Surface water floods in Switzerland : what insurance claim records tell us about the damage in space and time

Surface water floods (SWFs) have received increasing attention in the recent years. Nevertheless, we still know relatively little about where, when and why such floods occur and cause damage, largely due to a lack of data but to some degree also because of terminological ambiguities. Therefore, in a preparatory step, we summarize related terms and identify the need for unequivocal terminology across disciplines and international boundaries in order to bring the science together. Thereafter, we introduce a large (n= 63117), long (10–33 years) and representative (48 % of all Swiss buildings covered) data set of spatially explicit Swiss insurance flood claims. Based on registered flood damage to buildings, the main aims of this study are twofold: First, we introduce a method to differentiate damage caused by SWFs and fluvial floods based on the geographical location of each damaged object in relation to flood hazard maps and the hydrological network. Second, we analyze the data with respect to their spatial and temporal distributions aimed at quantitatively answering the fundamental questions of how relevant SWF damage really is, as well as where and when it occurs in space and time. This study reveals that SWFs are responsible for at least 45 % of the flood damage to buildings and 23 % of the associated direct tangible losses, whereas lower losses per claim are responsible for the lower loss share. The Swiss lowlands are affected more heavily by SWFs than the alpine regions. At the same time, the results show that the damage claims and associated losses are not evenly distributed within each region either. Damage caused by SWFs occurs by far most frequently in summer in almost all regions. The normalized SWF damage of all regions shows no significant upward trend between 1993 and 2013. We conclude that SWFs are in fact a highly relevant process in Switzerland that should receive similar attention like fluvial flood hazards. Moreover, as SWF damage almost always coincides with fluvial flood damage, we suggest considering SWFs, like fluvial floods, as integrated processes of our catchments.

before. In Sect. 3 we describe the data in detail and introduce the methodology to differentiate SWF damages from fluvial flood damages. Thereafter, in Sect. 4, we present general characteristics of the number of claims as well as associated loss caused by SWFs in comparison to fluvial floods. Furthermore, we present the spatial and temporal characteristics of damages caused by SWFs in Switzerland during the last decades and discuss the results in Sect. 5. Finally, by providing concluding remarks, we conclude the study (Sect. 6).

2 Terminology
As precipitation reaches the land surface, different runoff generation mechanisms determine whether surface runoff is generated (e.g., Fiener et al., 2013). The water may then take several routes towards the stream channels (Ward and Robinson, 2000), as depicted in Fig. 1. The flow path along the land surface is sometimes ambiguously referred to as "surface runoff", but is better defined by the widely-used term "overland flow" (Ward and Robinson, 2000). However, in the literature, this distinction 10 is inconsistently made, whereas either of the terms or even both are used. We adopt the term overland flow and, thereby, mean the transport of water downhill at the land surface as thin sheet flow or anastomosing braids of rivulets and trickles until the water reaches or is concentrated into recognizable streams (Chow et al., 1988;Ward and Robinson, 2000;Brutsaert, 2005). The propagation and accumulation (i.e. ponding) of overland flow can be considered as a flood, which in the glossary of Field et al. (2012) is defined as "the overflowing of the normal confines of a stream or other body of water, or the accumulation of water 15 over areas that are not normally submerged". However, we have to keep in mind that the term "flood" is sometimes implicitly used in the hydrological sense, but sometimes also in the context of "damaging floods" (Barredo, 2009). In the former case, any inundation of land is considered, while in the latter case, the flood necessarily interacts with the societal system causing adverse effects (Barredo, 2009). Obviously, damage data are inherently only reflecting damaging floods. Consequently, we have to bear in mind that by using damage data, we can only draw direct conclusions about damaging floods, and not hydrological floods 20 in general. Kron et al. (2012) noted the importance of consistently using well-defined terms for addressing the perils listed in disaster data bases. This is not just true for data, but for scientific research in general. For instance, Boardman (2010) has identified the lack of terminology as a possible reason for the accounts of specific flooding (i.e. muddy flooding) in some countries and the seeming non-existence in other areas. As this might also partly explain why SWFs have received attention in some countries 25 but remain unrecognized in others, we have identified the need for a short terminological elaboration, following hereafter.
SWFs are caused by intense rainfall that, due to whatever reason, cannot be drained by natural or artificial drainage systems and, thus, ponds in local depressions or propagates along the surface as overland flow (Pitt, 2008;Hurford et al., 2012), before it possibly, but not necessarily, reaches or is concentrated into regular watercourses (Fig. 1). The term is often used synonymously with "pluvial floods", although according to Falconer et al. (2009), SWFs have a broader meaning. Namely, the term does not 30 only include pluvial flooding, but also flooding from the sewer, small open or culverted watercourses as well as flooding from groundwater springs (Hankin et al., 2008;Falconer et al., 2009). Therefore, SWFs can be thought of as the most general form of rainfall-related (non-fluvial) floods. In this study, we analyze damages caused by SWFs, as defined above.  Figure 1. As precipitation reaches the land surface, it may directly fall into a stream or take different routes towards it, governed by the relevant runoff generation mechanism. Thereby, overland flow and ponding constitute a SWF that has not yet or will never reach a watercourse.
In contrast, a possible fluvial flood originates from the body of water itself, when the normal confinements are overflowed. Moreover, we note that all floods may or may not cause damages. Understandably, studies based on damage data analyze a subset of all floods, i.e. only those for which some sort of damage has been registered.

