Predicting environmental suitability and future spread range of An. stephensi in the Greater Horn of Africa using remote sensing and ensemble modeling

Abstract

Malaria, a life-threatening disease, remains a major global health challenge, particularly in Africa. While Anopheles gambiae sensu lato has long been the primary vector in Africa, the recent invasion of Anopheles stephensi—an urban malaria vector native to South Asia, poses a growing threat to malaria control and elimination efforts. Understanding An. stephensi environmental suitability and its spread dynamics is critical for designing effective surveillance and vector control strategies. Although previous studies have mapped potential environmental suitability for An. stephensi, most have focused on temperature or environmental variables, overlooking other critical factors affecting mosquito life cycles. Moreover, little is known about the species’ historical spread speed or projected expansion. While An. stephensi is already spreading in the region, this study aims to enhance predictive modeling of suitable habitats and identify areas at ongoing or future risk of invasion. Our approach integrates meteorological, environmental, geophysical, and socioeconomic variables, alongside an expanded occurrence dataset and ecologically constrained pseudo-absence sampling. The model achieved an accuracy of 0.93 in predicting An. stephensi locations during the 2021–2024 test period, outperforming previous studies in the region. We analyzed historical spread patterns, revealing a rapid increase in spread speed from 20 km/year in 2012 to over 120 km/year by 2024. Future spread was projected using environmental suitability, road connectivity, and population density, with the spread model achieving a temporal correlation of 0.66. Projections suggest continued expansion into western Ethiopia, southern Somalia, and southern Kenya, with climate change likely to increase environmental suitability in highland regions. This high-resolution, spatiotemporal framework provides actionable insights for current and future transmission hotspots and supports urgent, targeted interventions to mitigate the spread of An. stephensi under a changing climate.

1. Introduction

Malaria, a life-threatening disease transmitted by female Anopheles mosquitoes, remains a major global health challenge, particularly in Africa, with an estimated 263 million clinical cases reported in 2023—an increase of 14 million from the previous year (WHO, 2024b). The African continent suffers the highest malaria burden and accounts for 94 % of cases and 95 % of deaths in the world (WHO, 2024b). While Anopheles gambiae sensu lato is the primary malaria vector in Africa (Geissbühler et al., 2007; Pinto et al., 2013), Anopheles stephensi—an important malaria vector species native to south Asia, has rapidly expanded its range into the Horn of Africa, including Djibouti (Faulde et al., 2014), Ethiopia (Carter et al., 2018), Sudan (Abubakr et al., 2022), Somalia (Ali et al., 2022), Kenya **(Ochomo et al., 2023), Ghana(Afrane et al., 2023) and Eritrea (WHO, 2023b), and Nigeria in the past decade (Adeogun et al., 2023). The invasion of An. stephensi in Africa has been implicated in malaria outbreaks in Djibouti (Faulde et al., 2014) and Ethiopia (Balkew et al., 2020), and has posed serious threats to the malaria control and elimination efforts in the continent.

Knowledge of the environmental suitability of An. stephensi in Africa is valuable for planning malaria vector surveillance and designing preventive measures to limit its spread. An. stephensi prefers breeding in various artificial containers, which often have year-long water availability, allowing the vector population to survive during the dry season and extending malaria transmission. An. stephensi exhibits outdoor biting behavior, enabling it to evade the conventional malaria vector control methods like insecticide-treated bed nets and indoor residual spraying, which typically target indoor environments. Furthermore, An. stephensi shows high levels of resistance to four main classes of synthetic insecticides (Balkew et al., 2021). These ecological and behavioral characteristics make An. stephensi a challenging vector species to control. Therefore, it is crucial to prevent the spread and control the invasion of An. stephensi in the early stage while its density is low (Abubakr et al., 2022), and accurate An. stephensi environmental suitability mapping is highly desired.

Several studies have modeled An. stephensi environmental suitability in Africa. For example, using the published occurrence records of An. stephensi across Asia, Arabian Peninsula, and Horn of Africa from 1985 to 2019, Sinka et al. (2020) developed an An. stephensi environmental suitability map based on climate, vegetation index, and human density for Africa. The study estimated that if An. stephensi continues to spread across the continent, an additional 126 million people could be at risk of malaria. Villena et al. (2022) analyzed the impact of temperature on malaria transmission by comparing the invasive An. stephensi and native An. gambiae mosquitoes. This was done using mechanistic trait-based models, which accounted for the nonlinear responses of temperature-dependent mosquito and parasite characteristics. The study indicated that An. stephensi exhibits a broader thermal tolerance range compared to An. gambiae, suggesting that a greater portion of Africa could be more suitable for malaria transmission by An. stephensi than by An. gambiae. Liu et al. (2024) modeled the future distribution of An. stephensi under future climate change scenario, and found global climate change increases the areas suitable for An. stephensi globally. Ahn et al. (2023) assessed the risk of An. stephensi invaded Africa using the marine cargo traffic data and ranked the African countries at greater risk for An. stephensi invasion based on maritime connectivity. Although previous research has predicted potential locations where An. stephensi could colonize if its spread continues unchecked, the spread speed and future expansion and distribution of An. stephensi has not been examined.

This study aims to enhance predictive modeling of suitable habitats and assess the risk of ongoing and future An. stephensi invasion in the Greater Horn of Africa. First, we developed an ensemble machine learning model to predict environmental suitability, integrating multi-source Earth observation data, updated An. stephensi occurrence records, and refined pseudo-absence sampling strategies. Next, we constructed a dynamic spread model that incorporates predicted environmental suitability, estimated past spread speeds, road network connectivity, and population density, and used it to project potential expansion over the next two decades. Additionally, we simulated long-term expansion through 2100 under SSP2-4.5 climate and population growth scenarios to assess shifts in habitat suitability. Overall, this high-resolution environmental suitability and spread modeling framework support the development of targeted surveillance and vector control strategies in the near term, while highlighting potential future hotspots under continued climate and environmental change.

