Developing a framework for the assessment of current and future ﬂood risk in Venice, Italy

. Flooding has been a serious struggle to the old-town of Venice, its residents and cultural heritage and continues to be a challenge in the future. Despite this existence-deﬁning condition, limited scientiﬁc knowledge of ﬂood hazard and ﬂood damage modelling of the old-town of Venice is available to support decisions to mitigate existing and future ﬂood risk. Therefore, this study proposes a risk assessment framework to provide a methodical and ﬂexible instrument for decision-making for ﬂood risk management in Venice. It uses a state-of-the-art hydrodynamic urban model to identify the hazard characteristics inside the 5 city of Venice. Exposure, vulnerability, and corresponding damages are modelled by a multi-parametric, micro-scale damage model which is adapted to the speciﬁc context of Venice with its dense urban structure and high risk awareness. A set of individual protection scenarios is implemented to account for possible variability of ﬂood preparedness of the residents. The developed risk assessment framework was tested for the ﬂood event of 12 November 2019. It was able to reproduce ﬂood characteristics and resulting damages well. A scenario analysis based on a meteorological event like 12 November 2019 was 10 conducted to derive ﬂood damage estimates for the year 2060 for a set of sea level rise scenarios in combination with a (partially) functioning storm surge barrier MOSE. The analysis suggests that a functioning MOSE barrier could prevent ﬂood damages for the considered storm event and sea level scenarios almost entirely. It could reduce the damages by up to 34% for optimistic sea level rise prognoses. However, damages could be 10% to 600% times higher in 2060 compared to 2019 for a partial closure of the storm surge barrier, depending on different levels of individual protection.


