Satellite-derived longitudinal evidence of flood effects on community water and sanitation resilience in coastal Kenya

Abstract

Background

Flooding can disrupt water, sanitation and hygiene services in low-income rural and peri-urban settings, yet there is limited quantitative evidence on how repeated flood exposure affects safe water and sanitation over time at the level of Community Health Units (CHUs), which are the operational units for primary health care in Kenya.

Methods

We analysed repeated CHU-level data from ten Community Health Units in Kilifi County, Kenya, across eleven surveillance rounds from 2017 to 2024. For each round, we calculated the percentage of households with safe drinking water and with access to a functional latrine, and linked these outcomes to flood exposure derived from Sentinel-1 radar imagery at 10 m resolution during the relevant CHU-specific survey window. We fitted linear mixed-effects models with CHU random intercepts, adjusting for elevation and land cover composition, and constructed comparative flood-adjusted resilience scores.

Results

We analysed 220 CHU-round observations. The flood-by-outcome interaction was negative and statistically significant: a one standard deviation increase in flood frequency was associated with a 5.6% point greater decline in functional latrine coverage than in safe water coverage (flood-by-outcome interaction coefficient = -5.63; 95% CI: -9.33 to -1.93; p = 0.003). The rural versus peri-urban contrast also differed by outcome. Rural CHUs had substantially lower safe water coverage than peri-urban CHUs, whereas the rural versus peri-urban difference in latrine coverage was much smaller. Comparative resilience scores showed a clear gradient, with peri-urban CHUs generally ranking highest and several rural CHUs clustering at the lower end of the distribution, particularly for safe water resilience.

Conclusions

Repeated flooding was associated with greater disruption of functional latrine access than safe water access, and rural disadvantage was concentrated mainly in safe water coverage. Although the resilience scores are comparative CHU-level planning tools rather than absolute measures of system resilience, they can help identify low-performing units and guide targeted investments in protected water systems and flood-resistant sanitation in coastal Kenya.

Supplementary Information

The online version contains supplementary material available at 10.1186/s44263-026-00291-w.

Background

Climate change is intensifying the frequency and severity of extreme weather events, including floods, with major consequences for water, sanitation and hygiene (WASH) systems [1, 2]. Flooding can contaminate surface and groundwater sources, damage sanitation infrastructure, and disrupt basic services, thereby increasing the risk of diarrhoeal disease and other infections [3–6]. These impacts are particularly acute in low resource settings, where households often depend on shallow wells, surface water and pit latrines that are highly exposed to inundation and erosion [7, 8]. In many low and middle income countries, rapid population growth and unplanned urbanization have produced densely settled peri urban and informal settlements that are especially vulnerable to climate related shocks [9].

In coastal Kenya, recurrent seasonal floods interact with fragile WASH infrastructure and chronic poverty [10, 11]. Flood waters can collapse latrines, contaminate shared water points and impede access to health facilities, while rural communities face persistent underinvestment in safe water supply and sanitation services [12]. These differences in baseline access and vulnerability are likely to shape how peri urban and rural communities experience and recover from repeated flood events [13]. Kenya’s primary health care system is organized around community health units (CHUs), which are the frontline structures for health service delivery and community engagement [14, 15]. CHUs bring together community health volunteers, households, and link facilities within defined catchment areas, making them a natural decision making scale for planning and monitoring WASH services [16].

Routine CHU reporting on household characteristics and WASH indicators, when combined with satellite derived information on flooding and environmental conditions, offers a promising platform for monitoring resilience and guiding local action. The Kaloleni Rabai Demographic Health Surveillance System (KRHDSS) in Kilifi County provides one such platform through repeated household surveillance that captures community level WASH indicators alongside demographic and health outcomes [17]. Despite growing recognition of climate related risks to WASH, three important gaps limit the use of existing data for resilience planning. First, most studies rely on cross sectional surveys or short time series, which restrict the ability to examine how floods and WASH outcomes co-vary over time at the community level [18]. Second, water and sanitation are often treated as a single composite construct rather than distinct outcomes, even though flooding may damage pit latrines, contaminate sanitation infrastructure and affect safe water access through different physical pathways [19]. Third, few analyses explicitly compare peri urban and rural community settings while jointly accounting for flood exposure, land cover composition and elevation, despite the likelihood that these factors influence both baseline access and resilience.

Therefore, the current study addresses three key gaps that constrain the use of evidence for resilience planning. Specifically, the primary objective of this study was to estimate the association between satellite derived flood frequency and access to safe water and functional latrines over time at the CHU level in Kilifi County. We hypothesized that flooding is negatively associated with latrine access than with safe water access, reflecting the distinct physical mechanisms by which flooding affects functional pit latrine access and safe water access. Secondary objectives included testing whether these associations differ between peri urban and rural CHUs and quantifying the independent effects of land cover composition and elevation after adjusting for flood exposure. We hypothesized that rural CHUs exhibit lower access to safe water and lower resilience to flooding compared to peri urban CHUs. To accomplish these objectives, we leveraged KRHDSS data collected over 11 surveillance cycles [17]. The KRHDSS provides detailed community-level WASH data that were integrated with contemporaneous satellite-derived datasets (see Methods for details). The longitudinal design, with outcomes and exposures aligned to the same decision relevant spatial units, is intended to produce estimates and choropleth maps that directly inform the prioritization of resilient infrastructure and service protection in flood prone coastal Kenya. Longitudinal analysis, which aligns WASH outcomes with flood occurrence, provides scientifically robust estimates and contextually relevant insights to inform targeted water and sanitation interventions, including the design of low income flood prone communities.