5
According to Andrieu et al. (2004) and Douguédroit (2008), the main difference between flash floods and "urban floods" is that in the former case the flood originates from watercourses, while in the latter case overland flow is generated within the urbanized area itself. Urban floods are also referred to as an "intra-urban floods" (Evans et al., 2004;Hankin et al., 2008) and flash floods affecting urban areas are sometimes called "urban flash floods" (e.g., Hankin et al., 2008;DWA, 2013;Zhou et al., 2013) or are even used synonymously to urban floods (Kron et al., 2012), owing to a broad definition of flash floods in the first 10 place.
Another non-fluvial flood type is "muddy flooding". According to Boardman (2010), such floods are formed by muddy runoff from agricultural fields that damage adjacent properties downslope. The term is well-established (Ledermann et al., 2010).
Overall, many of the terms used to address different flood types are either used ambiguously in the literature or are not well- 15 defined. Flooding is a complex interlinked system, affecting many aspects of the physical, economic and social environments acting at different spatial and temporal scales (Evans et al., 2004;Barredo, 2009). As such, flooding involves a wide range of 5 Nat. Hazards Earth Syst. Sci. Discuss., doi:10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License.
interconnected hydraulic subsystems and processes (Evans et al., 2004). Therefore, it is understandable that the classification of such a complex process like flooding is not simple, particularly in practice. It seem all the more important that the used terms and the corresponding definitions are well documented. This might reveal terminological ambiguities and, ultimately, make meaningful comparisons possible (Gall et al., 2009).

5
The compiled data set is based on flood damage claim records from 14 different PICB. In addition, we obtained similar records from the cooperative insurance company Swiss Mobiliar, which are not part of this study's data analyses. Yet, they are used to set up the claims' classification scheme alongside the data from the PICB (c.f. Sect. 3.1). As mentioned before, each PICB holds a monopoly position and, thus, insures virtually every single building within the respective canton against various natural hazards including flooding. Thereby, damages caused by water entering the building envelope at the surface are insured, while damages associated with direct intrusion of groundwater or backwater from the sewer, as well as flooding from dams or other artificial water structures are generally excluded. In consequence, water-related damages covered by PICB are either caused by SWFs or fluvial floods, whereas the insurance companies themselves do not differentiate the two processes (Imhof, 2011).
Therefore, similar to other studies (e.g., Spekkers et al., 2013Spekkers et al., , 2015Grahn and Nyberg, 2017), the data have to be classified first. However, in contrast to the aforementioned studies, the claim records were provided in a spatially explicit way, enabling 15 a classification based on each claim's geographical context.
Following the data processing procedure depicted in Fig. 2, we first describe the compiled data set as well as the harmonization and geocoding thereof (Sect. 3.1). Following, we introduce the methodology to differentiate claims associated with SWFs and fluvial floods (Sect. 3.2) and, thereafter, we discuss the necessary normalizations of the data (Sect. 3.3). Note that the classification scheme is described as generally as possible, to make its application to other contexts and countries as straightforward 20 as possible. However, it could not be prevented that the classification scheme is adapted to some national characteristics, in particular concerning the properties of the considered Swiss flood maps. The specific input data for each data processing step listed in Fig. 2 are described in detail in Table 1. Figure 3 gives an overview of the compiled data set and illustrates all 19 cantons with a PICB, while the 14 PICB that 25 provided data are highlighted additionally. As the cantons' borders have mostly administrative meaning, we adopted the natural landscape units from Grosjean (1975), while constraining the borders to hydrological catchment boundaries. In this study, the data are analyzed with respect to these regions (Fig. 3). Overall, 43-100 % of the buildings are covered by our data set, with the exception of the Western Inner Alps (0 %) and the Southern Alps (6 %). The low values of the latter two regions are owed to the fact that practically no buildings are insured by a PICB within these areas. Consequently, these areas are excluded from this 30 study's analyses, even though some claims provided by the Swiss Mobiliar covered this region. The provided data provided by the Swiss Mobiliar contain flood damage claim records of content and, additionally, of property in cantons with no PICB.  D1: 13 PICB, 1999D1: 13 PICB, 1999D2: 14 PICB, 1993 Process Swiss Mobiliar Figure 2. Illustration of the main data processing steps (boxes) as well as the required input data, which are further specified in Table 1. D0, D1 and D2 refer to the data (sub-) sets, which were used to produce the output, illustrated by this study's tables and figures. Note that D0 constitutes the complete data set including data from 14 PICB in addition to data from the Swiss Mobiliar, whereas D1 and D2 consist of PICB data only, limited to the indicated periods (c.f. Table 2). The empirical cumulative distribution function (ECDF), as well as the altitude constrained Euclidean distance (ACED) between each claim and the next river are abbreviated (c.f. Sect. 3.2).

