Crowd-based spatial risk assessment of urban flooding: Results from a municipal flood hotline in Detroit, MI

Abstract

Climate change is increasing the frequency and intensity of extreme precipitation events, raising the risk of urban flood disasters. This study uses a crowd-sourced municipal call database to characterize the spatial distribution of flood risk in Detroit, MI. Call data including dates and addresses were obtained from the City of Detroit Department of Public Works for 2021. Calls were mapped and aggregated to census tract counts and merged with neighborhood-level data. Associations of predictors with flood calls were tested using spatial regression models. Flooding calls were located throughout the city but were concentrated in specific areas. Multivariate models of census tract level call counts indicated that increased poverty and Black, immigrant, and older residents were positively associated with flood calls, while increased elevation was associated with protective effects. Longer distances from waste water interceptors were associated with higher risk for calls. Crowd-sourced flood hotline call data can be used for effective spatial flood risk assessment. Though flooding occurs throughout the city of Detroit, infrastructural, neighborhood, and household factors influence flooding extent. Limitations included the self-reported nature of calls. Future modeling efforts might include input from local stakeholders to improve spatial risk assessment.

1 |. INTRODUCTION

Climate change is increasing the frequency and intensity of extreme precipitation events (Easterling et al., 2017), dramatically raising the risk of flooding, which is already the most common and widespread type of weather-related disaster on Earth (National Oceanic and Atmospheric Administration, 2022). Nearly a sixth of the population of the United States is at high risk for flooding, and those in coastal cities are at particularly high risk (Rentschler, 2022). Flooding has deleterious impacts on physical and mental health (Alderman et al., 2012) in addition to having severe economic effects, including property destruction and interruptions to economic activities (Gray, 2008; Zhong et al., 2018). Flooding and extreme precipitation events in the US often occur suddenly and without warning and disproportionately impact minorities and marginalized groups (Adepoju et al., 2021; Linscott et al., 2021; Maldonado et al., 2016; National Academies of Sciences, Engineering, and Medicine, 2019; Rentschler, 2022; Zanocco et al., 2022). The problem of urban flooding in the United States will be among the greatest challenges the country will face in the coming decades.

Home flooding in the North American context can occur through one or more pathways during a sudden extreme precipitation event, including stormwater overland flooding, water infiltration or seepage through walls, or sewer backups (Irwin et al., 2018; Sandink, 2016). In the United States, the risks for flooding result from a confluence of climatic, environmental, infrastructural, social, and household factors (IPCC, 2012). Increased heat causes atmospheric moisture retention, which results in sudden and intense rain storms (Armal et al., 2018; O’Gorman, 2015; Trenberth, 2011). The subsequent inundation of water can overwhelm insufficient and aging local stormwater management infrastructure and put neighborhoods and homes at risk (Rosenzweig et al., 2018; Thorsby et al., 2020). For example, interceptor drains prevent the accumulation of stormwater on surfaces in urban areas, but they can become overwhelmed when the inflow of water exceeds the drains’ capacity or when the system becomes blocked (Xu et al., 2014). Social and racial disparities in the quality of local water infrastructure—along with historical segregation and “redlining”—increase risk for some groups while lowering risk for others (Bodenreider et al., 2019; National Academies of Sciences, Engineering, and Medicine, 2019; PR Newswire, 2021; VanDerslice, 2011). Finally, households may not be resilient to flooding: cracks in basements or uncapped sewer outlets put homes at even higher risk for groundwater infiltration or sewer backups, as was found in a previous study in Detroit (Irwin et al., 2018; Larson et al., 2021; Peirce et al., 2022). Urban areas characterized by poverty commonly have older homes and have historically lacked infrastructure to support water drainage (Green et al., 2021; Rosa & Pappalardo, 2020), and residents or landlords may lack financial resources to repair and mitigate potential problems (Larson et al., 2021). Additionally, the percentage of Black families who live in substandard housing—which can include leaky roofs, windows and basements—is more than three times higher than the percentage of white families (Jacobs, 2011), suggesting that housing is a major source of inequity in disaster resilience.

