Projecting the impacts of climate and land-use change on avian influenza suitability in Bangladesh

Abstract

Climate and land-use change are expected to influence the future dynamics of zoonotic disease outbreaks, including avian influenza, a prime example of a One Health challenge. Environmental and socio-economic conditions modulate the risk of virus spillover from wild birds to farmed animals and, subsequently, to humans through multiple transmission pathways. However, the extent to which changing environmental conditions may alter the spatial suitability for avian influenza transmission across regions and interfaces in the future remains uncertain. To address this gap, we developed a spatially explicit, integrated modelling framework that simulates wild and farmed bird distributions and incorporates this information into a spatial structural equation model of avian influenza exposure suitability. This model captures the complex interactions between the environment, wildlife, and poultry production and retail systems in Bangladesh. In addition, the framework includes an assessment of the population at risk living in areas considered suitable for virus exposure. Our approach allows us to project the spatial suitability for avian influenza exposure under current and future climate, land cover and population density, and to better understand the key drivers and underlying mechanisms of exposure suitability. Results indicate that suitability is expected to increase significantly in poultry farming areas, driven by growing chicken density. We find that this shift could increase the human population at risk of exposure to avian influenza by up to 79 million by 2050. This modelling approach provides an evidence-based decision support tool to help prioritise surveillance and preventive interventions in key transmission interfaces across the country.

1. Introduction

The health risks associated with zoonotic influenza is a prime example of a global One Health problem, whereby cross-species transmission is strongly influenced by the interactions of multisectoral factors. Bangladesh is a country that is particularly at risk of avian influenza virus (AIV) given its location on the East Asian-Australasian migratory bird flyway and its high density of poultry in poor biosecurity husbandry systems [1]. In addition, the presence of live bird markets (LBMs) in densely populated urban areas contributes to increase the proximity and interactions between wild birds, feral animals, poultry and humans, thus facilitating spillover events [2].

So far, few studies have investigated the impacts of changing environmental conditions on the suitability of exposure in different interfaces of exposure to AIV. Most research have focused primarily on uncovering correlates of seasonality and how these are associated with exposure indicators such as virus environmental recovery and disease notifications [3]. For example, in a study in Bangladesh's LBMs, weekly environmental virus surveillance was conducted to assess the role of climatic factors on AIV environmental recovery [4]. Temperature and relative humidity were found to be significantly associated with weekly AIV circulation. Another study investigating the impact of vaccination on AIV outbreaks in commercial poultry also found a strong seasonal effect [5]. The findings showed that some winter months had an approximately three times higher chance of outbreaks than summer months.

Despite growing evidence that most infectious diseases are sensitive to and can be exacerbated by climate change [6], few studies has been dedicated to understanding and projecting the impacts of climate change on zoonotic influenza. To date, most research focused mainly on vector-borne diseases [7]. However, in the context of zoonoses, the impacts of climate and land use change can affect non-human hosts, through shifting ecological niches, intensification of husbandry systems, increased animal density and increased proximity between farmed and wild animals [8]. Understanding how these vulnerabilities at the animal-environment interface shape downstream risks to humans is critical for anticipating future spillover events and informing cross-sectoral strategies for disease prevention and control.

Beyond current seasonal patterns, the effects of climate change on AIV exposure suitability across multiple interfaces in Bangladesh have yet to be modelled spatially, despite clear indications that global climate shifts are altering the ecology and physiology of AIV hosts [9]. The impacts of these changes are unlikely to be uniform across the country, as exposure interfaces are distributed heterogeneously across urban and rural areas, and across different climate systems. For example, urbanised settings may be particularly vulnerable due to the co-occurrence of LBMs, high-density poultry farming, and areas vulnerable to climate change, such as flood-prone settlements with inadequate biosecurity measures. Conversely, extensive poultry farming in rural areas may experience more frequent exposure to wildlife host due to shifting wild bird migration patterns. A deeper understanding of how changing climatic and land-use conditions influence cross-sectoral AIV exposure suitability in different regions is critical for developing and implementing targeted interventions to prevent future outbreaks. It is also crucial to be able to anticipate the population that may become at risk of this increased exposure.

In this study we aimed to address these gaps directly by developing a novel integrated spatial modelling approach that links changing climatic and land use conditions, wild bird and chicken hosts distribution, and exposure suitability across wildlife and animal husbandry interfaces. To achieve this, we set the following objectives:

• •Simulate suitability of exposure to AIV in four different interfaces such as waterbird nesting sites, urban bird habitat, poultry farming and LBMs,
• •Project future suitability of exposure to AIV under changing climate and land cover,
• •Assess the change in the potential human population at risk under the different scenarios, and
• •Identify hotspots across interfaces where suitability is most likely to increase

By addressing these knowledge gaps, our study provides a novel modelling framework and foundation for more targeted surveillance, biosecurity strategies, and climate-informed interventions to mitigate the risk of future AIV outbreaks.