Introduction
Flood events are among the most disastrous natural catastrophes, causing significant damages and fatalities all around the world. In Europe, coastal flood events are estimated to affect more than 100,000 citizens, causing losses of about EUR 1.4 billion annually (Vousdoukas et al., 2020). Under consideration of climate change scenarios, future flood damages are expected to increase due rising sea level (Hinkel et al., 2014).
In this context, hazard and flood risk assessment has been broadly implemented according to the 60/2007/EC directive in the EU (European Commission, 2007). According to the IPCC, flood risk is defined as the combination of a specific hazardous flood event, the exposure of human systems, and their vulnerability, meaning predisposition to be adversely affected (Field et al., 2012). It can therefore include adverse effects on human health, environment, cultural heritage and economic activi-25 ties. As such, outcomes of a flood risk assessment framework can support systemic and individual decisions to mitigate flood damages or adapt accordingly, increase preparedness, and strengthen coping capacities (Arrighi et al., 2018b;Molinari and Scorzini, 2017;Scorzini and Frank, 2017;Amadio et al., 2016;Merz and Thieken, 2009).
A flood risk assessment framework typically follows four steps: 1) hazard modelling, 2) assessment of vulnerability of ex-30 posed assets, 3) damage estimation and 4) flood risk estimation (Arrighi et al., 2018a). The application of 2D hydrodynamic models is currently the state of the art method for deriving information about coastal and urban flood events (Yin et al., 2020;Sai et al., 2020;Xing et al., 2019;Teng et al., 2017;Gallien et al., 2014). Damage modelling traditionally focuses on direct, tangible damages in terms of replacement costs related to structures, interior, and public infrastructure since the cost-benefit analysis of flood mitigation measures is straight forward and indisputable (Molinaroli et al., 2018;Scorzini and Frank, 2017; 35 Dottori et al., 2016;Merz and Thieken, 2009). The vulnerability of exposed assets is determined not only by the type of exposed structure, its construction material (quality), age, and level of maintenance (Huijbregts et al., 2014;Drdácký, 2010;Merz and Thieken, 2009), but also by the level of present awareness. Risk awareness influences the level of preparedness by means of physical measures (e.g. permanent or mobile water barriers, emergency works like sand bags) or behavioral adjustments (e.g. adapting the vertical distribution of goods and values). Vulnerability therefore varies highly spatially and temporally (Hudson 40 et al., 2016;Kreibich et al., 2011;López-Marrero, 2010).
This study focuses on the assessment of flood damage in Venice. The low-lying historic city has a long-lasting record of flood events (Battistin and Canestrelli, 2006) which is likely to extend in future mainly due to relative sea level rise and continuing subsidence Medugorac et al., 2020;Morucci et al., 2020;Tiggeloven et al., 2020;Jordà et al., 2012). different closure scenarios of the MOSE barrier consider flood risk implicitly by using a maximum safeguard water level at the city of Venice (Umgiesser, 2020;Cavallaro et al., 2017;Umgiesser and Matticchio, 2006).
To develop a better understanding of the existing and future risk due to damages to structures and cultural heritage in Venice, a risk assessment framework is developed in this study as shown in Fig. 1. High resolution flood hazard characteristics are 70 computed by means of a 2D-hydrodynamic model. They feed into a micro-scale damage model to estimate expected absolute direct damages of the exposed buildings (Dottori et al., 2016). The flood model is calibrated and partly validated using data from the storm surge of 12 November 2019. Additionally, a damage claim data-set for the the same event is used for performance analysis of the damage model. Finally, the framework is applied to a set of scenarios of varying sea level change and MOSE closure to analyse potential development of flood damage in mid-term future. 75 The paper proposes a methodical and flexible assessment framework for Venice that is useful to analyse existing and future flood damages for different meteorological storm events. It is methodical, as it uses a hydrodynamic model along with a damage model that can resolve physical damage modelling of separate building components. The framework is flexible because both models can be refined to consider additional elements of influence or additional elements at risk. This could be of partic-  The old-town of Venice covers an area of about 6 km 2 and is pervaded by more than 100 canals of depths between 1 and 5 meters (Madricardo et al., 2017). The old-town is located in the Venetian lagoon, the largest in the Mediterranean with an area 1 If not highlighted otherwise, all levels refer to the local chart datum in Venice, given as Zero Mareographic of Punta della Salute (ZMPS), corresponding to the mean sea level of the 1885-1909 period. Present mean sea level (2019 annual mean sea level) is today 0.34 m ZMPS.  On 12 November 2019, the second highest storm surge ever recorded flooded the old-town of Venice and other parts of the Venetian lagoon. The maximum measured water level inside the old town was 1.89 m ZMPS, measured by the tidal gauge station Punta della Salute at 22:50 on 12 November 2019. It was comprised of a tidal contribution of 0.36 m, 0.47 m of storm surge induced by a strong Sirocco wind over the Adriatic Sea, 0.35 m of long-term preconditioning, and 0.34 m mean sea level with regards to the local datum . At the same time, a secondary, local cyclone passed over the 95 Northern Adriatic Sea resulting in additional set-up by causing an inverse barotropic effect and very high wind speeds from south-westerly directions of about 70 up to 110 km/h. It is noteworthy that the secondary low pressure field was not forecasted properly which lead to an underestimation of the flood by about 0.40 m . Unlike a storm event that occurred in 2018 where an even higher tidal peak (1.56 m ZMPS) coincided with low astronomical tides (-0.10 m ZMPS), the extreme sea level of 12 November 2019 was the product of less extreme, thus more likely conditions (Morucci et al., 2020;100 Cavaleri et al., 2019).
As a response to the unexpected extreme meteorological event of 12 November 2019, financial support to the affected parties was provided in two rounds: 1) limited amounts for immediate response (up to EUR 5,000 for residents and EUR 20,000 for non-residential entities (companies, NGOs,...)) and 2) support for more extensive flood damages. Residents and entities could  apply for compensation for either one or both rounds. In total, 7,644 eligible claims were issued inside the study area with a total cost of EUR 56.2 million 2 .
For residents and entities which submitted only immediate response claims (3,728 claims covering EUR 26.99 million of damages), the physical addresses of the claimants are publicly available. It was possible to allocate 95% of the reported im-110 mediate response claims (EUR 25.73 million) to 2,778 structures inside Venice using a set of 33,096 addresses 3 . For claimants that submitted claims in both rounds or just for more extensive flood damages (EUR 29.21 million), the available information provided was aggregated by city-district for data protection reasons 4 .