In this study, resilience is defined pragmatically as relative CHU-level WASH performance under observed flood stress after accounting for environmental context. The resulting resilience score is therefore a comparative and operational prioritisation metric intended to support planning at CHU level. It is not intended to represent a validated absolute measure of broader system resilience, nor does it fully capture absorptive, adaptive, or transformative capacity as described in broader resilience frameworks.

Methods

We report this observational study in line with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations for observational studies [20]. A completed STROBE checklist is provided as Supplementary Material 1.

Study area

The analysis was conducted within the KRHDSS platform in Kilifi County, coastal Kenya, which has been described in detail elsewhere [17]. The KRHDSS covers Kaloleni and Rabai sub-counties (3°38′-3°59′S, 39°21′-39°39′E), predominantly rural areas characterized by high poverty, subsistence farming, informal trade, and recurrent flooding during the long and short rainy seasons [11, 21, 22]. Within the KRHDSS, ten CHUs were included in this study. These are administratively defined service units under Kenya’s community health strategy, each typically covering about five hundred to one thousand households [14, 15]. The ten CHUs are Buni, Kibwabwani, Kombeni, Kwale, Mnazimwenga, Mutsengo, Mwele Kisurutini, Tsangatsini, Viragoni and Vishakani. Based on local administrative classification and settlement density, three were categorized as peri urban and seven as rural (Fig. 1).

Fig. 1.

Spatial classification of Community Health Units in Kaloleni and Rabai, Kilifi County, Kenya. Map of the ten CHUs in the KRHDSS. Peri urban CHUs are shown in blue and rural CHUs in green

Kilifi County experiences a bimodal rainfall pattern, with long rains from March to May and short rains from October to December. Annual rainfall is approximately one thousand millimeters, with peak rainfall in April and May often leading to seasonal flooding of low lying areas [11, 21]. Despite ongoing investments, household access to basic water and sanitation in Kilifi County remains below national targets, with less than two thirds of adults reporting access to basic potable drinking water and about half reporting access to basic sanitation in recent national surveys [22, 23]. This context of high climatic exposure, limited WASH coverage and administratively meaningful CHU boundaries provides a natural setting for examining flood related WASH resilience.

Study design

We used a longitudinal observational design [24] that leverages repeated CHU level measurements from the KRHDSS between 2017 and 2024. The analytic unit was the CHU round combination, that is, a specific CHU observed in a specific surveillance round. During the study period, KRHDSS conducted repeated household surveys across the ten CHUs. Early rounds (2017 to 2019) were conducted semi-annually and later rounds (2020 to date) annually, yielding eleven surveillance rounds with WASH data. For each CHU and round, we calculated CHU level coverage for safe water access and functional latrine access, then linked these outcomes to satellite derived flood exposure and environmental covariates for the corresponding period. All ten CHUs in the Kaloleni Rabai surveillance catchment that contributed complete data between 2017 and 2024 were included, so no formal sample size calculation was undertaken.

This design was selected to serve three purposes. First, it allowed us to describe trajectories in WASH coverage over several years rather than a single cross-sectional snapshot. Second, it enabled estimation of how changes in flood frequency over time were associated with changes in WASH coverage within the same CHUs. Third, it provided a framework to compare peri urban and rural CHUs after adjusting for land cover composition, elevation and other covariates, separating the effect of flooding from underlying geographic and environmental context.

Data collection and measures

All ten CHUs located within the KRHDSS catchment and contributing at least one round of complete data on the main study variables were included. In each surveillance round, trained field workers visited all households within every CHU and administered structured questionnaires. The responses were aggregated to produce CHU level percentages for analysis. Two key indicators of WASH access were assessed, using standard drinking water and sanitation classifications from the WHO/UNICEF Joint Monitoring Programme as the reference framework [25]. The first was safe water access, defined as the proportion of households whose main drinking water source was improved or protected, such as piped water, protected wells and boreholes, protected springs, bottled water, or harvested rainwater. We note, however, that in flood affected settings bottled water may at times represent a temporary coping response rather than stable household water security. The second was functional latrine access, measured as the proportion of households reporting access to any working latrine on the compound, whether improved or unimproved, in line with how the surveillance tools captured latrine functionality. We treated safe water access and functional latrine access as two separate outcomes to allow for potentially distinct pathways linking flooding to water supply and sanitation infrastructure.

