Global Flood Exposure from Different Sized Rivers

There is now a wealth of data to calculate global flood exposure. Available datasets differ in detail and representation of both global population distribution and global flood hazard. Previous studies of global flood risk have used datasets interchangeably without addressing the impacts using different datasets could have on exposure estimates. By calculating flood exposure to different sized rivers using a model independent geomorphological approach, we show that limits placed on the size of river represented in global flood models result in global flood exposure estimates that differ by 10 greater than a factor of 2. The choice of population dataset is found to be equally important and can have enormous impacts on national flood exposure estimates Up-to-date, high resolution population data is vital for accurately representing exposure to smaller rivers and will be key in improving the global flood risk picture. Our results inform the appropriate application of these datasets and where further development and research is needed.

dedicated to the investigation of the human interaction with rivers of different size (Kummu et al., 2011). Understanding this interaction globally, particularly with respect to river flooding, will inform us about the completeness of current global flood exposure studies and identify where further study and development is needed.
A comprehensive understanding of flood risk requires information about the hazard, what or who is exposed, and their 35 vulnerability. Exposure could include damages (both direct and indirect), exposed gross domestic product (GDP), exposed assets, and most commonly: exposed people . Identifying flood exposed populations usually involves intersecting a flood hazard map with a population map. The methods and inputs used to produce population datasets differ, and so does their intended use (Leyk et al., 2019). Recently released population maps, which utilize commercial RS data and are an order of magnitude more resolved than existing population datasets (Tiecke, 2017) are already being used for disaster 40 preparedness and response (Facebook, 2019). However, our current understanding of global flood exposure is based on existing global population datasets, and these datasets have been used interchangeably in global studies (Tanoue et al., 2016, Jongman et al., 2012, Dottori et al., 2018 with little comment about their relative merit. The credibility of existing global flood exposure estimates in light of new, more detailed, population data and the implications of their interchangeable use in studies of global flood exposure needs to be explored. A recent study by Smith et al. (2019) reported large disagreement 45 between flood exposure estimates calculated in 18 developing countries using three different population datasets. The identification of population data as one of the chief sources of uncertainty in global flood exposure studies warrants further investigation at the global scale.
Improving global flood exposure calculations necessitates a better understanding of both GFMs and the population data their flood hazard maps are intersected with. Specifically, river network size in GFMs and its impact on flood exposure estimates 50 needs to be explored. The number of rivers modelled in GFMs differs by several orders of magnitude depending on the model used. The size of a GFM's river network is limited by the resolution of input data, such as the underlying digital elevation model (DEM) and climatology (Dottori et al., 2016); or by the computational efficiency of the model, as the introduction of smaller rivers exponentially increases the modelled domain. Differences in river network size mean potentially exposed people and infrastructure are missed in global flood risk assessments. This missed exposure has not yet 55 been quantified at the global scale. Advances in RS and computational capacity will enable more rivers to be represented in future iterations of GFMs. But what coverage should modellers be aiming for globally? Understanding where the representation of smaller rivers is most needed (areas of higher exposure) can inform model developers where to prioritize model development and inform model users about the limitations of currently available models. The population data used in global flood exposure calculations varies just as much as the GFMs do. Recent advances in population data, providing more 60 detail and employing new modelling techniques, have been shown to drastically reduce flood exposure estimates in developing countries (Smith et al., 2019). Understanding how both new and existing population datasets differ in their resulting exposure estimates, both regionally and within the hierarchy of the river network, can inform users about the most appropriate population dataset to use. https://doi.org/10.5194/nhess-2021-102 Preprint. Discussion started: 9 April 2021 c Author(s) 2021. CC BY 4.0 License.
To explicitly explore the impact of river network size on global flood exposure estimates, we use a geomorphological 65 measure of a river's flood susceptibility, which is independent from current GFMs and the additional uncertainties their different model structures bring. Fluvial processes contribute to the evolution of a landscape over time. The erosional action of flowing water has shaped the terrain of drainage basins to reflect the historical flow of water through them.