Data
These records have quite similar characteristics as the data provided by the PICB, but are not limited to certain cantons and, thus, extend over the whole of Switzerland. However, as pointed out in the introduction, records of insurance companies with certain (unknown) market shares are much more challenging to interpret. Nevertheless, the data are useful for setting up the 7 Nat. Hazards Earth Syst. Sci. Discuss., doi:10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci.    Table 2). (a) Cantons with and without a PICB, as well as an indication which PICB provided data. The latter are additionally marked with an asterisk (*) in the legend. (b) Natural landscape units based on Grosjean (1975), which are used to analyze the data on a regional scale. As almost no buildings are insured by PICB within the Western Inner Alps as well as the Southern Alps, these two regions are excluded from the data analyses, as indicated by the domain. classification scheme, because every additional claim generally increases the method's robustness (c.f. Sect. 3.2). The data from the Swiss Mobiliar are included in the data set D0, which is used for the classification scheme (c.f. Table 2).
The minimal information of each flood damage claim includes the damage date, the location of the damage (address or coordinates) as well as the associated direct tangible loss to the respective building. As the claim data stem from 15 different data sources (14 PICB and data from the Swiss Mobiliar), the provided raw data are heterogeneous and need to be harmonized 5 first, as indicated in Fig. 2 (c.f. Bernet et al., 2016, for details). During this procedure the data were quality checked, obvious errors were corrected, if possible, or removed otherwise. Moreover, the loss values were corrected for inflation as of 2013 using index values of the corresponding PICB, in case the source data had not been indexed already.
During the next step, each damage claim is geocoded (Fig. 2). The coordinates of each damaged building could be obtained by matching the corresponding address with a geocoded register of all Swiss postal addresses (c.f. Table 1). Notably, only the 10 claims with an unique match were analyzed later. As the data quality of the addresses varies among the different PICB, the 8 Nat. Hazards Earth Syst. Sci. Discuss., doi:10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License. amount of claims that could be localized at the building level varies as well (Table 2). Nevertheless, most of all PICB claims (79 %) could be localized. A summary of the compiled (sub-) data set is given in Table 2.

Classification
The basic idea behind the classification scheme is simple: In case a building (and/or its content) has been damaged by flooding and was located far away from any watercourse, it is very likely that the damage was caused by a SWF. The opposite is not 5 necessarily true: Overland flow is propagating over the land surface towards the watercourses and might cause damages along the flow path until it reaches the next watercourse (Fig. 1). Thus, for damaged objects close to a watercourse it is difficult to deduce the responsible flood type without studying each case in detail. Given the size of the data set, detailed manual classification is not practical, in addition to the fact that the data generally do not contain additional information about the responsible damage causes. 10 In order to classify the claims pragmatically, we exploit the damages' known locations as follows: We assume that the dominant damage process in known fluvial flood zones are fluvial floods and, thus, damaged objects located within such zones were likely affected by this process. As these damages are inherently clustered around watercourses, we make use of this characteristic by assessing the distance between these damages and the next river. We then classify the claims outside of known flood zones based on how their own distance to the next river relates to the typical distances obtained from fluvial flood 15 damages. Thereby, the question is how these distance should be measured and how a representative cut-off distance can be determined.
We tested different distance measures, whereas the Euclidean distance performed well for instance, but neglected topography altogether. For instance, a building on a ridge can be associated with a short Euclidean distance to the next river, in spite of being safe from river flooding due to the building's elevated location. We therefore chose the following approach to address 20 this issue, while at the same time making use of the Euclidean distance's simplicity: Before calculating the Euclidean distance to the next river, we first masked the river network with areas lower than the respective object's altitude using a digital elevation model (DEM , Table 1). Only then, the Euclidean distance to the masked river network is assessed. The obtained quantity is hereafter referred to as "altitude constrained Euclidean distance" (ACED).
Typical distances for all fluvial flood damages can then easily be obtained by analyzing the ACEDs of all claims located 25 within known flood zones. For that matter, we selected all claims within such flood zones and compiled the empirical cumulative distribution function (ECDF) of the ACEDs. Based on the large data set, we can be confident that the claims located farther away from the closest river than the 99th percentile of the respective ECDF were caused by SWFs. Considering that fluvial floods become generally more probable the closer we get to the rivers, we chose evenly spaced percentiles, i.e. the 25th, the 50th, 75th and the 99th percentile, respectively. To take regional geographical characteristics into account, the percentiles 30 are calculated for each region separately (Table 3).
Inherently, flood claims also include damages caused by overflowing lakes, which could not be distinguished easily from fluvial floods. Consequently, damages related to lakes will be associated with a certain distance to the next river, even though the corresponding river was not the cause of the damage. As this rather tends to shift the ECDF to the right, in addition to the   Table 3. low number of such cases, it is safe to assume that this influence is negligible. Moreover, the hazard of overflowing lakes is consistently considered in the fluvial flood maps. Consequently, the damages caused by overflowing lakes are located within mapped flood zones and are, therefore, directly and correctly classified as fluvial floods (c.f. Fig. 5).
Using the precompiled percentiles (Table 3) as well as fluvial flood maps (Table 1)  it is becoming gradually more unlikely that an object is affected by fluvial floods the farther away an object is located from a river.
As outlined in Fig Thus, to increase the coverage, we used the aforementioned map called Aquaprotect (Table 1) No Intersects hazard map perimeter?
ACED to the closest river (d) Intersects Aquaportect flood zone?
ACED to the closest river (d)  Table 1). In all other cases, the specific ACED (d) of each claim is compared to the typical ACEDs of fluvial flood damages (d25, d50, d75 and d99, c.f. Table 3). The classification scheme is further illustrated by Fig. 6.
Aquaprotect is only used for the territory not covered by the flood hazard maps. Namely, the hazard map perimeters have been extracted from the Aquaprotect layer using common GIS tools.
It should be noted that the areas not covered by flood zones, i.e. the hazard-free zones, have similar implications for the two different sources. The smallest rivers were not considered by Aquaprotect, but there is no objective way of knowing where the cut-off was set. This also holds true for the flood hazard maps, as the study of a few examples revealed. Moreover, the

