Guiding malaria elimination interventions: a data-driven approach to resource optimization in Benin, West Africa

Abstract

Background

Malaria remains a major public health challenge in Benin, where environmental conditions strongly influence its transmission dynamics. Understanding the spatial heterogeneity of malaria risk is essential for targeting interventions more effectively.

Methods

This study applied high-resolution environmental covariates and a robust Stochastic Partial Differential Equation (SPDE) model to investigate the distribution of malaria prevalence across Benin. Remote sensing-derived variables, including temperature and vegetation indices, as well as information about soil, built-up areas, and bare surfaces, were integrated into the model. Malaria prevalence was estimated using rapid diagnostic test (RDT) data from national Demographic and Health Surveys (DHS), and model performance was assessed through receiver operating characteristic (ROC) analysis.

Results

The findings reveal pronounced spatial dependence in malaria prevalence, with transmission patterns strongly associated with temperature and vegetation cover. Distinct hotspots were identified in the northern and central regions of the country, indicating areas of elevated risk. The model demonstrated satisfactory predictive accuracy (with a sensitivity of 0.619 and a specificity of 1.000), underscoring the utility of environmental covariates in capturing the spatial variability of malaria transmission.

Conclusion

Malaria in Benin exhibits marked spatial heterogeneity shaped by environmental factors. The identification of high-risk hotspots highlights priority areas for intensified intervention. Integrating spatial modelling with environmental data offers a powerful framework for refining malaria control strategies and accelerating progress towards elimination targets.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12936-026-06033-5.

Introduction

Malaria remains one of the most significant and persistent public health concerns facing sub-Saharan Africa, a region that continues to bear a disproportionate burden of the global malaria caseload and mortality [1]. Within this broader context, the West African nation of Benin is not immune to the devastating impacts of this vector-borne disease [2].

Malaria remains a major public health challenge in many countries, despite sustained national and international control efforts [1]. Health systems continue to be constrained by limited information on the true burden of malaria, the risk of Plasmodium transmission, and their geographical distribution [3]. According to the most recent data from the World Health Organisation (WHO), the estimated global number of malaria cases in 2023 was 263 million, corresponding to an incidence of 60.4 cases per 1000 population at risk [1]. The WHO African Region continues to bear the greatest share of the global malaria burden, accounting for 95% of estimated malaria deaths worldwide. In 2023, across 33 moderate- to high-transmission countries within the region, there were an estimated 36 million pregnancies, of which 12.4 million (34%) were affected by malaria. Compared with 2021, the number of malaria cases continued to rise, with an increase of approximately 14 million cases [4]. Children under 5 years of age remain the most vulnerable group, with the risk of mortality particularly concentrated among those under the age of three [5].

In Benin, despite the government's consistent efforts over the years, malaria remains a serious burden in terms of both morbidity and mortality, accounting for a staggering 95% of deaths [2]. In 2021, the reported annual figures are 3,163,648 malaria cases, with a slight increase of 1.4% between 2020 and 2021 and 2956 deaths [6]. Some provinces exhibit more than 50% of the prevalence of malaria in Benin. This is the case of Couffo, Atacora, Donga, and Borgou with cases of 51%, 50%, 47%, and 45% prevalence, respectively [7]. With these proportions, the rural areas have the highest prevalence of 43% and the urban areas come at the last position with 32%.

A range of interventions have been implemented in Benin to reduce the burden of malaria. These include seasonal malaria chemoprevention, the distribution of mosquito nets, both insecticide-treated nets (ITNs) and long-lasting insecticidal nets (LLINs), intermittent preventive treatment (IPT), indoor residual spraying (IRS), and, more recently, a pilot campaign introducing the RTS,S malaria vaccine. Identifying the spatio-temporal distribution of malaria prevalence across Benin is essential for the effective planning and deployment of interventions in areas where the need is most urgent. Such an approach would optimise the impact of control measures and ensure that high-risk areas receive focused attention. In 2023, seasonal malaria chemoprevention (SMC, targeting children aged 3–59 months) achieved a coverage rate of 68.4%, compared with 51.8% in 2021 and 52.1% in 2022. Moreover, the 2023 campaign for the distribution of LLINs reached 92.27% of households [8].

