The object-specific flood damage database HOWAS21

The Flood Damage Database HOWAS21 contains object-specific flood damage data resulting from fluvial, pluvial and groundwater flooding in Germany. The datasets incorporate various pieces of information about flood impacts, exposure, vulnerability, and direct tangible damage at properties from several economic sectors. The main purpose of development and 10 design of HOWAS21 is to support forensic flood analysis and the derivation of flood damage estimation models. This paper highlights exemplary analyses to demonstrate the use of HOWAS21 flood damage data in these two application areas. The data applications indicate a large potential of the database for fostering a better understanding and estimation of the consequences of flooding. HOWAS21 recently enlarged its scope and is now also open for international flood damage data.

of EU guidelines for recording disaster impacts with the central aim of translating the Sendai Framework for Disaster Risk Reduction into action (Corbane et al., 2015).
Forensic flood damage analyses as well as damage model derivation and validation predominantly require object-specific data which permit in-depth investigations of causal relations between hazard, exposure, vulnerability, and damage magnitudes (e.g. Downton et al., 2005;Jonkman, 2005). However, object-specific databases are still rare. The Flood Hazard Research Centre 100 (FHRC) from Middlesex University, United Kingdom (UK), maintains a (national) object-specific flood damage database comprising mainly synthetic damage data generated via expert estimations about expected damage in case of a certain flood intensity (what-if-analyses). This synthetic data is complemented, whenever possible, by empirical data sourced from e.g. onsite surveys or insurance companies. The corresponding absolute flood damage functions for the UK are published in the Multi-Coloured Manual (Penning-Rowsell et al., 2013), as well as in its predecessors (Penning-Rowsell and Chatterton, 1977;105 Parker et al., 1987). The Austrian Federal Railways (ÖBB) holds an object-specific flood damage database for railway infrastructure and operation in Austria (Moran et al., 2010;Kellermann et al., 2016). This database incorporates information about the affected infrastructure object and resulting service disruptions, the structural damage and corresponding repair costs, the hazard characteristics, and possible mitigation measures. The ÖBB Natural Hazard Management uses the detailed information internally as a basis for the development and implementation of both structural and non-structural risk reduction 110 measures (Kellermann et al., 2016). In Italy, a new national database was developed and recently integrated into the Italian Civil Protection system. The main objective of this database is the development of specific depth-damage curves for Italian contexts. For this, new procedures for data collection and storage were developed and applied at the local level for the residential and commercial sectors (Molinari et al., 2014). An extended overview of existing flood damage databases, including general characteristics and references for further information, is provided in the supplementary material of this paper (Table  115   S1). The examples given therein do not claim to be exhaustive but serve to illustrate the variety of global efforts to collect (and learn from) natural hazards damage data.
The predecessor database of HOWAS21 was the HOWAS database, which was developed and maintained by the German "Bund/Länder-Arbeitsgemeinschaft Wasser" (LAWA) (see e.g. Buck and Merkel, 1999;Merz et al., 2004). HOWAS contained professionally surveyed damage information of approx. 4000 properties affected by nine flood events between 1978 120 and 1994 in Germany (see also Sect. 3.3). Only direct tangible flood damage to buildings was considered for HOWAS, distinguishing between damage to the building fabric, the fixed inventory, and the movable inventory. Similar to the concept of HOWAS21 (see Sect. 3.1), the data was classified into six damage sectors, i.e. private sector, public infrastructure (e.g. fire station), service sector (e.g. restaurant), mining and building industry (e.g. carpentry), manufacturing (e.g. beverage industry), and buildings for agriculture, forestry and horticulture (Merz et al., 2004). The damage records represented repair costs 125 (building damage) or replacement costs (inventory damage), and all costs were given in absolute values. The object-specific flood damage data is further complemented by additional information about damage-influencing factors (e.g. floor space) whenever possible. Due to missing cost coverage, LAWA stopped the database maintenance in 1994. https://doi.org/10.5194/nhess-2019-420 Preprint. Discussion started: 21 January 2020 c Author(s) 2020. CC BY 4.0 License.