Flood exposure was derived from Sentinel 1 Ground Range Detected synthetic aperture radar imagery processed in Google Earth Engine [26, 27]. Synthetic aperture radar is useful for flood mapping because it can observe the ground surface through cloud cover. We used Vertical Transmit, Vertical Receive (VV) polarization only, meaning radar signals transmitted and received in the vertical orientation, and restricted the analysis to scenes acquired in Interferometric Wide swath mode. Backscatter refers to the strength of the radar signal returned from the ground surface, and dB refers to decibels, a logarithmic unit used to express radar backscatter intensity. Exposure was summarised at 10 m spatial resolution. For each study year, we first constructed an annual dry season reference image by calculating the median VV backscatter from all January to March Sentinel 1 scenes intersecting the study area. Flooding was then identified within each survey period using a relative thresholding approach, in which a pixel was classified as flooded when its filtered VV backscatter was more than 2.5 dB below the corresponding annual dry season baseline.

To improve classification stability, we applied a focal median filter to each Sentinel 1 scene before thresholding to reduce speckle noise. We also excluded areas likely to generate persistent false positives. Permanent surface water was masked using the Joint Research Centre Global Surface Water dataset, with pixels showing at least 90% long term water occurrence excluded. Terrain with slopes of ten degrees or greater was masked using the Shuttle Radar Topography Mission Global 1 Arc Second digital elevation model to reduce misclassification related to radar shadow and layover [28].

For each CHU specific survey period, we calculated a per pixel flood frequency surface as the proportion of valid Sentinel 1 scenes within that period that were classified as flooded. CHU level flood exposure was then defined as the mean flood frequency across all pixels within the CHU boundary. This metric therefore reflects the average proportion of valid observations during the survey window in which a CHU was inundated, rather than a simple binary flooded versus not flooded classification. Survey windows followed the surveillance schedule for each CHU and round and generally covered approximately one month immediately preceding data collection, although one surveillance window spanned 16 October to 15 November. The number of valid scenes contributing to each CHU round estimate was also recorded to support data quality checks.

Candidate contextual covariates included mean elevation and land cover composition variables aggregated to CHU level. Land cover classes were obtained from the ESA WorldCover product at 10 m resolution [29] and summarised as the percentage of CHU area occupied by water, rangeland, trees, crops, and bare ground. Mean elevation for each CHU was computed from the SRTM GL1 data. We also linked a time invariant Relative Wealth Index (RWI) derived from satellite based proxies of socioeconomic status at the community level [30]. Continuous predictors were standardised prior to modelling. To reduce dimensionality among correlated contextual variables in this very small longitudinal panel, we used mutual information as a pragmatic nonlinear screening tool. Unlike simple pairwise correlation thresholds, mutual information can capture nonlinear dependence between candidate covariates and the outcome. We retained covariates with mutual information values greater than 0.01, a pragmatic minimum-information-contribution rule rather than a formal significance threshold. Using this rule, the final contextual adjustment set included elevation, water, rangeland, trees, crops, and bare ground. Survey round was evaluated during screening but was not retained because its mutual information score was 0.0000 and did not meet the pre-specified inclusion rule. Rainfall and temperature were not included as modelled covariates; rainfall is described only as part of the study setting. The retained environmental variables were treated primarily as contextual adjustment covariates rather than primary exposures or effect modifiers. Although the small number of candidate covariates (six) meant that screening had limited practical impact, the procedure was applied as a pre-specified step to guard against including noise covariates.

Statistical analysis

Descriptive analysis

We first described temporal and spatial patterns in CHU level safe water access, functional latrine access, flood frequency and environmental covariates. We computed summary statistics overall, by CHU and by peri urban or rural setting. We produced maps to visualize CHU level patterns of flood exposure, WASH coverage and derived resilience scores.

Mixed effects modeling

For inferential analyses, we restructured the dataset into long format so that each row represented a combination of CHU and survey round for one of the two outcomes. The dependent variable was the CHU-level percentage of households with either safe water access or functional latrine access, expressed on a 0 to 100 scale. Two effect-coded indicators were introduced: outcome type (-0.5 for safe water access and + 0.5 for functional latrine access) and CHU setting (-0.5 for peri-urban and + 0.5 for rural CHUs). This coding ensures the intercept reflects the overall mean, main effects capture average differences, and interactions quantify how flood impacts and rural versus peri-urban gaps vary between water and latrine access. To facilitate interpretation and improve model convergence, all continuous predictors were centred and scaled. Flood frequency, mean elevation, and the land cover percentage variables were standardised to z-scores (mean zero, unit variance).

We estimated associations between flood exposure and WASH coverage using a linear mixed-effects model with CHU-specific random intercepts to account for repeated observations within CHUs [31–33]. The dependent variable was CHU-level coverage, expressed as the percentage of households with safe water access or access to a functional latrine in CHU_i at survey round_j. We retained a linear mixed-effects specification because the outcomes were aggregated CHU-level percentage measures and the primary objective was to estimate directly interpretable percentage-point differences associated with flood exposure while accounting for within-CHU clustering over time. Fixed effects included standardised flood frequency, effect-coded indicators for outcome type and CHU setting, their interaction terms, and the selected contextual covariates. Random slopes for flood exposure were evaluated but did not improve fit, so the final model retained random intercepts only. The final model converged successfully. The model can be expressed as:

    

where denotes the CHU-level random intercept and the residual error term. Under this specification, represents the average association between flood frequency and WASH coverage across both outcomes, while captures differential flood effects between safe water and latrine access. The term reflects how the contrast between outcomes varies by CHU setting. Random slopes for flood frequency were evaluated but did not improve model fit, so the final model retained random intercepts only [34].    

