Understanding Flood Damage To Economic Activities In Italy From Post-Event Records

The evaluation of ood impacts to economic activities is characterized by obstacles like the paucity of available data to characterise the enterprises, the lack of high-quality data to derive new models or validate existing ones, and the high variability of activity types which hampers generalisation. This study contributes at improving knowledge about types and extent of impacts of ood events on economic activities through the analysis of empirical data, focusing on direct damage and with specic reference to the Italian context. The collaboration among different research groups allowed to collect around a thousand of observed damage records after four ood events in Italy: the inundations that occurred in the town of Lodi (Lombardia Region) in 2002, in Sardegna Region in 2013, and in the Emilia-Romagna Region caused by the Secchia (2014) and Enza (2017) Rivers. The analysis of these data lead to a better knowledge of the types of losses suffered by economic activities, to a representation of damage composition and to the computation of reference damage values. Such a result supports the identication of the more vulnerable elements within the business sector as well as the estimation of the order of magnitude of potential damage, orienting modellers’ and decision-makers’ choices.


Introduction
This paper presents an analysis of empirical ood damage data to Italian economic activities, with the nal aim of improving modelling capability and knowledge of damage mechanisms, towards more reliable and comprehensive ood risk assessment. The study of ood impacts on the different sectors that compose the built environment and the society is crucial to implement actions of risk prevention, protection, mitigation and risk-aware planning; nonetheless, the capability to perform a comprehensive ood damage assessment, including all potentially affected elements and kinds of damage, is still partial.
In particular, despite the businesses sector assumes a critical role, for both its importance for the economic well-being of the society and the high losses it suffers in case of inundations, methods to assess damage to economic activities are much less developed and affected by higher uncertainty compared to other sectors, as the residential (Gissing and Blong, 2004). In fact, studying damage to the business sector means facing a complex problem, involving interconnections among several systems, e.g., the society, the reference market, the nancial system. Indeed, oods may have devastating effects not only on the businesses survival but also on the economic and social fabric (Menoni et al. 2016, Wedawatta et al. 2014).
This paper focuses on a piece of the problem, i.e., the physical impacts of oods on the economic activities, de ned as direct and tangible damage; the scale of analysis refers to the single economic activity, i.e. the micro-scale level. Examples of microscale models for assessing direct ood damage to economic activities can be found both in Europe and Italy (see section 2).
However, on the one hand, existing damage models in Italy are few and scarcely validated, on the other hand, the literature is unanimous in de ning the transferability of damage models as an action that should be done with caution, as it implies signi cant uncertainty, especially in data-scarce regions. Moreover, it is recommended to use models developed in regions similar to that of the initial application, because characteristics of ood, vulnerability or exposed elements and relative values are strongly context speci c (Smith, 1994, Cammerer et al. 2013).
The analysis presented in this paper contributes to ll the gap existing in ood damage modelling capability to economic activities in Italy by investigating empirical damage data collected in the aftermath of ood events. In detail, data refer to four different oods that affected the Italian country in the last twenty years, all characterised by the common feature of being riverine and low-velocity oods.
The aims of the analysis are (i) to obtain a characterisation of the physical damage and the composition of damage to economic activities, (ii) to study the relation between damage and its main explanatory variables, and (iii) to get a preliminary, although rough, estimation of ood damage for different categories of activity.
Results give a clearer idea of the order of magnitude of damage to economic activities due to a ood event in Italy, providing reference damage values according to the activity size, the water level and the type of activity. Estimates supplied by reference damage values have been compared to the observed damage data of the analysed case studies to test their reliability and consistency.
The paper is organised as follows. A description of available damage models for the business sector is rstly provided, in Europe and Italy, along with a brief discussion of their limitations (section 2). Sections 3 and 4 describe the data, the methods and the results of the analysis, which are critically investigated in section 5. Further steps and studies that are required to develop a reliable model for assessing damage to the business sector in Italy are nally discussed in section 6.