2. Methodology

2.1. Modelling framework

The modelling framework of the study was conceptualised using a directed acyclic graph (DAG) co-designed with local experts to identify major interfaces of exposure to AIV and risk factors, including climatic variables and land cover [10]. This approach aligns with previous research on the identification of causal pathways and salient cross-sectoral risk factors for zoonotic influenza exposure through expert elicitation in the South Asia region [11]. This study goes further by 1) incorporating projections of climate and land use change impacts on AIV suitability, 2) explicitly modelling habitat suitability of wild birds and distribution of farmed bird density, 3) determining the importance of risk factors through machine learning and statistical approaches rather than through expert elicitation, and 4) project the change in the population at risk of AIV exposure.

The DAG depicts circular nodes representing interfaces of exposure to AIV and the rectangular nodes represent risk factors within a particular interface (Fig. 1). Edges indicate the influence of a risk factor on the suitability of an interface. In addition, hexagonal nodes show sub-models of species distribution of wild bird hosts species and chicken density distribution, and parallelogram nodes represent the reported occurrence of AIV in each interface.

Fig. 1.

Directed acyclic graph representing the modelling framework of AIV exposure suitability in Bangladesh.

The DAG was implemented by 1) generating suitability maps of wild birds (Section 2.2.1) and chicken density (Section 2.2.2), 2) simulating AIV suitability in various interfaces with a structural equation model (SEM) (Section 2.3), and 3) assessing the change in human population at risk in different interfaces (Section 2.4).

The interfaces and risk factors composing the DAG were selected based on evidence from the literature [1,2,11], a workshop with local experts from Bangladesh [12] and data availability to inform the model and parameterise the nodes. Crop density, chicken density and presence of transport infrastructure have been found to significantly explain variation in the number of high pathogenicity avian influenza (HPAI) H5N1 outbreaks across Bangladesh [5]. Thus, we considered spatial layers of land cover fractions [13], chicken density [14] and travel time to cities [15] as potential risk factors in the model framework. More information on input data used to parameterise the models is included in section 2.5.1.

The modelling framework allows for the incorporation of projections (see section 2.5.2) by integrating scenarios of bioclimatic variables and land cover as predictor variables in the species distribution sub-models of wild and farmed bird hosts. Climate and land cover projections under different socio-economic and emission scenarios can be used to assess how shifting conditions may alter the distribution of wild and farmed bird hosts, subsequently influencing potential exposure interfaces for AIV and population at risk. Bioclimatic variables are often used to explore shifts in habitat suitability for reservoir species, which in turn affects viral maintenance and spillover risks [16], and land-cover change can also significantly affect birds' habitats [17]. By integrating these climate and land cover variables as predictors in the DAG, the model enables scenario-based assessments of future AIV transmission risks under changing environmental conditions.

2.2. Modelling distribution of AIV hosts

2.2.1. Wild birds

The spatial distribution of natural reservoirs of AIV was simulated for a range of waterbird and crow species (hexagonal nodes in Fig. 1) using random forest (RF) modelling. Crow abundance in garbage dumping places and presence of migratory wild birds within villages have been associated with higher odds of H5 and H9 seropositivity in past studies in Bangladesh [18]. Thus, the distribution of these species was characterised as a potential risk factor in this study. To simulate suitable habitat distribution, climatic, land cover and environmental variables were used as predictors. Separate models were created for the spatial distribution of house crow (Corvus splendens) and for 64 different species of waterbirds within the Ardeidae, Anseriformes and Charadriiformes orders present in Bangladesh [19], considered as the most competent hosts for HPAI [9]. The resulting layers of spatial suitability of waterbird habitat were then combined by averaging the suitability, weighted by the total number of reported occurrences of each species in the country. This step resulted in two layers; 1) suitability distribution of house crow and 2) suitability distribution of waterbirds. These two layers were then used as risk factors in the next modelling step (section 2.3). Suitability units range from 0 to 100, with a suitability of 0 indicating no suitability and 100 indicating highest possible suitability of species occurrence.

For each RF model, we first obtained occurrence coordinates for a species within Bangladesh. To avoid the risk of spatial sampling biases and autocorrelation, each occurrence dataset was geographically thinned, ensuring a minimum distance of 1 km between occurrence points. Species with less than 20 occurrence points in the country after thinning were removed from the analysis following Benavides Rios et al. [20] recommendation to use at least 20 occurrence locations to fit species distribution models.

In the second step, we used Boruta algorithm for predictor selection. The algorithm is a robust feature selection method that identifies which of the initial set of 27 bioclimatic, land cover and environmental variables are deemed important to predict occurrence by comparing them with randomly shuffled duplicates, called shadow features [21]. It trains a RF classifier on both the real and shadow features and calculates feature importance scores. If a real feature consistently outperforms the shadow features across multiple iterations, it is then classified as important. We used a maximum number of 100 iterations and 1 % as a p-value threshold of significance to determine whether a feature is found to be important. The result of the algorithm is a classification of predictors as important, tentative or unimportant to inform feature selection. Only important or tentative predictors for a given species were retained for each RF model.