The modelling framework
As visualized in Fig. 1, the modelling framework consists of a combination of a hydrodynamic and a damage model which is presented in this section.

Hydrodynamic model
In the study area, hydrodynamic models have been used frequently but they do not account for the urban area of Venice Ferrarin et al., 2015;D'Alpaos and Defina, 2007;Umgiesser et al., 2004;Roland et al., 2009). Studies looking into the distribution of flood depths in Venice have used a static model, also called bathtub model (Cellerino et al., 1998)   . Furthermore, it provides additional modules that can be used for a better physical repre-140 sentation of the system. Only 2D flow was considered in this study, but the model allows to account for additional processes like wave action or 1D flow of the sewage system 6 .
An offline grid nesting framework was chosen, consisting of a parent model covering the study area and seven sub-models of higher resolution covering the area of the old-town of Venice. The parent model used 2.73 million elements covering the 145 study area with an average grid size between 2.6 m in the old-town and 200 m at the Adriatic shelf. In the seven nested models, grid size was increased to an average of 1.3 m to reproduce the narrow street system in Venice. Water level time-series from the parent model simulation were extracted at 168 locations inside and around the old-town of Venice and used as inputs for the nested sub-models, as shown in Fig. 3. For every nested model, the maximum water level at each grid point was extracted.
All grid points inside a 4m buffer around each structure were used to derive an average water level.

150
Most recent information on the depth of the lagoon flood plains, channels, and the elevation of the islands of the old-town were accessed from various sources. Table 2 presents an overview of all the elevation data used. All altimetry data were corrected to refer to ZMPS, the local chart datum in Venice.  Constant standard values were used for the viscosity, diffusivity, and density as the flow in the Venetian lagoon is relatively well mixed without stratification (Ferrarin et al., 2010). Roughness was added as Manning-type n. A standard roughness value  Similarly, the wind-induced shear stress, by means of drag coefficient, was used as a calibration parameter. It was implemented based on a linearly increasing relation between wind speed and wind drag developed by Smith and Banke (1975).
However, their relation was derived for wind speeds between 6 and 21 m/s, but extreme wind speeds for the 12 November 2019 165 reached up to 27 m/s. Therefore a higher drag coefficient of 0.00876 (for 100 m/s wind speed) was used. A comprehensive analysis of commonly used wind drag formulations confirmed that the chosen drag coefficient is within the range of available estimates (Bryant and Akbar, 2016). In addition, it was confirmed that the chosen values are in line with other Delft3D-FM studies of the Venetian lagoon 7 .

170
The barrier system was modelled by means of a set of three simple weirs with a crest height defined by a time-series. It is assumed that the barrier crest height increases at constant speed from the bottom of the respective inlet up to a height of 3.00 m ZMPS and closes within 30 minutes . For the considered meteorological storm conditions, the MOSE barrier starts closing when the tidal gauge station of Punta della Salute reaches a water level of 0.65 m ZMPS (Zampato et al., 2016). This threshold is assumed to be constant for all analysed scenarios. The starting time of closure was determined by 175 modelled tidal gauge information from Punta della Salute for the different scenarios without a closing MOSE barrier, see Tab. 4.