State Of The Art
Examples of micro-scale models for assessing direct ood damage to economic activities can be found both in Europe and Italy (Bombelli et al. 2021). For instance, in the European context, FLEMOcs ) is an empirical model based on data collected in Germany, which estimates the loss ratio of buildings, equipment and goods, products and stock for four sectors: public and private services, industry, corporate services and trade. According to the model, losses depend on several hazard and vulnerability variables being water depth, economic sector, company size, precautionary behaviour taken by company owners and the level of contamination of water. In the UK, the Multicoloured Manual (Penning- Rowsell et al. 2005) includes synthetic damage functions to estimate the damage to different categories of non-residential properties, including economic activities (e.g., retail, o ce, distribution/logistics, manufacturing). Susceptibility functions proposed by the model relate damage to water depth, ood duration and activity surface. Finally, France developed national damage functions to assess damage to equipment, stock and structure of economic activities, as a function of water depth and ood duration (Bremond et al. 2018, Grelot andRichert 2019). Functions vary with the category of the activity, identi ed by the NACE code, which is the Statistical Classi cation of Economic Activities adopted in the European Community (Eurostat, 2008). Although properly conceived and calibrated, foreign models can be hardly transferred to the Italian context due to the lack of data for their validation and the absence of databases supplying suitable input data for their implementation (Galliani et al. 2021). In addition, in Italy, tools to assess damage to enterprises are scarce. Three main studies propose methods to assess the  Sardegna. The Sardegna Region (Insular Italy) suffered severe impacts and losses due to oods occurred in November 2013 following heavy thunderstorms and bad weather conditions. Collected data refer in particular to the city of Olbia (Northern-East Sardegna).
Secchia. The ood occurred in the province of Modena (Northern Italy), in January 2014 (Carisi et al. 2018), and was caused by an embankment breach along the river Secchia. Data refer to three municipalities: Bastiglia, Bomporto, and Modena.
Enza. The territory of the Emilia Romagna Region (Northern Italy) was affected by severe weather conditions in the period 8-12 December 2017; in particular, the Enza river breached the embankment in the municipality of Brescello, ooding the town of Lentigione (Regione Emilia Romagna, 2018). Table 1 summarises the available information for the various case studies. In detail, data referring to hazard and damage were collected, for each affected company (i.e., micro-scale level). The former come from the hydraulic simulation of the ood events, providing the perimeter of the ooded area and the spatial distribution of the water depth, for each case study. The latter derive from the declarations lled in by citizens after the ood, claiming for national compensation. In this regard, the information collected for the various case studies is different, for two main reasons: (i) case studies refer to different years and regions, with different regulations and standards for data collection, and (ii) collected data were previously preprocessed by the different authorities (from local to regional) in charge of damage compensation. In Italy, although the recording of damage suffered in case of any natural disaster by public and private structures is required since 1992 (Law Data regarding the case studies of Enza and Sardegna were instead obtained in tabular form, the so-called "C tables", that are standardised tables developed by the National Civil Protection Authority for damage survey to economic activities. For these reasons, only for the events of Lodi and Secchia it was also possible to examine more in detail the types of occurred damage, as shown in Appendix. The main information reported in the claims is the cost incurred by the activities owners to x the damage. Depending on the case studies, this value is indicated for different elements characterizing the activity, which we aggregated into three main damage "components": structure, equipment, and stock. Structure identi es the building with the internal systems necessary to its functioning (e.g., electrical or heating system). Equipment refers to machinery, furniture, vehicles, and tools necessary In this regard, Table 2 shows the types of activities present in the dataset. In Lodi and Sardegna ooding events most of the affected activities were commercial, i.e., G code. In Secchia, the manufacturing sector (C code) was the most affected (27% of the total) followed by the commercial sector with 21% of activities. In Enza event, 33% of affected activities belonged to the construction sector, while 17% of the activities were in the commercial and manufacturing sectors. The analysis of the entire dataset reveals instead that 30% of the ooded activities were commercial, 17% construction, 15% manufacturing, 10% accommodation and food service. Still, data analysis focused only on some NACE categories (Table 2). In detail, category A (agricultural, forestry and shing) that refers to the agricultural sector was not investigated as agriculture is often considered a separate eld in risk analysis . Similarly, the categories D (electricity, gas, steam, and air conditioning supply), E (water supply, sewerage, waste management and remediation activities) and H (transporting and storage) were assigned to the sector of infrastructures, while the categories O (public administration and defence, compulsory social security), P (Education), Q (human health and social work activities) and partly S (other services activities) were considered part of the public sector and strategic infrastructures, and then neglected. The remaining categories (i.e., B, C, F, G, I, J, K, L, M, N, R) were included in the economic activities sector. Despite of the multiplicity of activities categorised as F (Construction), this category was neglected from a detailed investigation, because of its complexity compared to other activities. Indeed, it includes both o ces, warehouses and building sites, that cannot be associated to a "typical" economic activity con guration of one/some premises with contents. The category of the real estate activities (L) was also neglected, because the damage and the exposed value could refer to the several properties owned or managed by the business, that could be spread in the territory. The NACE categories B and R have very few data, so we decided to not investigate them for statistical reasons. The categories J (information and communication), K ( nancial and insurance activities), M (professional, scienti c, and technical activities) and N (administrative and support service activities) were aggregated in a unique macro-category named "O ce", since it was assumed the o ce con guration as prevalent. The categories G (wholesale and retail trade), C (manufacturing) and I (accommodation and food service activities) were considered as distinct macro-categories renamed for simplicity "Commercial", "Manufacturing" and "Restaurant", respectively. It is worth noting that damage to the investigated categories (Manufacturing, Commercial, Restaurant and O ce) is responsible for about 65% of the overall damage observed in the case studies. Table 3 shows the average damage and the standard deviation for the analysed categories of activities, to supply an idea of the order of magnitude of observed damage and of data dispersion.
Other information reported in the claims relates to the address, which was used to geo-localise the activities, the number of employees (only in the Secchia claims), and the surface of the activity for the majority of claims. In addition, GIS based-tools were applied to derive the footprint area of the building in which the activity is located, by processing information on localisation included in the regional topographic geodatabases (i.e. the footprint areas of the buildings).