Geomorphological approaches to mapping river flood susceptibility rely on the concept that the cumulative hydrogeomorphic effect of past flood events, evident in topography data, is indicative of a river's propensity to flood. Such 70 approaches to flood mapping have been applied over a number of scales: from local (Nardi et al., 2006, Nobre et al., 2016, Dodov and Foufoula-Georgiou, 2006, to national (Jafarzadegan et al., 2018), to regional (Lugeri et al., 2010) and global . The computational efficiency of hydrogeomorphic flood mapping, coupled with its reliance on only terrain data as input, make it useful for a 'first look' global scale analysis; intended to inform future development of higher accuracy hydrological flood mapping (Di Baldassarre et al., 2020). 75 Our geomorphological approach to mapping a river's flood susceptibility, herein referred to as the River Flood Susceptibility Map (RFSM) is based on new topography data (Yamazaki et al., 2017), which incorporates crowdsourced information to better represent the locations of rivers and streams (Yamazaki et al., 2019). Validation of our calibrated methodology (outlined in detail in the Supplementary Material) shows that the RFSM better replicates GFM hazard maps in Africa than an existing global geomorphological approach . We also show that the RFSM performs similarly to the best 80 GFMs (Dottori et al., 2016, Yamazaki et al., 2011 when validated against historical flood events (Bernhofen et al., 2018b). The RFSM allows us to easily discretize the flood map into different river sizes (independently of GFMs). We investigate the human interface with these different size rivers using three population datasets. Facebook's High Resolution Settlement Layer (HRSL), (https://dataforgood.fb.com/docs/high-resolution-population-density-mapsdemographic-estimates-documentation/) (1 arc-second, ~30 m resolution at the equator) (Tiecke, 2017) which is currently 85 only available in 168 countries globally, and two population datasets used extensively in previous studies of global flood risk: the Global Human Settlement Population (GHS-POP) (http://doi.org/10.2905/0C6B9751-A71F-4062-830B-43C9F432370F) (9 arc-second, ~250 m resolution at the equator) (Freire et al., 2015) and WorldPop (https://dx.doi.org/10.5258/SOTON/WP00645) (3 arc-second, ~90 m resolution at the equator) (Stevens et al., 2015, Lloyd et al., 2019. We present a global picture of flood exposure to different size rivers, both in the present day, and how it has 90 changed over the past 40 years. We then compare the flood exposure calculated using different population layers, exploring the implications this has on national level flood exposure estimates and examine the impact that river size has on any disagreement. Finally, we address the size of rivers represented in GFMs specifically and investigate how their chosen river network size impacts both global and national flood exposure estimates and what implications this has for previously published global flood risk assessments.

Mapping River Flood Susceptibility
We use a geomorphological approach to mapping river flood susceptibility, which is independent from the global flood models (GFMs). Previous GFM comparison studies found that multiple aspects of model structure contributed towards disagreement (Trigg et al., 2016b, Bernhofen et al., 2018b, Aerts et al., 2020. Using a geomorphological approach, we are 100 able to explore just one aspect of disagreement: river network size. This approach allows us to explore all stream scales as drainage paths can be identified from the terrain alone. It is not influenced by the structure of the different GFMs and does not have the same computational restraints as a global hydrodynamic model. This approach is different from the GFMs in that it does not measure the flood extent for a given return period flood, but rather a river and surrounding location's static susceptibility to flooding. 105 The method used for delineating a river's flood susceptibility is based on the Height Above Nearest Drainage (HAND) methodology developed by Nobre et al. (2011). This hydrogeomorphic approach, requiring only terrain data as input, is computationally efficient, and can be easily modified to produce auxiliary data layers.
Our method, referred to as the River Flood Susceptibility Map (RFSM) (Bernhofen et al., 2021), is illustrated in Figure 1 and takes three gridded datasets as input: a digital elevation model (DEM), its derived drainage directions, and its upstream 110 drainage area (UDA). We use MERIT hydro data (Yamazaki et al., 2019), a hydrography dataset based on the error improved SRTM DEM: MERIT DEM (Yamazaki et al., 2017). MERIT Hydro is an improvement on previously available global hydrography datasets such as HYDROSHEDs (Lehner et al., 2008) in terms of both spatial coverage and its representation of small streams. Its improved representation of small streams is enabled by its incorporation of global water body data and crowdsourced Open Street Map river data. This makes it particularly suited to this study, where we are 115 interested in examining the flood susceptibility of rivers down to the smallest streams.