Model adequacy was assessed using residual-versus-fitted plots, normal Q-Q plots, scale-location plots, inspection of random intercept distributions, the intraclass correlation coefficient (ICC), and likelihood-based comparison of alternative random-effects structures [31, 32, 34]. The final model showed substantial clustering at CHU level, with an approximate ICC of 0.705. Because the number of CHUs and repeated rounds was limited, we did not fit a more complex serial correlation structure, and we acknowledge this as a limitation. Bounded-outcome sensitivity analyses using logit-transformed and beta regression specifications were conducted separately and are reported in Supplementary Material 2: Table S5. Because outcomes were expressed on the original percentage scale and continuous predictors were standardised, coefficients can be interpreted as the absolute change in WASH coverage (percentage points) associated with a one standard deviation increase in the predictor.

Construction of resilience scores

Consistent with the study’s pragmatic framing, the resilience scores were constructed as comparative CHU-level summaries of expected WASH performance under observed flood exposure and environmental context, and are intended to support relative ranking and prioritisation rather than provide an absolute or externally validated measure of system resilience. Resilience scores were calculated as CHU-level summaries using mean safe water access, latrine access, flood frequency, setting, and elevation across eleven rounds. Flood-adjusted baseline coverage estimates from a mixed-effects model were normalised (-1.00 to 1.00), following a regression-based framework that defines resilience as performance better than expected under adverse exposure [35]. For each CHU_i, we predicted coverage for each outcome (water or latrine) at its observed mean flood frequency (F_i, expressed as the percentage of CHU area flooded) and at one percentage point higher flood frequency, holding other covariates at their sample means. The difference between these predictions approximated the marginal effect of flooding:

    

representing the CHU-specific sensitivity of coverage to flood exposure, analogous to metrics used in climate resilience studies [19].

We then computed flood-adjusted baseline coverage by removing the expected loss due to flooding:

where is the observed mean coverage for outcome. These adjusted values estimate performance under zero flood exposure, like residual-based resilience scores in other fields [35].

To enable comparison across CHUs, we normalized adjusted coverage to the range using a symmetric min–max transformation:

where and are taken over all CHUs. This mirrors normalization procedures widely used in composite resilience indices [18]. We defined an overall resilience score as the simple average of water and latrine resilience:

consistent with composite WASH resilience indices developed for sanitation infrastructure in African settings [12].

All analyses were conducted in Python (version 3.10) using statsmodels for mixed effects modelling and scikit learn for mutual information screening [36, 37]. Two sided p values and 95% confidence intervals were reported for fixed effects. The Google Earth Engine code used to derive flood exposure, the Python scripts used for modelling, resilience scoring and sensitivity analyses, and the aggregated CHU level analytic dataset are available in the project repository [38].

Results

The analytic dataset comprised 220 CHU–round observations from 10 CHUs in Kilifi County, spanning eleven survey rounds between 2017 and 2024. CHUs were classified into three peri urban units (Buni, Mwele Kisurutini and Vishakani) and seven rural units (Kibwabwani, Kombeni, Kwale, Mnazimwenga, Mutsengo, Tsangatsini and Viragoni) (Fig. 1). Each CHU contributed CHU level data on two outcomes in each round, the percentage of households with access to safe water and the percentage of households owning a functional latrine. Descriptive statistics for each CHU, including mean WASH coverage, flood frequency, land cover composition, elevation and wealth, are presented in Supplementary Material 2: Table S1.

Flood exposure and environmental context also varied sharply across CHUs. Across CHU–round observations, flood frequency ranged from about 0.2% to 30.5% of CHU area, with peak exposure concentrated in rural Tsangatsini (up to 30.5%) and Mnazimwenga (up to 29.4%) (Fig. 2). Land cover patterns followed a clear urban rural gradient. Peri urban CHUs had a higher proportion of built-up areas, with mean values between about 24.1 and 47.6%, whereas most rural CHUs had built up areas below 20% and were dominated by rangeland, which exceeded 85% of the area in Kibwabwani, Mnazimwenga and Tsangatsini. Elevation ranged from about 118 m above sea level in Buni to about 280 m above sea level in Kibwabwani, reflecting modest but non-trivial topographic variation within the catchment.

Fig. 2.