Since only information on occurrence of bird species was available for the RF models, pseudo-absence data were generated to fit models for each species. The same number of pseudo-absences was generated as the number of presences, as recommended for classification algorithms [22]. Each model was set up with ten sets of pseudo-absence, generated using the random method. We used a spatially stratified cross-validation approach to create 20 partitions of the cross-validation dataset, 10 along gradients of longitude and 10 along gradients of latitude to ensure transferability in geographic space [23]. Pseudo-absence points were not restricted to a specific area given the relatively small area and landscape homogeneity of the country.

The dataset was partitioned so that 75 % was kept for calibration while 25 % was set aside for validation. Each RF model was calibrated with 500 trees to grow, ten variables randomly sampled as candidates at each split to estimate variable importance and five minimum terminal nodes. This process generated up to 200 models for each species (10 pseudo-absences sets and 20 cross-validation partitions). We combined the models for individual species into a weighted-mean ensemble model, where the weight attributed to a single RF model is proportional to its true skill statistics (TSS) value. For each species, we used the five models with the highest TSS score to build the ensemble model to retain the best models while at the same time limit computation time requirement. The final evaluation metrics of the models composing each species' ensemble model included the area under the receiver's operating characteristic curve (AUC), the TSS and the Boyce index, which is a measure of presence-only based predictions [24].

2.2.2. Chicken distribution

Chicken density is a significant risk factor of AIV transmission [14]. Moreover, as boththe population and per capita annual consumption in Bangladesh are projected to rise, the total demand for poultry meat is expected to increase significantly over the next decades [25]. However, there are no projected spatial distribution of chicken currently available. The gridded livestock of the world (GLW) [14] and other gridded livestock maps [26] only attempt to represent the current animal population. To remediate to this gap, we created a new model of chicken distribution based on a comparable approach to the GLW and more recent gridded livestock products [27], bespoke to the context of Bangladesh.

We developed three models to predict chicken density across Bangladesh: 1) a generalized linear model (GLM), 2) RF, and 3) eXtreme Gradient Boosting (XGBoost) model. Since no observed bird numbers are available at a grid level, we first aggregated gridded predictors at district level to fit the models with the reported number of chickens in the 2019 Bangladesh Agriculture Census [28]. We then selected important features using the Boruta algorithm using a maximum of 500 iterations and a p-value threshold of 5 % to identify important features and trained the three different models. The GLM was tuned with five repetitions of 10-fold cross-validation, while the RF and XGBoost models were fitted using the leave-one-out cross-validation approach, which is preferable when working with small datasets [29]. We used the models to project chicken density on a 5-arcminute grid and scaled density so that the predicted total population at national level matched the total population in the Agriculture Census. To evaluate the models, we aggregated the predicted gridded chicken density at district level to compare the three models with the chicken density from the Agriculture Census and kept the best fitting model based on the root mean square error (RMSE), the coefficient of determination (R-squared) and the correlation coefficient. In addition, we compared these evaluation metrics to the chicken density map from the GLW aggregated at the district level. Having a model tailored to the Bangladesh context enabled the projection of chicken density under climate, land-cover and human population density change.

2.3. Modelling exposure suitability

We used a SEM to simulate the suitability of AIV exposure in different interfaces. This statistical method is well suited for modelling complex epidemiological relationships between risk factors as it can incorporate various latent variables which are not directly observed but inferred from multiple risk factors in one single model [30]. In this study, we modelled suitability in four AIV exposure interfaces including 1) waterbird nesting sites, 2) urban-dwelling birds, 3) poultry farms and 4) LBMs. In a SEM, each latent variable, in this case representing suitability in an interface displayed as circular nodes in Fig. 1, is estimated from a regression. For instance, the AIV exposure suitability in waterbird nesting sites is regressed on several risk factors, including the distribution of waterbirds and bioclimatic variables, among others. Each latent variable is also measured by an observed variable to inform the fitness of the model (parallelograms in Fig. 1). For instance, the spatial suitability of AIV exposure in poultry farms is measured by the geolocated reported occurrence of AIV cases in poultry farms. The layers of wild bird distributions obtained from the habitat suitability distribution models described in section 2.2.1 were used as a risk factors for the waterbird nesting interface in the SEM, along with other risk factors displayed in Fig. 1 as rectangular nodes. Similarly, projected chicken density distribution was considered as a risk factor for poultry farms exposure suitability. Predictors were not scaled before running the SEM model, so that absolute difference in predictor values could be accounted for in projections. However, the resulting exposure suitability units were scaled to range from 0 to 100 for reporting, whereby 0 indicates no suitability and 100 indicates the highest possible suitability of AIV exposure. The SEM was built in R using the lavaan package [31].

