Selection of species for a cohort model

Our model of the relative suitability of habitat for mosquito vectors was based on occurrence data for selected species in the genus Culex with some demonstrated potential for a role in JEV transmission namely: Cx. annulirostris, Cx. bitaeniorhynchus, Cx. gelidus, Cx. palpalis, Cx. quinquefasciatus, Cx. sitiens, and Cx. tritaeniorhynchus [22,28]. There is uncertainty about the vectorial capacity of these species in different Australian climate zones (see map S1 Fig), plus potential for taxonomic uncertainty (viz. Cx. tritaeniorhynchus, Cx. sitiens subgroup), so distinguishing between the distributions of these species seemed unlikely to contribute to our aim of national scale mapping of habitat suitability for JEV transmission. Therefore, occurrence data for these species (see S2 Fig) were combined to produce a cohort model; our mosquito vector habitat suitability model predicts areas where conditions are likely to be suitable for the presence of any one of these species. However, due to published weight of opinion about the primacy of Cx. annulirostris as a vector in Australia [15,22,28,48], we also developed a separate model restricting presences to Cx. annulirostris, using the same method as described below. The results are included in the Supplementary information. Given the broad similarity between the single-species model and the Culex cohort model, we focus our presentation on the cohort vector model in defining the potential habitat suitability for JEV transmission in Australia.

We also adopted a species cohort model approach for wildlife hosts. The target group for constructing a model of habitat suitability was narrowed to herons and egrets (Family Ardeidae): Ardea alba (great egret), A. intermedia (intermediate egret), A. pacifica (white necked heron), Bubulcus ibis (cattle egret), Egretta garzetta (little egret), and Nycticorax caledonicus (Nankeen night-heron) (see S3 Fig). These species were identified on the basis of published information and expert judgements about their: habitat associations; breeding ecology and movement patterns (e.g., [49]); viraemia; and evolutionary similarity to species implicated in JEV elsewhere in Asia through antibody prevalence or infection competence (e.g., [13]).

Data for mosquito vector species

All data preparation, model development and implementation described below was performed in R v4.2.2 [50]. After filtering and cleaning we retained 15876 occurrence records (unique sampling locations and months) for one or more of the selected Culex species of interest. The majority of the records (98%) were mosquito surveillance data obtained from Australian state and territory health authorities (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of vector occurrence datasets for the period 2000–June 2023 obtained to support habitat suitability modelling. Trap events refer to unique combinations of location (1 km raster grid cell) and year and month of sampling.

https://doi.org/10.1371/journal.pntd.0014127.t001

Mosquito surveillance datasets

Arbovirus surveillance programs, including mosquito surveillance activities, are implemented independently by each state jurisdiction [35,66] with varying degrees of devolved or centralised responsibility for data management, specimen identification, and reporting. While we were able to collate mosquito trap data for each jurisdiction, the resulting dataset does not represent a comprehensive picture of Australian mosquito surveillance over space and time. We obtained records for between 1–22 years depending on the jurisdiction (Table 1), and we know for example that records from commercial premises were removed from some jurisdictions’ data. Records were typically generated from carbon dioxide-baited encephalitis virus surveillance (EVS) or Centers for Disease Control (CDC) light traps, or similar. Some trap locations were represented by a single trap event but others had been sampled regularly for many years resulting in hundreds of sampling events at the same location. Trapped mosquitoes had been identified and counted by trained entomologists. We discarded records for traps that were marked as having malfunctioned, samples that had not been resolved to species level or where species identification was not conducted, and records that could not be assigned a location with the required spatial precision of ≈1 km. For jurisdictions where an abundance value for individual species was provided, records with 0 abundance for all species of interest were categorised as a non-detection and added to the background pseudo-absence data.

Public repositories and other published datasets

We augmented the data from jurisdictions’ surveillance with occurrence data for the same period published in literature or in publicly accessible databases (Table 1). Our literature extraction effort was not uniform, focussing on remote areas in the north-eastern jurisdiction of Queensland where routine trapping is difficult and sampling was usually undertaken in a single location at a single time-point (e.g., [53,55]). Those published records were all from CO₂-baited light traps, except Sebayang et al, who used CO₂- and octenol-baited POD traps. Additionally, occurrence data for mosquito species in Australia from the observation protocols ‘human observation’, ‘machine observation’, and ‘pickled specimen’ were obtained from the Atlas of Living Australia (ALA) [51] and Global Biodiversity Information Facility (GBIF) [52] (Table 1), derived from miscellaneous sampling methods. GBIF records were filtered to exclude those with geographic uncertainty of ≤ 1 km² or flagged as having potentially erroneous or unreliable spatial location attributes. ALA records were filtered to include only those flagged as appropriate for Collaborative Species Distribution Modelling. These databases include observations and data from citizens and professional scientists; subsuming the records of casual observations from iNaturalist, returns from licensed scientific collection, and other specific projects and databases auspiced by biodiversity and ecological monitoring programs of Australian states and territories.

