Urbanization and floods share a wicked interaction: as cities expand towards natural drainage systems, the flood hazard intensifies, exposing more urban areas to risk. This dynamic interaction is context-dependent and particularly complex in Sub-Saharan African (SSA), one of the regions most affected by floods globally¹. The development of SSA cities has been shaped by colonial planning practices, which continue to shape their spatial and social structure of urbanization today. These historical practices concentrated resources and infrastructure in central parts of the cities, thereby serving a limited demographic while marginalizing large segments of the population²-. As a result, many contemporary SSA cities are characterized by rapid unplanned urbanization, leading to the emergence of Deprived Urban Areas (DUAs)-neighborhoods which face high levels of deprivation, limited infrastructure, tenure insecurity and frequent exposure to hazards^2,3. In SSA, these areas are estimated to house around 60% of the urban population⁴. However, urban planning and disaster risk reduction programmes remain mainly siloed initiatives, focused on formal, planned districts, leaving DUAs outside despite facing great vulnerabilities⁵.

Previous work acknowledged that certain social groups are likely to experience greater flood exposure and risk than others^6–8. This differential exposure is far from random. In SSA cities, increased flood exposure is often concentrated in areas historically marginalized by planning provision, where settling in floodplains is one of the few options for the urban poor⁵. Global assessments estimated that low-income populations are disproportionately affected, where 74 million people are exposed both to flooding and extreme poverty in SSA countries⁹. Future projections also indicate that flood exposure will be concentrated in low-GDP countries, further amplifying impacts in underserved communities¹⁰. Most flood exposure studies still lack localized assessments capable of distinguishing between DUAs and non-DUAs. Such information, yet scarce, is highly needed in both research and policy, particularly given that exposure and vulnerability assessments are the foundation for equitable disaster risk reduction and adaptation measures^11,12.

To date, locally accurate flood modelling tools in data-scarce environments remain a challenge. Global flood models, such as GloFAS¹³ or Fathom¹⁴, though increasingly sophisticated and largely used for global and national assessments, show large inconsistencies at the city scale, with broad modelling assumptions and limited validation in many regions^15,16. These models primarily capture large-scale fluvial and coastal flooding, with inadequate resolution of urban pluvial dynamics¹⁷. Conversely, traditional hydrological models like HEC-RAS¹⁸ and OpenLISEM¹⁹ can simulate localized flood dynamics, including pluvial and flash floods. However, their reliance on extensive calibration, high-quality input and technical capacity often renders them unavailable or unaffordable for local authorities in SSA⁸. Consequently, in data-scarce contexts, such methods typically serve as conceptual representations rather than practical instruments for local urban planning.

Compounding this gap is the dominance of top-down flood risk assessment approaches in scientific approaches. The modelling outputs of these approaches are typically assessed through remotely sensed or instrumental data²⁰, with little or no involvement of residents, particularly not from those living in deprived settlements²¹. As a result, the rich, experiential knowledge that local communities have about the complex local flood dynamics is commonly ignored. This disconnection between top-down and on-the-ground realities contributes to the systematic overlook of DUAs¹⁰. In response, this study positions citizen science as a core method to ground our flood model to the lived experience and knowledge of the residents of DUAs.

Recognizing both the limited evidence of differential flood exposure in SSA-deprived areas and the technical gaps of existing models, we present an alternative approach. Our study integrates a fast simulation flood model, open data and citizen science data to evaluate flood exposure across six SSA cities. Specifically, we aim to: (1) analyze flood exposure patterns, providing spatial evidence of the alarming flood impacts suffered by deprived urban settlements in both primary and secondary SSA cities; (2) assess the findings at local scale through citizen science, bringing the perspectives of deprived communities into the center of disaster risk assessments. We ask: How do flood exposure patterns vary across and within SSA cities? And to what extent does the proposed approach capture local realities? In the next sections, we unpack these questions through spatial, statistical and community-based evidence in an effort to frame flood exposure through the lens of urban deprivation.

We examined flood exposure in six SSA cities, specifically in one primary and one secondary city in Kenya (Nairobi and Kisumu), Ghana (Accra and Tema) and Mozambique (Beira and Chimoio). After modelling the hazard, flood exposure was derived from two built-up density datasets: the Global Human Settlement Layer (GHSL)²² and a processed version of Google Open Buildings (GOB)²³. In the absence of reliable reference data, particularly in deprived contexts, we used both datasets to explore how flood exposure estimates vary depending on the input source. The comparative analysis revealed the influence of input resolution, classification methods and flood depth thresholds. The analysis adopted a 50x50m resolution grid to capture intra-urban variations. We assessed city-wide exposure across three flood depth thresholds (1cm, 10cm and 50cm) and compared results by built-up density categories (low, medium and high) and layers of delineated deprived settlements.