2. Materials and methods

2.1. Study area

The study area covers the Greater Horn of Africa (European Commission, 2025), including countries Djibouti, Eritrea, Ethiopia, Somalia, Kenya, Sudan, South Sudan, Uganda, Rwanda, and Burundi, spanning approximately 5.2 million square kilometers (Fig. 1). An important geological feature of this regions is the Great Rift Valley, which contains depressions with lakes and water bodies that serve as breeding sites for malaria-transmitting Anopheles mosquitoes (Amimo, 2023). West of the Rift Valley lie the East African Highlands, extending across Ethiopia and Kenya. Due to their higher elevations and consequently lower temperatures, these highland areas are generally less favorable for mosquito proliferation. Nonetheless, malaria prevalence in the region exhibits spatial and temporal variability, largely influenced by seasonal weather patterns and climate fluctuations (Afrane et al., 2012; Githeko et al., 2006). The region exhibits substantial climatic diversity: hyper-arid deserts dominate parts of Djibouti, Eritrea, Somalia, and northern Sudan; warm semi-arid lowlands are found in South Sudan and northern Uganda; while humid highland and equatorial zones occur in Rwanda, Burundi, and southwestern Ethiopia. Precipitation is concentrated in the spring and summer, particularly in the Ethiopian Highlands and southern Kenya, while Sudan receives minimal rainfall during these seasons (Supplementary Fig. S2). Annual mean temperatures range from 15°C in the Ethiopian highlands to 26°C in the Ethiopian Rift Valley and to 36°C in northern Sudan’s desert. The Lake Victoria basin, an active malaria transmission zone, maintains an annual mean temperature of 23°C.

Fig. 1.

The range expansion of An. stephensi distribution in the greater Horn of Africa.

In 2020, the study area had an estimated population of about 300 million inhabitants (WorldPop, 2023). The Greater Horn of Africa faces a significant malaria burden, with Uganda reporting over 13 million cases in 2022, while Ethiopia, Kenya, Sudan, and Burundi each recorded more than 3 million cases (WHO, 2024a). Malaria parasite prevalence remains notably high in Ethiopia (13.6 %) and Uganda (9.5 %) (Alegana et al., 2021; Kendie et al., 2021). The primary native malaria vector species in the region are Anopheles gambiae s.l. (including Anopheles gambiae s.s., Anopheles coluzzii, and Anopheles arabiensis), Anopheles funestus s.s., Anopheles coustani, Anopheles pharoensis, and Anopheles nili (Al-Eryani et al., 2023; Altahir et al., 2022; Sinka et al., 2012). By 2024, 424 confirmed locations of the invasive An. stephensi were identified within the study area, with more than 90 % of detections occurring after 2020 (Fig. 1; see Supplementary Information S2).

2.2. Data collection

2.2.1. An. stephensi occurrence data

An. stephensi occurrence data were obtained from the WHO invasive vector species map (WHO, 2023a) and supplementary public reports and surveys (Martin James Donnelly, 2024), spanning from 1984 to 2024. After removing duplicates and records with missing survey dates, 867 confirmed locations were retained. For An. stephensi environmental suitability modeling, these locations were rasterized to 30 arc-second resolution (approximately 1 km at the equator) using GDAL (OSGeo, 2024), yielding 649 unique positive grid cells.

2.2.2. Input variables

A suite of meteorological, environmental, geophysical, and socioeconomic variables (Table 1) was compiled to model An. stephensi environmental suitability and its potential future spread in the Greater Horn of Africa. These variables were selected based on their relationships and relevance between mosquito ecology, breeding habitat availability, microclimatic conditions, and dispersal.

Table 1.

Collected input variables for An. stephensi environmental suitability modeling.

Variable name	Spatial Resolution	Temporal Resolution	Source
Meteorological factors
Seasonal max. temperature	30 arc-seconds	1970–2000	WorldClim (Fick and Hijmans, 2017)
Seasonal min. temperature	30 arc-seconds	1970–2000	WorldClim (Fick and Hijmans, 2017)
Seasonal wind speed	30 arc-seconds	1970–2000	WorldClim (Fick and Hijmans, 2017)
Seasonal precipitation	30 arc-seconds	1970–2000	WorldClim (Fick and Hijmans, 2017)
Seasonal relative humidity	0.25°	1984–2022	ERA5 (Hersbach et al., 2020)
Socioeconomic factors
Population density	1 km	2000–2020	WorldPop (WorldPop, 2024)
Night light	30 arc-seconds	1992–2021	Earth Observation Group (Ghosh et al., 2021)
Land use	300 m	1992–2020	ESA (Harper et al., 2023)
Road types and network		2018	World Bank Group (World Bank Group, 2024)
Environmental factors
Vegetation Index	0.05°	2000–2022	NASA MODIS (NASA, 2023)
Geophysical factors
Elevation	1 km		HydroSHEDS (HydroSHEDS, 2023)

Meteorological variables including minimum and maximum temperature, precipitation, relative humidity, and wind speed, were incorporated, because mosquito growth, survival, biting rate, and parasite replication are strongly temperature and humidity. Temperature extremes regulate larval growth and adult longevity; precipitation influences larval habitat formation; and humidity affects desiccation risk and adult flight activity (Mordecai et al., 2019; Shapiro et al., 2017). Elevation, derived from HydroSHEDS, influences temperature and hydrological patterns that shape environmental suitability gradients across the region (Smith et al., 2020). Monthly minimum and maximum temperature, precipitation, and wind speed were obtained from WorldClim v2.1 at 30 arc-second resolution (Fick and Hijmans, 2017). Monthly relative humidity was sourced from ERA5 reanalysis (Hersbach et al., 2023). Vegetation indices were obtained from the MODIS Terra MOD13C2 dataset (NASA, 2023), as vegetation affects microhabitat shading, humidity, and local temperature buffering relevant to mosquito survival (Beck-Johnson et al., 2017). To reduce data volume while retaining seasonal variability, all monthly meteorological variables were aggregated into seasonal (quarterly) means: December–February (Quarter 1), March–May (Quarter 2), June–August (Quarter 3), and September–November (Quarter 4).

Socioeconomic factors were included because An. stephensi is strongly associated with urban environments, artificial containers, and human mobility (Tatem et al., 2008; WHO, 2024b). Nighttime lights capture urbanization intensity and infrastructure development (Ghosh et al., 2021). Population density (WorldPop, 2024) reflects host availability and container abundance. Annual land cover data (Defourny et al., 2017) infers the spatial physical environments (Shamaei and Jafarpour Ghalehteimouri, 2024; Shamai and Jafarpour Ghalehteimouri, 2024), were used to extract cropland ratios, which relate to irrigation-based breeding habitats. Road network data (World Bank Group, 2024) were incorporated because cargo transport and human movement facilitate long-distance dispersal of An. stephensi (Lessani et al., 2024). All datasets were reprojected to WGS84 and resampled to 30 arc-second resolution using GDAL to ensure spatial harmonization before modeling.