Concept and structure 130
HOWAS21 is a relational database hosted and administrated by the GFZ, which is also responsible for compiling, reviewing, and maintaining consistency of data, assigning access rights and verifying user requests. HOWAS21 does not benefit from funding for data collection or updating and, thus, is largely relying on voluntary data contributions from e.g. surveys and data acquisition campaigns. The use of HOWAS21 follows a community-based concept and is organized in three user groups with different levels of data usage : The World user group is designed for the interested public and grants 135 access to a range of general information and evaluations. Data selection options comprise structured queries, filtered by catchment area, regions (provinces and municipalities), periods (event year), sectors, data acquisition methods, and combinations thereof. Additionally, the HOWAS21 interface offers two registered user groups. Registered user group I is reserved for any kind of institution that provides a defined amount of data to HOWAS21. Full access to the entire database is granted to this user group. Users from academia or non-commercial projects, who do not provide data, can apply for the 140 registered user group II. Those users are granted permission for a restricted project-specific use of the database. In return, a feedback on project results based on HOWAS21 data is requested. Moreover, in case that flood damage data is collected at a later stage, these shall be provided to HOWAS21. The scope of use, the reporting requirements, and the prohibition of data dissemination are regulated via a user contract. 61 users from science, insurance, authorities, and engineering consultancy registered to HOWAS21. Among these, 12 organizations provided data to HOWAS21 (user group I). This ratio indicates that, 145 as yet, the majority of organizational users are mainly interested in extracting data, but hardly willing or able to contribute data to the database.
The data structure for HOWAS21 was derived from a multi-step online expert survey based on the Delphi-approach. The central idea of this approach is to reach a consensus among the respondents by having a questionnaire filled several times, after receiving feedback of earlier responses of all participants. Complementing the HOWAS21 database, a manual outlining 150 The attributes of individual damage cases are grouped into three (partly sector-specific) database tables as shown exemplarily in Table 1. Such attributes include information about the flood characteristics at the location of the affected object (i.e. hazard), object characteristics (i.e. vulnerability), and the extent of damage and damage mitigation (i.e. consequences). Moreover, additional meta-information is provided for each damage case, including information about the flood (e.g. event year, catchment name, seasonality, flood type) and the data acquisition campaign (e.g. survey type, period of the survey, sample 165 description). Minimum data provision requirements for damage cases to be incorporated into HOWAS2l were defined as follows: • economic sector of the affected object year (month) of the event • spatial location of the affected object at least on the level of zip codes or municipalities These requirements are set based on the rationale of ensuring the possibility to link flood damage to hydraulic impact, whereby water depth was found the most important explanatory variable for flood damage in a variety of studies (e.g. Merz et al., 2013;Vogel et al., 2018). 175 In total, HOWAS21 incorporates a broad range of hazard variables (e.g. flow velocity, flood duration, and contamination), vulnerability parameters such as building characteristics (e.g. building shape, year of construction) and precautionary measures (e.g. warning time, type and effectiveness of measures), and flood consequences (e.g. absolute and relative damage of floodaffected objects, economic damage due to business interruption in the commercial sector. The HOWAS21 concept further includes a procedure to determine the general quality of flood damage data. The approach is 180 based on the hierarchical framework of Wang and Strong (1996) and assesses the quality via scores ranging from 0 (poor quality) to 4 (very good quality). More detailed information and examples of the data quality assessment concept applied in HOWAS21 can be found in Kreibich et al. (2017). HOWAS21 data are in an anonymous format respecting personal rights according to data privacy regulations.