The river network is extracted from the upstream drainage area dataset by specifying a minimum threshold river size (in units of UDA). Identifying the headwater of a river is no trivial task, with regional and climatic factors playing a part (Montgomery andDietrich, 1988, Tarboton et al., 1991). Previous work exploring optimal initiation thresholds for geomorphological floodplain mapping found that DEMs with a resolution of 1 arc second (~30 m) could use initiation 120 thresholds less than 10 km 2 UDA. In the same study, a 3 arc second (~90 m) resolution DEM was used with a 100 km 2 UDA threshold . The MERIT Hydro data we use in this study has a resolution of 3 arc seconds (~90 m). But its incorporation of crowdsourced river data has optimized its representation of small streams and rivers. As such, we use a globally consistent river initiation threshold of 10 km 2 UDA for the RFSM. This is a large assumption, as in some locations globally there will be no visible channel at this location. However, we argue that removing areas of potential exposure to 125 avoid overprediction in some areas goes against the premise of this study, which is to explore and identify 'missed' areas of exposure. The exposure calculations for small streams should therefore be interpreted with these limitations in mind.  Once the river network has been extracted, the rivers in the network are classified based on their Strahler stream orders (Strahler, 1957). The Strahler stream order is a dimensionless indicator of the magnitude of the river based on its hierarchy within the drainage basin. Our method requires the user to assign each order river a maximum height above nearest drainage, Hn (see Figure 1a). Each order's Hn value is calibrated against reference flood maps across the world. To account for 140 climatic variability in a river's flood susceptibility , we split the globe into five simplified Köppen-Geiger climate zones ( Figure 2): Tropical, Arid, Temperate, Continental and Polar. Polar regions are excluded from our analysis as these regions are dominated by glacial not fluvial processes (Chen et al., 2019). The RFSM has uniquely calibrated Hn values in each of the four climate zones. We use 19 reference flood maps to calibrate the RFSM. These maps span 5 different continents across all four climate zones considered. Reference flood maps are a mixture of national, continental, and global 145 flood hazard maps. To maintain consistency across the calibration data, we use 100-year return period flood hazard maps.
We use a combination of national, continental, and global flood hazard maps for calibration in each climate zone. This is to ensure that there is sufficient calibration data for each Strahler order river, as only the national flood hazard data captures flooding for low order rivers. The final Hn values for each climate zone are shown in Figure 2. More detailed information on the calibration of the RFSM on each of the calibration basins can be found in the Supplementary Material. 150 Once Hn values for each order have been assigned, each stream order is processed separately (Figure 1b), and then merged together. In areas of overlap, the highest order stream retains the values. Two datasets are produced as output: a map of the flooded river's upstream drainage area, and a map of the flooded river's Strahler stream order. Illustrations of these two outputs are shown in Figure 1c. The RFSM was validated against both existing GFMs and observed flood events. Validation against GFMs was carried out 160 for the whole of the African continent using the aggregated output of 6 GFMs from a previous model intercomparison study (Trigg et al., 2016a). To assess the credibility of the RFSM, it was also validated alongside an existing global geomorphological floodplain map . The results of the GFM validation show that the RFSM produces credible flood extents when compared with existing GFM outputs in Africa. The RFSM correctly captures over 90% of high agreement flood zones (where at least 5 out of 6 GFMs agree) in 7 of the 8 major drainage basins in Africa. In the East 165 African basin, the RFSM captures 87% of this high agreement flood zone. Comparing commonly used measure of fit scores for the RFSM and the existing global geomorphological floodplain map, the RFSM scores higher in all the major drainage basins in Africa except for North Africa (where both maps score poorly due to the Sahara desert). The RFSM was also validated against observed flood events in Nigeria and Mozambique. The 2012 flooding in Nigeria and the 2007 floods in Mozambique affected four million people and over one hundred thousand people respectively (Bernhofen et al., 2018b). 170 Validation data for both these flood events used in a previous GFM validation comparison study (Bernhofen et al., 2018a) was also used to validate the RFSM. Validation of the RFSM against these historical flood events show that it performs similarly to the best performing GFMs. Further detail about the validation of the RFSM can be found in the Supplementary

Material.