2.2.3. Future climate data

To evaluate the future An. stephensi environmental suitability in the Greater Horn of Africa, we selected climate model demonstrating strong performance in simulating regional climate patterns. Three general circulation models (GCMs) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) were chosen based on their reported skills in reproducing temperature and precipitation patterns in the study region (Ayugi et al., 2021; Berhanu et al., 2023; Makula and Zhou, 2022): the Max Planck Institute for Meteorology Model (MPI-ESM1-2-HR), the Institute Pierre-Simon Laplace Model (IPSL-CM6A-LR), and the Hadley Centre Global Earth Model (HadGEM-GC31-LL). Studies show that multi-model ensemble mean (MMEM) projected climate model output can better capture both spatial and temporal trends of the temperature and precipitation (Ayugi et al., 2021; Makula and Zhou, 2022). Therefore, the MMEM was computed and used as input in this study.

We employed the Shared Socioeconomic Pathway (SSP) 2–4.5 scenario, which represents a plausible future trajectory (Hausfather and Peters, 2020). Under this scenario, CO₂ emissions are projected to remain around current levels until 2050, subsequently decreasing but not reaching net zero by 2100 (Masson-Delmotte et al., 2021). The temperature is expected to rise by 2 °C and 2.7 °C for 2041–2060 and 2081–2100, respectively. Future climate variables (monthly minimum and maximum temperatures and precipitation) from the selected GCP model and SSP scenario were obtained from WorldClim v2.1, which provides bias-corrected and downscaled CMIP6 projections (Fick and Hijmans, 2017).

2.3. Modeling framework

This section outlines the overall framework of our study, which includes three major components: (1) multi-source data harmonization, (2) An. stephensi environmental suitability modeling, and (3) An. stephensi spread modeling. The workflow is summarized in Fig. 2.

Fig. 2.

The overfall framework of the An. stephensi environmental suitability and spread modeling.

First, we cleaned and resampled all datasets to a 30 arc-second spatial resolution. The gridded data were indexed by latitude and longitude and stored in an SQL database to support efficient spatial querying. Second, we developed an ensemble machine learning model to predict environmental suitability. The model integrates meteorological, environmental, geophysical, and socioeconomic variables, as well as the most recent An. stephensi occurrence records. Pseudo-absence sampling was refined using ecologically informed physical constraints to better capture the species’ environmental niche. Third, we conducted a geospatial analysis of historical spread patterns to estimate past spread speeds and developed a dynamic spread model incorporating road network connectivity, predicted environmental suitability, and population density. Finally, we used this framework to project short-term expansion over the next two decades and long-term shifts in habitat suitability under climate and population scenarios extending to 2100.

2.3.1. Multi-source data harmonization

All collected datasets were initially obtained at their initial spatial resolutions, and then resampled to a 30 arc-second grid using bilinear or average methods depending on the data type. Each grid cell was uniquely encoded by its row and column based on its latitude and longitude. The harmonized data were then stored in an SQL database, allowing us to efficiently retrieve variables from for downstream modeling tasks.

2.3.2. An. stephensi environmental suitability modeling

Pseudo-absence data:

Environmental suitability models can be categorized presence-only or presence-absence approaches (Brotons et al., 2004). Comparisons of various species distribution models show that presence-absence models generally outperform presence-only models (Elith et al., 2006); therefore, we chose the presence-absence model to predict An. stephensi environmental suitability.

Field surveys have reported some absence locations of An. stephensi in western Ethiopia ((Balkew et al., 2021). However, these absence records could not be directly incorporated into environmental suitability modeling, as the observed absence does not necessarily indicate environmental unsuitability. There are two key reasons. Firstly, as an invasive species, An. stephensi may not have spread to the surveyed locations during the surveyed period. Secondly, the scope of the surveys may have been limited, and the species might be detected with more comprehensive sampling efforts. Therefore, we generated pseudo-absence data based on environmental constraints on the mosquito life cycle to enable presence-absence model (Stokland et al., 2011).

Villena et al. (2022) demonstrated that An. stephensi mosquitoes have the greatest temperature range for transmission suitability, with a range of 15.3 to 37.2 °C for P. falciparum and 15.7 to 32.5 °C for P. vivax, using a mechanistic model. Furthermore, Santos-Vega et al. (2022) indicated that relative humidity (RH) of at least 60 % as optimal for mosquito longevity and malaria transmission. We thus selected pseudo-absence locations based on: (1) monthly minimum temperature (Tmin) < 15.3 °C, (2) maximum temperature (Tmax) > 37.2 °C, or (3) monthly RH < 60 %. However, these criteria still resulted in relatively high false negative rates (Supplementary Table S1), indicating the need for additional or combined selection criteria (see Supplementary Information for details). While temperature and RH thresholds for pseudo-absence sampling could be adjusted to reduce false negative rates, setting more extreme thresholds would reduce the generalizability of pseudo-absence categories and could lead to overestimation of environmental suitability in model predictions.

Dataset Preparation:

For model development, we used 548 positive grid cells of confirmed An. stephensi occurrences from 1984 to 2020 for training and validation, while reserving 101 positive grid cells from 2021 to 2024 for testing. Following the recommendation (Barbet-Massin et al., 2012) of using an equal number of pseudo-absences as available presences, with 182 grid cells from each of the three pseudo-absence categories, totaling 546 pseudo-absences. Each An. stephensi positive grid cell was linked with their corresponding input variables based on both their geographical locations and year of detection. For An. stephensi occurrences extending beyond available input dataset timeframes, it was mapped to the nearest available year. For example, while we have An. stephensi observations from 2024, the collected vegetation index and relative humidity are only available up to 2022. In such cases, we used the 2022 vegetation index and relative humidity data for the 2024 observations. Locations with missing input variables were excluded from modeling. Variables were standardized using z-score normalization (Freedman et al., 2020). To address potential multicollinearity among input variables, we generated a correlation matrix and performed Principal Component Analysis (PCA) using Scikit-learn (Pedregosa et al., 2011).