5
Reported increasing trends of flood losses (e.g., Kron et al., 2012;Grahn and Nyberg, 2017) might be misleading. In fact, there is evidence that increasing flood losses are mainly owed to socio-economic development rather than trends in the flood processes itself (Barredo, 2009). Increasing losses caused by natural hazards such as flooding can, thus, mostly be attributed to increasing population and expansion into hazardous areas (e.g., Cutter and Emrich, 2005;Barredo, 2009;Bouwer, 2011;Kundzewicz et al., 2014), as well as increasing property values and diminishing awareness about such hazards (Kundzewicz 10 et al., 2014) and, additionally, better documentation of damages in the more recent past (Gall et al., 2009). Consequently, the loss data need to be normalized with regard to such effects, if the natural process rather than the product with the socioeconomic background is of interest. The most fundamental normalization is to adjust past losses to the current values (Kron et al., 2012). However, the more difficult part is to remove the influence of socio-economic development on the observed number of damages as well as the associated loss. In addition, the consideration of a change in the exposed objects' vulnerabilities 15 is even more difficult (Bouwer, 2011).
In this study, the values are adjusted for inflation during the harmonization procedure (Sect. 3.1). Furthermore, the absolute damage data are normalized in space by relating them to the number of buildings and the sum insured as of 2013 (Appendix A1).
Finally, by normalizing the data over time (Appendix A2), we obtain a time series of normalized SWF damages. Thereby, we assume that the buildings' vulnerabilities with regards to SWFs have remained constant within the last decades. To test whether Nat. Hazards Earth Syst. Sci. Discuss., doi: 10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License. the damages have increased or decreased over time, we apply the seasonal Mann-Kendall test (Hirsch et al., 1982) and use a significance level of 0.1 for the resulting p-value.

Results
In Switzerland, there are few quantitative studies about SWFs. One example is a case study based on damage records stemming from the PICB of the canton of Aargau (Aller and Petrascheck, 2008), undertaken in the aftermath of the devastating August 5 2005 flood. The study indicated that on average at least half of the flood damages to buildings were caused by overland flow (Aller and Petrascheck, 2008), i.e. by SWFs. However, the study is neither comprehensible in terms of applied methods, nor the underlying data, and covers only a small part of Switzerland. Therefore, after a rather qualitative validation of the methodology (Sect. 4.1), we quantify, characterize and compare damages caused by SWFs with damages caused by fluvial floods (Sect. 4.2).
Following, we present the spatial distribution of SWF damages (Sect. 4.3) and show how these damages evolved within the

Validation
There are few data sets available with which the claims' classification or normalization could be validated. A few possibilities are elaborated hereafter.
In 2016, the canton of Lucerne published an overland flow depth map stemming from hydrodynamic simulations based on the methodology described by Kipfer et al. (2012). However, the map indicating categorized flow depth polygons is not suitable 20 for a quantitative validation of the claims' classification. The polygons all indicate a minimal flow depth of 0.015 m and are very dense. In fact, 67 % of all building footprints of the canton of Lucerne intersect such a polygon, whereas only 6.5 % of the footprints are farther than 10 m away from the closest polygon. Consequently, neither quantitative nor a visual relationship could be found between each claim's class and the categorized flow depths.
Two of the cantons covered by our data set provide hazard indication maps concerning overland flow, i.e. the canton of 25 Basel-Landschaft and Aargau, respectively. However, the hazard of overland flow was not assessed comprehensibly judging from the technical reports that are publicly available. In some sub-regions, the hazard was assessed by means of GIS analysis and/or based on known past events, or the hazard was not considered at all. Consequently, these maps did not allow for a direct quantitative validation either. Namely, many claims associated with overland flow were far from any overland flow hazard zone, likely because the hazard was not assessed or no events have been registered so far. Nevertheless, the indicated hazard 30 zones were mostly located in the vicinity of SWF claims. This might highlight that the corresponding claims were the cause 13 Nat. Hazards Earth Syst. Sci. Discuss., doi: 10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License.
Overall, a systematic validation of the classification was not feasible, owed to the large number of claims and the lack of suitable data. Nevertheless, the classification was checked visually, drawing from the input data including flood maps, the river network, the DEM, for example (c.f. Table 1). This manual assessment indicated that the classification scheme rendered 5 reliable and plausible results.
Unlike for the classification, it was possible to verify the overall performance of the applied normalizations. Specifically, we could compare our normalized data set with virtually the same source data that had been normalized with the corresponding property data. The reference data root from a subset of the data shown in Imhof (2011). The reference data show aggregated flood damages per number of insured buildings (Fig. 7a), as well as the loss per total sum insured (Fig. 7b). Thereby, the 10 reference data consist of (almost) the complete records of the 14 corresponding PICB, whereas our data set contains less and less records, as we move back in time (c.f. Table 2). As we are looking at relative numbers, the comparison is still valid, but the different data coverages have to be kept in mind. In fact, Fig. 7 highlights that before 1989, the data sets are badly matching, but have very similar patterns thereafter. Together with the fact that after 1993 all regions are satisfactorily represented, these are the reasons why we have limited the time series of SWF damages to the period from 1993-2013 (c.f. Table 2 and Fig. 14). 15 The clear bias of the localized claims in comparison to the reference data can mainly be attributed to the 21 % of the claims that could not be localized, i.e. the curves aline much better, when also considering the claims without a precise geocode ( Fig. 7). However, a small bias persists, to a larger degree for the number of claims and to a smaller degree for the loss values, respectively. The remaining deviations are probably due to the coverage that becomes increasingly different in earlier years as well as the applied normalization procedure using auxiliary data. Notably, given the simple applied methods, the normalization works exceptionally well. sum in the end, as suggested by Kron (2009), for instance. For that matter, we have stratified the data according to the total number of claims per day using five categories ranging from single (1-5 claims per day) to vast (>501 claims per day). We defined an event as a day with at least one claim of any class (A-E), which amounts to a total of 1'490 events in the period of 1999-2013. Obviously, this is a pragmatic definition of an event, but it serves the purpose of a first simple analysis.