Flood frequency across Community Health Units, 2017 to 2024. Percentage of CHU area classified as flooded at least once in each year based on Sentinel 1 radar, summarized across the study period

There were marked spatial and socioeconomic disparities in WASH coverage between peri urban and rural CHUs. Across all CHU–round observations, safe water access ranged from 0.3% in rural units such as Viragoni and Tsangatsini, to 99.9% in the peri urban unit Mwele Kisurutini. Peri-urban CHUs consistently maintained high safe water coverage, with CHU–round values between about 45.9 and 99.9% and a mean of about 86.4%, whereas rural CHUs had much lower and more variable coverage, with a mean of about 24.3% and most CHU–round values below 20%. Within the rural group, Kombeni and Kwale were exceptions, with mean safe water coverage of about 87.8 and 52.7% respectively, while Kibwabwani, Mnazimwenga, Mutsengo, Tsangatsini and Viragoni all had mean coverage below about 8% (Supplementary Material 2: Table S1). In contrast, functional latrine ownership was high across settings, ranging from 20.3% in Tsangatsini to 98.3% in Kwale. Peri-urban CHUs had a mean latrine coverage of about 86.9% and rural CHUs a mean of about 63.6%, and several rural CHUs, including Kibwabwani, Kombeni and Kwale, attained latrine coverage comparable to or exceeding peri urban CHUs. These patterns suggest that latrine infrastructure has diffused more widely than safe water infrastructure (Figs. 3 and 4, Supplementary Material 2: Table S2). Peri-urban CHUs had positive wealth scores, with a median around 0.43, whereas rural CHUs had negative wealth scores, with a median around − 0.31. Vishakani and Mwele Kisurutini were the wealthiest CHUs (wealth index about 0.46 and 0.43 respectively) and Mutsengo the poorest (about − 0.48). (Supplementary Material 2: Table S1).

Fig. 3.

Safe water access across Community Health Units, 2017 to 2024. CHU level percentage of households with safe drinking water for each survey round

Fig. 4.

Functional latrine access across Community Health Units, 2017 to 2024. CHU level percentage of households with access to a functional latrine for each survey round

Based on 220 CHU-round observations, the model estimated a total variance of 666.7, with 470.3 attributed to between-CHU differences and 196.4 to residual variation, corresponding to an ICC of approximately 0.705. This indicates that approximately 70.5% of the total variance in WASH coverage was attributable to persistent differences between CHUs rather than temporal change within CHUs over time. This high ICC implies that the data were dominated by structural heterogeneity across CHUs, such as differences in infrastructure, settlement form, and environmental context, and that the flood associations are estimated in a setting characterised by strong between-CHU differences. Full fixed and random effect estimates and model fit diagnostics for the final mixed-effects model are presented in Supplementary Material 2: Tables S3 and S4. Specifically, flood exposure was more strongly associated with functional latrine access than with safe water access. In the reference configuration (peri-urban CHUs at mean land cover, elevation and survey round), there was no clear association between flood frequency and average WASH coverage (coefficient = 0.86; 95% CI: -1.13 to 2.85; p = 0.40). However, the strong negative flood-by-outcome interaction (coefficient = -5.63; 95% CI: -9.33 to -1.93; p = 0.003) indicates that a one standard deviation increase in flood frequency was associated with a 5.6%-point greater decline in functional latrine coverage compared with safe water coverage. These associations should be interpreted as reflecting short-term or acute flood-related disruption within the survey windows, rather than cumulative or lagged effects.

The main effect of rural setting on average WASH coverage was large in magnitude but imprecisely estimated, with confidence intervals that included zero (coefficient = -32.18; 95% CI: -73.06 to 8.70; p = 0.12). This indicates that there was no uniform overall rural deficit across both outcomes combined. Instead, the strong interaction between outcome type and rural setting (coefficient = 38.90; 95% CI: 30.82 to 46.99; p < 0.001) shows that the rural disadvantage differed markedly by outcome and was concentrated primarily in safe water coverage. Evaluated at mean covariate values, the model implied that safe water coverage was about 52% points lower in rural CHUs than in peri-urban CHUs, whereas latrine coverage was only about 13% points lower. This pattern is consistent with the descriptive findings that rural disadvantage is driven primarily by deficits in safe water access, while many rural CHUs have latrine coverage that approaches peri-urban levels. Among environmental covariates, there was weak evidence that CHUs with a larger share of surface water had lower WASH coverage (coefficient = -4.16; 95% CI: -8.58 to 0.26; p = 0.065) and that bare ground showed a similar pattern (coefficient = -3.89; 95% CI: -8.24 to 0.45; p = 0.079). Elevation also showed a suggestive negative association (coefficient = -15.94; 95% CI: -33.43 to 1.56; p = 0.074), but the confidence intervals for all three included zero. Rangeland, trees, and crop cover were not materially associated with WASH coverage in this model. These environmental associations should therefore be interpreted as suggestive rather than conclusive.