The structure of the SEM was co-designed with One Health experts from Bangladesh [12] and the initial model structure included a relatively large number of risk factors. The final risk factors were selected through a backward elimination approach whereby insignificant risk factors were removed iteratively until only statistically significant risk factors remained, i.e., with p-value lower than 0.05. Validation of the SEM was conducted using the comparative fit index (CFI), root mean square error of approximation (RMSEA) and Standardized Root Mean Square Residual (SRMR), three metrics often used to evaluate the performance of SEMs [32]. Recommended cutoff values to assess the goodness-of-fit of a model according to these indices are greater than 0.95, lower than 0.05, and lower than 0.08, respectively [33]. Although the model calibration was based solely on these three metrics, we also present the fitness of the model to simulate AIV suitability for each individual interface in comparison to reported occurrence using AUC.

Due to the lack of spatially-explicit reported human cases of AIV, the suitability in the human interface (blue circular node in Fig. 1) could not be simulated spatially with the same approach as for animal interfaces using the SEM in the current study. Instead, we assessed the area and population at risk of exposure to AIV.

2.4. Projected population at risk

We estimated the change in the population at risk by overlaying areas considered at risk of AIV exposure and the change in population density in these areas under the different scenarios. To do so, we first identified a threshold of suitability in all four interfaces to dichotomize model predictions and divide grid cells into unsuitable and suitable classes, in line with similar studies estimating population at risk of zoonotic diseases [34]. We used Youden's J statistics, that is the sum of sensitivity and specificity minus 1, to identify optimal thresholds in each interface based on the current predicted suitability and reported cases. We then masked cells with predicted suitability below the thresholds and calculated the population in areas where predicted suitability was found to be above the threshold. We repeated this step and calculated the area and population at risk for all SSP-GCM scenario combinations.

2.5. Data sources

2.5.1. Current predictions

The environmental predictors used in the RF models of wild bird distribution were terrestrial ecoregions, land cover fractions on grid cell, distance to wetlands and shorelines, elevation and net primary productivity (see Table S1 in Supplementary data). In addition, 19 bioclimatic variables were used as climatic predictors from the WorldClim database [35]. Occurrence data in Bangladesh for each bird species were obtained from the Global Biodiversity Information Facility (GBIF, https://www.gbif.org/) database to train and test each model. In addition to the aforementioned datasets, the chicken density distribution sub-model included population density [36] and travel time to the nearest city [15] in the initial set of predictors and the reported chicken density as the response variable [28].

Geolocated reported cases of AIV were obtained from the World Animal Health Information System (WOAH-WAHIS, https://wahis.woah.org). The 649 AIV occurrence points in this dataset were split into reported cases in domestic birds to inform the poultry farms interface (n = 642) and cases in wild birds (n = 7) to inform the urban-dwelling bird interface, since all cases in the latter dataset were in birds of the house crow (Corvus splendens) species. Occurrences of AIV in waterbirds (n = 8) were used to inform the suitability of exposure in waterbird nesting sites (Pers. Comm. Mohammad Mahmudul Hassan). Finally, a dataset of AIV cases in LBMs across the country (n = 74) was used to inform suitability of exposure in LBMs interface (Pers. Comm. Sukanta Chowdhury and Mohammad Mahmudul Hassan). A summary of input data and sources for all modelling steps is shown in Table S1. The measured variables of reported cases are considered as a binary variable in the model, i.e., presence or absence of AIV, and thus the results of the model indicate suitability rather than the incidence.

To estimate the population at risk, population density [36] was used in combination with the AIV suitability results of the SEM. Except for the district-level chicken distribution from the Bangladesh Agriculture Census, all spatial layers were resampled to 5 arcminutes (approximately 8.8 km in Bangladesh). This resolution constitutes a grid of 73 and 56 columns aligned with the extent of the country. Once the areas outside of the country's border were masked, approximately 2000 pixels were used as simulation units.

2.5.2. Future projections

To project the change of AIV exposure suitability under future climate projections, three combinations of shared socioeconomic pathways (SSPs) and representative concentration pathways (RCPs) where considered: SSP 1 with RCP 2.6, considered a sustainable pathway with low greenhouse gas (GHG) emissions, SSP 2 with RCP 4.5, defined as the “middle of the road” scenario with intermediate GHG emissions, and SSP 3 with RCP 7.0, characterised by meat-heavy diets and high GHG emissions [37]. In addition, we used bioclimatic variables from three downscaled general circulation models (GCM): 1) IPSL-CM6A-LR [38], 2) MIROC6 [39] and 3) ACCESS-CM2 [40], since these three GCMs had results available for all SSPs scenarios. These variables were generated from downscaled and bias-corrected monthly temperature and precipitation using a change factor approach [16]. Projected climate data were obtained at 5-arcminute resolution to align with and modify the dataset of predictor variables, which was then used as input in the models to project AIV suitability.