Method
Data analysis aimed at understanding what happened in terms of physical damage to economic activities during the oods, to provide a characterisation of the impacts and to investigate damage to economic activities exploring the main damage explanatory variables. Three kinds of analysis were performed, on different samples of the dataset, according to the information available for each case study and the objective of the analysis. In any case, outliers were removed before implementing the analysis; in particular, damage values higher than the 75th percentile+1.5 (75th percentile -25th percentile) were neglected. The percentiles were computed on the damage value in €/m 2 differentiated per category of activity.
In the rst analysis, we intersected the information about the activity type with information about damage to the three main components previously identi ed (i.e. structure, equipment and stock) in order to observe if there were similar trends in the different case studies and to compute the average portions of each damage component in the total damage, for activity category (Fig. 1). The latter was computed as the weighted average on the amount of data per case study.
Regarding the relation with water depth, data were divided per components and classes of water level. The latter were de ned to choose signi cant intervals for the expected damage mechanisms. For instance, the de nition of a unique water level class between 0.20 m and 1 m is not useful, as a lot of materials (both equipment and stock) might be positioned at different levels in this range, and then most of the damage would be concentrated in this class. On the other hand, choosing too small intervals means considerably reducing the amount of data in each class. Indeed, classes are limited at 1.5 m because there are little data for higher water levels. Finally, the relation between damage and activity size was investigated for each damage component (Fig. 3), whereas activity size refers to the activity surface. Fig. 1 shows the results of the analysis performed on the activity types and damage components, highlighting that damage composition, for each activity type, shows some common peculiarities in the case studies. The graphs for Commercial show a very similar behaviour in all events: about half of damage is related to stock, 30% to equipment and 20% to structure. A similar behaviour is described in Gissing and Blong (2004) for the ood that occurred in the commercial district of Kempsey (Australia), where damage to building was accounted for 15% and to contents 85% of the total. Even though damage to stock often represents the main portion of damage to commercial businesses, assessing its value is one of the most demanding challenges. The appraisal of the exposed value and the potential damage to stock must consider the variability of goods constituting the stock, the variability of their costs, and the variability of the amount of stock in different periods of time (during the day, the week, or in speci c seasons). However, based on the achieved results, if damage to structure and equipment is known, it is possible to estimate the order of magnitude of damage to stock as well.

Discussion
With respect to the other categories, also the Manufacturing shows similarities in all case studies. Damage to equipment is the most important component: it covers 44% of the total, stock covers 40% and structure covers the smallest part (16%). The category Restaurant shows that costs mainly derive from damage to equipment while damage to stock is less important. For this category, Lodi and Enza case studies are however very unrepresentative because of the low number of data available for the analysis. Concerning O ce, most of costs are caused by damage to equipment with a mean damage of about 50% of the total. Lodi and Enza show a different trend of damage, due to the low amount of data.
For the "O ce" category, considering that most of o ces are located in civil buildings, a suitable strategy to assess damage to structure could be adapting models developed to assess damage to residential buildings. Indeed, damage assessment to the residential sector is based on more developed and well characterised models if compared to those applied to the economic activities ( With respect to the relation of damage with water depth, Fig. 2 shows an increasing trend of damage, in particular when it is not classi ed per activity component (i.e., total damage). Weakest trends observed for damage components may be linked to the low quality of data, related to both damage and activity surface. The former was derived from citizens' declarations that could use different criteria to assess the costs; the latter was partly derived by citizens' declarations, and partly computed as the entire area of the building where the activity is located, which may not coincide with the real activity area. Moreover, many activities declared a null damage (or did not declared damage at all) for some damage components (i.e., structure, equipment and stock). These null values were taken into account in the calculation of the average damage, as we could not speculate on their justi cation (e.g., whether it was actually null or, for example, paid back by insurance companies). We also tried to investigate the relation between damage and water depth per activity category (not shown here), but no trend was visible in this case as this operation implies signi cantly reducing the number of data available in each class. Similarly, the analysis of Gissing and Blong (2004) on the relationships between water depth and damage to contents and buildings did not highlight any signi cant correlation between the variables, although in that case the total damage was considered. Other studies showing increasing trends, as Kreibich et al. 2010, are characterised by a high number of data or do not distinguish among activity types or damage components.
With respect to the relation with the activity size, graphs in Fig. 3 con rm the expectation that damage increases with the activity surface and that the activity size is a signi cant variable in computation of damage magnitude.