Technical design 185
HOWAS21 consists of two major components: the flood damage database and a web application. The database contains and manages access to all available flood damage records. The HOWAS21 web application complements the database by providing a user-friendly, internet-based data access interface. This interface is directly accessible using a standard web browser following the URL http://howas21.gfz-potsdam.de and can be used to visualize, analyze, and download HOWAS21 data. Users are provided query functionality in the database on selectable criteria, such as catchments, damage sectors, or year of event.
The technical design of HOWAS21 aims to take account of the complexity and heterogeneity of flood damage data in multiple sectors. Corresponding metadata is comprehensive and allows for putting the damage information into context. A variety of attributes (e.g. river catchment, flood event year, damage sector) can be used to filter and analyze the data.
Users can register for a HOWAS21 registered user group (see Sect. 3.1) via the web application. If access is granted and a data usage contract is signed, registered users can access the database, analyze, and download data. 195

Data sources
The Flood Damage Database HOWAS21 is designed for empirical flood damage data. A significant part of the data origins from the predecessor database of HOWAS21 -the HOWAS database (see Sect. 2). It was developed and maintained by the German Working Group on water issues of the Federal States and the Federal Government (LAWA) from 1978to 1994(Buck and Merkel, 1999Merz et al., 2004). Damage data of HOWAS were collected via on-site expert surveys by damage surveyors 200 of insurance companies and used as a basis for financial compensation, wherefore these damage estimates are considered to be reliable.
Further, an essential portion of HOWAS21 damage data result from computer-aided telephone interviews (CATI) with private households and companies who suffered flood damage in the years , 2011(e.g. Kienzler et al., 2015Thieken et al., 2016). Potential participants of CATI were identified by compiling lists of affected streets with the 205 help of e.g. flood masks derived from radar satellite data, or publicly available information such as official reports and press releases (e.g. Kreibich et al., 2007;Thieken et al., 2007;Kreibich et al., 2011). The interviews were mainly carried out by pollsters and, for every interview, it was consistently sought to consult the person with the best knowledge about the flood event as well as the resulting object-specific damage.
For the damage sectors "public thoroughfare" and "watercourses and hydraulic structures", only a few damage datasets are 210 implemented in HOWAS21 so far. All these datasets are collected via on-site expert inspection after the 2002 food in Dresden.
Damage data for public thoroughfare comprises 246 inundated sections of road infrastructure in the City of Dresden affected during the Elbe river flood in 2002. The dataset includes physical road characteristics (e.g. length, width, sidewalks), the road classifications, and additional object features. With respect to damage the data was collected in two steps: First, the absolute monetary damage was derived from reports of the city administration providing the reconstruction costs of affected road 215 sections, whereby no unit length is defined for the road sections. Second, the magnitude of structural damage was quantified on a six-point scale, and the condition of the road before the flood was quantified on a five-point scale by experts from the city administration (Kreibich et al., 2009). A similar procedure was applied for the 525 damage cases along watercourses and hydraulic structures in the City of Dresden also affected in the course of the 2002 Elbe flood.

Exemplary analyses 220
HOWAS21 aims at compiling comprehensive flood damage data (i.e. including object-specific hazard, exposure, and vulnerability characteristics) to support forensic flood damage analyses as well as flood damage model derivation. In this section, available HOWAS21 flood damage data are characterized and used for exemplary analyses to demonstrate the potential of the database in both of these application areas.
The up-to-dateness of the data used for the analyses is November 01, 2019. Damage values distinguish between building 225 damage, contents damage, and damage to goods and stock, whereby damage to goods and stock is only defined for the commercial and industrial sector. Further, damage in the sectors public thoroughfare and water courses and hydraulic structures is by definition classified as building damage. All costs are given in Euros and the reference year of an individual cost value is the year of the related flood event occurrence. Thus, in order to achieve comparability, all costs were converted to the year 2018. Conversion factors are the price indexes for construction works on residential buildings and the consumer price indexes 230 for replacement costs of household contents as well as commercial and industrial goods and stock, both published by the Federal Statistical Office Germany.