Relevance of surface water flood damages
The stratified number of claims (Fig. 10a, Total) confirms that smaller events are more frequent than larger events, i.e. 1'100 25 events of the smallest category (single) are opposing 11 events of the largest category (vast). Interestingly, days with single and few claims only account for a small share of SWF and fluvial flood claims, although for SWFs the shares are larger. Strikingly, 11 events within the last 15 years with more than 500 claims each, account for almost half of the claims caused by fluvial floods, but only to one-quarter of the claims associated with SWFs. In contrast, the same 11 events accounted for 45 % of the loss caused by SWFs, and even 76 % of the loss caused by fluvial loss, respectively (Fig. 10b). Based on this analysis, we can  Table 2). The numbers indicate the shares in %, while n represents the sample size.
-The largest events cause most of the losses, whereas small events only account for insignificant losses in comparison. The most severe floods within the last 15 years within the study domain are highlighted in Fig. 11, which highlights that these flood events are also associated with high numbers of SWF damages, even though these events are mostly known for 5 being devastating fluvial floods. Thus, our analyses show that fluvial flood damages generally coincide with SWF damages.
This has been noted before (e.g. Blanc et al., 2012) and can be explained by the fact that both flood types are generated by the same rainfall input. Particularly, during extreme rainfall events, we can expect fluvial flood damages, as well as SWF damages.
However, the shares of SWF damages in comparison to fluvial flood damages are different, which might be linked to the type of rainfall. For instance, the event on the 21-22 June 2007 was caused by high intensity rainfall (Hilker et al., 2008) and is 10 associated with a larger share of SWF damages (Fig. 11). All other highlighted extreme flood events were triggered by longduration rainfalls and, at the same time, larger numbers of fluvial flood damages. This could be an indication that the type of rainfall, and in particular the rainfall intensity, is an important driver of SWF damages, as noted for instance by Spekkers et al.   Table 2). (a) Each claim was categorized according to the total number of flood damages that occurred on the same day. For instance, all claims that occurred on 21 June 2007 fall into the category vast, since 1'162 damages were registered for that day in total. Thus, all these claims belong to one of the 11 largest events within the period of 1999-2013. As each claim was classified (Sect. 3.2), we can further group the data as claims related to SWFs (class A and B) or fluvial floods (class D and E), respectively. For the lowest two categories, i.e. single and few, the number of events of SWFs is larger than the number of fluvial flood events. This is due to the fact that some of these events consist of claims categorized as SWFs only. For all other categories, the event numbers match, indicating that for each of these days, some of the claims were classified as SWFs, while some were classified as fluvial floods. (b) The same stratification is applied to the associated loss.
Note that the indication of the number of events for the smallest two event categories, i.e. single and few, were omitted for better readability.
However, the values are identical to the values shown in panel a.

Spatial distribution
Thanks to the spatially explicit input data, we can get a good overview of SWF damages in space, as shown in Fig. 12. In general, it can be observed that the Swiss Plateau (2 and 3) is exposed most to SWFs, both in relative and absolute terms. Also in the Jura (1), many buildings are affected by SWFs. In contrast, the alpine regions of Switzerland, i.e. the Northern Alps (4) and also the Eastern Inner Alps (5) are exposed the least.

5
The visualization of the damage densities has advantages. For instance, in Bernet et al. (2016) low inundation rates by overland flow were reported for Grisons, i.e. the Eastern Inner Alps, and high values for Fribourg, which lies mostly in theWestern Plateau. Fig. 12 supports these findings, but presents a more differentiated picture, as differences within the mentioned regions q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq q q q q q q q q q q q q q q q q q q q qq q  Fig. 13). Thereby, all dates that belong to the same event are connected with lines, and severe events of the same year are shown in the same colors. The event dates are based on Hilker et al. (2008Hilker et al. ( , 2009 can be grasped, as well. In particular, we can see that the damage densities are not evenly distributed in space. The most affected regions are certainly those with high relative, as well as absolute number of damages, such as the areas indicated by the solid and dashed ellipses in Fig. 12. In addition, we see that such areas do not necessarily coincide with the most densely populated areas (dashed ellipse), but may lie in less populated areas (solid ellipses). Moreover, we can also identify areas that suffer from high absolute number of damages but are exposed less in relative terms (dotted ellipse).

Temporal evolution
To obtain an idea about the distribution of the damages throughout the year, we have plotted the number of claims as well as associated losses against the month in which they occurred in form of spider plots (Fig. 13). In relation to SWFs, by far the most damages occur in the summer months from June to August in all regions, except in the Eastern Inner Alps. In the latter region, the maximum number of damages were registered in November, which can be attributed to a single event that occurred 10 on 14-16 November 2002 (Romang et al., 2004), which is highlighted in Fig. 11, as well. The remaining damages occurred 19 Nat. Hazards Earth Syst. Sci. Discuss., doi:10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License. Overall, the number of claims are elevated in the last month in spring, i.e. May, and to a smaller degree in the first month of fall, i.e. September, for most regions. During the rest of the year, i.e. from October to April, very few damages occur, except for the Eastern Inner Alps in November, as discussed before.
Analogous to the number of damages, SWFs cause most of the associated losses in the summer months (Fig. 13c). Interestingly, the losses in the Eastern Plateau and the Northern Alps have larger shares in August, compared to the other regions, but 5 also compared to the corresponding number of claims (Fig. 13a). This can be explained by the particularly high losses during the August 2005 flooding, as indicated in Fig. 12.
The number of claims and associated losses of fluvial floods are highly concentrated in August in all regions ( Fig. 13b and   13d). The event in November 2002 that affected the Eastern Inner Alps is also showing up prominently for fluvial floods, as elaborated before. Furthermore, the data do not exhibit any trends of SWF damages in the period of 1993-2013 based on the seasonal Mann-(p = 0.006). In contrast, the relative losses in Jura do not exhibit such a trend (p = 0.52). The absence of any increasing trend might be a surprising result, as increasing damage trends are often reported (e.g., Kron et al., 2012;Grahn and Nyberg, 2017).
However, it is important to note that in this study we are talking about normalized, relative values, while in the aforementioned publications, the trends of the absolute numbers are considered.