Income inequality and natural hazards are closely linked (Howell & Elliott, 2018), and the level of disaster vulnerability varies with race, class, and income (Cutter, 1996; Fussell et al., 2009). Detroit, MI experiences a wide range of converging challenges, including population loss, demographic change, and decades of financial and political neglect (Sugrue, 2014). As of 2020, Detroit’s population is ~670,000 people, down from a peak population of 1.6 million people in 1960 (United States Census Bureau, 2020). Due to rapid “white flight” of the middle to late 20th century (Jackson & Leary, 2016), nearly 80% of Detroit residents are African-American. Detroit is among the poorest of large US cities, with a median household income of $31,000, less than half of the Michigan state median income of $72,000 (United States Census Bureau, 2020). The city struggles to provide and maintain public services as a result of declining tax revenues, large debt obligations, low levels of investment, and declining population density. This requires Detroit to provide services using antiquated systems appropriate for a much larger population (Dewar & Thomas, 2013; Jackson & Leary, 2016). As a result, Detroiters are at high risk for flooding and household water inundation. This is confirmed by surveys showing that more than half of homes experience regular flooding (Larson et al., 2021; Nassauer et al., 2019) and qualitative studies confirming the lived experiences of people living under the threat of water entering the home (Sampson et al., 2018).

Accurate knowledge of the true extent of flooding within urban areas is essential to the development of efficient and comprehensive flood disaster response. Flood risk zones as classified by the Federal Emergency Management Agency (FEMA) are often used to determine local flooding risk. However, this measure has been shown to be influenced by economic and political concerns (Hino & Burke, 2020) and might not effectively reflect true flood risk or extent, particularly in areas with aging water diversion infrastructure. Further, the FEMA flood risk zones might fail to adequately and objectively assess spatially granular and individual household flooding risk.

One strategy to assess flood risk is to use model-based predictions based on available measures such as elevation and rainfall (Smith et al., 2011). However, the multilevel nature of urban flooding presents unique challenges in assessing the extent and determinants of home flooding. Geographic factors such as elevation and proximity to water might determine baseline risk, but water and sewer infrastructure may protect against flooding inconsistently, diverting water away from some areas while increasing water levels in other areas. Neighborhood factors such as blocked drains might prevent or slow drainage, contradicting risk estimates based on elevation or proximity to water. Further, poor housing quality and associated cracks in basements, holes in roofs, and uncapped sewer drains might allow water inundation from both above and below and thus exacerbate individual flood risk (Larson et al., 2021). Lack of data on any or all of these levels—environmental, neighborhood, or household—might complicate the ability for statistical and mathematical models to accurately predict local flood risk.

Crowd-sourced data offers an alternative means of determining flood risk. Examples from the literature include research using Twitter streams to identify locations of flooding and non-flooding related emergency events (Ao et al., 2014; de Bruijn et al., 2018; Sakaki et al., 2013). One study retroactively analyzed cell phone activity to detect anomalous human behavior that might indicate disruptive events (Dobra et al., 2015). Studies have used mobile phone data to track displacement and changes in population after disasters, such as a hurricane in Puerto Rico (Acosta et al., 2020) and earthquakes in Nepal (Wilson et al., 2016), Japan (Yabe et al., 2019), Haiti (Bengtsson et al., 2011), and China (Chaoxu et al., 2019).

However, crowd-sourced data from social media and cell phone location records present important limitations for assessing the spatial extent of flooding impacts in urban areas, which may be highly localized. Inaccuracies in geographic positioning systems may not accurately measure the location of the flooding event (GPS gov, n.d.) to determine the exact source of the flooding. This limitation complicates strategies to mitigate street-level flooding or sewer backups that rely on precise, real-time spatial information. Without the accurate home location, investigators might miss important neighborhood and household factors that have been found to be important determinants of flooding risk (Larson et al., 2021; Woodruff et al., 2021).