Usability of results
Results shown in the previous sections represent a valuable knowledge for a rst estimation of damage to economic activities in the Italian context. Table 4 summarises the main ndings of the analysis as reference damage values that can be used to assess the order of magnitude of damage in case the modeller knows, or not, the surface of the activities. Indeed, Table 4 provides reference damage values expressed in €/unit or €/m 2 . The reference values "1.a" were computed as the ratio between the damage (sum of equipment, stock and structure) and the number of activities for the categories Manufacturing, Commercial, Restaurant and O ce, while the value "1.b" as the mean damage for all the economic activities in the dataset (NACE Code B, C, F, G, I, J, K, L, M, N, R, see section 3). Analogously, damage values "2.a" and "2.b" were computed as the ratio between the damage and the sum of the surface of the activities. Damage per component (i.e., structure, equipment and stock), in both cases "1.a" and "2.a", was derived by multiplying the damage by the percentages in Fig. 1. The reference values "3" can be used in case both the surface of the activity and the water depth are known. For this scenario, the table supplies the expected damage for different ranges of water depth, without differentiating by component and activity type, because a clear relation between damage and water depth was visible only for the total damage (Fig. 2). Table 4 represents, in practice, a solution for a rough estimation of ood damage to economic activities in the Italian context.
We implemented a simpli ed approach for verifying the consistency of results and the estimation bias, applying the model to the original case studies: Lodi, Secchia, Sardegna and Enza. Only Manufacturing, Commercial, Restaurant and O ce are considered for the simulation and only activities with information about water depth and surface were used in the analysis. Table 5 shows the results of the analysis, comparing simulated and observed damage in total terms. Table 5 shows that knowing the surface of the activities improves the damage estimation for most of the case studies. In fact, damages simulated with values "2" are more similar to the observed damage than those with values "1". Damage of Sardegna is underestimated using reference damage expressed in €/m 2 , while that of Enza using reference damage in €/unit. For Sardegna, the average damage values in €/m 2 , calculated with data used for the validation, are signi cantly higher than the reference values, probably due to the low value of surface that characterizes the activities in this case study. For this reason, damage of Sardegna activity is underestimated using the reference values "2". Differently, the case study of Enza contains data with high values of damage and surface, but they are considerably fewer if compared to the other case studies.
For this reason, they have a lower weight in the computation of refence values and the damage is strongly underestimated using the reference value "1", that does not account the surface. Table 5 shows that the information about the activity type (reference model "a") does not improve damage simulation in all case studies (e.g., for Lodi event, the simulations "2.b" is better than simulations "2.a"). Still, Table 6, that compares observed and simulated damage per activity category, without distinguishing among the case studies, shows that the information about the activity type tends to improve the results. Damage in the scenarios "1.b" and "2.b" was computed using the same reference values per activity type. The reference values "3", which consider the water depth, provide similar results to the values "2", since both models consider the activity surface. However, reference values "3" could be useful when mitigation measures must be de ned/designed as they enable to assess the damage for different ranges of water levels. Fig. 4 compares instead point values of observed and simulated damage, for two reference simulations, to give a picture of estimation uncertainty. Scatter plots in Fig. 4 show that, for the single activity, the damage could be signi cantly over or under-estimated. It implies that reference values can be used for estimating the overall damage to all affected companies, but they are not recommended for the estimation of damage at the microscale.