Land cover projections were obtained from results of the future land use simulation (FLUS) model, a global cellular automata land use model that provides land cover projections for the same three scenarios mentioned above [41]. We used population density projections for the same SSP-RCP combinations from Wang et al. [36] who projected future population density using a RF model trained on the WorldPop dataset [42].

For each scenario, we updated the predictor datasets for the wild bird suitability and chicken distribution with climate, land cover and population density projection layers. For the chicken density distribution sub-model, density was scaled so that the national density matched national projected density under the different scenarios. To obtain this information, we ran the Global Biosphere Management (GLOBIOM) model, a partial equilibrium model of agriculture and land use that projects future demand of agricultural commodities in different world regions based on changing population, per capita income, trade and preferences for different SSP-RCP scenarios [43]. We defined the scaling factor as the ratio between projected consumption in 2050 in the region that includes Bangladesh in the model (‘Rest of South Asia’) for each scenario to official consumption in 2020 [44]. Since imports and exports of live chickens and chicken meat is negligible in the country, i.e., only 1 % of chickens and less than 1 % of chicken meat are imported [44], we assumed that the proportion of domestic demand produced within the country would remain the same.

Once the predictor datasets were updated with these projections, the host distribution sub-models (section 2.2) and the SEM (section 2.3) were run again to evaluate the new suitability for a given SSP-RCP and GCM combination scenario in each AIV exposure interface and the population at risk (2.4) was re-assessed accordingly.

2.6. Hotspot analysis

In addition to gridded suitability results, we further analysed the change in suitability in different interfaces resulting from climate change through a hotspot lens. Using the projected suitability described in section 2.3, we first calculated the mean difference between projected and current suitability across all pixels within each upazila administrative region in the country, and across all SSP-RCP scenarios and all climate models in the four interfaces of transmission. We then computed the local spatial statistic Gi [45] of this difference in mean suitability across all scenarios for each upazila. The Gi values were discretised into an ordinal scale ranging from very cold to very hot for ease of interpretation, whereby very cold spots indicate upazilas with the most negative changes in suitability under climate change and very hot spots show upazilas with the highest increases in suitability.

2.7. Sensitivity analysis

Given the small number of reported cases available to fit the SEM model for the interfaces of waterbird nesting sites and urban-dwelling birds, we conducted sensitivity analysis to assess the impacts of limited availability of reported cases on the resulting coefficient values in the model. We used a leave-one-out (LOO) approach whereby the model is fitted with one reported case removed from the dataset, repeating this process iteratively for all cases. This means that the model was fitted 8 times with 7 different water bird cases and 7 times with 6 different urban bird cases. We then assessed the robustness of the model coefficients resulting from using different combinations of cases.

3. Results

3.1. Validation of models

On average, RF models retained to build the ensemble models of wild bird suitability predicted species occurrence well with mean AUC and TSS values of 0.91 and 0.74, respectively, while the Boyce index showed moderate ability of models to predict presences with a mean score of 0.58. Box plots of model fitness showing the ranges of AUC, TSS and Boyce index values for all bird species are shown in Fig. S1 in Supplementary data.

The best of the three models tested to simulate chicken distribution was the RF model, which yielded a RMSE of 8.5 heads per ha, a Pearson's correlation coefficient of 0.72 and a R² of 0.41. These metrics were superior to the two other models tested and provided a better fit against the official census data compared to the GLW, which resulted in a RMSE of 10.08 heads per ha. Table S2 in Supplementary data provides a comparison of goodness-of-fit measures for each model.

The SEM of avian influenza exposure suitability resulted in a CFI value of 0.996, a RMSEA value of 0.038 and a SRMR value of 0.03, all three indicating an excellent fit, well within the recommended cutoff values. For the accuracy of individual interfaces, the predicted exposure suitability in LBMs and waterbird nesting sites had the higher ability to discriminate between occurrence and absence of cases with excellent AUC values of 0.95 and 0.9, respectively (Fig. 2). The predicted suitability of exposure of AIV in poultry farms and urban birds interfaces had lower AUC values, 0.76 and 0.73 respectively, albeit considered as acceptable fit [46].

Fig. 2.

Results of simulated suitability of exposure in four interfaces under current climatic conditions. Validation metric (AUC) is indicated for each interface.

3.2. Variable and risk factor importance

The importance of each predictor in the wild bird species distribution models was aggregated to present a summary figure of variable importance across all models and species. Land use predictors, in particular the fraction of cropland, grassland and urban land cover on a grid cell, were identified as highly influential in explaining wild bird distribution suitability (Fig. S2 in Supplementary data). Urban fraction was considered an important predictor for a fewer number of species (56 out of 64), but the large variability and outliers suggest significant relevance for some bird species. Regarding climatic predictors, annual mean temperature, precipitation seasonality and mean diurnal range were found to be the most important on average. Annual mean temperature was also the most important predictor of chicken distribution, along with the travel time to the nearest city and population density (see Fig. S3 in Supplementary data).