1.1 |. Research goals

There were three main goals of this research. First, we sought to assess the spatial extent and determinants of flooding in a large, flood-prone metropolitan city using a unique crowd-sourced data set of calls to a municipal flood hotline. Second, we sought to test the applicability or usefulness of using crowd-sourced, hotline-based data as a means of determining flooding extent. Finally, we explored infrastructural, social, neighborhood, and household determinants of flooding to better understand possible predictors of the spatial distribution to identify populations at risk.

This research tested three main hypotheses. First, we hypothesized that there would be distinct spatial patterns of flooding within the city of Detroit. Second, we expected that flood risk would be positively associated with increased numbers of minority residents, poverty, and rental properties. Third, we believed that flooding would be correlated with a combination of household, environmental, and infrastructural factors.

2 |. METHODS

2.1 |. Flood call and census tract level data

The data set comprised calls to a municipal hotline operated by the Detroit Water and Sewage Department (DWSD). The time and date of each call are recorded electronically. Callers are requested to provide the address of the incident, and operators record addresses manually. Residents of the city of Detroit are allowed to call the hotline to report basement flooding, as well as issues that may cause basement flooding, such as catch basin blockage and street flooding. Public media announcements encourage all Detroit residents to call the hotline as soon as possible during or after a water-related event. DPSW internally identifies major and important flooding events within the city limits of Detroit. DPSW then assigns calls to each event based on the date of a major precipitation and/or flooding event. Calls may be on the same day as the flooding event or up to several days following. Callers can be homeowners or renters; there is no requirement that landlords alone report basement flooding or water inundation. Though the hotline is intended to respond the concerns of people in residential units, no distinction is made between residential and commercial callers. There may be repeated calls from the same household. In that case, only one of the calls is included in the dataset. No information on the type of flooding (pluvial flooding, water leakage, sewer backups, etc.) was available. We expected there to be variation in the number of calls by flooding event. During the analysis phase, we explored temporal patterns of hotline calls by event date. However, for the statistical models we opted to select the date with the largest number of calls, since the goals of the research were spatial in nature. Latitude and longitude coordinates for each household were added to the data set using the home address listed in the data set and the geocoder tool in ArcMap ver. 10.6.1 (Environmental Systems Research Institute, 2011). Points outside of the outer boundary of the city limits of Detroit were excluded from the analysis. Census tract level measures of poverty, age, race and ethnic composition, and types and age of housing were obtained from the American Community Survey (ACS), a sample-based yearly household survey maintained by the United States Census Bureau (United States Census Bureau, 2020). Data for administrative boundaries and locations of waterways were downloaded from the State of Michigan’s GIS Portal (GIS Open Data, n.d.). One-meter elevation data were obtained from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Digital Elevation Model from the National Aeronautics and Space Administration (NASA) (NASA, METI, AIST Japan Space systems & US/Japan ASTER Science Team, n.d.). Composite indices of affluence and disadvantage were obtained from the National Neighborhood Data Archive (NaNDA). NaNDA is a publicly available data archive containing measures of the physical, economic, demographic, and social environment at multiple levels of spatial scale (e.g., census tract, ZIP code tabulation area, county) (National Neighborhood Data Archive (NaNDA), 2022). Data from all data layers were extracted to the flooding data set based on the latitude longitude location of the reported address of the call. We also obtained data from the City of Detroit for water diversion infrastructure, including the locations of all stormwater interceptors. Using simple Euclidean distance (“as the crow flies”), we calculated the distance from each household to the stormwater interceptor and the nearest point on the Detroit River and added these to the database.

2.2 |. Analytic methods

We took three approaches to analyzing the data. First, we performed a descriptive analysis of the individual flood calls to assess the spatial distribution of calls in Detroit with the aim of identifying obvious spatial patterns. To do this, we started by mapping the latitude longitude locations of all calls for all event days combined. To assess possible spatial patterns or clustering of call “hot spots,” we visualized the distribution of flooding using the kernel density. The kernel density estimates the density of points (in this case calls to the flood hotline) within a moving window around each point. It is a non-parametric method widely used for risk mapping and hotspot detection (Yin, 2020). Further, we explored each of the individual flooding events (by date) to identify which one might have the best sample size to perform a deeper analysis of the data.