Climatic and environmental factors play a crucial role in shaping the transmission dynamics of malaria in Benin. Temperature strongly influences the development of mosquito larvae, thereby determining the survival and reproduction rates of Anopheles vectors [9, 10]. Vegetation conditions also contribute to malaria transmission, as areas with dense vegetation provide suitable resting sites and microclimates that favour mosquito survival [11]. Rainfall is another key determinant, creating breeding habitats conducive to mosquito reproduction and larval development [12].

Beyond these direct effects, climatic variability and extreme weather events, such as droughts and floods, can alter the ecology of mosquito populations and disrupt established transmission patterns. For instance, prolonged rainfall may increase vector abundance by expanding breeding sites, whereas unusually high temperatures may accelerate the parasite’s sporogonic cycle, thereby enhancing transmission intensity. Environmental factors, including land-use change, deforestation, and agricultural practices, further interact with climatic drivers, creating complex spatio-temporal patterns of malaria risk [13, 14].

Understanding the combined effects of these climatic and environmental determinants is therefore essential for designing more accurate predictive models of malaria transmission. In Benin, there is a notable lack of comprehensive studies examining the relationship between malaria prevalence and environmental factors. Integrating climatic and ecological factors into malaria elimination strategies could accelerate the control of mosquito ecology and address the seasonality of malaria transmission. Developing malaria prevalence risk maps by incorporating environmental conditions and forecasting seasonal patterns of transmission would provide valuable insights for policymakers, supporting the design of more targeted and effective interventions in high-risk areas.

Methods

Study area

Benin is a West African country located along the Gulf of Guinea, within the intertropical zone between the Equator and the Tropic of Cancer, specifically between latitudes 6°30′ and 12°30′ North, and longitudes 1° and 3°40′ East (Fig. 1). It covers a land area of 114,763 km², bordered to the south by the Atlantic Ocean, to the west by Togo, to the north by Burkina Faso and Niger, and to the east by Nigeria. The population of Benin was estimated at 13,712,828 inhabitants in 2013 [15].

Fig. 1.

Situation of Benin

Benin’s climate is generally hot and humid throughout the year, with mean annual temperatures of approximately 28 °C (historically ranging from about 27.5 °C along the southern coast to 28.5 °C in the northernmost departments of Alibori and Atacora during the period 1991–2020). Annual national precipitation averages around 1050 mm [16]. The southern region experiences two rainy seasons, from March to July and from September to November, separated by two shorter dry periods. In contrast, the northern tropical savanna zone has a single rainy season, extending from May to October, followed by a marked dry season from December to March.

For this study, malaria data were drawn from the Demographic and Health Surveys of Benin (DHSB). The DHSB sample is nationally representative, as well as representative at the level of the twelve departments, the city of Cotonou compared with other urban centres, and both urban and rural settings. The national territory was stratified into twelve study domains corresponding to the administrative departments. Within each domain (with the exception of Littoral, which contains no rural areas), two strata were defined: urban and rural.

Sampling for the DHSB was undertaken within each stratum using a two-stage stratified cluster design. In the first stage, 750 clusters and 555 clusters were selected for DHSB-IV (2011–2012) and DHSB-V (2017–2018), respectively, using systematic random sampling with probability proportional to cluster size. In the second stage, a complete listing of households was conducted within each selected cluster, from which a systematic random sample of 24 households per cluster (for DHSB-IV) and 26 households per cluster (for DHSB-V) was drawn, with equal probability in both urban and rural areas.

Public health data

Malaria-positive cases identified through Rapid Diagnostic Tests (RDTs) were extracted from the Demographic and Health Survey (DHS) databases for DHS-IV (2012) and DHS-V (2018). A total of 17,999 households were selected for DHS-IV, and 14,435 for DHS-V. All women aged 15–49 years who were usual residents of the sampled households and present the night before the survey were eligible for interview. In each household, children under 5 years of age were weighed and measured to assess their nutritional status (wasting, underweight, and stunting). In a subsample of every second household, all children aged 6–59 months were tested for anaemia and malaria (using RDT and thick blood smear), while all women aged 15–49 years were assessed for anthropometric indicators and anaemia.

In this study, malaria-positive cases among children aged 6–59 months, as determined by RDT, were obtained from DHS-IV (2012) and DHS-V (2018) (Table 1).

Table 1.