Damage Modelling
While general damage drivers are broadly acknowledged (Patt and Jüpner, 2013;Kelman and Spence, 2004), the exact effect of hazard characteristics on an exposed structure is still poorly understood as it also heavily depends on the material and its 180 quality (Huijbregts et al., 2014;Merz and Thieken, 2009). This is particularly relevant for cultural heritage sites built by materials which have deteriorated by centuries of existence (Drdácký, 2010). Consequently, the chosen model was selected with special care to allow for an inclusion of differing exposure and vulnerability characteristics.
Various approaches and post-flood data analysis have been conducted to develop relations between the flood hazard char-185 acteristics and corresponding tangible, direct damages. Several comparative studies have looked into the characterization and performance analysis of some frequently used damage models (Molinari et al., 2020;Gerl et al., 2016) 8 . In general, loss estimates reflect high uncertainties and disparities because of the inaccuracy of the models and the lack of knowledge about the system in which they have been applied (Scorzini and Frank, 2017;Gerl et al., 2016).  (Molinari et al., 2020;Scorzini and Frank, 2017;Dottori et al., 2016). As such, it is ideal to be extended to include new building types, e.g. cultural heritage sites like churches etc., with specific hazard-structure responses. The INSYDE model also makes use of categorization into building types to account for differences in the exposure or vulnerability characteristics between typical buildings in a study area. As a result, the absolute damage, D, per structure is calculated as the 200 sum of a set of damage components summarized in Tab. 5: where j represents the damage component and i describes the considered activity, e.g. cleaning, removal, and replacing. up i,j is the unit price per damage component for for a given activity, ext i,j is the extent of exposed component and E damage mediating factor since flow velocity and flood duration were found to be too low to add an additional source of damage (Dottori et al., 2016;Penning-Rowsell et al., 2005) 10 . The fragility functions allow not only for a deterministic multi-parametric consideration of the flood-structure interaction, but also to account for uncertainties in the flood-structure interaction in a probabilistic framework. An example is shown in Fig. 4: damage to partition walls occurs if the partition walls absorb too much water to be dried up, i.a. if water depth exceeds a certain threshold (Dottori et al., 2016). The fragility function can be used to 215 determine an expected damage ratio or expected share of damaged partition wall for a given flood depth. However, damage to partition walls due to a certain water depth could range from 'no damage' to 'full damage', depending on factors such as the quality of wall (material). In the probabilistic framework, a large set of realizations for each component is drawn to derive the 5-and 95-percentiles expressing an optimistic and pessimistic estimate of the absolute damages. Even though the probabilistic framework was not used in this study, it may be useful in case of extending the framework to explicitly cover cultural heritage 220 sites in Venice, which may be more sensitive to varying flood characteristics.
Information on the individual building area and extent were derived from cadastral data of the city of Venice 11 . A total of 14,460 structures were considered. Information on the structural properties, the year of construction and the maintenance 9 More details regarding the background and set up of the INSYDE model is provided in the supplementary material of this study. 10 Results of the hydrodynamic model suggest that flood velocities are generally lower than 0.3 m/s and the flood duration is between 2 and 4 hours. 11 Accessible here: http://geoportale.comune.venezia.it(accessed 05/07/2021)  level were accessed from census data from year 2011 by the Italian National Institute of Statistics (ISTAT, 2020). The census data is not building-specific but aggregated in census blocks covering multiple buildings. As a consequence, the most frequent characteristic was applied to all buildings within a census block 12 .
GoogleMaps StreetView was used to gather visual information about typical house fronts, size and number of windows along with information about possible elevations of the entrance at ten random locations in different districts of the old-town.