Next, we aggregated calls by census tract to allow for the statistical testing of the numbers of flood calls by tract with census tract level factors that might be predictive for flooding locations (United States Census Bureau, 2020). One of the major goals of this research was to test for statistical associations of neighborhood-level and environmental predictors and flooding. We tested for crude, bivariate associations of the number of homes flooded using an offset of the total numbers of residential units within the census tract with all census tract level variables included in this analysis (percent in poverty, percent African-American, number of older homes within the census tract, etc.) We produced estimates and tested for the statistical significance using linear regression (OLS) models. We took this crude approach to allow for easy interpretation of the magnitude and direction of the association of census tract level predictors with the number of calls, even though OLS models—with their assumption of normality in the outcome—are not appropriate for this type of highly skewed, count based outcome.

Finally, we used Poisson regression to better model the association between predictors and flood related call counts within the census tract level aggregate units. We created a full model using all of the chosen predictors for this analysis. To identify the magnitude and direction of the association of each variable while also controlling for all other predictors, we produced parameter estimates and confidence intervals of the effect of a unit change in the predictor on the log expected counts of calls. Of course, one of the major assumptions of linear regression modeling is independence between observations. A model which does not take spatial autocorrelation into account might lead to spurious correlations or biased effect estimates (Ploton et al., 2020). As a first step to assess spatial autocorrelation, we calculated Moran’s I for the chosen flooding event. Moran’s I is a simple measure that assesses the amount of statistically significant clustering within a spatial data set (Moran, 1950). To account for possible spatial autocorrelation, we created a spatial lag model, including spatial terms into the model based on principal coordinates of neighbor matrices (PCNM). PCNM is a method that creates spatial predictors that can be easily incorporated into regression. The method diagonalizes a spatial weighting matrix and then extracts the eigenvectors that maximize the Moran’s index of autocorrelation (Dray et al., 2006). We present both types of models (non-spatial and spatial) in a single table and plot the residuals of both models to determine the presence of statistical “hot” and/or “cold” spots, that is, areas where counts are significantly higher or lower than the expected number of calls in that tract, given the predictors.

3 |. RESULTS

The data set comprised six flooding events in Detroit, MI that occurred between January and July 2021. A flooding event starting on June 25 resulted in the most calls (25,361), followed by an event on July 16 (4455) (Table 1). Events before June 25 resulted in relatively few calls overall and very few per census tract. The June 25 event, however, saw a mean of more than 73 calls and a median number of 38 calls per census tract (see Table 3 and Figure 1).

TABLE 1.

Number of calls related to major flooding events in Detroit, MI 2021.

	Flooding date	Total calls	Mean calls by census tract	Median calls by census tract
1	January 1	578	1.68	1.00
2	January 13	303	0.88	0.00
3	March 30	230	0.67	0.00
5	May 19	238	0.69	0.00
5	June 25	25,361	73.17	38.00
4	July 16	4,455	12.91	9.00

TABLE 3.

Census tract level summary statistics for all predictors used in this analysis. Means and standard deviations are presented along with the number of non-missing observations out of all 345 census tracts in Detroit.

	Mean (SD)	N
Poverty (%)	39.64 (14.20)	339
Black (%)	75.46 (32.40)	342
Older units (%)	34.87 (21.81)	39
Rental units (%)	50.60 (18.15)	339
Older residents (%)	66.37 (12.07)	339
Ethnic immigrants (%)	6.04 (11.98)	339
Affluence (0–1 scale)	0.18 (0.14)	339
Disadvantage (0–1 scale)	0.38 (0.12)	339
Distance to nearest interceptor (km)	3.44 (2.49)	345
Elevation (m)	189 (6.82)	345

FIGURE 1.

Calls to Detroit Department of Public Works by internally determined incident date. Daily mean precipitation for Wayne County is shown in grey.