We observed large differences across cities in the relative flood exposure, and a high variability depending on the used exposure dataset (Fig. 1). Using the GHSL, the proportion of built-up areas exposed to flooding found is 29% in Kisumu and 23% in Nairobi. In contrast, GOB reduced this estimate to 17% and 13% respectively. While GOB generally reported less exposed areas, due to its exclusion of roads, bare lands and open areas -features often classified as built-up by GHSL- this does not necessarily infer greater accuracy. Road infrastructure, for example, is also an element at risk and can exacerbate flooding by limiting water infiltration. Additionally, the omission of buildings in densely built-up areas in the GOB data²⁴ also plays a role in the smaller estimation.

Chimoio presents a distinct profile: nearly 80% of its built-up areas are estimated to flood at 1cm flood depth, declining to 10% at 10cm. This spatial distribution points to widespread shallow ponding depressions and systematically poor run-off conditions. A notable trend in all cities was the sharp decrease in flood exposure rates between 1 and 50 cm flood heights, suggesting that deeper inundations are more localized. For relative exposure at 50cm, Nairobi, Beira and Accra each presented approximately 8% of their urbanized territory exposed to floods. On one hand, this might appear modest, but in absolute values, the size of the built-up areas exposed to floods were substantial, i.e., using GHSL, we calculated flooded areas with 50cm floods to be 58 km² for Nairobi, 41 km² for Beira and 16km² for Accra. Considering that large parts of these areas intersect with high density areas, the estimated impacts on urban population can have massive consequences. On the other hand, contrary to expectations, absolute exposure was uniformly highest in the primary cities. Aside from Nairobi, Beira (83km²) and Kisumu (60km²) registered the highest exposure at the 10cm flood depth.

The association between exposure and the built-up density categories analysis revealed important patterns. Nairobi, Beira, Accra and Tema (which is also part of Greater Accra) consistently presented higher relative flood exposure (medium and high-density areas that are most flood prone). The two Ghanaian cities stood out with the largest flood exposure rates. In Accra, the GHSL dataset estimated around 80% of the flooded areas as medium and high built-up density areas across all flood depth thresholds. Tema followed with approximately flooded areas of 52% at 10cm flood depth and 40% at 50cm flood depth. Certainly, the urbanization patterns influence the exposed areas, as Accra and Tema portrayed a spread-out urbanization connecting multiple built-up density cores. In contrast, secondary cities with more fragmented urbanization patterns showed a dominant exposure in low-density zones, regardless of the used built-up dataset.

The comparison between the two built-up datasets furthermore revealed consistent spatial classification biases in the representation of built-up density. GHSL generally overestimated the density across all cities, overrepresenting peri-urban and transitional areas, where sparse and heterogeneous land uses are misclassified as high-density fabric. For example, at the 10cm flood depth threshold, GHSL estimated 35% of Accra’s flooded areas as high density, whereas GOB estimated only 12%. In Nairobi, however, the pattern was reversed; GHSL underestimated dense informal settlements, likely due to its lower sensitivity to complex morphologies²⁵.

The general pattern that the percentage of exposed areas shrinks as flood depth increases emerged but was less pronounced than expected. The data suggested that the same neighborhoods are repeatedly affected, regardless of the flood depth thresholds (see Supplementary Information). These persistent exposure patterns were especially clear in high-density areas, which might be connected to the highly dense DUAs, the limited drainage capacity from imperviousness, inadequate infrastructure and management capacity²⁶.

Analyzing the intra-urban patterns by overlaying flood exposure with the DUA’s boundaries, this study highlights sharp disparities: our estimates showed that in high- and medium-density zones, DUAs are up to 47% more exposed to floods above 10 cm when compared to non-DUAs. When considering only high-density zones, this ratio rises to 60%. However, we report 47% as a more representative comparison across cities with different urban forms. In secondary cities, few areas meet the high-density threshold, yet many medium-density DUAs are functionally equivalent to high-density contexts in primary cities. Including medium-density areas provided a more consistent measure of intra-urban assessment, avoiding underrepresenting flood exposure in smaller cities.

