2 Not
determined Not
determined No No No Yes Dropped pixel
20480 No No No No No No Clear Terrain
20484 No No No No Yes No Clear Terrain with Terrain Occlusion
20512 No No No Maybe No No Water terrain
23552 No No Yes No No No Snow/Ice terrain
28672 No Yes No No No No Cirrus clouds
31744 No Yes Yes No No No Cirrus clouds or Snow/Ice terrain
36864 Maybe No No No No No Mid-confidence cloud
36896 Maybe No No Maybe No No Mid-confidence cloud or Water terrain
39936 Maybe No Yes No No No Mid-confidence cloud or Snow/Ice
terrain
45056 Maybe Yes No No No No Mid-confidence cloud with Cirrus
clouds
48128 Maybe Yes Yes No No No Mid-confidence cloud with Cirrus
clouds or Snow/Ice terrain
53248 Yes No No No No No High -confidence clouds
56320 Yes No Yes No No No High -confidence clouds or Snow/Ice
terrain
61440 Yes Yes No No No No High -confidence clouds with Cirrus
clouds
64512 Yes Yes Yes No No No High -confidence clouds with Cirrus
clouds or Snow/Ice terrain
Table 12 . Examples of pixels values and their meaning for QA band (Landsat 8) ( USGS, 2018 )

Fig. 28. Landsat 8 satellite image, band 1
The first step was to process the QA bands ( Fig. 29 ) for each image, using the Raster
Calculator from Gdal plugin (QGIS software). Thus, the values below the value 53248 (which
represent the smallest value of the pixels covered by clouds – Table 6) have been replaced by 1
and the higher values have been replaced by 0. In this way a raster with two class of values was
obtained:
– 1 – for the surfaces that are not covered by clouds
– 0 – for the surfaces covered by clouds ( Fig. 29 , b)

Fig. 29. QA band processing

Therefore, multiplying each band of one imagine with the processed QA band (with values
of 1 and 0), the pixels covered by clouds will be excluded. The result is represented in Fig. 30.

Fig. 30 . Satellite image, band 1, with no clouds covered areas
Given the fact that the panchromatic band (band 8) has a resolution different (15 m) from
the rest of the bands, these bands were resampled to 30 m in order to be used for further analy sis.
For the reclassification the r.resamp.stats. function from GRASS, QGIS was used.
Moreover , the 286 raster files have been organised into categories of bands (band 1, band
2, etc.). Thus, each band contains 26 raster files which will be overlapped and united in order to
form an image of the entire surface of Romania. Due to the extended surface of Romania, the raster
files have different coordinate reference system as following:
– 11 images have the EPSG: 32634 coordinate reference system
– 14 images ha ve the EPSG: 32635 coordinate reference system
– 1 image has the EPSG: 32636 coordinate reference system
The images with the same coordinate reference system have been merged using the SAGA
(System for Automated Geoscientific Analyses) tool from QGIS software, obtaining three images

with three different coordinate reference system. This images have been re -projected in the EPSG:
3035 coordinate reference system and merged in one image. This process was repeated for each
band, resulting 11 images of the Romanian territory (Annex 3).

4.3.2.2. Image segmentation
In this process, basically the p ixels are grouped into segments that are further analysed. For
this purpose, an algorithm , called graph -based segmentation algorithm, is used in order to
determine the boundary betwe en two segments (Felzenszwalb and Huttenlocher , 2004). For this
purpose the algorith m compares the features values of the pixels within one segment with the
features values of pixels outside of the segment. The segmentation parameters have been adjusted
using reference data, selecting the parameters that had the best performance. This ste p is important
in order to define the scale of analysis. A good quality segmentation implies that the image
properties within a segment are homogeneous having a low variability, while the image properties
between segments are heterogeneous.

4.3.2.3. Fe ature -based description and classification
The image features are used to describe the segments, the classification process being done
based on these features. In this study 44 image features were used, categorised on three types:
spectral features, textu ral features and spectral band indexes.
The spectral features were represented by:
– mean spectral value in the image band (mean_b1, mean_b2, mean_b4, mean_b5,
mean_b6 and mean_b7);
– standard deviation of the brightness value in the image band (sdt_b1, sdt_b 3);
– mean value of the Normalized Difference Vegetation Index (mean_NDVI);
– minimum brightness value in the image band (min_pixel_b1, min_pixel_b3,
min_pixel_b9 and min_pixel_b10);
– maximum brightness value in the image band (max_pixel_b1, max_pixel_b4,
max_pixel_b6, max_pixel_b9 and max_pixel_b10);
– Weighted brightness (WB).
The textural features were represented by:

– Homogeneity derived from the GLCM* in the band (homogeneity_b1,
homogeneity_b4 and homogeneity_b6);
– Contrast derived from the GLCM* in the b and (contrast_b1, contrast_b2, contrast_b4
and contrast_b7);
– Dissimilarity derived from the GLCM* in the band (dissimilarity_b1, dissimilarity_b2
and dissimilarity_b6);
– Angular Second Moment derived from the GLCM* in the band (asm_b7);
– Mean derived from the GLCM* in the band (glcm_meani_b1, glcm_meani_b3,
glcm_meani_b4, glcm_meani_b5, glcm_meani_b6, glcm_meani_b7, glcm_meani_b9
and glcm_meani_b10);
– Variance derived from the GLCM* in the band (glcm_variancei_b10).
*GLCM (Gray -Level Co -occurrence M atrix)
The spectral band indexes are represented by:
– Mean value of soil adjusted vegetation index, SAVI (savi_mean);
– Mean value of normalised difference vegetation index, NDVI (ndvi_mean);
– Mean value of modified normalised difference water index, MNDWI (mndwi_mean) ;
– Mean value of normalised difference built -up index, NDBI (ndbi_mean);
– Standard deviation value of normalised difference built -up index, NDBI (ndbi_std)
(Wieland and Pittore, 2016).
For the classification of land use classes using the Landsat 8 satellite imagery data, an
Artificial Neural Network (ANN) called Multi -layer perceptron (MLP) was applied, using as
training data the available Urban Atlas land use maps for Romania. The MLP is using three types
of layers, each of them containing processing nodes i nterconnected to each other but not within
the same layer. The MLP used in this study contains one input layer, two hidden layers and one
output layer. The input layer is represented by the bands of the imagery; the hidden layers are used
for computations purposes; the output layer contains different codes, representing the land use
classes. The ANN calculated the weights of each layer node. Each node contains the weighted
input from all of the nodes connected to it from the previous layer. The selection of the features is
based on the weights assigned by the algorithm to this features.
In order to be able to recognize the land use classes, the MLP algorithm must be trained.
For this purpose, the back -propagation learning algorithm is used. First, based on t he features

weights, the training data are introduced in the algorithm in order to identify the pattern. In the
second step, the error, which is the difference between the known and the estimated value, is send
backwords through the network. In this way th e weights of the nodes are adjusted, in order to
minimize the error. This process is repeated until the error become acceptable/minimal ( Rumelhart
et al., 1986).
In this study 8 neural networks were used, each of them receiving the 44 features. One
cascade binary classification is done:
– 1: built -up area – non built -up areas
– 2: urban – non urban
– 3: vegetation – water
– 4: agriculture – forest
– 5: industrial – infrastructure
– 6: continuous urban – discontinuous urban
– 7: water – non water
– 8: built -up area – vegetation

4.3.2.4. Training data. Land use data processing
In order to train and calibrate the algorithm for the recognition of land use classes, the
Urban Atlas land use data were processed and used as training data.
Urban Atlas database
For this st udy, the available data for 14 cities of Romania (Bucarest, Cluj -Napoca,
Timișoara, Craiova, Brăila, Oradea, Bacău, Arad, Sibiu, Târgul Mureș, Piatra Neamț, Călărași,
Giurgiu, Alba -Iulia) were acquired form European Environment Agency site. The Urban Atlas
database has a large number of land use classes providing detailed data, however this data are
available just for large urban areas.
For the homogeneity of the data, the Urban Atlas classes have been reclassified, reducing
their number ( Table 13 ). Furthe r, this vector files have been converted into raster files with a
resolution of 30 meters. In Fig. 31 the reclassified Urban Atlas land use for Cluj -Napoca city of
Romania is represented in vector and raster format.

Table 13 . Reclassification of Urban Atlas classes
Urban Atlas Code Description Reclassification
code Description
11100 Continuous Urban Fabric (S.L. >
80%) 111 Continuous urban
fabric 11210 Discontinuous Dense Urban
Fabric (S.L.: 50% – 80%)
11220 Discontinuous Medium Density
Urban Fabric (S.L.: 30% – 50%) 112 Discontinuous urban
fabric
11230 Discontinuous Low Density
Urban Fabric (S.L.: 10% – 30%)
11240 Discontinuous Very Low
Density Urban Fabric (S.L. <
10%)
11300 Isolated Structures
12100 Industrial commercial public
military and private units 121 Industrial,
commercial, public,
military, private units
12210 Fast transit roads and associated
land 122 Infrastructure
12220 Other roads and associated land
12230 Railways and associated land
12300 Port areas
12400 Airports
13100 Mineral extraction and dump
sites
13300 Construction sites
13400 Land without current use 200 Agricultural and
green areas 14100 Green urban areas
14200 Sports and leisure facilities
20000 Agricultural + Semi -natural
areas + Wetlands
30000 Forests 300 Forests
50000 Wetlands 500 Water