Conclusions
The collection, analysis and comparison of data performed in this study provide new signi cant knowledge on damage to economic activities in Italy, which allow increasing the real capability of de ning potential damage scenarios. The study supplies empirical evidence of the most frequent damage types and the most important components of damage per activity type, which are stock for Commercial, equipment for Manufacturing and Restaurant, and structure for O ce. Results of such an analysis can be used to infer the composition of damage in future events, according to the type of impacted activities, and then to identify which components to target for protection and mitigation actions. Moreover, the knowledge of damage composition allows to appreciate the portion of the total damage that can be assessed with available tools and the portion that cannot. For instance, if we have models to assess damage to structure and equipment for trade activities, we are aware that we are modelling only half of the expected damage.
The main result of this work is however represented by Table 4, which supplies reference damage values in two main implementation scenarios corresponding, respectively, to whether the activity size is known or not. The assessment of estimation error (Table 5) shows that the scenarios that consider the activity surface (2.a, 2.b and 3) give coherent results and lead to an improvement in estimation accuracy; however, when small enterprises are considered (as in the Sardegna case study), underestimation is possible. Table 4 can be used as a rst approach to assess the order of magnitude of potential damage, still, it is necessary to test obtained results in other real case studies, whose data are not used to compute the reference damage values, to evaluate their reliability/uncertainty, before actually delivering the results to decision makers. Future studies will be addressed to enrich the set of observations to verify the robustness of obtained results.
This study did not analyse damage as a function of all the signi cant variables (activity type, water depth, activity size) put together, due to the limited number of data with all the required information. Still, collected data could be processed by a different type of analysis (not implemented in this study), such as machine learning techniques, to identify new patterns or trends in the data. Collaboration with other research groups and further surveys could help to increase the richness of the dataset and to experiment with different data analysis techniques.
Results obtained from this research supply then key knowledge on ood damage mechanisms to economic activities in the Italian context as well as a simpli ed and quantitative representation of the damage. Such information can be used for a rst assessment of damage in quantitative terms, to de ne damage scenarios in risk mapping or cost-bene t analysis, and as the starting points of more sophisticated modelling tools.
Finally, the research leaves a legacy of many questions and propositions for the future, highlighting the need to continue investigating in this research area, towards more comprehensive and complete ood risk analyses.
For the events of Lodi and Secchia, it was also possible to examine more in detail the types of occurred damage. Damages reported in claims were assigned to the components structure, equipment, and stock and the frequency of each damage type was computed. With respect to Lodi, all the 89 claims were analysed (Molinari et al. 2019) ( Table 7). With respect to Secchia, only 28 claims were investigated reporting the partition of the costs per components and a description of loss types (Table 8).
For the other case studies, the individual claims were not available, but only the summary table with the reimbursement information, therefore this detailed analysis was not possible. Table 7and Table 8 list different types of damage documented in the claims for Lodi and Secchia case studies. The frequency is the number of claims in which that type of damage is declared, the percentage is the frequency divided by the total number of claims (89 for Lodi and 28 for Secchia).
The data analysis of Lodi (Table 7) highlighted that the most frequent type of damage to structure refers to internal plaster and walls (approximately 49%), followed by damage to doors (12%), while the most affected installation was the electrical one (approximately 33%). With respect to the equipment, 55% of activities suffered losses to furniture, 38% to instruments for sale or provision of services, and 31% to materials to carry out the activity. Finally, damage to goods mainly referred to nal products or products for sale (with 38%). Table 8 shows that in Secchia claims the most frequent damage to structure was to nishing. Regarding damage to equipment, 60% of activities needed to repurchase new tools and about 50% to repair and to dispose damaged items. With respect to stock, most activities suffered losses of semi-nished and nished goods (29%). Moreover, Secchia claims recorded costs related to technical expenses (as administrative costs, costs for professional consulting, etc.) in almost half of the activities. Such costs were not explicitly recorded in the Lodi claim forms. Still, the Secchia case study highlighted that they can be not negligible. This analysis can suggest self-protection behaviour for the enterprises. For instance, to reduce the frequency of damage to furniture it is suggested to use resistant material, such as metal, or to avoid damage to nished products, it is desirable to place them to elevated positions. Tables   Table 1-Information Figure 1 Damage composition for macro-categories of activities and average composition. n=number of data for each class. Mean damage against water depth classes for all activity types. n=number of data elements for each water depth class.

Figure 3
Mean damage against activity surface classes for all activity types. n=number of data elements for each class.

Figure 4
Scatter plot between the observed damage and the simulated damage by the reference values "2.a" and "2.b" for the categories Manufacturing, Commercial, O ce and Restaurant aggregated