Ensemble Model:

Single learning algorithms risk overfitting and poor generalization to unseen data. Ensemble models mitigate this risk by combining multiple models, improving overall predictive performance (Sagi and Rokach, 2018). We selected six robust machine learning models for the ensemble approach, including Random Forest (RF) (Breiman, 2001), Extremely Randomized Trees (ERT) (Geurts et al., 2006), XGBoost (XGB) (Chen and Guestrin, 2016), Multilayer Perceptron (MLP) (Goodfellow et al., 2016), Bernoulli Naive Bayes (BNB) (Singh et al., 2019), and Logistic Regression (LR) (Hosmer Jr et al., 2013). These models offer complementary strengths: the tree-based models (RF, ERT, and XGB) enhance prediction stability through different strategies—RF uses bagging, ERT increases randomness (Friedman, 2001; Geurts et al., 2006), and XGB employs gradient boosting for accuracy optimization (Chen and Guestrin, 2016). MLP captures complex non-linear relationships (Goodfellow et al., 2016), while BNB and LR provide efficient computation and robust performance on linearly separable data (Hosmer Jr et al., 2013; Singh et al., 2019). All the models were implemented using the Scikit-learn package (Pedregosa et al., 2011).

Model settings:

The model predicts environmental suitability for each grid cell on a continuous scale from 0 to 1, with higher values indicating greater environmental suitability. A threshold of 0.5 was used to classify predictions, whereby values above 0.5 were considered suitable (positive cases). Model performance was evaluated using accuracy, defined as the proportion of correctly predicted grid cells out of the total number of test cells (Metz, 1978), with values ranging from 0 to 1.

The dataset was randomly split into 80 % for training and 20 % for validation. Each sub-model’s performance was assessed using the validation dataset, and their predictions were aggregated into an ensemble model weighted by individual accuracy, meaning that sub-model with higher accuracy will be given greater weight. The accuracy of the ensemble model was then calculated. Given the uncertainty introduced by pseudo-absence sampling, as sampled locations may incompletely represent unfavorable environmental conditions for An. stephensi, we performed 500 ensemble runs with different pseudo-absence samples, and selected only those runs achieving an accuracy over 0.95 to build the final model output. The selected runs were averaged to calculate test dataset accuracy and generate the An. stephensi environmental suitability map. Model uncertainty was quantified as the relative standard deviation across selected ensemble runs for each grid cell.

2.3.3. An. stephensi spread modeling

Estimate past spread speed:

To estimate past spread speed and analyze historical spread patterns, we applied established methods (Kraemer et al., 2019; Tisseuil et al., 2016) for species spread speed measurements. First, we interpolated the invasion year ($I_{\mathit{year}}$) across the grid cells in the study areas using existing occurrence locations. For grid cells with multiple detections of An. stephensi in different years, the earliest invasion year was assigned. Universal Kriging (Kitanidis, 1997) was then applied to interpolate the potential invasion year across the study area, based on $I_{\mathit{year}}$ and their coordinates (x, y), using the PyKrige package (Benjamin Murphy and Müller, 2024), shown in Eq. (1).

$$I_{year}=UniversalKriging\left(x,y,year\right)$$ (1)

Since An. stephensi initially invaded Africa through Djibouti before spreading to other regions in the Greater Horn of Africa, coordinates were adjusted as relative distances from Djibouti ($x_{\mathit{Dijibouti}},y_{\mathit{Dijibouti}}$). Additionally, spread speed is not uniform, and it may take longer time for An. stephensi to occupy or cross areas that are less suitable for its growth. The coordinates were further weighted by the predicted environmental suitability within the grid cell, shown in e.q. (2) and (3). The constant term of 0.1 was added to the denominator to prevent excessively large values if the value of suitability was too small.

$$x=\frac{x-x_{Dijibouti}}{x_{suitablity}+0.1}$$ (2)

$$y=\frac{y-y_{Dijibouti}}{y_{suitablity}+0.1}$$ (3)

Secondly, after deriving the interpolation of invasion year of An. stephensi, the local slope ($S_{\mathit{local}}$) of the surface was measured using a 30 × 30 km moving window convolution filter, and smoothed with an average 110 × 110 km filter to prevent local null values (Kraemer et al., 2019), shown in shown in e.q. (4).

$$S_{local}=Convolution\left(I_{year,}30\times 30\right)$$ (4)

Lastly, the local spread speed ($S_{\mathit{speed}}$) was then obtained by taking the inverse of the friction value, shown in Eq. (5). The past spread speed of each positive grid cell was then extracted for further use and analysis.

$$S_{speed}=\frac{1}{S_{local}}$$ (5)

An. stephensi spread modeling:

To model the spread of An. stephensi, we considered two main pathways: local dispersal via the mosquito flight and long-distance dispersal via road networks through human and cargo movement. Previous studies indicate that African Anopheles mosquito can disperse approximately 4.7 km per year (Carlson et al., 2023). Based on this, we assumed a local spread rate of 4.7 km/year for An. stephensi. An. stephensi occurrence data (Fig. 1) shows that most occurrence locations were near major roads, suggesting that the spread of An. stephensi is closely tied to human and cargo movement along road networks. We tested three increasingly complex spread hypotheses:

1. Uniform-speed scenario: In this baseline scenario, An. stephensi spreads along the road network due to human and cargo movement from already invaded areas (Swan et al., 2022). A constant spread speed is applied across all road types, assuming no differences in movement efficiency or traffic volume between major and secondary roads.

2. Road-weighted scenario: Here, the spread speed varies based on the road types. Major roads are assigned faster spread rates, while secondary and tertiary roads have progressively lower spread speeds. This reflects the real-world variation in traffic volume and connectivity between road types.

3. Road- and environmental suitability-weighted scenario: This advanced scenario builds upon the second by incorporating both road-based speed and environmental context. Spread likelihood is further modulated by local environmental suitability and population density. This assumes that An. stephensi is more likely to establish in areas where the environment is favorable for mosquito survival and where dense human populations increase opportunities for introduction and establishment (Tatem et al., 2008).

For long-distance spread, spread speed were sampled from prior estimated data, and road-based dispersal ranges were calculated accordingly. Local spread was then simulated from road-adjacent grid cells into neighboring unoccupied areas. To reflect the differing spread potentials across road types, we introduced a road type factor $\left(F_R\right)$. For example, Major roads, which connect urban centers, were assigned higher $F_R$ values due to their greater facilitation of long-distance. In contrast, local roads were assigned with lower $F_R$ values due to their limited connectivity and slower travel speeds.