Malaria prevalence from RDT results (DHS 2012 and 2018)

Year	Total_Tested	Positive	Prevalence	CI_Lower	CI_Upper
2012	53,232	9427	17.71	17.39	18.03
2018	58,527	17,992	30.74	30.37	31.12

Covariate data

Yearly climatic and environmental data for each district of Benin were collected using Google Earth Engine from 2009 to 2024. Daytime and nighttime Land Surface Temperature (LST_Day, LST_Night) and Diurnal Temperature Difference (LST_DiurnalDiff) were obtained from the MODIS MOD11A1 (1 km) product. Tasseled Cap Brightness (TCB) and Wetness (TCW) at 500 m resolution were computed from the MODIS MOD09GA surface reflectance product (Red, NIR, Blue bands). Two-metre air temperature and relative humidity were retrieved from the ERA5-Land reanalysis dataset (ECMWF/C3S). Annual precipitation was obtained from the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) dataset at 5 km spatial resolution (5 km), and elevation was computed using SRTM elevation data (30 m). Vegetation was assessed via the Enhanced Vegetation Index (EVI) from MODIS Terra/Aqua (500 m), while annual Normalized Difference Water Index (NDWI; 500 m) was derived from MOD09A1 (Green and NIR bands). Annually evapotranspiration (ET; 500 m) was obtained from MODIS MOD16A2GF. All datasets cover 2009–2024 at the district level across Benin (Table 2). Data from 2012 and 2018 were used to model malaria prevalence, whereas data from 2024 were employed to predict future prevalence. All years were included to visualise the spatial distribution of the covariates (Fig. 3).

Table 2.

description of the covariate data

Covariates	Source/Product	Temporal Coverage	Spatial Resolution
Daytime Land Surface Temperature (LST_Day)	MODIS MOD11A1	2009–2024	1 km
Diurnal Difference Land Surface Temperature (LST_DiurnalDiff)	MODIS MOD11A1	2009–2024	1 km
Nighttime Land Surface Temperature (LST_Night)	MODIS MOD11A1	2009–2024	1 km
Tasseled Cap Brightness (TCB)	MODIS MOD09GA	2009–2024	500 m
Tasseled Cap Wetness (TCW)	MODIS MOD09GA	2009–2024	500 m
Air Temperature (2 m)	ERA5-Land (ECMWF/C3S)	2009–2024	9 km
Relative Humidity (RH)	ERA5-Land (ECMWF/C3S)	2009–2024	9 km
Precipitation	CHIRPS (UCSB-CHG/CHIRPS/DAILY)	2009–2024	5 km
Elevation	SRTM	2009–2024	30 m
Enhanced Vegetation Index (EVI)	MODIS Terra/Aqua (MYD09GA)	2009–2024	500 m
Normalized Difference Water Index (NDWI)	MODIS MOD09A1	2009–2024	500 m
Evapotranspiration (ET)	MODIS MOD16A2GF	2009–2024	500 m

Fig. 3.

Spatial distribution of the covariates used

Data processing

The study adopted a Bayesian geospatial modelling framework in which spatial dependence was represented using the Stochastic Partial Differential Equation (SPDE) approach of Lindgren et al. [17], which approximates a continuous Matérn Gaussian random field by a Gaussian Markov Random Field (GMRF) defined on a spatial mesh. Inference for the resulting latent Gaussian model was performed using the Integrated Nested Laplace Approximation (INLA), an efficient approximate Bayesian method for latent Gaussian models.

Multicollinearity within the regression model ($y=B_0+B_1{x}_1+B_2{x}_2+B_n{x}_n+ϵ$) was assessed using the Variance Inflation Factor (VIF), where $y$ denotes the dependent variable and $x_1,x_2,x_n$ represent the independent covariates [18]. In accordance with standard guidelines from the statistical literature, covariates exhibiting VIF values greater than 10 were excluded from the analysis [19]. A VIF value greater than 10 is widely considered to indicate severe multicollinearity that may inflate standard errors and destabilise coefficient estimates [20]. This approach reduced multicollinearity, thereby strengthening the stability of the estimated regression coefficients (β0, β1,…,βn) and enhancing the validity of statistical inferences. The term ϵ corresponds to the residual error of the model.