3.1 |. Spatial distribution of flooding calls

Flooding calls were located throughout Detroit but were concentrated on the east side bordering the suburb of Grosse Pointe and in the southwest bordering the city of Dearborn (see Figure 2). Spatial patterns of calls to the hotline varied by flood incident but were mostly concentrated in northeast Detroit—bordering Grosse Pointe— and in southwest Detroit, north of East Dearborn (see Figure 3). There were few calls in areas along the Detroit River, which is partially explained by the presence of a parkway and fewer residential units. There was evidence for spatial clustering for all dates with the exception of March 30, which suggests that calls are not randomly distributed in space but follow some possibly predictable pattern (see Table 2). The highest amount of spatial clustering was seen on the June 25 event, followed by the July 16 event.

FIGURE 2.

Locations of all calls for all days within the city limits of Detroit and in relation to neighboring communities.

FIGURE 3.

Kernel density of locations of all flood-related calls.

TABLE 2.

Moran’s I index values and significance for all flooding dates to assess the level of evidence that might suggest spatial clustering or a non-random spatial distribution of flooding events.

	Date	Moran’s I	p
1	January 1	0.26	≤0.0001
2	January 13	0.20	≤0.0001
3	March 30	0.07	0.069
4	May 19	0.22	≤0.0001
5	June 25	43	≤0.0001
6	July 16	0.31	≤0.0001

3.2 |. Overview of census tract level measures

We used data from the ACS for the percentage of homes in poverty and the percentage of households where residents were Black, ethnic immigrants, and elderly. We also used census tract-level measures of affluence and disadvantage as an indicator of general social resilience. To test associations of flooding to important points of water diversion infrastructure, we calculated the distance to the nearest stormwater interceptor from each household location. Descriptive statistics of census tract-level measures are presented in Table 3. Nearly 40% of Detroit households are below the federal poverty level. Nearly 40% of homes within census tracts in Detroit are occupied by Black residents. More than a third of homes were built before 1960. Nearly two-thirds of homes are occupied by at least one elderly resident. Detroit ranks high on the disadvantage metric and low on the affluence metric used for this analysis. On average, homes are located 3.44 km from a water interceptor. Elevation in Detroit is 189 m above sea level with very little variation (standard deviation 6.82 m) (see Table 3). Census tract level poverty and Black residents were weakly correlated (r = 0.22), presumably due to the high percentage of Black households throughout the city (results not shown).

The spatial distribution of the census tract level variables varies by measure (Figure 4). Poverty and elderly residents are spread across the map. All areas outside the southern area that bounds East Dearborn are overwhelmingly Black and disadvantaged, while the southern area has the largest percentage of ethnic immigrants. Rental units and affluent areas are concentrated in the downtown area close to the Detroit River. The only areas not proximal to a water interceptor are the middle region in the northern part of Detroit and the area adjacent to the eastern area that bounds Gross Pointe. The lowest lying areas are along the Detroit River. We found that census tract level percent of homes that were in poverty, occupied by Black families, and were older constructions were positively associated with the number of calls made to the hotline on June 25. The composite disadvantage and affluence measures were significantly positively and negatively associated with calls. The percentage of homes identified as rental units, as being headed by elderly residents, and having one or more ethnic (non-European) immigrants and the distance to the interceptor was not significantly associated with calls in the bivariate models.

FIGURE 4.

Spatial distribution of all census tract level predictors.

3.3 |. Bivariate models of census tract counts of calls given tract level predictors

We created individual linear regression (OLS) models for each predictor as an exploratory step that would allow us to more easily interpret the direction and magnitude of association of census tract level measures on aggregated counts by tract. A unit increase in the fraction of households in poverty was associated with a mean increase of 113 flood-related calls. Similarly, disadvantage (140), percentage Black (30), older units (70) all had positive and significant associations with flood-related calls. Increased affluence was significantly associated with a decreasing number of calls. Rental units and the number of homes with elderly residents were not significantly associated with calls. Increased distance from interceptors and higher elevation were both associated with a lower number of calls, but these relationships were not significant.