General descriptive statistics
HOWAS21 comprises a total number of 8558 object-specific flood damage records from flood events between 1978 and 2013 in Germany. The geographical distribution of the damage data is depicted in Fig. 1. The private households sector accounts 235 for the by far largest data fraction (57.1%), followed by the commercial and industrial sector (33.9%). The sectors water courses and hydraulic structures (6.1%), and public thoroughfare (2.9%) are as yet rather underrepresented (see Table 2). No data are yet available for the sectors agricultural and forested land, and urban open spaces. In fact, the commercial and industrial sector does contain a small number of flood affected agricultural buildings.
Both in respect to the number of damage records and the level of detail of information, i.e. the number of different hazard, 240 exposure, and vulnerability variables, HOWAS21 is the most comprehensive flood damage database for empirical data worldwide. Although the current set of damage records provides data for a number of variables exceeding the minimum requirements of HOWAS21 (see Sect. 3.1), most of the records are far from exhaustively filling the defined sector-specific variable space. For example, for any damage record in the private households sector, there is currently a 42% chance of data availability for non-mandatory variables. This chance even decreases to around 22% for the commercial and industrial sector. 245 In turn, for certain other non-mandatory variables such as building type or building shape (in the variable space for the private households sector) the data availability is close to 100%. Generally, the availability of detailed flood damage data is often limited due to the facts that damage data collection in the aftermath of a flood is not mandatory, sufficient and properly trained personnel is mostly not available, and collection standards do not exist (Menoni et al., 2016;Thieken et al., 2016).

Gaining insights into flood damage processes
Forensic analysis techniques attract growing interest in science and risk management since they help to uncover the complex underlying causes and effects of disasters (Wenzel et al., 2013;Dolan et al., 2017). More specifically, such analyses are performed e.g. to quantify the relative contribution of damage drivers such as exposure, vulnerability and coping capacity to the overall damage. Other applications include the assessment of interdependencies of damage drivers, and the change of the 265 correlation between a specific damage driver (e.g. water depth) and the resulting consequence (damage) over time (e.g. between different events in the same region).

How do potential flood damage-influencing variables interact?
An important task for a better understanding of flood damage processes is the investigation of potential damage influencing variables and their interactions. For this, the correlation structure of such variables and their contribution to the total variance in the HOWAS21 data (related to building damage in the private households sector) is examined by means of a PCA. A Principal Component (PC) is a normalized (z-transformation) linear combination of the original variables capturing the 280 maximum variance in a dataset. Hence, each PC explains a certain percentage of the total variance in the dataset, whereby the first PC by definition explains the largest share of total variance, the second PC explains the second largest share, and so forth.
Moreover, all PCs are uncorrelated (i.e. perpendicular) to each other and, altogether, they reflect the underlying structure in the given dataset.
All damage influencing variables for which non-categorical data are (currently) available in HOWAS21 were considered for 285 the PCA, and variables quantifying flood consequences (e.g. absolute building damage in Euros) were neglected. The data sample was furthermore centered and scaled to avoid bias in the variance of the data due to variable scale mismatches and, consequently, false estimations of the PC.
According to the Kaiser criterion as well as the scree plot, four significant PCs can be identified in the HOWAS21 data sample.
They explain around 59.9% of the total variance, whereby approx. one third of this variance can be attributed to the first PC 290 (21.4% of total variance). In order to facilitate the interpretation of the variable contributions to each PC and, thus, to gain insights into the interaction of variables a varimax rotation was applied. Table 3  In order to supplement the investigation of variable interactions with an estimation of the influence of the PCs on flood damage, the correlation between the factor scores of each PC and the absolute building damage was analysed (Table 3). Results show that absolute building damage correlates best with PC1, i.e. the component driven by the flood impact variable water depth.
Lower (but still statistically significant) correlation is further given to PC2, PC3, and PC4.
The PCA shows that flood impacts variables, particularly water depth, are the factors with highest influence on absolute 305 building damage. These are, however, closely followed by a variety of exposure and vulnerability characteristics, in particular lead time, year of building construction, and building type. When looking at the explained variance of only around 59.9% for the extracted PCs, the results further indicate that a large number of variables is needed to sufficiently explain the total variance in the HOWAS21 data for fluvial flood damage to private households. This supports the results of e.g. Schröter et al. (2014) showing that flood damage processes are intrinsically complex and, thus, can be better described using a variety of explanatory 310 variables representing different flood damage processes. Consequently, the findings basically underline the importance of a comprehensive damage data collection approach as followed by HOWAS21.