235
It was found that typical characteristics of residential buildings do not differ significantly from the implemented characteristics in INSYDE. One major difference related to the external wall perimeter exposed to floods was detected and incorporated as a new parameter EP ef f : most buildings in Venice are attached to other buildings reducing the exposed perimeter. Additionally, a new building type 'buildings with economic activities on the ground floor' (BEA) was added to account for observed differences in the exposure and vulnerability characteristics from typical residential buildings: the windows are generally larger 240 (increased from 1.4m x 1.4m to 2m x 2m), the window sills are lower (new sill height of 0.5m instead of 1.2m), and many shops are on ground level without any steps of elevation. Additionally, the internal perimeter (reduced from 2.5 to 1.5 time the external perimeter) and number of doors is smaller (reduced to 3 per 100 m 2 ).
It was detected that many buildings had installed mobile protection systems, mainly bulkhead protections, at doors and 245 windows to protect the interior from flooding during the 12 November 2019 storm event. Other protection measures were not commonly installed and therefore not incorporated in the damage model. A new parameter BuHe, representing the bulkhead protection height, was implemented to mediate the water level inside the buildings. Due to lack of data on the spatial distribution and protection height of mobile protection systems, three conceptual individual protection scenarios (IPS) were characterized and applied: expected IPS, risk averse IPS and risk-taking IPS. For the risk taking IPS, it was assumed that no bulkhead 250 protection was installed at all. For the expected IPS, it was assumed that residents would install bulkheads protecting their building against the forecasted maximum water level (F C) at Punta della Salute incremented by a safety margin of 10 cm.
For a risk averse IPS, the protection height also refers to the forecasted maximum water level at Punta della Salute but is 12 More detailed information on the census block data can be found in the supplementary material of this study. 13 accessed from: https://www.aquagrandainvenice.it/en/welcome incremented by a safety margin of 50 cm. The water level h inside the buildings is consequently calculated as where h e is the water level outside the buildings, GL is the ground floor level of the considered structure and BuHe is the bulkhead protection height as visualized in Fig. 5. F C was set to 1.50 m ZMPS for 'SLR0-allopen' and to 1.10 m ZMPS in all other scenarios given that a functional MOSE barrier is expected to keep the water level below a threshold of 1.10 m ZMPS.   To evaluate the performance of the model, the Pearson R coefficient and the Root-Mean-Square-Error were used. Results for the three runs are compiled in Tab. 7 and suggest that measured data can be reproduced well, including the storm surge peaks for the wind calibration and validation run. Accuracy of the maximum flood peak lies within a margin of ±5cm. For San Nicolo, Malamocco and Murano, the observed water level data were partly corrupted or not available. 17 .

285
The nested models were used to derive the flood depth estimates inside the city. Analysis of the difference in water depth estimates inside the old-town of Venice from the parent and nested model domains suggest that the grid resolution of the hydrodynamic model has significant impact on the flood characteristics inside the city. As Fig. 6b shows, a coarser grid tends to provide lower flood depth estimates. A coarser grid may fail (more often) to resolve possible flow paths in the very narrow    shows. Additionally, it is visible that the hydrodynamic model gives high flood depths for some buildings while the bathtub models suggests that those structures are not affected by water levels at all (or to a much lesser degree). This unexpected result was linked to grid instabilities of the nested models. In total, higher water levels were suggested by the hydrodynamic model at 383 buildings. Additionally, grid instabilities of the nested sub-model 'Castello' (refer to Fig. 3) could not be resolved, resulting in missing flood depth data based on the hydrodynamic model for 2,098 buildings (14 % of the total number of buildings).

300
For buildings affected by instabilities, flood depth estimates from the bathtub model were used for the damage modelling of these buildings.