5
Our results show that SWFs caused almost the same number of damages as fluvial floods. For the first time, these numbers are based on a large data set including more than 30'000 damage claims covering 15 years and 48 % of all Swiss buildings.
Notably, SWFs only account for roughly one-quarter of the total loss, which is inline with results from the pilot study (i.e. Bernet et al., 2016). Nevertheless, the associated yearly loss is highly significant, as the following numbers exemplify: The median of total yearly losses to buildings caused by fluvial floods within the considered regions is even slightly lower (5.0 mio 10 CHF y -1 ) than the median of SWF damages (5.9 mio CHF y -1 ) based on the data set D2 covering the period of 1999-2013 (c.f. Table 2). However, the mean yearly loss of fluvial floods is more than 3 times the loss caused by SWFs (i.e. 31.3 mio CHF y -1 versus 10.1 mio CHF y -1 , respectively). The difference between the maximum yearly losses caused by each flood type is even more pronounced: While the maximum loss of SWFs amounts to 38.3 mio CHF y -1 in 2005, fluvial floods caused 234.3 mio CHF y -1 in the same year, which corresponds to a factor of roughly 6. 15 These observation concerning annual flood losses are supported by the characteristics of the individual losses. Their exploration ( Fig. 9) expressed that the range of loss per claim is much narrower for SWFs than for fluvial floods. As SWFs are expected to be associated with significantly lower flow depth than fluvial floods, this might be one of the main reasons for the lower associated loss, since water depth is among the most significant single impact parameters for structural damages to residential buildings (e.g. Kreibich et al., 2009;Merz et al., 2013). Interestingly, the median loss of each claim associated with 20 fluvial floods is also rather low, although significantly higher than the median loss of claims related to SWF. Yet, the highest losses per claim are caused by fluvial floods during the most severe events within the study domain (Fig. 10). As during extreme events, larger areas are affected and the associated shares of objects inundated by large water depths are higher (Elmer et al., 2010), higher losses per claim can be expected. Along the same lines, Hilker et al. (2009) report that the most severe events contribute to more than half of the estimated total loss and Barredo (2009) found an even higher share for flood losses in the 25 whole of Europe. Undoubtedly, loss ratios are higher during more extreme events (Elmer et al., 2010). Although, this probably also holds true for damages caused by SWFs, such damages certainly seem less influenced by the severity of the event (c.f. Fig. 9 and 10). Consequently, SWFs may rarely cause the total destruction of a building, and associated loss ratios may, thus, mostly be well below 1.
As outlined in the introduction, this study is limited to direct tangible damages to buildings. Therefore, the absolute loss 30 values are low in comparison to other loss estimations that include other losses, as well. For instance, Hilker et al. (2009) report a mean financial loss of 317.2 mio CHF y -1 between 1972-2007, which is roughly 7 times higher than the mean of all flood losses to buildings, as represented by our data set. For one, the data published by Hilker et al. (2009) cover the whole of Switzerland and consider a longer period. More importantly, however, these estimates also include damages to infrastructure, forestry and agricultural land, in addition to damages to buildings and their content. Therefore, the associated losses are inherently higher than the numbers presented in this study. This exemplifies that one has to be careful when comparing values from different data sources (Kron et al., 2012). Moreover, it highlights the fact that the damages to buildings are just a small fraction of the total loss caused by SWFs for the society. Nevertheless, these data serve well for assessing the relevance of SWF 5 damages in Switzerland, especially when considering relative values.
The spatial distribution of damages caused by SWFs can be deceiving: Obviously, an area with a higher building density will likely result in a larger number of damages compared to an area that is less populated (Fig. 12b versus 12c). Therefore, it is important to have a look at relative values, as well (Fig. 12a). Thereby, the effect of higher values caused by a denser number of buildings is considered. However, the relative values are quite sensitive in sparsely populated areas. A damaged 10 house with virtually no other houses in the vicinity will produce a high relative value or a low value, respectively, if the same is not affected. In contrast, in more populated areas, the relative value will not change much in case a building is more or less damaged. Thus, to obtain a complete picture, the relative and absolute values should be considered alongside the building density. In that way, the most exposed areas can be identified, like the two highlighted areas in the Western Plateau that are associated with high relative and absolute numbers of damages (Fig. 12). 15 Furthermore, it is important to keep in mind that in case an area has no registered damages, it does not necessarily mean that the area has not been affected by any floods at all. It just indicates that either no buildings were in the vicinity of the flooded area or the buildings were properly protected against such floods. Therefore, damage records can only indicate floods that lead to some sort of damage and never to the occurrence of floods in the hydrological sense, as discussed in the introduction (c.f. Fig. 1). However, understanding the characteristics of damaging floods can open the stage to understand the process in a 20 broader context, as well.
The temporal distribution of claims related to SWFs exhibits a distinct seasonality ( Fig. 13 and 14). Similar to the flood losses reported by Hilker et al. (2009), most damages clearly occur in summer, with a few exceptions. Thereby, thunderstorms associated with short but intense rainfall are certainly an important driver of SWF damages. Nevertheless, long duration rainfall events are also responsible for a large share of SWF damages, highlighted by the most severe events that are mostly associated 25 with long duration precipitation. In contrast, much fewer damages are caused in spring and fall, and virtually no damages are caused in winter. Damages in winter can likely be attributed to rather local events coinciding with conditions promoting overland flow generation such as rain on frozen soils. Overall, these observations have important implications for assessing the hazard of SWFs. In particular, simply focusing on high intensity rainfall events may lead to an underestimation of the risk of SWFs.