The degree of the exposure inequity varied considerably across the cities. Nairobi exhibited the highest overexposure, with DUAs being two times more exposed than non-DUAs, followed by Kisumu (52%), Accra (46%) and Beira (17%). In contrast, Chimoio and Tema showed near parity in exposure levels between DUAs and non-DUAs. This might indicate that, in secondary cities, limited overall investment in infrastructure and disaster preparedness leaves the entire city exposed, narrowing the intra-urban differences². Figure 2 presents the proportion of flooded areas in each city and underscores the divergences between DUAs and non-DUAs: flood-prone DUAs are consistently characterized by higher built-up densities. In most cities, flooded non-DUA zones coincide with non-built-up land, whereas flooded DUAs are predominantly located in moderately and highly built-up environments. This pattern is evident across both built-up datasets - GOB and GHSL - though GHSL consistently provides larger estimates in both DUA and non-DUAs, with the exception of the city of Nairobi. This outlier is likely due to the differences in dataset completeness: as a primary megacity and pilot site for the GOB initiative, Nairobi benefits from more detailed and accurate coverage. Compared to other building footprint datasets, GOB reports the smallest average building area per structure, indicating greater sensitivity to detecting smaller structures which is particularly relevant for high-density consolidated settlements such as the ones in Nairobi²⁷. By contrast, previous work has proven that GHSL tends to underestimate highly built-up areas and overestimate peri-urban areas in secondary SSA cities²⁸.

The mapped results indicated that flood exposure is concentrated in specific areas, where both geography and urbanization play a critical role, i.e., highly exposed areas clustered close to or along riverine areas, confluence of rivers or delta regions, and often home to DUA populations (Fig. 3). In Nairobi, systemic spatial inequalities were spotted, where the highly dense and deprived settlements situated along the major tributaries of the city are the ones with higher exposure. In Kisumu, the eastern side is considerably more exposed than the western part of the city, with the largest ratio of high flood exposure for the districts of Manyatta B, Nyalenda and Kolwa Central, closely located to the Nyamasaria River. In Accra, high flood exposure is concentrated along the Odaw River, where many industrial hubs and DUAs are located. These areas were originally demarcated as buffer zones, but due to economic drivers, especially proximity to trade and unaffordability of land, many migrated there, and, the lack of enforcement of planning regulations resulted in the growth of the communities. In Tema, the Ashaiman district, close to the Ashaiman Dam, notably had the highest exposure levels. These areas, encroached on the farmlands around the dam, were originally planned for farming, unsuited for residential use and deprived from adequate infrastructure and services. In Beira, many DUAs are located in the surroundings of the Central Business District (CBD) and are estimated as highly exposed. The districts of Ndunda (Fig. 3) located close to the sewage station showed great exposure with GHSL estimates. The CBD, at the coast, has undergone redevelopment projects and showed minimal to no exposure. The same was seen in the central part of Chimoio, where flood exposure is largely riverine and spatially concentrated south of the CBD, most prominent in the districts of 25 de Setembro and Centro Hípico. These patterns reinforce how flood exposure is shaped both by geography and uneven urban development.

A closer inspection highlighted considerable differences between GHSL and GOB datasets (Fig. 3). Insights from local workshops - conducted with residents and city planners - played a critical role in identifying these mismatches. In Nairobi, for example, GHSL underestimated built-up areas along the tributaries (Kibera settlement) and adjacent to industrial areas (Mukuru settlement), leading to the omission of flood spots reported in the local workshops. In Accra, a similar issue emerged in Old Fadama, a dense informal settlement surrounded by riverways and a major road. While GOB classified the surrounding areas as low-built-up, GHSL smoothed them out as highly built-up. This difference matters: our analysis did not differentiate among types of elements at risk, such as roads, buildings, open areas. So, higher built-up density is directly interpreted as higher potential impact. Yet this assumption can mislead. With GOB, the roads and the river were classified as areas as moderately built-up, while GHSL classified them as highly built-up, potentially overstating risk where no structures exist. Given that damage to roads and open areas as less severe than damage to buildings (often life-threatening), the estimates from GOB seemed more aligned with ground realities. That said, the pattern of underestimation by GOB also downplays exposure in suburban or early-stage development zones and overestimates exposure in highly dense areas (see Fig. 3). Therefore, at first sight, the GHSL’s broader completeness and higher estimates seem to be the best alternative at the city scale, but at the community scale, the GOB can provide a spatially more accurate representation of differential exposure, especially in dense DUAs.