Damage model performance
To analyse the performance of the transferred model, the total modelled damages for the old-town were compared against the 305 total sum of the eligible 7,644 damage claims. Additionally, a structure-wise analysis was conducted for the sub-set of 2,778 structures with 3,728 immediate response claims. reported immediate response claims corresponds to a individual protection scenario between 'risk averse' and 'expected', the total volume of all reported damages is closer aligned with a risk averse IPS.
Additionally, a structure-wise comparison was conducted for 2,778 structures. As shown in Tab. 9, correlation and average relative error, computed as the ratio of the reported damage and the estimated damage per building, suggest limited alignment 315 of the modelled damages with the reported claims. Both indicators suggest that the damage claims might be slightly better estimated based on a expected IPS or risk taking IPS for the majority of buildings. At the same time the RMSE, which gives more weight to extreme variations due to its definition, is lower when assuming a risk averse IPS. Moreover, the Kernel density plot gives insight in the relative frequency of damages as shown in Fig. 7. It can be seen that in a risk averse IPS, the number of structures with rather low damages is overestimated, meanwhile larger damages are underestimated. The opposite applies to 320 risk neutral and risk taking scenarios.
According to the INSYDE model, the most affected building components are external and internal plaster removal (R6, R7), replacement (F1, F2) and painting (F3, F4), followed by costs for the replacement of electrical (P3) and plumbing systems (P4), as shown in Fig. 8. It can be seen that the model suggests no damage for many damage components as hazard characteristics are 325 below thresholds for which damage is reported to occur. It can be seen that the expected IPS leads to limited damage reduction  regarding plaster, but a strong reduction for the building systems. In a risk averse IPS, no damage occurs inside the buildings.
It is worth mentioning that damage estimates based on flood depth information from the bathtub model generally give similar damage estimates for both sets of considered structures; deviations for risk averse and risk taking IPS is between 1.5 and 6.3%.  It is noteworthy that for the 'allclosed' scenarios, SLR2 results in a slightly lower flood peak estimate than the other two scenarios. A possible explanation is that for SLR2 the closure of the MOSE barrier occurs about 24 hours earlier relative to the flood peak, while for SLR0 and SLR1 it is closed about 4 hours before the flood peak. As the barrier is closing during flood, 350 the part of the tidal wave that propagated into the lagoon before the full closure has more time to evenly spread out across the lagoon, resulting in a slightly lower average flood depth in the centre of the lagoon than for the other two scenarios. This ultimately influences the wind effect and maximum water levels at Punta della Salute.
Analysis of the implications of the different scenarios on the average inundation depths concludes that a partially functioning 355 MOSE barrier would significantly reduce the expected average flood depth for 90% of the buildings for sea level rise scenarios of SLR0 and SLR1. In SLR2 the increased sea level dominates over the dampening effect of the partial closure as visualized in Fig. 9b. This analysis also shows that for the storm surge of 12 November 2019, 50 % of all structures in Venice experienced a flood depth of 0.55 m or higher. Only 10% of buildings experienced flood depths lower than 0.10 m and only 5% of buildings were not exposed to floods at all.
360 Figure 9. Flood depths for scenarios. a: Modelled flood peaks at Punta della Salute. b: Share of buildings exposed to certain average flood depths for the risk averse IPS is less apparent given that for SLR0-allopen, damages only occurred at the external walls, but for SLR0lidoopen also partly on the inside due to lower protection levels. Results are compiled in Tab. 10.
An interesting observation can be made when comparing the damage estimates of SLR0-allopen to those of SLR2-lidoopen.
Despite an approximately 0.21 m higher flood depth for SLR2-lidoopen, the effect on damage estimates for risk taking IPS and 370 expected IPS are smaller than expected even though protection heights are on average also 0.40 m lower than in SLR0-allopen.
Analysis of the formulations for vulnerability and exposure implemented in INSYDE provide a possible explanation: not only the part of external and internal plaster in direct contact with the water has to be replaced, but also an additional height of one meter. Given that cost for plaster removal is independent of the required removal height, this implies that for a small flood depth, higher replacement costs occur already which are only incremented linearly for higher flood depths. As extreme flood 375 depths are frequently lower than one meter, the influence of the additional height weights heavier compared to the difference for higher water level scenarios.

Discussion
Venice is a city with a long history of flooding that is likely to extend into future despite the presence of the MOSE barrier.

380
Until now, limited methodological approaches exist which provide estimations of future flood risk to structures and particularly to cultural heritage. This study developed a flood risk assessment framework that can be used for assessment of direct, tangible damages to residential and economic buildings, and can be extended in future research to account for the special conditions of cultural heritage as well. The framework performs well compared to available damage claim data and gives some indications about possible future flood risk for extreme storm surges under a partially failing MOSE barrier system.