30
Although the time series is relatively short, the data do not exhibit any increasing trends of SWF damages in the period of 1993-2013. Obviously, the general increase of absolute damages in time, which can be found in our data as well, is eliminated when the data are normalized. Thus, as suggested for instance by Kundzewicz et al. (2014), the increase in loss can be mainly attributed to the socio-economic development. However, we did not consider further aspects that could have an influence on such trends, such as a change in vulnerability (Bouwer, 2011). Moreover, insurance or local governmental policies that might Nat. Hazards Earth Syst. Sci. Discuss., doi:10.5194/nhess-2017-136, 2017 Manuscript under review for journal Nat. Hazards Earth Syst. Sci. Discussion started: 10 April 2017 c Author(s) 2017. CC-BY 3.0 License.
have changed over time were not taken into account either. Nevertheless, it is important to note that increasing absolute losses are most likely not attributable to climate change, but to socio-economic factors. Consequently, the major associated risks related to SWF damages is not climate change, but the increased exposure due to population growth and increasing wealth.
This has implications for decision and policy makers, as well as for insurance companies and similar stakeholders.
The key to the exploitation of the insurance data with regards to SWFs lies in the classification of the damage claims example is the statistical model applied by Spekkers et al. (2013) in order to differentiate rainfall-related damages clustered around wet days from non-rainfall related damages occurring throughout the year. Thereby, the classification of each claim is not independent anymore, but is depends on how many other damages occurred on the same day. 15 Although, the classification scheme presented in this study has striking advantages, it has the following short-coming: As overland flow propagates over the land surface, it may eventually reach a watercourse. Areas alongside watercourses, where the overland flow joins the river, may be a flood hazard zone. If so, all claims in that specific area are classified as a fluvial flood, even though, possibly, the claim might have been caused by incoming overland flow. However, a more qualified classification entails likely event-specific, time-consuming manual assessments. In fact, it is extremely difficult to disentangle the different 20 flood types, even more so for events in the more distant past and if no data with that particular focus are available. In contrast, for claims that are located far away of any watercourse, it is very unlikely that they are affected by watercourses at all. Therefore, our methodology renders a lower boundary of claims associated with SWFs, in essence. In reality, the numbers are likely higher, but as mentioned before, disentangling the flood types within their overlapping domains is difficult.
Indeed, the flood processes are a complex interlinked system, as Evans et al. (2004) stated. In fact, the insurance data 25 illustrated that damages caused by SWFs occur (almost) always alongside claims caused by fluvial floods (c.f. Fig. 11). Be it a short and intense thunderstorm or a long duration event, rainfall is the main trigger of every SWF, as well as (almost) every fluvial flood. Understandably, if there is enough rainfall to cause a SWF, it may as well cause or at least contribute to a fluvial flood, once part of the water reaches the next watercourse. Undoubtedly, severe events that include hundreds or thousands of damages entail a combination of flood processes, while, of course, some local events may be associated with a single flood 30 process only.
In this study, we have presented a simple and pragmatic approach of how spatially explicit insurance data records can be exploited to investigate damages caused by SWFs. The methodology provides a robust lower estimate of SWF damages. Using the presented percentile values (Table 3), the methodology is applicable for classifying any claim in Switzerland, except in the Western Inner Alps and Southern Alps, where data were lacking. For these regions, appropriate values could be approximated.

5
Moreover, the methodology is transferable to other regions and countries, but has to be adapted to locally available flood hazard maps.
There seems to be a consensus among Swiss practitioners and experts that SWFs are responsible for a large share of all flood damages. However, this perception does not stem from quantitative research, but rather from single case studies or practical experience. With the study at hand, we are able to quantify the striking relevance of SWFs in Switzerland based on a sound In general, the spatial and temporal distribution of SWF damages is complex. Different factors might be responsible for high damages within certain areas or during certain periods. For instance, the meteorological forcing differ spatially and temporally, the predisposition due to unfavorable soils or landuse practices play a role, past human interventions such as the installation of 25 drainage and the removal of small natural rivulets can have an influence, but also slightly differing practices by the insurance companies or different rules applied for buildings to be built might be relevant. Undoubtedly, we stand at the beginning of better understanding SWFs in Switzerland, and also on an international level. Thereby, a common terminology is the base to strengthening and extending the science within this field across the countries' borders.
This study highlights the fact that SWFs are a highly significant flood process in Switzerland. Unlike for fluvial flood hazards,

30
there is no publicly available information about the hazard of SWFs up to date, in spite of the process's obvious relevance.
Since SWFs can occur practically anywhere in the landscape, the more paramount it seems to have detailed information about local SWF hazards. Such information can help to make well-founded decisions by all different stakeholders, e.g. planning and installing appropriate property protections by house owners, applying measures to reduce overland flow generation on agricultural fields by local farmers, providing surface retention ponds by municipalities or amend regulations to prevent SWF damages by the federal government. However, as a first priority, SWFs in general and the influencing factors of SWFs in particular should be further studied, and, ultimately, better understood.
As a first step in this direction, we propose that SWFs should not be regarded as an isolated process by itself. A better way is probably to extend our focus from rivers and lakes alone to hidden rivulets, covered drains, the sewer system, impervious 5 areas, agricultural fields and headwaters, which all contribute to the generation of SWFs. Therefore, we should regard overland flow and ponding as an integrated part of our catchments. In this manner we may start to understand the complex interlinked flood processes better in the future.