Given the stark disparities in exposure, we assessed whether the model accurately captured these patterns, particularly in the most exposed DUAs, where systematic assessments are absent, yet flood impacts are expected to be most severe^9,26. To do this, we integrated the flood model with two citizen science approaches: local workshops at city and community levels and participatory mapping surveys co-developed with and conducted by the residents of DUAs.

These participatory methods provided a range of findings. The workshops with city planners generally found that the FastFlood model²⁹ provided valuable estimates (in terms of flood extent and depth), despite limited input data availability and computation requirements. Participants agreed the model had an underestimation tendency in DUAs, indicating these areas as highly exposed, meanwhile suggesting that the model overestimated floods in formal areas, where improvement and investment in flood mitigation had taken place.

The participatory survey highlighted the underestimation of floods in DUAs. Results showed that 60% of ankle-level and 50% of the head-level floods were correctly predicted (Fig. 4). However, the model often simulated ankle-depth floods where residents reported knee or waist-level floods. The findings reflect the model's tendency to underestimate or even omit artificial flooding, caused by human-induced factors such as drainage issues, blocked waterways or overall poor infrastructure.

Comparing the data distribution across cities, we found that underestimation was also city dependent (Fig. 4). In Nairobi, simulated flood depths varied highly, with localized extremes and considerably lower medians, indicating systematic underestimation. Contributing factors might include: (1) the DEM limitations, e.g., vertical accuracy, even though proven to be better than ASTER, ALOS and NASADEM in SSA countries³⁰; (2) the steep topography of the city and the consequent increasing runoff estimates³¹, (3) and the deviations of the streams at the local level. The latter factor was largely brought up during the local workshops, since the model mistakenly captures points inside the streamlines (following the DEM), while these have been deviated on the ground. Statistical inspection showed that half of the simulated values for ankle-level floods reached only 1cm, confirming the structural underestimation by the model. Additionally, from the workshops we learned that the modelled extent focuses on the main tributaries, often neglecting secondary channels that also contribute significantly to natural drainage and flood impacts. This effect is likely related to the DEM resolution and the hydraulic assumptions of the model, prioritizing dominant flow paths while underrepresenting small or informal drainage structures that play a big role in flood dynamics in DUAs.

Kisumu and Beira followed the same underestimated orientation. In Kisumu, the model failed to capture backflow from the lake, while in Beira, although the overall agreement was high, the flood extent was still underestimated in the Munhava settlement which was subject to artificial flooding. Conversely, Accra and Chimoio presented overestimation patterns. Both had wide interquartile ranges similar to Nairobi, but with higher medians. In Accra, overestimation was mostly concentrated in the Old Fadama area that has gone through recent development interventions, not captured by the model. In Chimoio, the observed overestimation likely results from a combination of the DEM coarse resolution, local deviations of drainage pathways and field survey uncertainties inherent in complex dense urban environments which affect the validation through positional errors and reporting variability.

Tema had the highest underestimation rate, partially related to the artificial floods, frequently reported by the residents in the workshops, but also influenced by the limited number of surveyed points - the smallest among all cities (Fig. 4) Accra and Tema together had fewer than 50 surveyed points, whereas Kisumu alone had almost 100. As we co-developed the survey process with the residents, we asked them to map observed flood events representing the flood situation in their communities until they were satisfied. This process provided flexibility in terms of hours and points surveyed, but introduced variability in coverage. The level of commitment and support of the local leaders and the digital literacy of the participants certainly influenced the variability in the data.

A) Boxplots of simulated and observed flood depths across cities (capped at 1.5m for comparability). B) Confusion matrix of flood depth simulation (observed vs. simulated). C) Stacked bars of hit and miss evaluation across cities. D) Count of flood observation points across cities.

Up until here, the analysis shows that the model captures flood patterns well at city scale but significantly underestimates flood intensity in DUAs. This is problematic. However, further inconsistencies emerged when comparing the observed flood depths with the reported impacts. To explore this further, we conducted a thematic analysis of 310 community-reported flood points, identifying nine overarching impact categories which align with the sectoral effects from the Sendai Framework targets³² (see details on Methods).