How important are individual variables for multivariate flood damage estimations?
Building on the insights into variable interactions in the HOWAS21 data (see Sect. 4.2.1), a logical next step towards a better understanding of flood damage processes is to investigate the individual variable importance for multivariate damage modeling. To do so, the Random Forest variable importance algorithm was used to identify non-monotonic and multivariate relationships of the damage-influencing variables (i.e. the predictors) to estimate absolute building damage in the private households sector. However, in contrast to the variable selection used for the PCA (see Sect. 4.2.1) and due to the applicability in a Random Forest framework, also categorical variables (i.e. roof type and building shape) are included in this analysis.
The variable importance of a predictor is estimated by random permutation of values of this particular variable. The idea is 320 that this random permutation leads to an increase in the prediction error compared with the error generated by original values.
Accordingly, the mean increase in prediction error caused by permutation of a certain variable serves as a measure for the importance of this particular variable. Overall, similar to what could be inferred from the PCA with regard to variable interactions and variance in the data (see Sect.

Has the relation of flood impact and resulting damage changed over the years?
Another important step towards a more thorough understanding of flood damage processes is seen in the examination of the process dynamics. For example, by comparing the functional relationship between flood damage drivers and the resulting damage for different years of flooding, potential patterns or trends over time could be identified. Against this background, the 340 following exemplary analysis is aimed at identifying potential trends in the linear relation between water depth and absolute building damage to private households between different flood years using multilevel regression.
Multilevel models, also known as e.g. mixed models, hierarchical models or group-effects models, are useful for analyses involving hierarchical, nested, clustered, or longitudinal data. Hierarchical data, for example, consist of units (e.g. floodaffected objects) which can be grouped into other units (e.g. flood events), whereby the grouped units represent a distinct data 345 level (i.e. level 1), and the grouping unit forms a superior data level in the hierarchy (i.e. level 2). Multilevel models allow e.g.
to assess the amount of data variability due to each data level and, thus, to explicitly identify and investigate group effects. Thus, these model features are useful also for forensic flood damage analyses and, in particular, for investigating changes of relations between a predictor variable (e.g. water depth) and the response variable (e.g. absolute building damage) from one flood event to another (i.e. the group effect in our case). More detailed information about principles of multilevel models can 350 be found in e.g. Gelman and Hill (2006).
In a first step, the relevance of group effects in the HOWAS21 data sample is measured by means of the Intra-class Correlation Coefficient (ICC). The ICC quantifies the ratio of variance on the hierarchical data level 1, i.e. the level of flood-affected objects, being explained by data level 2, i.e. the flood events. For example, an ICC value of 0.2 indicates that 20% of the total data variance lies between the groups, and 80% within the groups accordingly. Common practice suggests to consider ICC 355 values of 0.05 or higher as an indicator for significant group effects (LeBreton and Senter, 2008). In such cases multilevel models should be favored over simple linear models. Since the ICC value of the damage data of absolute building damage to private households amount to approx. 0.08, a multilevel linear regression model is applied to further investigate the expected changes in the depth-damage relation between individual flood events.
The scatter plot of water depth and damage to building structure including multilevel linear regression lines is shown in Fig.  360 4. Looking at the lines reveals a more or less continuous increase of both the intercept and the slope with increasing flood event year. This trends can be further investigated when plotting the group effects of the model (see Fig. 4). Three main findings emerge: First, the high group variability of both regression parameters among the 13 flood years clearly confirms significant group effects as already suggested by the ICC. Second, the range of group-specific residuals indicate that, on the whole, flood damage data of years of flooding further back in time tend to scatter more compared with more recent events in 365 the 21st century (see Fig. 4). Third, the development of the group effects of depth-damage relations over time points to an overall increasing trend in the flood damage (see also Fig. 3). In other words, the majority of regression models derived from flood events, which occurred in the 21st century, show higher intercepts as well as higher slopes relative to the values of a simple linear regression model, whereas flood events of the 20th century mostly have lower parameter values, respectively. Summing up, based on water depth as the determinant and absolute building damage as the response variable, multilevel 375 regression results reveal considerable flood damage process dynamics between individual flood years, which is manifested by changes in both the intercept and the slope of regression lines from one flood year to another (see Fig. 4 and Fig. 5). These changes or, more specifically, the overall positive trend in the occurred flood damage with time, can have manifold reasons such as i) the increase in event severity in terms of depth and/or area, ii) the increase in exposure in terms of number of objects and/or asset values, iii) the increase in vulnerability of affected objects, and iv) changes of data collection methods. More 380 (forensic) analyses would be required to better attribute the observed trend to potential causes. For example, the current multilevel model is based on the entire HOWAS21 dataset for the private households sector, i.e. involving a region-unspecific series of flood years. In order to reduce the effects of changes of exposure characteristics and, thus, to focus more on hazard and vulnerability-related influences, the input data for regression could be limited to a series of flood events that occurred in the same region (e.g. river catchment or Federal State). More sophisticated approaches to assess spatio-temporal variability in 385 flood damage processes are presented in Sairam et al. (2019).