Global spatial autocorrelation of malaria rapid diagnostic test (RDT) positivity was first assessed using Moran’s I and Geary’s C statistics. Moran’s I quantifies overall similarity of RDT positivity between neighboring districts, while Geary’s C is more sensitive to local differences, providing complementary information on spatial clustering. Based on these assessments, Bayesian spatial models were developed to smooth observed data and quantify spatiotemporal variation in malaria risk [21, 22].

Model formulation

To model spatial dependence, the stochastic partial differential equation (SPDE) approach of [23] was used to represent the Gaussian random field on a triangular mesh constructed over the study area. The latent spatial field x(s) is defined as the solution to the following SPDE (Eq. 1):

$${\left(k^2-\mathrm{\Delta }\right)}^{\alpha /2}x\left(s\right)=W\left(s\right),s\in D$$ 1

where κ > 0 controls the spatial scale, Δ is the Laplacian operator, α determines the smoothness of the field, and W(s) denotes spatial Gaussian white noise.

The number of malaria-positive RDT results $y_i$ at location $\mathit{xi}$ is modeled using a binomial likelihood (Eq. 2) [24].

$$y_i=Binomial\left(n_{i,P_i}\right)$$ 2

where $n_i$ is the total number of individuals tested at location $x_i$ and p_i is the underlying probability of malaria positivity.The logit of p_i is modelled as a combination of fixed and random effects(Eq. 3)

$$log\text{i}t\left(P_i\right)=\alpha +X_i^T+S\left(x_i\right)+{\epsilon }_i$$ 3

where α is the intercept with a Gaussian prior, $X_i$ is the vector of standardized covariates at location $x_i$. $\alpha$ is the vector of regression coefficients with Gaussian priors (mean zero, precision 0.001), $S\left(x_i\right)$ is the spatial random effect modeled via the SPDE, and ${\epsilon }_i$ represents unstructured residual variation [24].

The model was fitted at the cluster (location) level, where the outcome $y_i$ represents the number of malaria-positive rapid diagnostic test (RDT) results among nin_ini individuals tested at location $x_i$. Thus, the likelihood is specified using aggregated binomial counts. For prediction and validation purposes, the estimated cluster-level probabilities $\underset{i}{\overset{\prime }{\rho }}$ were assigned to all individuals within the corresponding cluster, under the assumption of homogeneous risk within each location. This allowed comparison between observed individual-level binary outcomes and model-based predicted probabilities during the ROC analysis.

The spatial random effect S(x) is assumed to follow a zero-mean Gaussian process with Matérn covariance function (Eq. 4).

$$\text{cov}{S}_{\left(x_i\right),}{S}_{\left(x_i\right)}=\frac{{\beta }^2}{2^{\vartheta -1}\mathrm{\Gamma }\left(\vartheta \right)}{\left(k{x}_i-x_J\right)}^{\vartheta }{K}_{\vartheta }\left(k{x}_i-x_j\right)$$ 4

where ${\beta }^2$ is the marginal variance, Kν(⋅) is the modified Bessel function of the second kind of order ν > 0, and ∥xi − xj∥ is the Euclidean distance between locations x_i and x_j. The smoothness of the field is determined by ν, and the practical spatial correlation range is $\rho =\sqrt{8v}/k$, defined as the distance where correlation drops to approximately 0.1.

Gamma priors with parameters (1, 0.00005) were assigned to precision parameters associated with the spatial random effects and residual variation. Covariates were standardised to improve model convergence [25]

. Predictions of malaria RDT positivity were generated for unobserved locations by projecting the fitted SPDE onto a prediction mesh and back-transforming from the logit scale to proportion (%) values using the inverse logit function.

Model assessment

Models fit were evaluated using the Deviance Information Criterion (DIC) and Watanabe–Akaike Information Criterion (WAIC). The model with the lowest DIC and WAIC scores will be chosen for Bayesian analysis [17]. To quantify the uncertainty surrounding the posterior mean of model parameter estimates, 95% credible intervals has been calculated [24]. The best model has been used to develop the spatio-temporal malaria risk map for Benin.

To evaluate the discriminatory performance of the best fitted Bayesian spatial model, we compared the predicted probabilities of malaria infection against observed binary outcomes using a Receiver Operating Characteristic (ROC) analysis. The ROC analysis was performed post-estimation using posterior mean fitted probabilities from the INLA model and observed individual-level binary outcomes.

yi ∈ {0,1} denotes the observed outcome for individual i, where 1 indicates at least one positive rapid diagnostic test (RDT) result and 0 otherwise.