We also incorporated a population density factor $\left(F_P\right)$ to account for the role of human population in mosquito establishment and spread. Population density influences mosquito establishment through human mobility patterns (Marshall et al., 2018; Wesolowski et al., 2012), captured by a population density factor $\left(F_P\right)$. $F_P$ was scaled logarithmically, shown in Eq. (6).

$$F_{p(i,j)}=\frac{log10\left({Pop}_{(i,j)}\right)-log10\left({Pop}_{min}\right)}{log10(P)-log10\left({Pop}_{min}\right)}$$ (6)

where $Po{p}_{(i,j)}$ is the population density of each grid cell i, j, $Po{p}_{\mathit{min}}$ is the minimum population density, and P is the population density threshold.

Environmental suitability (S) was also integrated, as regions with low environmental suitability are less likely to support sustained mosquito populations, thereby reducing spread potential. The final effective spread speed for each grid cell is shown in Eq. (7)

$$U_{S(i,j)}=F_{R(i,j)}\times F_{p(i,j)}\times S_{(i,j)}\times V_{(i,j)}$$ (7)

where $F_R$ is the road type factor, $F_P$ is the population density factor, S is the environmental suitability, and V is the sampled past spread speed.

Moreover, higher suitability and population density increase the likelihood of mosquito establishment in newly invaded areas, while less suitable or low-density regions require multiple spread attempts for establishment. To simulate mosquito establishment in newly invaded regions, we applied a positive threshold (PT) condition. A grid cell $G_{(i,j)}$ becomes positive (i.e., colonized) based on the number of spread attempts exceeding a threshold adjusted by local environmental suitability and population density, shown in Eq. (8).

$$G_{(i,j)}\ge \frac{PT}{F_{p(i,j)}\times S_{(i,j)}}$$ (8)

This ensures that grid cells with higher environmental suitability and population density require fewer spread attempts for successful establishment, while less favorable areas demand more repeated introductions.

This approach assumes that once An. stephensi is established in a grid cell, it continues to spread to neighboring suitable areas. Model parameters (Supplementary Table S4) were explored over 1000 scenarios, and optimal configurations were selected using a grid search algorithm (Liashchynskyi and Liashchynskyi, 2019). The final model was chosen by maximizing the product of temporal accuracy (correlation between simulated and observed invasion years) and spatial accuracy (number of correctly identified positive grid cells).

3. Results

3.1. Variable correlation, PCA, and ensemble model composition

Correlation analysis (Supplementary Fig. S1a) revealed strong positive correlations between relative humidity and precipitation, with both variables strongly negatively correlated with temperature. Population density, cropland ratio, and vegetation index showed moderate positive correlations with relative humidity and precipitation, while negatively correlating with temperature and wind speed. More details of correlation between each input variable can be found in Supplementary Fig. S1a.

Principal Component Analysis (PCA) identified six principal components (PCs) that captured 91 % of the total variance across 25 input variables. PC1 explained 45 % of variance, showing negative correlations with temperature and wind speed and positive correlations with other variables. PC2 (19 % variance) correlated positively with wind speed, and PC3 (11 % variance) showed negative correlation with relative humidity and positive correlation with elevation. PC4-6 captured smaller variance proportions, representing distinct patterns in cropland ratio, population, and summer precipitation, respectively. Detailed correlation patterns between PCs and input variables are presented in Supplementary Fig. S1b and Table S2.

In the ensemble model (Supplementary Table S3), tree-based algorithms demonstrated strongest performance, with ERT, RF, and XGB contributing weights of 0.177, 0.176, and 0.176, respectively. MLP and LR contributed weights of 0.175 and 0.162, while BNB provided complementary probabilistic modeling with a weight of 0.135. The tree-based models’ dominance reflects their effectiveness in capturing complex, nonlinear relationships in environmental suitability patterns.

3.2. Predicted An. stephensi environmental suitability

Our model demonstrates improved performance in identifying suitable habitats for An. stephensi compared to existing approaches (Sinka et al., 2020; Villena et al., 2022). For the independent test period (2021–2024), our model achieved 0.93 accuracy, exceeding both Sinka et al. (2020) at 0.76, and Villena et al. (2022) at 0.87 (predictions are correct when environmental suitability is sustained for more than two months).

The predicted environmental suitability map (Fig. 3a) identifies highly suitable regions across in Uganda, South Sudan, Kenya, eastern Ethiopia, and southern Somalia. While the East African Highlands (western Ethiopia and southwestern Kenya) are predominantly predicted unsuitable, the model identifies suitable conditions in lower-elevation valleys. In Sudan, although the Sahara Desert creates widespread unsuitable conditions, the Nile River corridor maintains environmental suitability. The East African Rift system, from Ethiopia’s Red Sea Rift to the Kenyan Rift Valley, including major water bodies (Lakes Abaya, Turkana, and Victoria), shows high suitability. Coastal zones along the Indian Ocean, Gulf of Aden, and Red Sea also demonstrate high environmental suitability.

Fig. 3.

The ensemble machine learning model results. (a) Predicted environmental suitability map of An. stephensi; (b) Uncertainty calculated per pixel across the predicted range, indicating where the ensemble model provides the most reliable (lower uncertainty: light green) and least reliable (higher uncertainty: dark green) predictions. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3b illustrates the uncertainty of the predicted environmental suitability map, showing the ratio of standard deviation to mean across the selected model runs for each grid cell. A larger value indicates higher model uncertainty. Areas in northern Sudan and boundary regions between suitable and unsuitable zones in Kenya and Ethiopia exhibit relatively high uncertainties.

3.3. An. stephensi spread modeling

Analysis of invasion patterns shows that An. stephensi initially spread from Djibouti into eastern Ethiopia before rapidly expanding across Ethiopia and into neighboring countries including Sudan, Kenya, Somalia, and Eritrea (Fig. 1). Spread speed has increased substantially over the past decades, with median speeds rising from approximately 20 km/year in 2012 to 44 km/year in 2016, and exceeding 120 km/year by 2024 (Fig. 4).

Fig. 4.

Reconstruction of An. stephensi spread speed of An. stephensi in 2012–2024 based on the confirmed positive locations in the study area. The red line represents the average yearly spread speed across confirmed locations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Our spread modeling framework successfully reconstructed historical dispersal dynamics and projected future range expansion. We evaluated model performance under increasingly complex scenarios, as described in Section 2.3.2. The uniform-speed scenario, which assumes a constant spread speed across all road types, achieved a temporal correlation of 0.56 between observed and simulated invasion years, capturing 199/210 positive grid cells. The road-weighted scenario, which accounts for varying spread speeds by road type, improved the temporal correlation to 0.59 while maintaining spatial accuracy (199/210). The road- and environmental suitability-weighted scenario, which further incorporates population density, environmental suitability, and positive detection thresholds, achieved a temporal correlation of 0.66 and captured 203/210 positive grid cells (Supplementary Fig. S2a). Optimized parameters are detailed in Supplementary Table S4. Model outputs suggest possible undetected spread into western Ethiopia and southern Somalia (Fig. 5a).