The importance of each risk factor to explain each latent variable in a SEM can be evaluated using the path or regression coefficients. Precipitation seasonality was the most influential risk factor to determine exposure suitability in the waterbird nesting sites interface, followed by waterbird distribution (Fig. 3). Exposure suitability in this interface was the primary risk factor of poultry farms exposure, highlighting the significance of the natural AIV reservoir in affecting exposure in domestic poultry farms. The chicken distribution had the second highest coefficient to explain exposure suitability in poultry farms. The latent variables for urban birds and LBM exposure suitability were predominantly explained by the fraction of urban areas on a grid cell. The distribution of crows was initially included as a risk factor for the urban bird interface but was removed during the backward risk factor elimination process due to its negligible and insignificant effect on exposure suitability in this interface. All risk factors were significant with a p-value below 0.05. Coefficients, standard errors, p-value and standardized coefficients, showing relative importance of each risk factors for a given interface, are summarised in Table S3 in Supplementary data.

Fig. 3.

Graphical representation of the SEM including path coefficients. Risk factors are shown as rectangles, sub-models of host distribution as hexagons and latent variables, in this case exposure interfaces, are shown as circular nodes. Blue and red edges indicate positive and negative relationships, respectively. Asterisks represent significance level; *** p < 0.001, ** p < 0.01, * p < 0.05. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Most coefficients were robust to the LOO sensitivity test. The coefficients of water bird distribution and distance to shorelines risk factors for the nesting birds' interface were the most sensitive to the resampling of water bird cases, with the relative importance of coefficients ranging from 0.4 to 0.54 and 0.44 to 0.53, respectively. The sensitivity of all coefficients of the SEM to water and urban bird AIV cases resampling is displayed in Supplementary data, Fig. S4.

3.3. Projected bird species distribution

The impact of climate change on different bird species varied significantly across species, SSP scenarios, and particularly across climate models. On average the range of species tend to be reduced by climate change, particularly under MIROC6 GCM scenario, which predicts a reduction in ranges of 87.5 % on average across all species and all three SSP scenarios. However, the range of several species, particularly Ardeola grayii and Xenus cinereus, two species that were found to be seropositive for subtypes H5N1 and H9N2 in Southeast and East Asia region in the past^58,59, is expected to increase across all scenarios. The largest predicted range increase was for Ardea cinerea, with suitable area increasing by up to 325 % to 110,586 km² under SSP3-ACCESS-CM2 scenario combination (see Fig. S5 in Supplementary data for the individual range shifts for all bird species under all climate and SSP scenarios and section S4 for maps of change in wild bird suitability).

3.4. Projected suitability of AIV exposure in animal interfaces

3.4.1. Current suitability

The suitability of exposure in waterbird nesting sites under current climatic conditions was the highest around the northeast and the southeast parts of the country, and, to a lesser extent, around the capital city, Dhaka (Fig. 2). Suitability of exposure in poultry farms was high or very high in several parts of country with high population density and around urbanised areas. The suitability of exposure in urban birds and LBMs interfaces was highly concentrated around the largest cities, with high suitability in densely-population areas of Dhaka and Chattogram.

3.4.2. Projected suitability per district

Changes in suitability in the waterbird nesting sites interface show that districts in the north of the country, such as Mymensingh, Netrakona, Nilphamari and Joypurhat could be some of the most affected districts (Fig. 4A) by 2050. SSP scenario 3–7.0 generally had a greater impact on the change in suitability, however SSP scenarios do not significantly change the order of affected districts for this interface. Conversely, the suitability could vary substantially depending on the GCM used in the scenario (line ranges in Fig. 4). Averaged across Bangladesh, the change in AIV exposure suitability is expected to increase between 8.62 (SSP 3–7.0 and GCM IPSL-CM6A-LR) and 18.3 (SSP 3–7.0 and GCM ACCESS-CM2) in waterbird nesting sites depending on the SSP and GCM scenarios.

Fig. 4.

Change in suitability of AIV exposure under climate change in 2050 compared to current climate in A) waterbird nesting sites interface and B) poultry farms by district. The 10 districts with highest and lowest changes in suitability are shown, in addition to the average change in suitability for Bangladesh. Boxplots show the range of values for all pixels within a district and for the three GCMs selected: ACCESS-CM2, IPSL-CM6A-LR and MIROC6.

Suitability in the poultry farms interface is expected to rise more significantly. For this interface, GCMs have negligible effects but SSP scenarios have a large influence, particularly SSP3, which drives the increase in poultry demand at country level and therefore chicken density (Fig. 4B). The mean suitability in the poultry farms interface across the country is projected to increase by 57 for SSP1–2.6 to 100 under SSP 3–7.0 (Table S4 in Supplementary data). The spatial distribution of the change in suitability in the four interfaces for each SSP scenario and each GCM can be found in Supplementary data section Fig. S7 to Fig. S10. Mean suitability in urban bird and LBMs interfaces is expected to increase more moderately at the country level given the concentration of suitable areas in densely populated cities (Table S4 in Supplementary data). To account for the increase in suitability in densely populated areas, we estimate the population at risk in the next section.