${\widehat{P}}_i$ denote the posterior mean fitted probability of infection for individual i obtained from the INLA model.

The ROC curve was constructed by plotting sensitivity against 1-specificity1 across all probability thresholds t ∈ [0,1] (Eq. 5 and Eq. 6):

$$Sensitivity\left(t\right)=\frac{TP\left(t\right)}{TP\left(t\right)+FN\left(t\right)}$$ 5

$$Specificity\left(t\right)=\frac{TN\left(t\right)}{TN\left(t\right)+FP\left(t\right)}$$ 6

with TP(t), TN(t), FP(t), and FN(t) denote the numbers of true positives, true negatives, false positives, and false negatives respectively, when classifying ${\widehat{P}}_i$ > t as positive.

Overall model discrimination was summarised by the Area under the ROC Curve (AUC), defined by Eq. 7:

$$AUC=\underset{0}{\overset{1}{\int }}Sensitivity\left(u\right)\text{d}\left(1-Specificity\left(u\right)\right)$$ 7

which can be interpreted as the probability that a randomly selected positive case receives a higher predicted probability than a randomly selected negative case. To determine the optimal classification threshold, we employed Youden’s J statistic (Eq. 8).

$$J\left(t\right)=Sensitivity\left(t\right)+Specificity\left(t\right)-1$$ 8

The value of t that maximises J(t) was chosen as the optimal cut-off, and the corresponding sensitivity, specificity, and threshold probability were reported. Predicted malaria prevalence among children under five was aggregated and mapped to visualise spatial heterogeneity providing a basis for identifying high-risk areas and informing targeted interventions. All statistical analyses were computed using the R software version 4.5.1 (2025–06–13 ucrt).

Results and discussion

Results

Malaria spatial autocorrelation

Global measures of spatial autocorrelation revealed significant clustering of malaria rapid diagnostic test (RDT) positivity in both 2012 and 2018. In 2012, Moran’s I was 0.86 (p < 0.001), far exceeding the expected value under spatial randomness (−0.0017). By 2018, spatial dependence remained statistically significant, although weaker in magnitude, with Moran’s I decreasing to 0.62 (Table 2). These results indicate strong positive spatial autocorrelation, whereby districts with high malaria positivity rates were located near other high-prevalence districts, while low-prevalence areas were similarly clustered (Fig. 2).

Fig. 2.

Malaria prevalence based on RDT results among children aged 6–59 months, DHS 2012 and DHS 2018

Consistently, Geary’s C statistic was 0.14 (expected = 1.0, p < 0.001) in 2012 and 0.28 (expected = 1.0, p < 0.001) in 2018, providing further evidence of pronounced local similarity between neighbouring districts (Table 3).

Table 3.

Moran’s I and Geary’s C statistics for malaria RDT positivity

Year	Morans' I	Moran’s I p-value	Geary’s C	Geary’s C p-value
2012	0.86	3.10E-273	0.14	2.21E-215
2018	0.62	1.14E-103	0.28	1.41E-64

The spatial analysis thus highlights a clear pattern of non-random malaria distribution, with strong clustering in 2012 that weakened, yet persisted, in 2018. This temporal shift may reflect changes in malaria transmission dynamics, potentially linked to environmental variations (Fig. 3). The significant Moran’s I and Geary’s C statistics confirm that malaria positivity exhibits spatial dependence, thereby justifying the application of advanced geostatistical models such as the Stochastic Partial Differential Equation (SPDE) approach in INLA to better quantify spatial patterns, account for residual autocorrelation, and produce high-resolution risk maps.

Variable selection

Twelve covariates including Daytime Land Surface Temperature (LST_Day), Diurnal Difference Land Surface Temperature (LST_DiurnalDiff), Nighttime Land Surface Temperature (LST_Night), Tasseled Cap Brightness (TCB), Tasseled Cap Wetness (TCW), Air Temperature (2 m), Relative Humidity (RH), Precipitation, Elevation, Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI), Evapotranspiration (ET) aggregate at district and cluster level were used in this study (Fig. 3).