It is important to note, our methodology does not account for flood protection measures and cannot communicate the 175 probability of flooding in any location. It consistently represents a river's flood susceptibility based on the surrounding terrain alone. The method's intended use is as a global 'first look' analysis to inform future model development and use.

Measuring Exposure
We investigate the human exposure to river flood susceptibility. Human exposure is herein defined as the intersection of our 180 flood susceptibility map and a spatially distributed population layer. Three population datasets are used to measure exposure: Facebook's High Resolution Settlement Layer (HRSL) (https://dataforgood.fb.com/docs/high-resolution-population-densitymaps-demographic-estimates-documentation/) (Facebook and CIESIN, 2016) (Stevens et al., 2015). 185 These population datasets all use the same initial input census data, from GPWv4 (Center for International Earth Science Information Network -CIESIN -Columbia University, 2016), but their methods for allocating the population across gridded cells differ. Facebook's HRSL is the only dataset of the three lacking full global coverage (at the time of writing 168 countries have been mapped). It is also the most recent, with work ongoing to map the remaining countries. HRSL uses ultra-high resolution commercial satellite imagery (~50 cm resolution) and convolutional neural networks to detect 190 individual buildings at the country level (Tiecke, 2017). Subnational census data is then proportionally allocated to the identified buildings at 1 arc second resolution (~30 m at the equator).
Similarly to the HRSL in methodology, JRC's GHS-POP dataset identifies built up areas from Landsat imagery and proportionally allocates census data to the built up areas (Freire et al., 2015). In regions where no settlements can be identified, but where census data indicates there is a population, the population is evenly distributed across the census area 195 using areal weighting (Friere et al., 2016). This can occur in some rural areas, where small settlements are not captured by the Landsat imagery. Despite being coarser in spatial resolution at 9 arc seconds (~250 m at the equator), GHS-POP provides consistent multi-temporal population estimates (1975-1990-2000-2015) allowing for accurate analyses over time (Freire et al., 2020).
Unlike the other two population datasets, which evenly spread census data over identified settlements, WorldPop uses a 200 complex model to disaggregate population over an area (Leyk et al., 2019). It uses a random forest model and a number of ancillary datasets to dynamically weight the distribution of census data over a 3 arc second (~90 m at the equator) gridded area (Stevens et al., 2015).
Exposure calculations necessitate uniformity between the intersecting datasets in terms of spatial resolution. As such, the GHS-POP layer was resampled from 9 arc second resolution and the population evenly distributed to a 3 arc second 205 resolution grid to allow for analysis with a flood map of the same resolution. Conversely, for the HRSL exposure calculations the RFSM was resampled from 3 arc second to 1 arc second resolution. When comparing the exposure results between population datasets the epoch used for comparison was 2015. National population totals for the HRSL and WorldPop datasets were scaled relative to GHS-POP 2015 national population totals.
Flood exposure is first calculated using the GHS-POP layer. Globally, we find 1.94 billion people are susceptible to flooding from rivers with a UDA greater than 10 km 2 . Breaking this down by continent, Asia's flood exposure is 1.49 billion, Africa's is 203 million, Europe's is 104 million, North America's is 81 million, South America's is 59 million, and Oceania's is 3.5 million. Splitting global flood exposure by river size, of the total exposed: 18.2% are from streams, 26.4% from small rivers, 220 23.7% from medium rivers, 17.2% from medium-large rivers, 8.4% from large rivers, and 6.1% from huge rivers. Asia makes up over 75% of the total global flood exposure, the majority of this amount coming from India and China, which are by far the two most exposed countries (see Figure3a). Roughly half of India's flood exposure is from streams and small rivers. Comparably, in China, this figure is closer to a third. This is likely due to the degree of urbanisation in both countries; the percentage of China's urban population is double that of India's (WorldBank, 2018). Urban areas are disproportionately 225 located on large rivers due to the historical tendency for settlements to form in areas fertile for farming and convenient for transport (McCool et al., 2009). As such, a greater proportion of flood exposure in China comes from larger rivers, whereas in India, a greater proportion comes from rural exposure to smaller rivers. Rivers classified as 'huge' are only found in some countries, but often they pose a large proportion of the national flood risk. For example, the Brahmaputra in Bangladesh and the Nile in Egypt and Sudan are responsible for just under half of the national flood exposure in their respective countries. 230 To identify countries with the most acute flood risk, exposure was normalized against total national population ( Figure 3c).