3.4 |. Multivariate models of census tract level flood call counts and tract level predictors

A multivariate model was created using all predictors included in this analysis. The full model was compared with all other submodels using Akaike’s Information Criterion (AIC) (Bozdogan, 1987) and was found to be the best model (Table 4).

TABLE 4.

Bivariate linear regression models (OLS) of the census tract number of calls for each predictors included in the analysis.

	Predictor	Estimate	Standard error	p
1	Poverty	113.1	16.31	≤0.0001
2	Black	30.37	7.42	≤0.0001
3	Older units	70.97	10.67	≤0.0001
4	Rental units	3.97	13.62	0.771
5	Elderly	−8.63	20.48	0.674
6	Disadvantage	140.19	19.96	≤0.0001
7	Affluence	−127.23	16.82	≤0.0001
8	Distance to interceptor	−5.92	9.88	0.549
9	Elevation	−5.38	3.58	0.135

Note: Estimates for poverty, Black households, older units, rental units, elderly residents, and ethnic immigrant households represent the change in the number of calls given a 10% increase. Disadvantage and advantage represent the change in the number of calls given a 0.1 change in either. Estimates for distance to interceptor and elevation on the predicted change in flood calls are given calculated given a 1 km and 1 m increase, respectively.

After comparing several models, the best model included all variables initially selected for the analysis. In the non-spatial (OLS) model, census tract-level calls to the hotline were positively and significantly associated with poverty, percentage Black, older units, elderly residents, the distance to the nearest interceptor, and the affluence index. Rental units and ethnic immigrants were negatively associated with calls. When accounting for spatial auto-correlation in the residuals, poverty was no longer significantly associated with hotline calls, and the effects of older residents were diminished. See Table 5 for full results. Comparing residuals of the non-spatial and spatial models, the non-spatial model indicated spatial patterns of flooding risk (Moran’s I: 0.39, p = 0.001), with areas far away from the river having consistently lower numbers of calls than areas close to Grosse Pointe or near the Detroit River. The spatial model accounted for some of the spatial autocorrelation (Moran’s I: 0.18, p = 0.001) but indicated that census tracts close to the Detroit River had higher numbers of calls than elsewhere (see Figure 5).

TABLE 5.

Multivariate Poisson regression models of census tract level counts of calls to municipal flooding hotline, accounting for overdispersion in the outcome.

Dependent variable: Number of flooded homes in census tract
	Poisson regression models
	Non-spatial model	Spatial model
Poverty	1.210*** (1.087, 1.333)	1.135** (1.017, 1.254)
Black	1.376*** (1.297, 1.454)	1.408*** (1.325, 1.491)
Older residents	1.144*** (1.046, 1.242)	1.018 (0.922, 1.115)
Ethnic immigrants	1.254*** (1.110, 1.398)	1.466*** (1.319, 1.613)
Old units	1.016 (0.961, 1.071)	1.009 (0.955, 1.064)
Rental units	0.841*** (0.768, 0.914)	0.841*** (0.769, 0.912)
Disadvantage	0.790 (0.542, 1.038)	0.882 (0.636, 1.129)
Affluence	0.789*** (0.651, 0.927)	0.782*** (0.644, 0.919)
Distance to interceptor	1.127*** (1.085, 1.169)	1.117*** (1.076, 1.159)
Elevation	0.905*** (0.890, 0.919)	0.886*** (0.870, 0.901)
Observations	345	345

Note: A non-spatial model is presented along with the same model but includes terms accounting for spatial autocorrelation between census tracts. Estimates represent the mean change in expected counts of flooded households given (1) a 10% increase in the percentage of all homes within each census tract for poverty, percentage Black, older residents, ethnic immigrants, old units, and rental units; (2) a 0.1 increase in the disadvantage and affluence metrics (on a 0–1 scale); and (3) and 1 km increase in distance to the nearest interceptor and a 1 m increase in elevation.