The distribution of the impacts across the observed flood depth categories revealed that shallow floods – up to ankle height – are associated with substantial impacts. 33% of the reported impacts at ankle-level floods were linked to access disruption issues, while over 50% of the impacts reported property damage, disease outbreak and infrastructure damage (Fig. 5A). These findings challenge the prevailing assumptions that flood depths below 15 cm are non-hazardous, as if shallow flooding only impairs mobility of daily activities. In contrast, lived experiences in the DUAs indicate that shallow flooding can result in serious impacts, extending far beyond impaired mobility of daily activities²¹.

City-level patterns further reinforced these insights (Fig. 5B). Disease outbreaks, notably prominent in Beira, Kisumu and Accra, were concentrated in low-lying settlements near large water bodies such as the Indian Ocean and Lake Victoria. These locations are flood-prone and suffer from poor drainage systems, and consequently, stagnant water, increasing vulnerability to health-related impacts. The larger economically central cities – Nairobi, Tema and Accra – showed correspondence regarding high-severity impacts, such as property destruction and loss of life, suggesting an association between the scale of urbanization and the concentration of flood risks. Another pattern emerged from the secondary cities – Chimoio, Kisumu and Tema – with larger shares of impacts linked to environmental issues (soil erosion, environmental contamination). Workshop discussions revealed a regular practice of landlords in DUAs, where they used floodwaters to decongest the latrines, intentionally worsening water and soil pollution. Our field observations confirmed these reports, revealing how poor environmental and housing infrastructure can make even shallow floods highly disruptive and damaging (Fig. 5C). An additional and unexpected result was that, in Accra and Beira, displacement was not reported as an impact. From the workshops we learned that is partially true, as in the case of destruction of property, the affected residents opt to reside with friends and relatives close by, while they will rebuild their homes on the original spot after the rainy season.

Our study provides new evidence of flood exposure across and within six SSA cities, offering one of the first comparative analyses that systematically - with help of citizen science - centers DUAs in flood risk assessment. By examining cities of varying size and morphological and geographical context, our results show that DUAs are consistently and disproportionately more exposed to flooding, particularly at medium and high built-up densities. In Nairobi, the most extreme case, DUAs were found to be over 200% more exposed than non-deprived areas. Disparities were evident not only in primary cities like Nairobi and Accra, but also in secondary ones such as Kisumu and Beira, highlighting the systemic nature of these inequalities. Given that exposure rates in low-income urban areas highly exposed to floods are expected to keep rising¹⁰ and that many of the residents have no viable relocation options³³, these results underscore an even more pressing and complex situation.

To address technical constraints in flood exposure analysis, our methodology builds on open data and a lightweight flood model (FastFlood) that proved useful for producing city-wide flood exposure estimates and capturing intra-urban differences. Albeit not the first to map flood exposure in SSA cities using available remote sensing data^8,17, the use of FastFlood helped us tackle the complexity of traditional hydrological models, which often demands technical expertise and computational resources that are inaccessible to most local authorities. In many low- and middle-income countries (LMICs), hydrological modelling is constrained by the absence of flood risk or hydrology experts in local governments^2,16 and by the high costs of commercial tools (financially and computationally)⁸. These factors result in a limited number of comparative city-scale flood exposure assessments in SSA cities. FastFlood is computationally efficient due to the steady state assumptions while also user-friendly and low-cost, making it well-suited to low-resource environments, where flood risk assessments are urgently needed yet sparsely conducted.

In parallel, we addressed data limitations by analyzing two built-up datasets (GOB and GHSL), which revealed clear variations in flood exposure estimates. This comparison was highly motivated by the absence of flood exposure reference datasets across our study areas²⁴. The consistent overestimation by GHSL-especially in peri-urban and transitional zones- and the finer-grained representation of dense DUAs by GOB illustrate how exposure results are shaped by the structure and assumptions of the “elements-at-risk” datasets. Yet, this study only focused on the spatial distribution of built-up areas as a proxy for exposure. Future research could enhance the analysis by incorporating differentiation of building types to enable more granular insights within flood-prone zones and better targeted adaptation measures. Recent research on Volunteered Geographic Information (VGI) considering building functions from OpenStreetMap tags are alternatives to look for^34,35. In addition, given that data quality directly influences flood risk assessments and subsequent policy decisions, priority should be given to validation and triangulation of data sources to ensure more accurate results.