Model derivation and validation
The following section presents examples of different flood damage models that can be derived on the basis of HOWAS21 data, and briefly evaluates their performance. These include 1) a variety of univariate, i.e. linear, polynomial, and square-root, depth-damage curves for all four economic sectors as well as all damage types available in HOWAS21 (i.e. damage to 390 buildings, damage to contents, damage to goods and stock), and 2) a multi-variate Random Forest regression model for absolute damage to building structures in the private households sector.
Univariate depth-damage curves as well as underlying data are plotted in Fig. 6. It appears that for most of the combinations of economic sector and damage type the different regression types result in similar depth-damage curve progressions. An exception is, however, the public thoroughfare sector, for which the polynomial regression curve is noticeably undulating. 395 Generally, the visual evaluation of the curves fitted to the considerably scattered data samples already indicates that univariate depth-damage relations, i.e. the use of water depth as a sole predictor, only partly (and often insufficiently) explain the complexity of flood damage processes.
The suitability of the models to estimate absolute flood damage was evaluated by means of three different error measures: the Mean Bias error (MBE), the Mean Absolute Error (MAE), and the Root Mean Square Error (RMSE). These are calculated 400 based on one third of the respective original data sample as validation, whereby the remaining two thirds were used for the model derivation.
The error statistics for all models are summarized in Table 4. Generally, according to RMSE and MAE, the three univariate regression types perform similarly in estimating absolute flood damage for all damage types. Based on the MBE, the model performances in some cases differ significantly, whereby, however, none of the models consistently performs best (or worst) 405 across all sectors and damage types. Also, all univariate models except the models for damage to contents in the private households sector show a MAE being in a comparable order of magnitude as the observed mean damage of the sector (see Fig.   2 and Table 4).
When used for regression, Random Forests are ensembles of a (large) number of regression trees. Each regression tree is constructed by recursive binary splits of a bootstrap sample from the original data called bagging, and each binary split can be 410 related to any predictor variable at any value. The data not included in the bootstrap sample to train a regression tree, the socalled Out-Of-Bag observations (OOB), are used for the calculation of model performance measures as well as estimations of predictor variable importance (see Sect. 4.2.2). A more detailed description of the Random Forest algorithm can be found e.g.
in Breiman (2001). Random Forests are generally seen useful for flood damage modeling, since they are applicable to both categorical and continuous data, they allow for non-linear and non-monotonous input data, and they are able to capture 415 predictor interactions (e.g. Merz et al., 2013;Schröter et al., 2014;Kreibich et al., 2016;Sieg et al., 2017;Sultana et al., 2018).
Due to the low importance of the predictor variables flow velocity class, number of floors and roof type (see Sect. 4.2.2,Fig. 3), they are excluded from the Random Forest model derivation. Consequently, the Random Forest model is learned on the basis of eight predictor variables, of which two variables (water depth and flood duration) are hazard characteristics, five variables (building type, building shape, building area, year of construction, equipment class) represent exposure 420 characteristics, and one variable (lead time) addresses damage mitigation.
In comparison with the univariate modeling approaches, the multivariate Random Forest model shows better results, although the estimation errors are still very high when viewed in relation to the observed mean damage (see Fig. 2 and Table 4). Overall, the findings of this model derivation and validation exercise suggest a limited capacity of univariate models to explain the complex flood damage processes. The error measures indicate large estimation uncertainties for all univariate models, whereby 425 the regression type (i.e. linear, square-root, and polynomial) has only marginal influence on the individual model performance irrespective of the economic sector. The largest errors are observed in the commercial and industrial sector, which can be explained by the strong heterogeneity of this sector (see Fig. 2, Sieg et al., 2017). Already in 1999, using damage data from the predecessor database HOWAS, the Department of Water Resources Management and Rural Engineering of the University of Karlsruhe (IWK) showed that the derivation of generally valid damage functions is difficult, in particular for damage 430 categories being poorly represented. The lack of comprehensiveness of flood damage data also led Merz et al. (2004) to the conclusions that the HOWAS database is not totally representative for flood damage in Germany, and that the use of empirical flood damage data involves considerable uncertainties. They therefore claimed that i) flood damage data should be collected at the object level whenever possible to better support the development and validation of flood damage models, and ii) the data base should be enlarged with regard to the number of variables to follow a more comprehensive and systematic data collection 435 approach. Both of these claims were also directly addressed in the successor database HOWAS21. Indeed, adding new predictors significantly improves the flood damage estimates as can be clearly seen from the example of the Random Forest performance in the private households sector (see Table 4) -whereby, although the Random Forest estimates are still subject to considerable uncertainties. Again, this result is in line with Schröter et al. (2014) showing that complex models better capture the multidimensional nature of flood damage processes. 440