Resilience scores summarise each CHU’s average performance across eleven rounds, adjusted for flood exposure and environmental context. Normalised scores (-1.00 to 1.00), based on flood-adjusted coverage estimates, reveal a strong but nuanced gradient (Figs. 5, 6 and 7). Safe water resilience showed a clear divide. The highest safe water resilience scores were observed in Mwele Kisurutini (1.000, peri-urban), Buni (0.921, peri-urban) and Kombeni (0.786, rural), the latter being a notable exception among rural CHUs. By contrast, rural CHUs such as Tsangatsini (-1.000), Mnazimwenga (-0.972) and Viragoni (-0.928) occupied the lowest ranks. Latrine resilience scores were high and less sharply separated by setting. All CHUs achieved moderate to high latrine resilience, with Kwale scoring highest (1.000) and Tsangatsini lowest (-1.000), and several rural CHUs performing comparably to peri-urban units. Overall resilience, calculated as the average of safe water and latrine resilience for each CHU, confirmed a persistent rural disadvantage (Fig. 7). The three highest scoring CHUs were Mwele Kisurutini (overall resilience 0.956, peri-urban), Kombeni (0.758, rural) and Buni (0.744, peri-urban), while the five lowest scoring CHUs were all rural (Tsangatsini − 1.000, Mnazimwenga − 0.843, Viragoni − 0.640, Mutsengo − 0.363 and Kibwabwani − 0.166). This pattern highlights that even after accounting for flood exposure and environmental context, most rural CHUs remain substantially less resilient than peri-urban CHUs, although Kombeni demonstrates that strong safe water resilience is achievable in a rural setting. CHU-level resilience rankings are summarised in Table 1.

Fig. 5.

Safe water resilience index by Community Health Unit. Normalized safe water resilience scores from − 1.00 to 1.00, where higher values indicate greater resilience to flood related disruption of safe water access

Fig. 6.

Latrine resilience index by Community Health Unit. Normalized latrine resilience scores from − 1.00 to 1.00, where higher values indicate greater resilience to flood related disruption of functional latrine access

Fig. 7.

Overall WASH resilience index by Community Health Unit. Overall resilience score for each CHU, calculated as the mean of safe water and latrine resilience indices. Peri urban CHUs cluster at high resilience values and rural CHUs cluster at low resilience values

Table 1.

Resilience rankings of community health units to flooding in Kilifi County, Kenya

Rank	CHU	Setting	Safe water resilience	Latrine resilience	Overall resilience
1	Mwele-Kisurutini	Peri-Urban	1.00	0.91	0.95
2	Kombeni	Rural	0.79	0.72	0.76
3	Buni	Peri-Urban	0.92	0.57	0.74
4	Vishakani	Peri-Urban	0.42	0.91	0.67
5	Kwale	Rural	0.06	1.00	0.53
6	Kibwabwani	Rural	-0.94	0.60	-0.17
7	Mutsengo	Rural	-0.89	0.17	-0.36
8	Viragoni	Rural	-0.94	-0.34	-0.64
9	Mnazimwenga	Rural	-0.98	-0.71	-0.84
10	Tsangatsini	Rural	-1.00	-1.00	-1.00

Scores range from − 1.00 to 1.00 for safe water, latrine and overall resilience, where higher values indicate greater resilience to flood related service disruption. Overall resilience is the average of safe water and latrine resilience scores

Discussion

This longitudinal analysis of ten CHUs in coastal Kenya showed that flooding destabilises sanitation more than water supply and that the rural disadvantage in WASH resilience is concentrated in safe water coverage rather than distributed uniformly across both outcomes. Three main findings stand out. First, increases in flood frequency were associated with larger declines in functional latrine access than in safe water access, supporting the primary hypothesis that sanitation systems are more fragile under flood conditions. Second, the rural versus peri-urban contrast differed sharply by outcome: rural CHUs had chronically low safe water coverage across the study period, but many maintained latrine coverage comparable to peri-urban units. Third, there was suggestive evidence that land cover characteristics, particularly a higher proportion of surface water and bare ground, are linked to reduced WASH coverage, indicating that environmental context may compound flood-related vulnerability even though confidence intervals for these associations include zero.

The stronger association between flooding and declines in functional latrine access than in safe water access is consistent with how different technologies respond to inundation [7, 8]. Basic pit latrines and other onsite sanitation systems are highly sensitive to rising groundwater, erosion and physical damage of superstructures, which can lead to pit collapse, overflow or abandonment during and after floods [12]. In contrast, protected water sources such as boreholes, piped networks and storage tanks can be elevated, lined or otherwise engineered to withstand short term flooding when they have received adequate investment [19]. In this study, a one standard deviation increase in flood frequency was associated with an estimated 5.6% point greater decline in functional latrine coverage than in safe water coverage, quantifying the differential sensitivity of sanitation and water systems to flooding. This pattern is consistent with national and county level evidence showing that rural residents in Kenya report lower access to basic drinking water and sanitation than urban residents and rely more heavily on onsite sanitation and informal water sources [22, 39]. Together these findings support treating water and sanitation as distinct systems in climate resilience planning because they are exposed to different physical failure modes when flooding occurs.