The analysis of the collinearity test results, summarised in Table 4, indicates that some covariates are highly correlated. Covariates with a Variance Inflation Factor (VIF) greater than 10 are highly correlated and problematic. All covariates with VIF > 10 were removed from the model but we decided to keep Air temperature (VIF = 10.59) and Enhanced Vegetation Index (VIF = 15.96) due to their importance in malaria transmission [10, 26, 27]. In total, five covariates were included in our models.

Table 4.

Multicollinearity assessment using VIF

Variables	VIF
Precipitation	1.49
Elevation	19.47
Evapotranspiration	29.07
EVI	15.96
LST_Day	10.97
LST_Night	5.98
NDWI	16.97
Relative_Humidity	47.44
TCB	9.84
TCW	10.23
Air Temperature	10.59
Diurnal Difference Land Surface Temperature	30.83

A trade-off was considered between reducing multicollinearity and retaining epidemiologically important predictors, prioritising predictive performance and interpretability over strict variable exclusion.

Model selection

Table 5 presents the results of the model selection process. A total of 32 Bayesian Hierarchical Spatiotemporal (BHST) models were constructed, comprising 31 models that included covariates and one model without covariates (non-spatial model) (see supplement materials).

Table 5.

Selection of the best model

Models	DIC	WAIC
Model 1: Intercept-only	6272.488	6763.018
Model 31: Spatial model with covariates (EVI, LST_Ngh, TCB and Temp)	5754.703	5811.397

Model fitting was first conducted using a coarser mesh to improve computational efficiency during the exploratory phase. Model selection was based on DIC and WAIC, and the best-supported model was identified as the model including EVI, LST_Ngh, TCB, and temperature (Model 31) (Supplementary materials: Table S1). This model was then refitted using a finer mesh to improve spatial resolution and confirm robustness (Table 5).

Analysis of the results indicates that the spatial model (Model 31), with DIC = 5754.703 and WAIC = 5811.397, outperforms the non-spatial model (Model 1), which has DIC = 6272.488 and WAIC = 6763.018.

These findings highlight the importance of integrating environmental and climatic covariates in predicting malaria prevalence. Model 31’s superior performance demonstrates that the inclusion of relevant covariates substantially improves the predictive accuracy of the response variable. This model could be applied for malaria risk prediction following further analysis and validation.

Validation of the best model

The predictive performance of the model was assessed using the Receiver Operating Characteristic (ROC) curve analysis, from which the optimal probability threshold was derived. The final Bayesian spatial model demonstrated good discriminatory performance with an AUC of 0.812 (95% CI: 0.748–0.876), indicating a strong ability to distinguish malaria-positive from malaria-negative cases.

At the threshold of 0.232, the model achieved a sensitivity of 0.619 and a specificity of 1.000 (Fig. 4). This indicates that the model correctly identified approximately 62% of true malaria cases while perfectly classifying non-cases. The high specificity demonstrates the model’s strong ability to avoid false positives, ensuring that areas predicted as low risk are indeed unlikely to experience high malaria prevalence.

Fig. 4.

Receiver Operating Characteristic (ROC) curve for the selected spatial model (Model 2)

Estimated effects of climatic and environmental covariates on malaria

Figure 5 presents the posterior estimates from model 31, showing the relationship between several covariates and the outcome of interest. The intercept (− 0.282; 95% CI: − 0.410, − 0.141) represents the baseline log-risk of malaria when all covariates are at zero, indicating a low underlying risk in the reference condition. In practice, this baseline provides a reference point that underscores the importance of including the covariates in the model. Without accounting for factors such as EVI, LST_Ngh, TCB, and Temp, the model would rely solely on the intercept, which is negative and may not accurately reflect the variations in malaria transmission across different environmental conditions.

Fig. 5.

Effect of Environmental and Climatic Covariates on Malaria Risk

Enhanced Vegetation Index (EVI) shows a positive association with malaria risk (0.164; 95% CI: 0.075, 0.254). This suggests that areas with higher vegetation cover have increased malaria prevalence, likely due to more favourable habitats for mosquito vectors.

Night-time land surface temperature (LST_Ngh) also exhibits a significant positive association (0.157; 95% CI: 0.013, 0.304), indicating that warmer night temperatures increase malaria risk. Tasseled Cap Brightness (TCB) has a non-significant effect (0.020; 95% CI: − 0.119, 0.163), indicating an uncertain influence on malaria risk in this model.