3.5. Projected population at risk

The total area considered suitable for AIV exposure across the country, and the population living in these areas, are expected to increase across all interfaces. The area and number of people at risk of exposure in waterbird nesting sites vary depending on the GCM used to simulate wild bird habitat suitability. For instance, under SSP scenario 2–4.5, between 29,500 and 76,400 km² could be considered suitable for AIV exposure in the waterbird nesting sites, compared to 20,300 km² estimated under current conditions (Fig. 5). The large increase in suitability for AIV exposure in the poultry farms, combined with the projected increase in population density, is expected to substantially increase both the area and population at risk of exposure. Overall, the population living in areas considered suitable for AIV exposure in at least one of the four interfaces could increase by up to 79 million, from 107 million to 186 million by 2050 under SSP3–7.0 scenario.

Fig. 5.

Area and population at risk under current and projected AIV suitability. Error bars show variability across GCM scenarios.

3.6. Hotspot analysis of climate change impacts

The hotspot analysis in waterbird nesting sites shows large concentrations of highly suitable upazilas in the Rangpur and Mymensingh divisions in the north of the country (Fig. 6). The divisions of Dhaka and Rajshahi both show hot and cold spots. Poultry farming hotspots are concentrated around largest cities, particularly Dhaka, Rangpur, Rajshahi and Chittagong, with cold spots in the less densely populated upazilas in the south and southeast parts of the countries. Upazilas considered very hot spots in the two interfaces include Gangachara and Kaunia in the Rangpur division, Roypura in the Dhaka division, and Nabinagar in the Chittagong division. Hotspots maps for all interfaces along with maps of p-values for hotspot significance can be found in Supplementary data, Fig. S11 to Fig. S14.

Fig. 6.

Hot and cold spots at upazila level of change in suitability under climate change in 2050 compared to current climatic conditions in nesting bird and poultry farming interfaces.

4. Discussion

4.1. Spatial heterogeneity of projected suitability

The suitability of exposure to AIV in Bangladesh is expected to change differently across the country and in different interfaces based on the risk factor importance. Suitability of exposure in waterbird nesting sites is expected to change primarily according to shifts in wild bird distribution, driven partly by changes in precipitation seasonality in the north and southeast part of the country. The suitability of exposure in poultry farming, urban bird and LBM interfaces was shown to increase more significantly around the large cities of the country, driven by an expected growing chicken density and urban expansion. Although changes in suitability in poultry farms and waterbird nesting sites overall show different spatial patterns, some areas were considered hotspots in the two interfaces, which could highlight risk areas of spillovers from wildlife to farmed animals.

4.2. Potential improvements in the model structure

We presented in the paper a novel approach to model the impacts of climate and land-cover change on AIV exposure across multiple host types including wild birds, animal production sector and a retail interface. The modelling framework could be further improved by incorporating trade networks and transport of live poultry and wild birds between farms and markets. Tracing the origin location of animals sold in LBMs [47], and the associated AIV transmission suitability could improve the representation of the causal links, particularly between poultry farms and LBM interfaces. Furthermore, including information on biosecurity measures in poultry farms [48], vaccination rate [5] and trading practices in LBMs [49] would likely improve the predictive ability of the model and allow to simulate the impacts of interventions on AIV suitability.

The model in this study was fitted without an explicit interannual temporal dimension. Further work should focus on improving the understanding of seasonality impacts on residence time of migratory birds, virus evolution and environmental persistence [9] in waterbird nesting sites. Seasonality can also influence the prevalence of avian influenza, particularly in LBMs [3,4]. Although the bioclimatic variables used in this study provide indicators of seasonality, explicitly considering climatic variables at monthly intervals could reveal distinct causal links across seasons. This approach could enable decision-makers to not only target interventions on areas of the country and interfaces of transmission, but also specific periods of the year.

4.3. Data availability improvement

A main source of uncertainty in this study was the limited data availability on reported cases of AIV. Although many occurrences of AIV outbreaks in poultry farms were available from the WAOH-WAHIS database, the data availability on reported cases in wild birds was much more limited. This highlights the need for accrued surveillance and improved environmental detection among migratory waterbird nesting sites to gain a better appraisal of AIV prevalence in this interface, particularly in areas at higher risk with greater bird density. Although the sensitivity analysis showed the SEM coefficients were robust to the resampling of cases, a larger number of geolocated data points on seropositive wild birds would improve the calibration of the model. The framework presented here is flexible and the reported cases could be updated dynamically over time to improve the goodness of fit of the model as new input data become available.