However, our modelling approach is not without limitations. While validated as effective at estimating macro-scale flooding patterns, the flood model systematically underestimates flood intensity and extent in DUAs, which has been seen in literature^36,37. The discrepancies stem, in part, from the simplified hydrological assumptions of the model, which cannot account for backflow from lakes (in Kisumu) or from coastal surges (in Accra). Discrepancies happen because the model assumes hydraulic connections between waterways are steady, such as, instant runoff, missing the lake and ocean response³⁸. Also, dealing with averaging terrain parameters at shallow flood depths likely concentrated the water in small, localized spots. However, a large portion of this underestimation stems from the inability of the model to capture artificial flooding factors since detailed information, such as waste disposal and blocked drainage locations, is scarce^39,40. This information is not easily represented in hydrodynamic flood models⁴¹, especially considering the informality of the systems in place. As a result, the same conditions that make DUAs highly flood-prone also undermine their representation in flood models, reflecting a wicked problem: the most exposed areas are also the hardest to model with technical tools alone²¹.

To address the methodological blind spots, we integrated citizen science (CS) methods to the flood modelling process to co-produce locally relevant spatial flood data and community-based insights^42,43. Through participatory mapping and workshops, residents of DUAs assessed the model outputs while challenging core assumptions in standard risk frameworks. One of the most significant insights was the mismatch between observed flood depth and experienced impacts: the dominant logic that greater depths equals greater impact is challenged. In fact, our thematic analysis reveals that even shallow flooding (below 15 cm) leads to severe disruptions, including disease outbreak, damage to properties and soil pollution. This contradicts traditional thresholds of water depth (one of the most commonly used parameters in flood assessments) and suggests a systematical exclusion of vulnerable DUAs. Considering broader planning implications, previous work indicated that most highly dense areas are affected by low frequency high magnitude flood events⁴⁰. Yet, our results show that in deprived contexts, floods in DUAs with low magnitude but high frequency also pose high risks, with compounding effects in public health, urban infrastructure and livelihoods.

The integration of CS in this study was not merely a methodological choice, but also an ethical and epistemological one. As we engaged with the communities, it became evident that the technological frameworks driving urban science - however promising - often rely on assumptions disconnected from lived realities. While AI, remote sensing, and open data are reshaping how we model and map cities, many residents in our study areas remain structurally excluded from these transformations². For many, participation in science - and politics - must be negotiated alongside daily struggles for food, housing, and safety⁴³. Certainly, the legacy of colonialism still reverberates⁴⁴, and in this context, urban science has been shaped by the values, needs and resources of the Global North³⁷, highlighting a deep asymmetry in how knowledge is produced, valued and mobilized globally. This can have serious consequences in decision-making by reinforcing existing inequalities in the flood risk assessments we attempt to carry out today. As technology continues to shape the urban research agenda, we must ensure it does not deepen existing divides⁴⁴. In this research, we tailored a people-centered approach to refrain from a dominantly technocratic model of assessment and we argue that participatory validation processes should not be treated as one-off exercises, but should rather be integrated into flood risk policies for more realistic models and more just outcomes^20,45.

Our methodology involves certain considerations. Firstly, in our framework, high exposure occurs from both high built-up density and a positive hazardous cell coinciding spatially. This removes the biases from index building⁴⁶ and provides a clear procedure, but it also limits the analysis to human exposure, overlooking exposed ecological systems. Secondly, as did not rely on spatially distributed global rainfall⁴⁷, instead using data from local gauge stations. While this choice offered better spatial accuracies comparatively, the completeness of the local datasets were limiting, where we may have missed short duration rainstorms that can significantly affect urban flooding⁴⁸. Thirdly, the overall quality (and resolution) of the globally available datasets used also added uncertainties. From the comparison between GOB and GHSL layers, our study indicates considerably higher flood exposure estimates using GHSL compared to GOB, which can be related to GHSL’s overestimation tendencies in peri-urban areas and to its modelling assumptions (e.g., including other built-up structures aside from buildings)²⁸. Unfortunately, these uncertainties are difficult to quantify in the absence of high frequency ground truth data⁴⁹. Additionally, the datasets with the extent of DUAs were from different sources using different methodologies (with and without local validation) and conceptual framings (slum-like morphologies to broad ranges of deprivation). Thus, the delineated boundaries are subject to omissions and inconsistencies, often influenced by the subjective confidence of the mappers. Finally, we did not apply return period-based scenarios, a decision that emerged from the fieldwork, where stakeholders found the concept abstract. Thus, our focus remained on the current flood status, in a more actionable entry point for communities and local authorities. Still, incorporating temporal dynamics and climate projections represent a valuable direction for future research, especially considering that extreme events cause irreparable effects for all¹⁰.