The resilience scores revealed a pronounced and policy relevant gradient across CHUs. Peri urban CHUs, together with one relatively better off rural CHU, exhibited high safe water resilience and high overall resilience, combining high baseline coverage with modest declines in coverage when flood exposure increased. In contrast, several rural CHUs had very low safe water resilience and low overall resilience. In these units the combination of chronically low safe water coverage and greater sensitivity of sanitation to flooding produced the lowest resilience scores. These rural CHUs are characterized by limited built up area, higher poverty and more exposure to floodplains and erosion prone landscapes, which makes it harder to develop and maintain protected water infrastructure and flood robust sanitation where investment has historically been limited.

Sanitation resilience was more nuanced. Many rural CHUs achieved high levels of functional latrine coverage and scored well on latrine resilience, sometimes matching or surpassing peri urban units. This reflects the cumulative impact of national sanitation efforts and local programmes such as community led total sanitation, which emphasized latrine construction and use through strong community mobilization [40, 41]. These efforts appear to have produced real gains in reported basic sanitation even in the absence of equivalent gains in safe water coverage. However, the analysis shows that these gains are fragile in the face of flooding [42]. Floods undermine latrine functionality by saturating pits, washing away slabs and superstructures and contaminating the immediate environment [12]. Rural CHUs that appear successful when judged by static latrine coverage alone are therefore still vulnerable once chronic safe water deficits and flood induced sanitation failures are considered. This points to a need for a two-pronged strategy in rural settings, with one arm focused on closing persistent safe water gaps and the other on upgrading sanitation technologies and siting practices to designs that are more resistant to inundation and structural damage.

The CHU level resilience scores and maps (Figs. 5, 6 and 7) provide an operational tool for decision makers. They summarize how each CHU performs relative to others given its flood exposure and environmental context and they do so at a scale that aligns with Kenya’s community health system [14]. In Kilifi County the resilience maps highlight clusters where chronic safe water deficits intersect with high flood frequency and fragile sanitation infrastructure. County managers can use these outputs to target investments in resilient water systems such as deep protected boreholes, elevated treatment and storage facilities and robust distribution networks, and to introduce sanitation options better suited to flood prone environments, including raised or sealed pits, toilets with lined vaults and improved drainage around facilities [43].

The high ICC is substantively important. Approximately 70.5% of the total variance in WASH coverage was attributable to persistent differences between CHUs rather than temporal change within CHUs. This indicates that the data were dominated by structural heterogeneity across CHUs, such as differences in infrastructure, settlement form, and environmental context. Accordingly, the estimated flood associations should be interpreted as longitudinal associations observed within a setting characterised by strong between-CHU heterogeneity, rather than as sharply identified within-CHU causal effects. The small number of CHUs further limits statistical precision, constrains the complexity of random-effects and temporal correlation structures that can be estimated reliably, and reduces power to detect heterogeneity across units.

It is also worth noting that if bottled water was used as a temporary coping strategy during or after flood events, the differential vulnerability of safe water relative to sanitation could be underestimated. Some of the apparent resilience of safe water coverage may reflect unsustainable substitution rather than durable access to protected sources. Future studies could examine whether bottled water use increases in the immediate aftermath of flooding and whether this transient coping masks underlying vulnerability in piped or borehole-based water systems.

The observed associations between land cover, flood exposure and sanitation performance further support the need for integrated planning, although they should be interpreted with caution. CHUs with a higher share of surface water and bare ground tended to sit in landscapes that are more flood exposed, erosion prone and difficult to service. In such settings latrines are harder to construct on stable foundations and more likely to fail during heavy rains, and surface water and bare ground influence flood pathways and settlement choices. These associations are best viewed as indicative signals of compounded vulnerability rather than definitive causal effects, but they are consistent with the view that land use, water resource management and WASH programming should be coordinated [44]. Measures that reduce runoff and erosion, stabilize soils and avoid siting critical facilities in the most exposed micro terrains could complement investments in infrastructure and behaviour change.

This study has several strengths. The longitudinal design follows the same CHUs across eleven surveillance rounds, allowing separation of persistent differences between CHUs from within CHU changes that coincide with variation in flood exposure. This improves internal validity compared with a single cross-sectional survey, although causal relations cannot be established definitively in an observational study. The use of high-resolution satellite derived flood frequency as the exposure metric avoids reliance on self-reported flood histories and reduces recall and reporting bias [26]. Integrating these Earth observation data with routine CHU level surveillance and an explicit resilience index provides a practical way to rank communities by WASH resilience while accounting for both hazard and context. Using the CHU as the unit of analysis also links the evidence directly to the structures responsible for community health and WASH service delivery in Kenya [14], which increases the likelihood that county governments and implementing partners can use the findings to prioritize investments and monitor progress towards climate resilient water and sanitation services.