Data availability
The data, on which this study is based, were provided by 15 different insurance companies. Each record contains confidential 10 information such as the location (address and/or coordinates), claim date, associated loss etc. Due to privacy protection, the data are subjected to strict confidentiality and, thus, cannot be made accessible.

Appendix A: Normalization
Obviously, it would be best to normalize the damage data with the corresponding property data of the respective insurance company. However, property data are generally even more difficult to obtain than damage data, as the former contain additional 15 sensitive and confidential information. Therefore, ancillary data are required to estimate the number of insured buildings, as well as the replacement value of each corresponding building. Moreover, as these values change over time, we need additional ancillary data to take these temporal changes into account. As outlined in Sect. 3.3, the spatial normalization as well as the temporal normalization of the damage data are described in detail in the following sections.
After all, we divide the number of claims by the estimated number of insured buildings, while the losses are divided by the 20 corresponding total sum insured, respectively. For that matter, all quantities have to be spatially aggregated. For this study, we aggregated the data to regular grids and tested various resolutions. We chose a resolution of 3 by 3 km, which seemed like a good compromise between level of detail and smoothing. The point of origin of the corresponding rasters is chosen arbitrarily.
Thereby, we note that the choice may change the absolute values of each cell, but in general does not change the larger picture.
A1 Spatial normalization 25 As property data were not available, we inferred the number of buildings using ancillary data. For this purpose, we made use of the terrain model swissTLM3D (Table 1). From this data set, the number of buildings represented by their footprints can easily be extracted. However, the data needed to be preprocessed: Invalid geometries had to be corrected and overlapping polygons were dissolved into single polygons in order to obtain a homogeneous data set as of 2013.
The definitions of a building are quite similar among the PICB (Imhof, 2011). Nevertheless, the number of footprints does not match the number of insured buildings, since a row house might be represented by one footprint, while it constitutes several buildings as defined by the respective insurance company, for instance. To consider this, we referred to publicly available annual reports of 2013 and thereby obtained the total number of insured buildings for each PICB. We then divided the obtained values by the number of footprints, resulting in a simple multiplication factor (f n , Table A1). By multiplying the aggregated number 5 of buildings with the factor f n , we obtain the approximated number of insured buildings as of 2013. For each grid cell, the aggregated number of claims is then divided by the aggregated number of buildings to obtain spatially normalized damage numbers.
To normalize the loss, we need to relate the loss values to the total sum insured. There are few published methodologies to assess building values in detail, but these can be too time-consuming for applications in large study areas (Kleist et al., 10 2006). Given the large data set, we chose a simple approach similar to the methodology shown by Grünthal et al. (2006), who used the product of mean insurance values and the number of buildings to estimate the replacement costs of residential buildings. However, instead of the buildings' footprint area, we considered the buildings' volume, which we expect to be a more representative measure for estimating building values.
Specifically, we first assessed the mean altitude of each building's footprint by using common zonal statistic functions of a 15 GIS as well as a digital elevation model as input (Table 1). The top of each building was then assessed by the same methodology, but using a digital surface model, instead. The approximated building height resulted from the difference of the two values.
Implausible results were corrected, i.e. values below 3.5 m or above 100 m were set to the standard building height of 3.5 m.
Thus, a standard height of 3.5 m is assigned for buildings that might have been built after the last update of the DSM in 2008 (c.f. Table 1). Following, the building volumes are obtained by multiplying the building's footprint area with the mean building 20 height. The total building volume for each canton is assessed and divided by the respective total sum insured, in order to obtain the insurance value per cubic meter (ρ v , Table A1). The product of each building's volume and ρ v finally results in each building's value as of 2013. Analogous to the number of buildings, the loss is aggregated to regular grids and divided by the aggregated sum insured.
A2 Temporal normalization 25 As the considered terrain model itself does not include attributes for such considerations, we used another auxiliary data set, i.e the buildings and dwellings statistic of the Swiss Federal Statistical Office as of 2013, from which the number of newly built residential buildings can be inferred (Table 1). The data are regularly updated, whereas the number of residential buildings can be assessed at any time by linear interpolation between the sampling points. Normalizing with the number of buildings per canton as of 2013 (Appendix A1), we obtain a dimensionless factor (f t , Table A2). With the assumption that the 30 residential buildings are representative for the development of all buildings, we obtain the temporal development of the number of buildings and the total sum insured. To that end, we multiply the interpolated factor f t for each time step with the number of buildings and the total sum insured as per 2013, respectively. data harmonization process. Also, we would like to thank the canton of Lucerne for providing the overland flow map. Last but not least, we thank Markus Mosimann for his support of harmonizing the insurance data, Veronika Röthlisberger for the estimation of the buildings' values as well as her and Andreas Zischg for the many valuable inputs.