Conclusions
The Flood Damage Database HOWAS21 incorporates object-specific data about flood impact, exposure, vulnerability, and direct tangible damage in various economic sectors resulting from fluvial, pluvial and groundwater flooding in Germany. Its strengths include data quality features, the compliance with strict minimum requirements for data entries, the integration of sectors such as public thoroughfare or watercourses being widely underrepresented in damage data collections, and the 445 consideration of a multitude of damage-influencing variables.
These features are also essential and integral components of the HOWAS21 concept to support forensic flood damage analyses as well as the development of flood damage models. The exemplary analyses presented in this paper give a hint on the large potential of this database for such application fields. They further confirm two central findings of other relevant studies, i.e. the fundamental role of the hazard variable water depth to estimate flood damage on the one hand, and the need for a variety 450 of different explanatory variables to better understand and describe the intrinsically complex flood damage processes on the other hand.
The HOWAS21 database has been further developed and optimized for the hosting of object-specific flood damage data with a global scope. This extension of scope of HOWAS21 includes, inter alia, the incorporation of a globally valid spatial identifier and the international standard classification for economic activities. 455 However, it depends on the cooperation and commitment of the (scientific) community to provide flood damage data to HOWAS21, so that the empirical flood damage database continuously growths and as such increases its value for the whole community. The higher the amount of data and the more diverse the contained data is (e.g. from different flood types and regions, covering various sectors), the better it can support forensic flood damage analyses as well as the development of flood damage models. Therefore, if flood damage data is or becomes available, we expressly encourage data owners to include it 460 into HOWAS21 for their own benefit (i.e. getting access to all data contained in the database) but even more important for the benefit of the whole community. Check out HOWSAS21 via http://howas21.gfz-potsdam.de/howas21/ or e-mail howas21@gfz-potsdam.de.