Fig. 5.

Projected An. stephensi distribution for 2024–2050. Predicted range expansion shown for (a) 2024, (b) 2030, (c) 2040, and (d) 2050, based on environmental suitability, transportation networks, population density, and historical occurrence data.

Future projections indicate potential expansion into southern Somalia, Kenya, Uganda, Rwanda, and Burundi by 2030, particularly around Lake Victoria (Fig. 5b). By 2050, further spread is projected into South Sudan through road connections with Sudan and Kenya, potentially reaching northeastern Democratic Republic of the Congo (Fig. 5d).

3.4. Future climate and environmental suitability change

Supplementary Figs. S3–S5 present baseline and future seasonal meteorological conditions under SSP2-4.5 scenarios for the mid- (2041–2060) and late-century (2081–2100) periods, based on the MMEM of three selected CMIP6 models. Climate change projections for 2081–2100 (Supplementary Fig. S6) show maximum temperature increases ranging from 1.8 °C to 4.5 °C, with moderate warming in Kenya and Ethiopia but more increases in Sudan. Minimum temperatures are projected to rise by 1.9 °C to 4.7 °C, particularly in eastern Ethiopia and northern Sudan. Despite the warming trend, precipitation is projected to increase across most of the study region, with total seasonal changes ranging from −15 mm to + 50 mm. Ethiopia shows consistent rainfall increases across all seasons, while Somalia, Kenya, and Uganda display seasonal variability, with decreasing trends during summer. In Sudan, spring and winter precipitation is projected to decline, whereas summer and autumn are expected to see increases.

Future An. stephensi environmental suitability under the MMEM projections (Supplementary Fig. S7) remains broadly consistent with the 2024 baseline, but exhibits an expansion of high-suitability areas, particularly across the African highlands. Comparison of the 2081–2100 projections with the 2024 baseline (Fig. 6) shows more pronounced changes in environmental suitability than those projected for 2041–2060. Most countries within the Greater Horn of Africa are expected to experience increased suitability, with only limited areas showing declines. In Ethiopia, 66 % of the land area is projected to experience environmental suitability increases greater than 0.1 by 2081–2100, with 24 % of the land transitioning from unsuitable to suitable conditions. Only 1 % of the area is projected to experience decreased environmental suitability, primarily in the northeastern region where elevated temperatures and increasing aridity reduce habitat suitability. In Kenya, 37 % of the land area shows environmental suitability increases of over 0.1, with 8 % transitioning from unsuitable to suitable. Across the entire study region, 47 % of the area is projected to experience environmental suitability increases exceeding 0.1, and 10 % is expected to transition from unsuitable to suitable conditions (see Table 2 for country-level statistics). Despite projected precipitation increases in northern Sudan, these areas remain unsuitable for An. stephensi due to persistently high temperatures and arid conditions (Supplementary Fig. S5).

Fig. 6.

Projected changes in environmental suitability for An. stephensi comparing (a) 2041–2060 and (b) 2081–2100 to baseline conditions in 2024. Changes are shown as absolute differences between future and current environmental suitability values, where positive values indicate increased suitability and negative values indicate decreased suitability.

Table 2.

Projected changes in An. stephensi suitable habitat by country in the Greater Horn of Africa. Environmental suitability trends and habitat transitions were determined by comparing environmental suitability values between 2081–2100 and baseline (2022). Directional changes are expressed as positive (increasing) or negative (decreasing). Habitat transitions are defined by crossing the 0.5 suitability threshold in either direction.

Country	Suitable areas (%)			Environmental suitability trends (%)		Habitat transitions (%)
	2022	2041–2060	2081–2100	Increasing trend	Decreasing trend	Non-suitable to suitable	Suitable to non-suitable
Ethiopia	59	75	82	66	1	24	1
Kenya	85	90	92	37	1	8	0
Sudan	36	38	40	43	1	5	1
Uganda	89	96	97	72	1	8	0
South Sudan	98	99	99	39	0	1	0
Rwanda	46	66	76	84	0	29	0
Burundi	34	63	73	90	0	40	0
Somalia	86	95	97	90	0	11	0
Eritrea	88	95	96	40	7	10	2
Djibouti	91	99	99	37	3	9	1
Total	63	70	73	47	1	10	1

Population projections for the Greater Horn of Africa (Gao, 2019) estimate a total population of approximately 600 million by the end of the century. Based on future environmental suitability projections, around 500 million individuals are expected to reside in areas suitable for An. stephensi, a significant increase from approximately 200 million in 2024. This projected expansion of the mosquito’s range, coupled with rapid urbanization and population growth, substantially heightens the risk of malaria transmission across the region. Country-specific estimates of population exposure to An. stephensi-suitable environments are provided in Supplementary Table S6.

4. Discussion

4.1. The importance of improving environmental suitability mapping

Accurate environmental suitability mapping is essential for public health planning and modeling the potential spread of An. stephensi. While confirmed occurrence locations provide valuable insights into suitable conditions, presence-only approaches face significant limitations for invasive species modeling. The Maximum Entropy (MaxEnt) model (Phillips et al., 2006), which relies solely on presence data, appears to be a suitable choice. MaxEnt identified high environmental suitability in regions that correspond to confirmed occurrence locations during the 2012–2020 training period, as indicated by the red circles (Supplementary Fig. S8). However, it reveals lower suitability for occurrence locations during the 2021–2024 test period locations that are spatially distinct from training data. The underestimation occurs due to MaxEnt’s reliance solely on presence data. Since An. stephensi is an invasive species, the confirmed occurrence locations may not fully capture its complete ecological niches and adaptive capabilities. As a result, the model struggles to accurately predict environmental suitability in areas beyond those captured in the confirmed occurrence locations.