Several limitations should be considered. First, the analysis was conducted at CHU level, and conclusions therefore pertain to CHU level service performance rather than individual household outcomes. Aggregation to CHU level aligns with how community health services are organised and how decisions are typically made, but it can mask important within-CHU heterogeneity. The study is therefore susceptible to ecological fallacy and to the Modifiable Areal Unit Problem (MAUP) [45], since exposure, infrastructure vulnerability, and coping capacity may vary substantially within CHUs. Second, the number of CHUs was small. Although the longitudinal design provides repeated observations over time, the small number of clusters limits statistical precision, constrains the complexity of random-effects and temporal correlation structures that can be estimated reliably, and reduces power to detect heterogeneity across units. Third, the flood exposure metric relied on a relative thresholding approach applied to Sentinel 1 VV backscatter [26]. Although this method improves temporal comparability by anchoring classification to an annual dry-season baseline, the resulting flood estimates remain sensitive to the choice of threshold, the quality and representativeness of the baseline period, residual radar noise, and the number of available scenes within each survey window. We reduced false positives by applying speckle filtering and by masking permanent water and steep slopes, but some misclassification of inundation is still possible. In addition, CHU level exposure was defined as the mean flood frequency across pixels within each CHU, which may obscure finer spatial heterogeneity in flood extent within individual units. Fourth, the outcomes were bounded percentages, and Gaussian mixed-effects models can in principle generate fitted values outside the feasible 0 to 100 range, particularly near the boundaries. We therefore interpret the estimated coefficients as approximate marginal percentage-point differences at CHU level; bounded outcome sensitivity analyses are reported in Supplementary Material 2: Table S5 and confirm the direction and significance of the key interaction. Fifth, safe water access was defined using standard improved-source classifications and included bottled water. In flood affected settings, however, bottled water may also reflect a temporary coping strategy rather than durable household water security, which could overstate sustained resilience in some contexts. Sixth, the sanitation outcome captures reported access to any functioning latrine on the compound rather than comprehensive safe sanitation, which would also consider safe excreta management, environmental contamination, and hygiene conditions. Finally, although the models adjusted for contextual environmental covariates, unmeasured time varying factors, such as local infrastructure projects, behaviour change campaigns, facility management changes, and broader economic shifts, may have influenced WASH coverage and resilience.

Conclusions

Flooding poses a differential threat to water and sanitation resilience in coastal Kenya. It disproportionately affects latrine functionality during and after flood events, while safe water coverage, though less acutely sensitive to flooding, remains chronically low in rural CHUs. The rural disadvantage in overall WASH resilience is driven primarily by persistent deficits in safe water coverage rather than by a uniform rural deficit across both outcomes. The CHU-level resilience scores and maps provide a practical, comparative tool for identifying the most vulnerable units and can guide targeted investments in protected water systems and flood-resistant sanitation. Priority should be given to rural CHUs with the lowest safe water resilience, where closing safe water gaps and upgrading sanitation technologies and siting practices to designs that withstand inundation would deliver the largest gains. Future studies should examine cumulative and lagged flood effects, incorporate household-level data to disaggregate within-CHU heterogeneity, and evaluate whether targeted investments improve resilience trajectories over time.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

WASH

Water, Sanitation and Hygiene

CHU

Community Health Unit

KRHDSS

Kaloleni Rabai Health and Demographic Surveillance System

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

SRTM GL1

Shuttle Radar Topography Mission Global 1 Arc Second

DEM

Digital Elevation Model

RWI

Relative Wealth Index

VV

Vertical Transmit, Vertical Receive

ICC

Intraclass Correlation Coefficient

ODF

Open-Defecation Free

AKU

Aga Khan University

Author contributions

AKN and FO conceived the study, FO developed the methodology, curated and managed the data, performed the formal analysis, and drafted the original manuscript. RIO, FG, AKN, and RO contributed to the study design, assisted with interpretation of the results, and provided critical revisions of the manuscript. JD, ZB, and AKN contributed to the conceptualization of the study, offered methodological and policy guidance, and critically revised the manuscript for important intellectual content. All authors read and approved of the final manuscript.

Funding

This study was supported by the operations of the KRDHSS through a hybrid model consisting of internal Aga Khan University (AKU) core funding and external sources from nested studies. The surveillance system was initially set up and supported through funding from Global Affairs Canada (GAC) between 2017 and 2020 (Grant Number D-001983). The funding bodies had no role in the study design, data collection, analysis, interpretation, decision to publish, or preparation of the manuscript.

Data availability

The external geospatial datasets used in this study are publicly available from the original providers as cited in the manuscript. The aggregated CHU level analytic dataset that supports the findings of this study, together with the Google Earth Engine code used to derive flood exposure and the Python code used for modelling, resilience scoring and sensitivity analyses, are available in a public GitHub repository archived in Zenodo [38].

Declarations

Ethics approval and consent to participate

Ethical approval for the Kaloleni Rabai Demographic Health Surveillance System and for secondary analyses of de-identified surveillance data was obtained from the Aga Khan University Institutional Ethics Review Committee (Reference 2013/IREC 65, versions v4 to v13). The present analysis uses only aggregated CHU level data without individual identifiers. Permission to use the de-identified surveillance dataset for this study was granted by the KRHDSS governance committee. Written informed consent was obtained from participating households at enrolment into the surveillance system and individual consent was not required for this secondary analysis of aggregated data. The study conformed to the ethical principles of the Declaration of Helsinki and applicable institutional and national requirements.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.