^*p < 0.1;

^**p < 0.05;

^***p < 0.01.

FIGURE 5.

Plots of the residuals of the multivariate non-spatial and spatial Poisson regression models of census tract level counts of calls from flooded homes during the June 25th flooding event given predictors for poverty, Black and older residents, ethnic immigrant households, older and rental housing units, measures of disadvantage and affluence, distance to the nearest interceptor, and elevation. These plots allow the identification of “hot” and/or “cold” spots where the numbers of flooded homes significantly deviates from expected counts.

4 |. DISCUSSION

Using a municipal, phone-based data set of flooding reports, we have characterized the spatial distribution of flooding events in a flood-prone, large city in the midwestern United States. We have identified specific risk factors associated with flooding reports and, by proxy, flooding risk. Our study is one of only a handful of studies which use crowd-sourced data based on municipal hotline calls to assess flooding extent and/or risk. In this research, we found that census tracts with increased poverty and numbers of Black occupied homes were associated with higher numbers of flood reports. Thus, these census tracts might be at higher risk for flooding than census tracts with a smaller percentage of homes occupied by Black and low-income residents. This would agree with other studies on flood risk that suggest communities of color and low-income residents are at disproportionate risk for urban flooding (Keenan et al., 2019).

As expected, we found that increased elevation was associated with a lower risk for flooding. However, in the multivariate model we found a positive relationship of distance to storm-water interceptors and flood risk. This could indicate that areas far from interceptors are underserved for water diversion, suggesting a shortfall in water-related infrastructure in Detroit. We found that calls were less likely to come from areas with large numbers of rental units. This could be explained by newer multiunit housing developments that are more resistant to water entering the home. It could also suggest that renters are less likely to call the hotline; presumably they would call their landlords, and their landlords might not utilize the hotline. We expected flood risk in areas with large numbers of older homes to be elevated, but—though it remained in the multivariate model—the relationship was non-significant. This could be a statistical feature, but it could also be associated with older and worse homes being demolished or recent efforts to upgrade and improve older homes. However, we found a U-shaped association of home age with flood risk in another study (Larson et al., 2021) that suggested flood risk was highest among homes built in the mid to late 20th century, compared with pre-war and 21st century homes. More work needs to be done to assess how housing quality influences flood risk.

These results have implications for flood risk assessment and response planning for flood related disasters in urban areas. First, our results should indicate that flood risk exists throughout the city of Detroit and not just a few areas indicated by, for example, FEMA flood risk maps. Calls were made to the hotline on multiple days, and calls were not simply restricted to the water front, even on days where there were fewer calls. We found that the determinants of flood calls (and thus flooding) were not restricted to elevation or even water diversion infrastructure, but rather comprised numerous social and neighborhood factors. Our study was not designed to assess the impact of individual housing conditions on flood risk, but previous research has indicated that the end experience of water in the home is the result of a cascade of factors. This would suggest that researchers and policymakers broaden definitions of flood vulnerability when assessing flood risk, re-evaluate methods of communicating risk to include highly localized factors, and create new means of prevention and mitigation before and after heavy precipitation events.

Detroit, MI suffers from a dual risk of river and pluvial flooding. Its proximity to the Detroit River and low elevation puts homes at risk during sudden river rises, and its aging combined sewer outflow system insufficiently diverts water to the river during extreme precipitation events. We found that flood risk was inversely associated with distance to the nearest stormwater interceptor, but this relationship was non-significant. Both of these results might first suggest that our geographic range was insufficient to capture variation in risk based on distance to the river or water diversion infrastructure. These results might also suggest that risk for river and pluvial flooding is shared among all residents of Detroit, MI. The totality of our results suggests that flood extent in the city of Detroit might be determined by neighborhood and household factors, rather than simply geographic or infrastructural factors.