In contrast, presence–absence models generally outperform presence-only models by leveraging both presence and absence information (Elith et al., 2006). For example, Sinka et al. (2020) used an approach that samples pseudo-absence locations from records of Aedes and Culex species, where An. stephensi was not found. However, given its invasive nature, the absence of An. stephensi at surveyed locations does not necessarily imply a lack of preference for those habitats. Instead, it might indicate the limited spread or constrained survey coverage of the species, Thus, more refined strategies are needed to generate biologically meaningful pseudo-absence data. In this study, we addressed this limitation by generating pseudo-absence locations based on physical constraints—specifically, thresholds in temperature and relative humidity that affect mosquito survival. Nevertheless, our results suggest that An. stephensi exhibits adaptability to more extreme conditions than these thresholds (Supplementary Table S1), necessitating combined criteria for more effective pseudo-absence sampling.

Additionally, robust environmental suitability modeling requires comprehensive occurrence data and a diverse set of input variables. Our model outperforms existing approaches that rely on fewer occurrence records and limited variable sets (Sinka et al., 2020; Villena et al., 2022). During the independent test period (2021–2024), our model achieved 0.93 accuracy, exceeding both Sinka et al. (2020) at 0.76, and Villena et al. (2022) at 0.87 (predictions are correct when environmental suitability is sustained for more than two months). A greater number of occurrence locations better represent the ecological niche of An. stephensi, enhancing the accuracy of environmental suitability maps. For instance, Sinka et al. (2020) found that incorporating more An. stephensi occurrences in Africa in training dataset improved accuracy compared to using only Asian occurrences. Furthermore, as an urban malaria vector, An. stephensi is capable of adapting to localized environments and microclimates, particularly in urban and peri-urban areas where it may breed in artificial water containers. For example, breeding in human-made water containers provides shelter, helps maintain relatively stable temperatures compared to the surrounding environment, and ensures access to water during dry seasons when malaria transmission is generally low. Therefore, accurate environmental suitability mapping requires integrating comprehensive input variables, occurrence data, and refined pseudo-absence sampling strategies.

4.2. The continuous spread of An. stephensi

The spread of An. stephensi is inherently complex, driven by suitable environments that provide ideal temperature and blood source that support its life cycle, and human activities that facilitate its reach beyond natural dispersal limits. Our modeling demonstrates that incorporating multiple factors—road network types, environmental suitability, population density, and thresholds that determine when a grid cell becomes positive can improve the reconstruction of past spread patterns compared to simpler models. However, this approach assumes that the timing of detection closely reflects the actual timing of invasion, emphasizing the critical need for continuous surveillance to accurately monitor spread dynamics.

Initially, the spread of An. stephensi was relatively slow following its emergence in Djibouti. This early low spread rate was likely influenced by the sparse population density and limited transportation infrastructure along the Djibouti–Ethiopia border region, particularly within areas such as the Mile Serdo Wildlife Reserve and Yangudi Rassa National Park. Once the species reached major road networks in eastern Ethiopia, its spread accelerated rapidly, reaching urban centers like Semera and Dire Dawa through human and cargo movement.

The spread speed of An. stephensi varies depending on the road types, resulting in different spread patterns across countries. Supplementary Fig. S9 illustrates the distribution of transportation networks and city locations within each country. The proximity of cities to different road types differs by country. In Ethiopia, most cities are located near major roads, suggesting that An. stephensi is more likely to spread via high-capacity transportation corridors. In contrast, in countries such as Uganda and Sudan, tertiary roads are the most accessible road type for many cities, implying that dispersal in these regions is more influenced by lower-tier transportation networks.

An. stephensi is expected to continue its expansion across Africa via road networks and cargo transport. For instance, it has been hypothesized that the vector could spread from the Red Sea port in Sudan to inland regions, including eight Sudanese states bordering Chad and the DRC. Despite low environmental suitability in some desert areas, the vector may cross unsuitable zones and establish localized suitable environments. This highlights the risk of further spread into neighboring countries such as Chad, DRC, South Sudan, and Uganda.

This continued spread poses a significant public health risk, particularly in regions where the species can establish and reproduce. The threat is further amplified by rapid urbanization and population growth, which increase opportunities for human-mediated movement. Therefore, human efforts to prevent its future global dissemination will be most effective if focused on preventing-mediated spread and establishment (Kraemer et al., 2019). Furthermore, targeting suitable habitats with environmentally sustainable and non-insecticide-based vector control methods like larval source management, may offer a cost-effective strategy for integrated malaria vector control (Abubakr et al., 2022; Jiang et al., 2023).

There is a critical need for continuous surveillance to accurately monitor the spread dynamics of An. stephensi. The absence of confirmed records in some areas may reflect gaps in surveillance rather than true absence, as illustrated by the recent detection of An. stephensi in northeastern Nigeria (Moustapha et al., 2025). More importantly, Africa’s vector surveillance systems are lagged in monitoring changes in vector composition and distribution due to insufficient financial and workforce resources, advanced molecular and sequencing techniques, and training (Ahmed et al., 2022a), underscoring the urgent need to strengthen surveillance and early warning systems. To improve detection and control, multiple types of surveys should be strategically combined. Exploratory surveys aim to determine whether An. stephensi has invaded new regions and should prioritize urban centers near borders, as well as international and national points of entry—such as seaports, airports, dry ports, and major transport corridors—where the movement of people, animals, or goods may facilitate mosquito introduction. In parallel, bionomic surveys are essential to identify preferred hosts, assess the species’ role in malaria transmission, and determine insecticide susceptibility, thereby informing targeted surveillance and vector control efforts (Ahmed et al., 2022b).

4.3. Current and future suitability and range

The future distribution of An. stephensi will be shaped by a combination of environmental suitability and anthropogenic factors, with climate change playing a central role in altering habitat availability. Warming temperature and uneven changes in precipitation patterns will impact malaria transmission differently across regions. These shifts are influenced by the current and projected local climates concerning the optimal thermal conditions for disease transmission, resulting in changing malaria transmission hotspots (Mordecai et al., 2020). Our findings show a significant increase in relative environmental suitability within the Kenyan and Ethiopian highlands under SSP 2–4.5, a climate scenario which is most likely to happen (Hausfather and Peters, 2020). This suggests that warming may create newly suitable habitats for An. stephensi in previously inhospitable, high-elevation regions. These findings are consistent with previous studies: Mordecai et al. (2020) projected that malaria risk hotspots would shift towards high-elevation regions under climate change, and Colón-González et al. (2021) reported that an extended length of the transmission season in the Greater Horn of Africa across multiple climate models and scenarios. Similarly, Liu et al. (2024) projected that these regions will gain environmental suitability under SSP 1–2.6, although this climate scenario is considered less likely.