Over limited evidence, our findings reveal flood exposure patterns and how it is distributed and often underestimated in DUAs. By combining scalable models with participatory data collection, we offer an approach that is grounded in local realities while scalable across cities. This work supports global calls for context-sensitive and equity-driven climate adaptation and advances science-policy interface by showing that technical assessments and local knowledge can complement each other. Thus, we align our work under SDG targets on sustainable cities (SDG 11) and climate action (SDG13), insisting on just outcomes, but also methods and processes. In deprived areas facing recurring flood threats, the integration of local knowledge is not optional, but essential.

The study focuses on six cities across three SSA countries (Kenya, Ghana and Mozambique) that exhibit high human vulnerability to climate change yet face significant data scarcity¹². Three primary cities (Nairobi, Accra and Beira) and three secondary cities (Kisumu, Tema and Chimoio) were selected based on three criteria: the increasing threat of urban flooding, representation of diverse urban morphologies, and established collaboration with local partners (Fig. 6).

To contextualize the risk profiles of each city, we conducted a review of the most severe flood events reported over the past decade (see Supplementary Information). Sources included scientific publications, websites, and government records, which were cross-referenced with the Dartmouth Flood Observatory (DFO)⁵⁰ database for validation. This review highlighted the diversity of flood dynamics and the unique urban conditions across the case studies.

In Ghana, Accra and Tema share a coastal location and the same metropolitan governance but considerably differ in urban form and density⁵¹. In Kenya, the 4 million residents of Nairobi regularly experience intense flash floods, driven by rapid run-off in steep urbanized catchments. Conversely, Kisumu, a secondary mid-sized city on the shores of Lake Victoria, faces compounded risks from the lake backflow and riverine flooding, particularly during rainy season⁵². In Mozambique, Beira, located at the mouth of Pungwe river delta is often exposed to cyclones, while Chimoio, with half of the population of Beira, is located inland and surrounded by river valleys, illustrating the challenges of flood risk in secondary cities⁵³.

In this study, we used globally available datasets, open repositories and citizen-generated data. Table 1 provides an overview of each dataset, their respective properties and sources.

To achieve reliable flood hazard and exposure estimates this study consists of four main steps: (i) the development of a flood hazard model using FastFlood model and open data sources; (ii) the mapping and estimation of flood exposure using Google Open Buildings and Global Human Settlement Layer; (iii) the local validation with citizen science approaches; (iv) the thematic analysis of flood impacts reported by the residents from deprived areas.

Flood modelling was performed with FastFlood tool, a web application that uses steady-state assumption– and further mathematical corrections – providing high speed simulation, global data repository and minimal parameter tweaking²⁹. Catchment areas were delineated using a HydroSHEDS dataset, and aligned with GADM administrative boundaries to ensure complete spatial coverage. The catchment extents were chosen to balance hydrological accuracy and computation feasibility. Then, Stream Power Index (SPI) was computed to identify the areas most susceptible to water flow accumulation, thus more likely to be impacted by accumulated water⁵⁴. The identification of these critical points informed the selection of the deprived communities due to their hydrological exposure.

Rainfall analysis involved a series of tests for the pilot city. First, satellite-derived datasets (CHIRPS and GSMaP) were tested without downscaling techniques but results were found too coarse for city-scale application, leading to misconceptions of the precipitation distribution^6,55. Then we adopted TAHMO precipitation data from local gauge stations, which despite the shorter time series (2019 onwards), offered hourly resolution and reasonable coverage for all cities. The hourly precipitations were aggregated (daily and yearly) and analyzed to identify rainfall peaks and check their synchrony – suggesting the presence of large-scale weather systems⁷. We also prioritized weather stations based on data completeness and higher monthly averages. The latter, considering our interest in climate peaks or anomalies, can indicate higher sensitivity of the weather station to precipitation events⁵⁶. Finally, the rainfall peaks were cross-validated through media reports and insights from local partners.