Suriname has the highest normalized exposure, with 894 people exposed per 1000. The country's low elevation relief, and its capital city situated on the banks of the Suriname river near its outlet into the Atlantic Ocean, makes Suriname particularly vulnerable to flooding (WorldBank, 2019). Four of the top 10 most 'normally' exposed countries are in south or south east Asia. These include: Bangladesh, Cambodia, Thailand, and Vietnam. Flooding in these countries is severe and annual, 235 normally occurring each year during the monsoon season. In Europe, the Netherlands has a high normalized exposure, 738 exposed per 1000. The Netherlands has a long history of flooding due to its low elevation, flat terrain, and high population density. It also has the most advanced flood defence systems in the world, designed to contain river water levels with a probability of occurrence once every 1250 years (Stokkom et al., 2005). Geomorphological approaches to flood mapping,

Variation in Exposure
Exposure differences arising from the use of different population layers were calculated for the 168 countries where all three population datasets are available ( Figure 5) (see Supplementary Material for a list of the missing countries). In the countries examined, the exposure calculated with WorldPop data was the highest (270 exposed per 1000), followed by GHS-POP (256 exposed per 1000), HRSL exposure was the lowest (235 exposed per 1000). These findings correlate with a previous study 290 by Smith et al. (2019) which found WorldPop data overestimated flood exposure compared to HRSL data in each of the 18 developing countries examined.
Differences in calculated exposure across the river sizes are shown in Figure 5b. Exposure differences were most pronounced for smaller rivers (streams, small, and medium rivers), while there was almost no exposure difference for the largest river class (huge). The overall trend across all river sizes consistently shows that WorldPop estimated the highest 295 exposure, followed by GHS-POP, and then HRSL with the lowest.
The methods for producing the three population layers can go some way towards explaining the differences in calculated exposure, these methods are visualized in Figure 6, where we qualitatively compare the settlement distribution of the three population datasets along the Likuala-aux-Herbes river in the Republic of Congo. WorldPop's population distribution algorithm spreads some residual population across the grid in areas where no settlements have been identified. This is done 300 under the assumption that not all 'built up' areas will be picked up in the satellite imagery (TReNDS, 2020). When The resolution of the population layers should also be considered. GHS-POP's fairly coarse (9 arc second) resolution means that in some areas where the potential for flooding (or not) falls within the resolution of a 9 arc second grid cell, the settlement's avoidance (or not) of the flood risk cannot be accurately represented. This effect can be reduced by upsampling and proportionally reallocating the population to a grid that matches the resolution of the flooded data, as we have done in this study. Similarly, the spatial resolution of the underlying satellite imagery should be considered. Both GHS-POP and 315 WorldPop identify settlements using Landsat imagery at 30 m resolution, while HRSL identifies settlements using DigitalGlobe imagery at 0.5 m resolution. Previous work by (Tiecke, 2017) showed that HRSL was able to identify The use of different population datasets had a negligible effect on exposure estimates for the huge river class. Large settlements tend to form around rivers of this size, and on coastlines where rivers of this size drain. Large urban areas are easily identifiable from remote sensing data, which means the population distribution (and resulting exposure estimates) for these urban centres show less variation between the datasets. Conversely, non-urban flood exposure estimates to smaller rivers show greater sensitvity to the choice of population layer. This is because the approach to non-urban population 330 mapping between the three datasets differ. WorldPop, as mentioned previously, distributes administrative level census data across all 3 arcsecond pixels in order to mitigate the impacts of potential omission and commission errors in the settlement data. This approach leads to some overestimation in rural populations (Smith et al., 2019, Wardrop et al., 2018. GHS-POP, which distributes census data over Landsat identified settlements (and in non built-up areas distributes population at the census unit by areal weighting), tends to underestimate rural populations. (Liu et al., 2020, Leyk et al., 2019. HRSL's use of 335 ultra-high resolution sattelite imagery has been shown in previous studies to accurately identify rural settlements (Tiecke, 2017, Smith et al., 2019. However, there is still significant uncertainty in the underlying census data and the method of proportional allocation used to distribute the census data is relatively crude. Calculating the general trends of exposure between the population layers is useful for making broad conclusions about the suitability of a population layer. Understanding the variations of the data at the country level leads to more actionable 340 information about the appropriate use of different population layers. The disagreement between the population layer exposure estimates for each country varies significantly (Figure 5c). In the three countries with the highest exposure disagreement (Belize, The Republic of Congo, and Guinea-Bissau) WorldPop estimates of exposure are far greater than either HRSL or GHS-POP estimates. In Belize, a country with large areas of inundated wetlands, WorldPop estimates 135,000 people exposed, while GHS-POP and HRSL estimate 70,000 and 80,000 exposed, respectively. In the Republic of 345 Congo, a country with large areas of floodplain, WorldPop estimates1.3 Million people exposed and GHS-POP and HRSL estimate 810,000 and 780,000 exposed respectively. WorldPop's method of distributing the population over a large area results in significant overestimation compared with HRSL or GHS-POP in these rural inundated areas. This is evident in Figure 6, where the population distributions are qualitatively compared with manually identified settlements in a small region in the Republic of Congo. WorldPop's distribution of population results in a larger area of exposure compared with 350 the other two datasets and the manually identified settlements. This then leads to overestimation in regions with large areas of flood extent. This can be seen in greater detail in Figure 7 for Guinea-Bissau. In Guine-Bissau, GHS-POP and HRSL (which estimate exposures of 180,000 and 160,000 respectively) identify settlements largely situated outside the floodplains ('dry' cells in red). Comparatively, WorldPop's residual population spread leads to far more 'wet' population cells and an estimate of exposure (480,000) more than double that of the other two population layers. The exposure disagreement in these 355 three countries is compounded by the relatively large areas of inundation in each country. The percentage inundated area is 25%, 30%, and 26% for Guinea-Bissau, Belize, and The Republic of Congo, respectively. In comparison, the percentage of populated area defined by the population layers is less than 5% for GHS-POP and HRSL, but more than 95% for WorldPop in each of the three countries. As exposure in this study is defined as the intersection of the flooded area and the populated https://doi.org/10.5194/nhess-2021-102 Preprint. Discussion started: 9 April 2021 c Author(s) 2021. CC BY 4.0 License. area, it is understandable that WorldPop's exposure estimates are more sensitive to the area of inundation. This is evident 360 when examining a country with high exposure disagreement but with a comparatively smaller area of inundation. In Bosnia and Herzegovina (Figure 7), the percentage of flooded area is just 9% and the GHS-POP layer estimates far greater exposure (1 Million) than either WorldPop (680,000) or HRSL (610,000). Here, where much of the exposure occurs near the banks of the rivers, the coarse spatial resolution of GHS-POP is less able to precisely locate settlements situated just outside the floodplain. As a result, more populated cells are flagged 'at risk' compared to the higher resolution HRSL layer. 365 These results have shown that the use of different population layers can lead to vastly different flood exposure estimates because of inherent differences in their spatial resolutions, methods used, and assumptions made to produce them. Our comparative analysis has identified in which countries exposure calculations are sensitive to the choice of population layer and shed light on some of the reasons for exposure disagreement. However, there is a limit to the conclusions that can be drawn from comparative analyses alone, and there is an urgent gap for more studies which validate the accuracy of these 375 population layers using ground-truthed data.