Air temperature at 2 m (Temp) shows the strongest positive effect (0.331; 95% CI: 0.265, 0.398), highlighting temperature as a key environmental driver of malaria risk.

These results reveal that temperature and vegetation are the primary environmental factors influencing malaria risk in Benin, with night-time temperatures contributing additionally. Tasseled Cap Brightness does not appear to have a significant effect. It captures information about soil, built-up areas, and bare surfaces. Monitoring these climatic conditions, particularly air temperature, is crucial for accurately predicting malaria risk and guiding targeted public health interventions.

Predicted malaria risk in Benin

Figure 6 presents the predicted malaria risk map for Benin in 2024. Analysis of this figure indicates that the districts at greatest risk are predominantly located in the northern and central regions of the country. In the Alibori department, districts such as Karimama, Malanville, Segbana, Kalalé, Kandi, Gogounou, Banikoara, and Kérou exhibit particularly high risk, with predicted prevalence exceeding 70%. Similarly, certain districts in Atacora, including Materi and Cobli, show elevated risk levels.

Fig. 6.

Predicted malaria risk maps for 2024 in Benin

In the central region, districts such as Ouèssè, Bantè, Glazoué, Savè, Savalou, Dassa-Zoumé, Djidja, Kétou, Covè, Zagnanado, Zogbodomey, Abomey, Bohicon, Agbangnizoun, Ouinhi, and Adja-Ouèrè are identified as high-risk areas. Although Abomey and Bohicon exhibit substantial risk, the intensity appears somewhat lower than in the northern districts. Within the central region, particular attention should be given to Covè, Zagnanado, Zogbodomey, and Ouinhi, where predicted prevalence among children aged 6–59 months exceeds 80%, similar to Karimama, Malanville, and Segbana in Alibori.

Targeted interventions should prioritise these high-risk areas and include enhanced vector control measures, distribution of insecticide-treated bed nets, malaria vaccination for children aged 6–59 months, seasonal malaria chemoprevention, and community engagement initiatives to effectively reduce malaria transmission among children under five.

Discussion

This study examined the spatial distribution and environmental determinants of malaria risk in Benin using Bayesian hierarchical spatiotemporal modelling. The findings demonstrate strong spatial clustering of malaria rapid diagnostic test (RDT) positivity in both 2012 and 2018, although the degree of clustering weakened over time. High-risk districts were concentrated in the northern and central regions, particularly in Alibori, Atacora, and parts of the Collines and Zou departments. The observed spatial dependence, supported by both Moran’s I and Geary’s C statistics, underscores the non-random nature of malaria distribution and the necessity of incorporating spatial structure into epidemiological analyses.

The decline in spatial autocorrelation between 2012 and 2018 may reflect shifts in malaria transmission dynamics associated with changes in environmental conditions, population movement, and the scale-up of control interventions. Previous studies across sub-Saharan Africa have similarly reported that malaria distribution exhibits spatial heterogeneity and that the strength of clustering can vary over time in response to vector control measures and climatic variability [28, 29]. In Benin, the expansion of insecticide-treated net (ITN) coverage and indoor residual spraying campaigns during this period may have contributed to reducing malaria prevalence in some districts, thereby weakening spatial dependence.

The modelling results highlight the pivotal role of climatic and environmental covariates in shaping malaria risk. Air temperature emerged as the strongest predictor, with higher temperatures significantly increasing the probability of malaria positivity. This can be supported by the fact that temperature influences mosquito survival, biting rates, and parasite development within the vector [30, 31]. Similarly, the positive association observed with the Enhanced Vegetation Index (EVI) reveals that vegetation provides suitable breeding habitats for Anopheles mosquitoes, thereby facilitating transmission. Night-time land surface temperature was also positively associated with malaria risk, corroborating studies revealing that warmer nocturnal conditions prolong mosquito activity and enhance parasite development [32]. In contrast, Tasseled Cap Brightness exhibited no significant effect, which may be explained by its mixed representation of soil, bare surfaces, and built-up areas that are less directly linked to mosquito ecology.

The superior performance of the spatial model compared with the non-spatial alternative, as reflected in the DIC and WAIC values, underscores the value of incorporating spatial dependence and relevant covariates when predicting malaria risk [33]. The model achieved high specificity, ensuring reliable identification of low-risk areas, although sensitivity was moderate. This suggests that while the model is effective in ruling out areas unlikely to experience high transmission, further refinements may be necessary to enhance case detection in high-risk zones.