385
The developed hydrodynamic model provides reliable estimates of hazard characteristics inside the old-town. First, the validated hydrodynamic coastal model reproduces the flood peaks with an accuracy of ±5cm despite some simplifications of the lagoon system, such as applying uniform meteorological conditions over the entire domain and neglecting freshwater inputs and wave action. Second, the cross-model comparison suggests that the hydrodynamic model performs as expected and may 390 provide optimistic flood depth estimates inside the city compared to the presently used static model (Liu et al., 2018). A final confirmation of the flood depths inside the city by means of calibration and validation with flood depth records was not possible but should be a key focus in future studies as flood-enhancing components, such as the sewage system, water coming from the ground, or wave influence were neglected. In addition, following from the comparison of parent and nested model depth estimates, a grid convergence analysis should be conducted to find the optimal grid resolution for the city of Venice. Despite a 395 grid size of 1.3m near structures, which is already rather high compared with other hydrodynamic urban models (Xing et al., 2019), the specific setting of Venice with its narrow street system may require increasing the resolution even further.
Some modelling challenges of the hydrodynamic model have to be highlighted. Due to the complex urban structures and altimetry, some extreme local water levels occurred in the parent and nested models were likely caused by the complex grid 400 structure and the algorithm describing the wetting and drying process inside the model (Deltares, 2021). This led not only to incorrectly high flood depths at a few buildings but also prevented the consideration of one of the nested sub-models. Part of the instabilities can be solved by grid refinement, bathymetry alteration, or adjusting the modelled time periods. In accordance to previous studies (Scorzini and Frank, 2017;Arrighi et al., 2013), it was found acceptable to use bathtub flood depth estimates for the remaining structures instead, given the limited influence of flood depth variation on the damage estimate. Additionally, a 405 fully functioning hydrodynamic model may add additional benefits to the flood risk assessment framework as it can account for (changing) physical characteristics explicitly, allow for a proper calibration, and incorporate additional flow path-components such as a 1D sewage system.
The adjusted version of the INSYDE damage model is able to reproduce the total damage claim volume related to the storm 410 event of 12 November 2019 as shown in Tab. 8. Analysis of the sub-set of immediate response damage claims also confirm initial expectations of relatively high individual protections levels in Venice as frequent and intense experience of flooding have been reported to contribute to higher levels of individual flood preparedness (Kreibich et al., 2015). Moreover, results imply that the effect of protection measures has a strong influence on the estimated damages.

415
However, the poor structure-wise depth-damage correlation and the alignment of the two considered sets of reported damage claims with different (combinations) of IPS reiterate commonly faced challenges of flood damage modelling (Ahn et al., 2019).
Limited knowledge of the system introduces uncertainty in the damage estimates. As an example, about half of all damage claims ( 7,644) were linked to about 20% of the structures in Venice only. Meanwhile 90% of structures were found to be exposed to an average flood depth of at least 0.1 m according to the hydrodynamic model. Thus, it is questionable whether 420 exposure and vulnerability of the system are adequately represented given that modelled damages of external walls alone are almost as high as the reported damages. In addition, preparedness was simplified as perfectly functioning mobile barrier systems installed at all buildings, like in this study. However, protection levels have been reported to be very diverse and could also (partially) fail to provide the promised level of protection in reality. Additionally, more protection measures may be in place to reduce the flood damages. Moreover, many exposure and vulnerability relations of the synthetic damage model were 425 transferred unaltered, despite the possibility that they may not reproduce the present hazard-structure interaction processes in Venice.
At the same time, limitations of the available damage claim data-sets have to be accounted for as well. It can generally be questioned whether reported damages represent the full set of effective damages of a flood event. Potential claimants may 430 have opted to undergo significant bureaucratic efforts for (sometimes) limited financial support (Molinari et al., 2020). Alternatively, claimants may not have seen the need to replace (some) damaged elements, e.g. because of their experience with frequent flooding. Marks of previous floods at house fronts throughout the old-town support this hypothesis. Additionally, given that the available damage data are spatially and/or component-wise aggregated, limited conclusions can be drawn from the damage data analysis to address the mentioned limitations of the framework. Information from a detailed investigation When discussing the accuracy and reliability of the applied damage model, it is also worth considering that another study analysing exceptionally extreme flood events suggests much higher flood damages (Caporin and Fontini, 2014) Results on the effect of the MOSE barrier on the water level inside the lagoon align with previous studies, suggesting that a partial closure will still cause flooding of the old-town of Venice . The study adds to the existing knowledge as it considers the second most extreme flood event experienced, while previous studies have mainly investigated more frequent, less extreme flood events (Zampato et al., 2016;Vergano and Nunes, 2007). The present study adds new insights 455 suggesting that the damping effect of a partially closed MOSE barrier on the flood wave will reduce as sea level rises and may consequently amplify flood risk in future. To confirm this finding in future studies, some of the present's study limitations should be addressed: for the applied future scenarios, present conditions of the system were used. However, the sediment budget of the lagoon is negative, meaning that the lagoon currently deepens and may look significantly different in 40 years from now (Tambroni and Seminara, 2006). The same applies for local subsidence processes which have significantly contributed to 460 flood risk in the past and may continue to do so in future as well (Zanchettin et al., 2021). Also, variation in tidal amplitude due to changes in bathymetry and mean sea level as observed in the past, may continue in future as well (Ferrarin et al., 2015).
In addition, some inaccuracy regarding the flood levels is likely to be introduced as processes of seepage through the barrier and freshwater input in the lagoon have been neglected in the present study. This is particularly relevant for SLR2, where the 465 MOSE barrier would be closed for more than 36 hours. In previous studies it has been suggested that seepage through the fully closed barrier could result in water level increase between 0.27 cm to 2.1 cm per hour (Umgiesser and Matticchio, 2006 depending on both possible socio-economic and political developments and the reliability of the MOSE barrier to protect the old-town and its residents in the future.