It would be imprudent to definitively recommend one population dataset for use in flood exposure studies without extensive comparative global validation. However, previous studies have shown that HRSL performs better than existing population datasets at mapping reference building footprints, especially in rural areas (Tiecke, 2017, Smith et al., 2019. Our results also point to some of the benefits of using HRSL. Its settlement identification method for population distribution avoids exposure 380 overprediction common in other population data and its high resolution can better capture the accurate location of settlements. Despite this, HRSL shouldn't be considered a catchall dataset for flood exposure. Its high resolution may limit its use in certain situations due to computational restraints. Similarly, in studies of flood risk over time population data with multiple temporal epochs, such as GHS-POP or WorldPop, are better suited. The results we present in this section, and

Relevance to Global Flood Models
The minimum size of river represented in Global Flood risk Models (GFMs) varies (see Table 1), with minimum river size 395 thresholds ranging between 50-5000 km 2 UDA, three orders of magnitude. River network size can be limited by the granularity of input data such as rainfall (Dottori et al., 2016), or by the computational demand of modelling floods at the global scale.  (Pappenberger et al., 2012) and U-Tokyo (Yamazaki et al., 2011) Small (P), Medium, Medium-Large, Large, Huge 1000 km 2 CIMA-UNEP (Rudari et al., 2015) Medium, Medium-Large, Large, Huge 5000 km 2 JRC (Dottori et al., 2016) Medium (P), Medium-Large, Large, Huge 400 Differences in river network size between GFMs undoubtedly lead to differences in global flood exposure estimates. These differences can be even more pronounced at the national level, where GFMs have been used to inform disaster risk management (Ward et al., 2015). Flood exposure was calculated for the different GFM river thresholds using the GHS-POP layer. Globally, we found that exposure estimates between the river threshold which results in the largest river network (>50 km 2 UDA), and the river threshold which results in the smallest river network (>5000 km 2 UDA), differ by over a factor of 2. 405 If the size of the river network was further increased by reducing the river threshold to 10 km 2 UDA (below current GFM representation), the exposed population captured increases by 13%.
At the national level, in countries such as Suriname, The Republic of Congo, and Egypt, the greatest proportion of flood risk is posed by rivers with a UDA of 5000 km 2 or greater. In these countries, GFMs could be used interchangeably.
Understanding what size rivers pose a significant flood risk is key to accurately representing national flood risk. In Benin, 410 for example, the estimated flood exposure when a 5000 km 2 UDA threshold is applied is 0.49 million people. When the threshold is reduced to 1000 km 2 UDA, the estimated exposure increases to 1.8 million people. Some countries do not have large rivers flowing through them, and the flood risk will result entirely from smaller rivers. Often these are island nations, such as in Jamaica or Trinidad and Tobago, where all flood risk is from rivers smaller than UDA 1000 km 2 . However, in Andorra for example, a landlocked country, to capture any flood exposure, a 50 km 2 UDA threshold is needed. 415 To aid national level flood risk practitioners in their choice of GFM, we calculated the minimum river threshold required to capture a given percentage of the largest river network's (>50 km 2 UDA) national exposure. Exposure percentages ranging https://doi.org/10.5194/nhess-2021-102 Preprint. Discussion started: 9 April 2021 c Author(s) 2021. CC BY 4.0 License. from 10-90% were calculated for each of the three population datasets used in this study and mapped for each nation, globally. All 27 maps are included in the Supplementary Material. Figure 8, which shows the minimum river threshold required to capture at least 50% of possible GHS-POP exposure illustrates these results. The map shows that while in some 420 countries GFMs could be used interchangeably, in others, the size of the river network could significantly impact national flood exposure estimates. It is difficult to exhaustively compare global flood exposure estimates from previous GFM studies as often exposure is expressed differently (e.g. expected annual exposure (EAE) vs. exposure to a return period flood) and sometimes global exposure is not reported at all. In the comparable studies, there is significant variation in global flood exposure estimates. In The need for independent model comparison studies was met by Trigg et al. (2016b) and Aerts et al. (2020) who compared GFM output in Africa and China respectively. These studies compared the output of multiple GFMs, finding large disagreement between the modelled flood extents. Both studies also found large variations in calculated exposure. However, differences in exposure calculated by the GFMs were found to be influenced just as much by different model forcing and 435 resolution as by differences in river network size. Uncertainty in GFMs needs to be explored across the model cascade to identify where the models need to improve. Studies such as Zhou et al. (2020), which explores uncertainty in model forcing; and this study, which explores uncertainties in river network size, are important steps in directing future model development.
Granularity of input data is the main obstacle to increasing river network size in GFMs. The terrain data in all these models, which strongly influences their performance, is derived from the Shuttle Radar and Topography Mission (SRTM), a mission 440 over two decades old (Farr et al., 2007). New, 1 arc second resolution (~30 m at the equator) global DEMs have recently been released by both the National and Aeronautics and Space Administration (NASA) and the European Space Agency (ESA). The ESA DEM is particularly important as its elevation is based on newer satellite data from TanDEM-X. A new method for deriving an elevation map from satellite images has also been developed by Google, capable of generating DEMs at 1m resolution (Nevo, 2019). Whether its terrain or climatology data, new and improved methods are constantly being 445 developed and better datasets are being released. There is scope in the near future for increasing river network size in GFMs.