The predicted malaria risk map for 2024 identifies persistent transmission hotspots in northern and central Benin. Districts such as Ouèssè, Bantè, Glazoué, Savè, Savalou, Dassa-Zoumé, Djidja, Kétou, Covè, Zagnanado, Zogbodomey, Abomey, Bohicon, Agbangnizoun, Ouinhi, and Adja-Ouèrè in the central region, exhibit prevalence estimates exceeding 70–80% among children aged 6–59 months. In the Alibori department, the districts of Karimama, Malanville, Segbana, Kalalé, Kandi, Gogounou, Banikoara, and Kérou exhibit particularly high risk, with predicted prevalence exceeding 70%. Similarly, certain districts in Atacora, including Materi and Cobli, show elevated risk levels.

These findings align with national malaria reports, which indicate that the northern and central regions consistently bear the heaviest burden of disease [8]. They also echo findings from neighbouring countries where similar ecological conditions have been associated with high transmission intensity [27, 34].

This risk map for pinpoints specific districts in Alibori, Borgou and central Benin where over 80% of children are predicted to be at risk. And hence targeted interventions should prioritise the identified high-risk districts, with a particular focus on children under five who remain the most vulnerable group. Integrated approaches combining ITN distribution, seasonal malaria chemoprevention, vaccination, and enhanced community mobilisation are likely to be most effective. Furthermore, given the influence of climatic and environmental drivers, malaria control programmes should incorporate climate-informed early warning systems to anticipate seasonal and interannual variations in transmission risk.

Conclusion

In light of the findings above, the study highlighted that malaria in Benin exhibits strong spatial dependence shaped by environmental conditions, particularly temperature and vegetation. Alibori, Attacora and the Zou are the provinces with the highest malaria risk. These identified hotspots are priority areas for intensified intervention. Integrating spatial modelling with climate and environmental data offers a powerful approach for refining malaria control strategies and achieving greater progress towards elimination targets.

Limitation

The study has several strengths, including the use of high-resolution environmental covariates, robust spatial statistical modelling, and validation through ROC analysis. Nevertheless, certain limitations should be acknowledged. First, the reliance on RDT data may have led to an under- or overestimation of the true prevalence, as test sensitivity varies with parasite density. Second, although the model incorporated key environmental variables, socio-economic and behavioural determinants that also shape malaria risk could not be included owing to data constraints. A key limitation is the potential misalignment between the malaria data (point prevalence from DHS) and the environmental data (annual aggregates). The absence of intervention coverage variables (Insecticide-Treated Nets (ITNs), Indoor Residual Spraying (IRS), Seasonal Malaria Chemoprevention (SMC), and intermittent preventive treatment in pregnancy (IPTp)) may partly explain the observed spatial and temporal variation in PfPR. Finally, a key limitation concerns the potential misalignment between malaria data (point prevalence from DHS) and environmental data (annual aggregates), as seasonal peaks in transmission during the rainy season may be obscured by annual averaging.

Author contributions

Gouvidé Jean GBAGUIDI and Nikita TOPANOU designed the study. Gouvidé Jean GBAGUIDI collected the data, analysed, interpreted the results and wrote the manuscript. Nikita TOPANOU, Anges YADOULETON and Rock AIKPON verified the public health and the covariates data collected and validated the statistical methods applied. Nikita TOPANOU, Guillaume K. KETOH and Gabriel Hoinsoudé SEGNIAGBETO reviewed the manuscript. All authors had full access to all the data in the study, had the final responsibility for the decision to submit for publication, and approved the final version of this Article to be published.

Funding

This work didn’t receive any external funding.

Data availability

Data are available at the University of Abomey, Benin. You should contact the corresponding author for the data request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the Ministry of Health of Benin. All methods were carried out in accordance with relevant guidelines and regulations and all experimental protocols were approved by the Ethics Committee. Informed Consent was obtained from all subjects and/or their legal guardian(s). The data on malaria prevalence are anonymised and were extracted from the Demographic and Health Survey databases. The corresponding author had obtained authorisation to collect this data. Consequently, we were granted permission to use the data, but no ethics approval number was provided.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.