Then, our flood modelling strategy started with a systematic sensitivity analysis in Nairobi, the pilot city. In this phase, over 15 model runs were conducted with pre-defined variations of key parameters. In the absence of high-quality validation data, we defined a set of criteria that guided the assessment of the model: (1) minimized visual noise, often associated with the DEM and channel settings³⁰; (2) higher maximum flood depths for capturing damaging events⁵⁶; (3) and flood extents spread beyond natural channels, thereby effectively representing pluvial flooding patterns¹. The analysis demonstrated that increasing model complexity resulted in significant escalation of computation time without evident improvements in output differentiation. It became harder to assess the differences between results, either because further parameter complexity led to equifinality or to model limitations to properly capture physical processes⁵⁷. Thus, to ensure reproducibility across the other study areas, particularly in SSA municipalities with limited technical resources, a simplified setup was adopted. Further methodological details are provided in the Supplementary Information.

Flood hazard maps were generated using a 50x50m grid. This grid was selected based on three main parameters from previous intra-urban studies in LMIC^58–60: the size of the input parameters, the computational time of the analysis and the depiction of settlements of organic (or non-geometrical) shapes. We used three flood depth thresholds: 1cm (due to the underestimation of FastFlood), 10 cm (natural break of the dataset and in accordance with international benchmarks⁶¹), 50cm (where life threatening impacts were reported by the residents of the deprived communities).

Based on the same grid, we estimated built-up density derived from two global datasets: the GOB layer (after computing building area sum per grid cell) and the GHSL layer. While GHSL offers a consistent census-satellite hybrid product, it has known limitations in capturing complex urban environments^28,62. GOB, in contrast, provides an AI-derived footprint with one of the highest counts of buildings in African cities²⁷, but a comparative analysis of the buildings to estimate exposure is yet to be implemented. As the built-up areas are the elements at risk of this analysis, we defined density thresholds to classify the built-up categories (low, medium and high built-up density) that would determine the level of exposure. The thresholds were defined by a statistically driven approach, based on the highest level of agreement between the two datasets and their data distribution values (see Supplementary Information).

Flood exposure was then calculated by overlaying the flood hazard maps and the elements at risk datasets. We computed both absolute and relative exposure across cities and evaluated the results by flood depth category and built-up density class. Lastly, we analyzed the differential exposure between deprived and non-deprived areas, also computing relative exposure within the cities.

Given the absence of conventional validation data, such as flood extent binary maps or observed discharge records, we integrated a citizen science-based approach, combining local workshops and participatory mapping surveys. In each city we organized two sets of workshops: one at the city scale, involving city council representatives, NGOs and academia, and another at the community scale engaging with residents of selected DUAs. Participants were recruited taking into account diversity of age, gender and livelihood. The number of visited communities was refined through fieldwork insights. We started with four communities in Nairobi, but subsequent iterations showed that two contrasting areas (in terms of morphology, stage of development, economic functions and location) provided sufficient insights. During the local workshops, the participants were divided into small groups (3–5 people)⁶³ to assess printed flood maps, by identifying flood-prone and safe spots and discrepancies between model outputs and the community experiences (Fig. 7). The group findings were then discussed in plenary, where they evaluated the level of agreement across the groups and consolidated insights.

At the community level, and refraining from the limited perspective of citizens as simply data collectors⁶⁴, we engaged with leaders and residents from DUAs in Nairobi through a participatory process structured in three phases: (1) we codesigned a data collection protocol. Through a series of online focus groups, participants selected Google My Maps as the preferred tool (that can be used on mobile phones with limited mapping experience), and we co-created the survey content based on their information needs, ensuring the results were actionable for the communities⁶⁵; (2) we trained the residents to map flood observation points (Fig. 7), where they documented flood intensity using locally defined categories: ankle (up to 15cm), knee (up to 50cm), waist (up to 100cm) and chest (above 100cm), along with descriptions of the impacts in each category; (3) we compared the collected observation points to the FastFlood outputs using hit-and-miss evaluation approach to assess the model performance. A video documentary about the fieldwork in the pilot city is available here.

To better understand the reported impacts of shallow flooding (up to ankle height) – particularly where the model underestimated the flood intensity – we conducted a thematic analysis of the reported impacts from the 338 flood observation points. After excluding incomplete or ambiguous responses (Supplementary Information), 310 entries were coded. We derived 56 distinct impacts, which were grouped inductively and thematically into nine overarching impact themes based on the Sendai Framework (Table 2). To handle ambiguity and deal with more than one reported impact per point, the insights from the local workshops were used to guide the classification. Based on this thematic analysis, we had a solid basis to quantify the frequency of impact patterns across the flood depth categories.