Conclusions
In this study, a flood risk assessment framework has been developed. It was able reproduce the flood event of 12 November 505 2019 with an accuracy of ±5cm in the proximity of the old-town and provides damage estimates in accordance with available damage claim data. The implemented damage model can reproduce damage claim data but faces commonly acknowledged uncertainties due to limited knowledge about the system and damage processes.
Developing a methodical risk assessment framework for the cultural heritage city has provided some valuable insights into 510 expected flood exposure and damages in the old-town of Venice. While this study confirms the general appropriateness of the MOSE barrier to protect the city of Venice for extreme storm events for additional rising sea level up to 45 cm, it was also found that the damages in case of a partially closed MOSE barrier may still increase significantly for most considered scenarios. While an improved individual protection level in future could lead to a damage reduction of up to 78% for present sea level and 74% for an optimistic sea level rise prognosis, damages could be up to 1.08 to 5.92 times higher in 2060 in case of un-515 changed or decreased level of individual protection. Based on the findings of relative importance of individual flood protection in light of a potentially failing MOSE barrier, this study provides indication that a better understanding of presently applied flood protection is needed to identify realistic individual protection scenarios for future conditions. This would be helpful to identify possible areas of action to maintain (or advance) existing structure-wise flood protections and individual preparedness.
In addition, the influence of the MOSE barrier on the reported warning levels and the effectively installed protections was 520 identified as an important question to address in order to reduce flood risk in Venice until 2060. As such, the proposed flood risk assessment framework provides a methodical approach useful to support future decisions on flood risk management.
Additional studies should be done to improve the presented framework. Addressing some of the limitations, particularly the simplification of the system by excluding the sewage system, grid instabilities and lack of calibration data, may add additional 525 confidence to the exposure modelling. Moreover, incorporating information on future return levels of storm events as well as failure probabilities of the MOSE barrier should be addressed and incorporated in the present framework to allow for a proper flood risk assessment to support the efficient and effective allocation of (additional) resources to flood protection in Venice.
Also, a better understanding of the spatial distribution of protection measures and other exposure mediating characteristics within the districts of the old-town, ideally for each structure, is required for a better representation of the system. Additionally, 530 new building types in the damage model can be implemented to account for some characteristic cultural heritage buildings as proposed in the supplementary material. This would contribute to a better and multidimensional understanding of the present and future flood risk.
Code and data availability. Files and data used for the hydrodynamic and damage modelling are made available on the following repository