Another important source of uncertainty and variability in results, particularly for suitability in waterbird nesting sites, was the GCM used for climate projections. The WorldClim projections have been used extensively to project the impacts of climate change on species distribution and are known to be globally downscaled and bias-corrected through a factor change approach [16]. However, biases from the dataset have been observed in some regions of the world, for instance, inadequately representing mean annual precipitation [50]. Using regional bias-corrected climate models or a weighted ensemble model would help to reduce uncertainty associated with climate projections. To our knowledge, the only bias-corrected climate projections for South Asia are only available at a lower spatial resolution than the one used in this study [51], presenting a trade-offs between spatial resolution and quality of climate projections. Here we chose to use WorldClim 2.1 bioclimatic factor dataset at a higher spatial resolution and showed the results of three different GCMs to account for uncertainty range.

The water bird distribution models showed excellent discrimination capacity through a high average AUC value but more moderate TSS, and especially Boyce index value. Given the relatively small extent of the country, spatial thinning of the occurrence dataset could significantly reduce the number of presence points available to train the models. To overcome this issue, the models could be trained on a wider geographical area, for instance the South Asia region, to train the models with larger sets of occurrence and predictors. Furthermore, ensemble models could be built with individual models that maximise not only TSS, as it was the case in this study, but also the Boyce index and AUC. Furthermore, the predictive performance for individual interfaces based on the AUC value varied significantly, ranging from 0.73 for the urban bird interface to 0.95 for the LBMs. Although most urban bird cases were reported in C. splendens, the spatial suitability layer of C. splendens was found to be insignificant to explain AIV suitability in the urban bird interface. This could be due to the low number of cases of urban bird and missing important risk factors for this interface, which could also partly explain the moderate AUC value for this interface. Conversely, high AUC value for LBMs could be attributable to the concentration of LBMs in urbanised areas, which is easier to discriminate in the model.

4.4. Relevance for policy making

By identifying hotspot areas of increasing transmission suitability for AIV, this modelling framework can be leveraged to target efficient interventions in the most vulnerable areas of the country. This analysis enables users to identify not only specific upazilas with higher risk, but also concentrations of upazilas that may require more attention in coming years to prevent future outbreaks. Furthermore, by differentiating different interfaces of AIV transmission, the modelling framework can enhance the relevance of the results to specific sectors, such as wildlife and environmental health decision-makers for waterbird nesting interfaces and animal and human health decision-makers for poultry farm and LBM interfaces, and population at risk.

5. Conclusions

In this study, we presented a novel methodological approach to link 1) projections of bioclimatic variables, land cover and population density, 2) distribution models of wild birds and chicken distribution, 3) an epidemiological model of AIV exposure across multiple wildlife and poultry production and retail interfaces, and 4) an assessment of the population at risk of exposure. Using this integrated model, we identified key variables to predict the distribution of host species, specifically land cover and bioclimatic variables such as annual mean temperature and precipitation seasonality. We also determined the most important risk factors driving AIV exposure suitability across different interfaces. For example, precipitation seasonality and waterbird distribution can strongly influenced suitability at waterbird nesting sites, while this exposure suitability was the main risk factor for poultry farming suitability. We found that precipitation indicators played a crucial role in predicting suitability at waterbird nesting sites while suitability in other interfaces were not directly influenced by climate variables. Rather, projected growth in poultry density was a key factor driving the suitability in poultry farming interfaces. This, in turn, substantially increased the population at risk, particularly under SSP scenario 3–7.0. Additionally, the approach allowed us to predict hotspots of changes in AIV exposure, highlighting greater AIV suitability in the northern part of the country and around Dhaka in the future.

This One Health framework provides decision-makers with a valuable tool for prioritising interventions in high-risk areas to mitigate future AIV outbreak risks. It can be further adapted to simulate exposure suitability in human interfaces directly within the epidemiological model and implemented for any other infectious diseases under climate and land-use change in different world regions, subject to geolocated disease case data availability.

Funding

This research was funded by the World Bank Group (TF0B5942).

CRediT authorship contribution statement

Adam C. Castonguay: Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Sukanta Chowdhury: Writing – review & editing, Resources, Conceptualization. Ireen Sultana Shanta: Writing – review & editing, Conceptualization. Bente Schrijver: Writing – review & editing, Project administration, Conceptualization. Remco Schrijver: Writing – review & editing, Funding acquisition, Conceptualization. Mohammad Mahmudul Hassan: Resources. Shiyong Wang: Writing – review & editing, Supervision, Funding acquisition, Conceptualization. Ricardo J. Soares Magalhães: Writing – review & editing, Supervision, Project administration, Methodology, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Remco Schrijver and Bente Schrijver are employed by VetEffecT. This affiliation did not influence the study design, analysis, interpretation, or conclusions. All other authors declare no competing interests. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Declaration of generative AI and AI assisted technologies in the writing process.

During the preparation of this work, the author used ChatGPT (developed by OpenAI) to check for spelling and grammar errors. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the content of the published article.

Appendix A. Supplementary data

Supplementary material

Data availability

Data and R scripts will be made available on request.