This comes at a computational cost, however; whether it's the use of a higher resolution DEM or the exponential increase in number of rivers to model when the threshold river size is reduced. Understanding where the representation of smaller rivers is needed most, namely in areas of high exposure, would streamline the future development of GFMs, targeting improvements in areas where flood risk is highest. 450

Conclusions
This study has presented the first global picture of flood exposure categorised by different sized rivers. We introduced a simple geomorphological approach to delineating a river's flood susceptibility, which is suitable for global scale 'first look' studies such as this and importantly, allows an assessment of river network size independent of global flood model structural and computational limitations. We find that over 75% of the global flood exposure is in Asia, with China and India making 455 up a significant proportion of this total. Streams (UDA 10-100 km 2 ) and small rivers (UDA 100-1000 km 2 ) are responsible for over half of India's flood risk. At the global scale, these rivers contribute to 45% of total flood exposure, emphasizing the importance of the incorporation of these smaller rivers into global flood risk studies. We find that large increases and decreases in flood exposure over the last forty years are a result of urbanisation, either inside the flood risk zone or outside of it. The effect that the choice of population dataset had on exposure calculations differed between countries. Globally, this 460 effect was most pronounced on smaller rivers, suggesting future studies that incorporate these smaller rivers should be careful in their choice of population data. Global flood models, the current tools for examining global flood risk, differ significantly in the size of their river networks. We found that the global flood exposure estimates differed by greater than a factor of 2 when calculated using the GFM river threshold which results in the largest river network (UDA >50 km 2 ) compared to the river threshold which results in the smallest river network (UDA >5000 km 2 ). These differences were often 465 more pronounced at the national level.
The results of this study are intended to inform both the developers and users of global river flood models. Consideration of river network size, and how this relates to exposure, is imperative to having a comprehensive picture of flood risk. Increasing https://doi.org/10.5194/nhess-2021-102 Preprint. Discussion started: 9 April 2021 c Author(s) 2021. CC BY 4.0 License. the size of the river network comes with both data and computational restraints. Doubling the resolution of the models (from 1km to 90 m to 30 m) requires an order of magnitude increase in computing power. Finer resolution grids are imperative for 470 representing small streams accurately. This has big implications for models currently operating at coarse resolution.
Modelling smaller rivers requires not only detailed high-resolution data, but also efficient modelling structures capable of running at higher resolutions. Understanding where the representation of small rivers is needed most (areas of high exposure) can focus future model development. Similarly, accurate flood exposure estimates necessitate accurate population data. We have shown that the choice of population data used in exposure calculations can have an enormous impact on flood 475 exposure estimates and we have identified in which countries this disagreement is most extreme and have identified some of the reasons for this. Flood risk practitioners should use these results as guidance about which population layer is best suited for their locality and use. There is need for further research in this area, incorporating more population data, as these layers play such an integral role in flood exposure calculations. In addition to more comparative analyses, there is also an urgent need for this population data to be validated at the global scale with actual data collected on the ground. Only then can 480 definitive conclusions be drawn about the appropriate use of different population datasets. The selection of GFMs available to the end user is large and increasing. However, differences in the size of river networks between the models can have a significant impact on flood exposure estimates. While available GFMs could be used interchangeably in some countries, in others, discrepancies in river network size would lead to vastly different national flood exposure estimates. The results of this study should help to inform GFM users about the appropriate choice of GFM for their country of interest. 485

Author Contributions
MVB and MAT conceived of the study. MB designed and carried out the analysis. MAT PAS CCS AMS supervised the project. MVB drafted the manuscript. All authors contributed towards the discussion and editing of the manuscript.

Data Availability
All the data used in this study is freely available to download. The River Flood Susceptibility Maps are available from the University of Leeds at https://doi.org/10.5518/947. Facebook's High Resolution Settlement Layer can be downloaded by 500 following the instructions in this link https://dataforgood.fb.com/docs/high-resolution-population-density-mapsdemographic-estimates-documentation/. The GHS-POP data can be downloaded here http://doi.org/10.2905/0C6B9751-