There were a number of important limitations to this study. Calls may not be made or may be delayed when residents are away from the area during the flooding event. Some residents, particularly transient residents, might not be aware of the hotline. Yet others might not perceive the hotline as a useful means to obtain assistance. Others may not wish to use a telephone-based system, preferring to report through web-based portals. Further, not every flooded household will call the hotline with the same probability, and—as with crowd-sourced social media data—there might be biases related to gender or age (Garrote et al., 2019).

Studies using social media based data to assess impacts of natural disasters found that reporting events on Twitter during Hurricane Sandy, for example, was more determined by socioeconomic factors than damage to homes (Xiao et al., 2015). We might assume also that such biases exist in calls to municipal hotlines, but we note that Tweets—which are intended to simply speak to the public—differ from calls to municipal services, which are assumed to be calls for help. All of these factors might influence the reliability of the call data. Buy-in and participation from community members, particularly those from marginalized communities, is essential for timely and accurate assessment of the spatial extent of flooding and the identification of flood-impacted areas (Samaddar et al., 2022).

Another limitation of this study was that we could not distinguish between multi- and single-unit dwellings. Multiunit dwellings can often include rental units, so a single call from a landlord or building owner might actually represent a larger number of units affected. Future studies might use tax record data to make the distinction between multi- and single-unit dwellings and even better characterize associations of flooding with rental/owner occupied status.

Properly assessing flood risk should be a major priority for flood responders and policymakers. The approach used here assumes that reporting might be a sufficient proxy for flooding risk. Reporting is a response to the realization of a flooding event. An incident cannot be reported if it has not occurred. Reporting an event also represents access to a reporting means and its perceived utility. People will not report if they are unable to do so, and they are unlikely to report if they do not believe that it will result in a useful response. Risk is a function of the flood hazard at a location, and the level of exposure and vulnerability to the flood hazard depends on proximity to a flood plain or water outlets (Shabman et al., 2014). Thus, the two concepts—reporting and flood risk—are intrinsically different. Further, flood risk is dynamic. An extreme flooding event might result in a response which alters the nature of flood risk. Thus—as we have shown—reporting can be a useful method of assessing spatial flood risk based on who has been flooded in the past, but it might not be useful for forecasting where flooding might occur in the future. Regardless, we feel that reporting data can help answer questions about which demographic groups might be most impacted by flooding. This could help policymakers tailor response resources so they have the most impact for specific groups (low SES, minority groups) or individuals (those living in substandard housing) at risk.

We have taken an approach that expands on our previous work on flooding in Detroit (Larson et al., 2021). We would like to validate our results against other data sets, such as FEMA insurance claims databases, that would allow us to compare the spatial distribution of calls with other types of data for the same events that this research explored. Further, this research uses available census tract level data to explore factors that contribute to flood risk and/or reporting. An approach which takes the input of regional, community, and individual stakeholders into account would provide results that are more relevant to the needs of those stakeholders (Sampson et al., 2018). The input of people directly impacted by flooding and flood risk in the city of Detroit and insight into their lived experiences should be considered a priority for informing future modeling efforts (Prell et al., 2007). Participatory input on the possible sources of flooding and specific household factors which might elevate or mitigate flood risk would greatly improve the outcomes and applicability of model based flood risk estimates (Maskrey et al., 2022). This is particularly true in Detroit, where the risk of home water inundation is determined by a set of factors outside of precipitation or hydrology.

5 |. CONCLUSIONS

Crowd-sourced flood hotline call data can be considered for use as a tool to assess spatial flood risk, but care must be taken to account for possible biases due to socioeconomic and technological factors. Geolocated, crowd-sourced data of this kind creates opportunities for deeper analyses of flood risk and its determinants, both on a local level and to draw general conclusions about the root causes of home flooding. Using this data from Detroit, MI, we found that though flooding occurs throughout the city of Detroit, infrastructural, neighborhood, and household factors influence flooding extent. These results should prompt deeper discussion of the determinants of flood risk and should help policy makers and stakeholders design and implement strategies to mitigate risk.

DATA AVAILABILITY STATEMENT

Data can be made available on reasonable request. Code is available on the Github repository https://github.com/kambanane/Flood_calls.