While warming temperature may reduce the environmental suitability for malaria transmission by An. gambiae due to its relatively narrow thermal limits (Mordecai et al., 2020), there remains a risk of invasion and replacement by An. stephensi. It is also important to note the potential demographic changes, such as population growth, migration, and socioeconomic factors, may either amplify or counterbalance the effects of climate change on malaria transmission dynamics (Gething et al., 2010; Tompkins and Caporaso, 2016; Wesolowski et al., 2015). Additionally, malaria is not the only vector-borne disease to be concerned about in this area. Arthropod-borne viruses – or arboviruses, such as dengue (Bhatt et al., 2013), zika (Lessler et al., 2016) and chikungunya (Grabenstein and Tomar, 2023), are also widespread. An. stephensi is also known for its potential to transmit Rift Valley fever (Altahir et al., 2022), and zoonotic malaria (Ahmed et al., 2022a). While the temperature is too extreme for Anopheles, it might be just right for the others, such as Aedes aegypti. Aedes mosquitoes and arboviruses can thrives in warmer environments compared to Anopheles mosquitoes and malaria parasites (Mordecai et al., 2020). Further studies on co-existing diseases, particularly in rapidly urbanizing areas, are urgently needed.

4.4. Limitations

Our environmental suitability modeling incorporated seasonal meteorological data to capture the seasonal dynamics within year. However, discrepancies in seasonality may arise, particularly when working with large geographic areas or when the spatial characteristics of training and test datasets differ significantly. An alternative approach is to utilize bioclimatic variables derived from monthly meteorological data—such as the minimum temperature of the coldest month or precipitation of the driest month—which may provide more consistent and ecologically meaningful predictors than categorical seasonal variables. Additionally, although pseudo-absence locations were sampled using physical constraints derived from prior mechanistic models, these constraints may not fully reflect the range of environmental conditions in which An. stephensi can persist. Emerging evidence suggests the species may be more adaptable to extreme conditions than previously assumed. This may result in misclassification of potentially suitable habitats, particularly in marginal or novel environments. It is also important to note that the model’s predictive performance depends on the availability and quality of surveillance data. Incorporating data from West Africa would likely enhance the model’s accuracy and its geographic scope. As An. stephensi continues to expand into new areas, increasing the number of confirmed occurrence records will further improve the reliability of environmental suitability models.

While the publicly available remote sensing data used in our model is well-suited for assessing broad-scale environmental suitability, its spatial resolution is insufficient to capture fine-scale urban features such as human-made water storage containers. These microhabitats, particularly common in arid and semi-arid urban areas, are critical breeding sites for An. stephensi. For instance, our model predicts low environmental suitability across much of Sudan due to persistently high temperatures and aridity. However, confirmed detections of An. stephensi in these regions suggest that localized human practices, such as the use of cisterns, barrels, pots, water tanks, and water-dependent air conditioners, are facilitating its persistence. These anthropogenic factors create breeding environments that are not detectable through coarse-resolution environmental variables alone, highlighting a key limitation of remote sensing-based environmental suitability modeling and the need for complementary local data on human behavior and infrastructure.

To address this, future modeling could incorporate indirect estimates of container abundance using proxies such as population density, housing density, local income levels, and public health infrastructure. Additionally, recent studies have demonstrated that artificial container presence can be estimated using street-level imagery (Haddawy et al., 2019). Urban areas naturally exhibit greater heterogeneities due to active human activities. To enhance modeling with local temperature and land cover features, we can use data from Sentinel-2/Sentinel-3, which offers a set of broader spectral bands with high spatiotemporal resolution. This resource provides valuable information on surface temperature, land cover, vegetation, and water ponding information, offering deeper insights into the environmental suitability of An. stephensi at a local scale.

Our spread modeling aimed to fit overall past observations but may either underestimate or overestimate the spread speed at certain locations. For instance, the model inadequately captured the observed spread to western Sudan. This discrepancy is primarily due to the model’s reliance for environmental suitability as a key driver of spread speed. In western Sudan, environmental suitability was under-estimated—partly because coarse-resolution remote sensing data do not capture small-scale anthropogenic breeding sites such as man-made water containers. Moreover, human migration driven by the ongoing civil war in Sudan is also not accounted for in the current model. Future work should incorporate high-resolution urban infrastructure and behavioral data, including human-mediated transport and water storage practices, to better predict spread in such contexts. Moreover, our framework assumes that the year of detection reflects the actual year of invasion. However, An. stephensi might have already invaded these areas but went undetected due to insufficient surveillance, introducing temporal uncertainty and potential biases into the model.

5. Conclusion

This study advances the An. stephensi environmental suitability and spread modeling of Anopheles stephensi by integrating multi-source Earth observation data, including meteorological, environmental, geophysical, and socioeconomic variables. Our approach combines an expanded set of occurrence records, refined pseudo-absence sampling based on environmental constraints, and ensemble machine learning techniques to more accurately characterize the ecological niche of this invasive species. Our environmental suitability model achieved a high predictive accuracy of 0.93 for the 2021–2024 test period, significantly outperforming previous models applied to the region, which reported accuracies of 0.78 and 0.87. The spread model reconstructed historical dispersal patterns with a temporal correlation of 0.66 between observed and simulated invasion years at confirmed locations.

The results reveal that An. stephensi possesses greater environmental adaptability than previously recognized, with environmental suitability extending into highland regions such as the East African Highlands (western Ethiopia and southwestern Kenya), as well as rift zones and major water bodies, including Lakes Abaya, Turkana, and Victoria. Coastal areas along the Indian Ocean, Gulf of Aden, and Red Sea also exhibit high environmental suitability. The median spread speed of An. stephensi has increased markedly—from approximately 20 km/year in 2012 to over 120 km/year by 2024. Future projections suggest continued expansion into western Ethiopia and southern Kenya in coming decades.

Under the SSP2-4.5 climate scenario, large portions of the Kenyan and Ethiopian highlands—historically unsuitable for malaria vectors—are projected to become increasingly conducive to An. stephensi establishment due to rising temperatures. By 2100, approximately 500 million people in the Greater Horn of Africa could be at risk, driven by the intersecting pressures of climate change and population growth.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.jag.2025.105026.

Data availability

All the input variables are publicly available, as detailed in the Data Collection section. A summary of confirmed An. stephensi locations is provided in the Supplementary file 2.