Data for ardeid host species

Bird observation checklists were obtained from eBird [67] and Birdlife Australia [68] for the period January 2000–June 2023. Checklist submissions to both datasets are organised using unique identifiers attributable to an observer at a given location at a certain time (though from Birdlife Australia we could only access the observer identifier at the level of the data custodian). Contributors can mark checklists they submit as “complete”, warranting that they recorded all and only species that were seen or heard during an observation event. The observation events are defined by metadata tags indicating the observation protocol (e.g., [69]) and search effort (time, distance, etc). “Complete” checklists with suitable search effort attributes can therefore be used to model data as presences (detected) and absences (zero-filled records of non-detections) (e.g., [70]). We used the locations of complete checklists with detections of any of the selected ardeid species (presences) and the locations of all other complete checklists (absences) to generate the habitat suitability model for ardeid hosts (see Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of heron and egret occurrence data from selected field observation protocols obtained from Birdlife Australia (BA) and eBird (eB) for the period 2000–June 2023. The counts are unique, complete checklists (“lists”) recording detection of the target species (“present”), any one of the target species (“Any ardeid”), or asserting absence of all. Overall empirical prevalence (“prev.”) is also reported. To form the model dataset, for any cell in any month with records from across the included protocols, a single observation representing presence or absence was randomly drawn in proportion to prevalence per cell per month.

https://doi.org/10.1371/journal.pntd.0014127.t002

On examination of the data, it was noted that some individual observers had submitted many checklists for a given location within a short time-interval. Other observers may also have independently or knowingly submitted checklists for a given location in the same period. Therefore, the raw dataset contains many observations that are pseudo-replicates for our purposes. Spatial and temporal thinning of records from species occurrence databases has been recommended to ameliorate the biases that characterise such unstructured data [70,71]. However, these characteristics of the data were common to presence and absence checklists and therefore do not constitute an issue of imbalance in spatial sampling bias so much as spatio-temporal pseudo replication with respect to the resolution of our covariates (1 km grid cell with monthly averages). For a given 1 km grid cell, per month, we randomly selected a single observation of presence or absence in proportion to the empirical prevalence from observations made during daylight hours in that cell and month. We combined data from several observation protocols as equivalent on the basis of similar empirical prevalence (see Table 2) and overlap in definitions; in each case the submitter warrants that the primary purpose of activity generating the record was bird observation, and supplies the start time, duration, location and distance/area covered [72,73]. A “500m area search” is a Birdlife Australia protocol defined as a search with a maximum displacement of 500 m from the recorded starting location. A “stationary” search is eBird’s equivalent protocol, but with a 30 m limit on displacement. eBird’s “travelling” protocol is for activities with greater than 30 m displacement, though the trajectory and distance must be recorded. We excluded records where a search area was likely to have encompassed more than 1 km² in extent. The “2ha-20min search” is an Australian research protocol found in both databases that imposes a fixed effort in space and time. Preliminary analyses indicated that there was a reduced prevalence of the target species associated with the 2 hectare-20 minute protocol. We could not accommodate an effect of protocol in our geospatial model, so records obtained via that protocol were omitted from the model dataset.

‘Static’ land cover variables

We constructed 1 km grids summarising the fractional cover of eight ‘static’ land cover types denoting vegetation structure and with special emphasis on water features. Culex mosquitoes and ardeids are strongly associated with the availability of still or slow-moving surface water such as naturally provided by wetlands, estuaries, tidal flats and billabongs created by receding flow through anabranching watercourses. These associations can be complex; for example habitat suitability for some ardeids in the Northern Territory seems higher in the drying phase of the tropical wet [74]. Expansive artificial habitats can be important too, as seen in the increase of Culex habitat due to irrigated agriculture in Asia [75]. This process has its analogue in parts of Australia’s interior where cotton and rice farming has expanded, and other irrigation schemes with similar water demands [76]. Culex species, including C. annulirostris can also reach high abundance in freshwater habitats in urban and peri-urban contexts particularly those subject to eutrophication (e.g., [77]), whilst small artificial habitats such as containers can favour abundance of species like Cx. quinquefasciatus [78]. Though productive habitats can exist within urban and peri-urban areas we omitted the ‘built up’ land cover type from the model covariates since that category conflated urban areas with, for example, sealed roads and other surfaces in remote areas. To further expand on the representation of water features in the environment we included a layer summarising the long-term relative frequency of water observation for each pixel for the period 1987–2021 which captures areas in semi-arid zones with low precipitation that may periodically become and remain inundated under certain circumstances (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Metadata for spatial covariates.

https://doi.org/10.1371/journal.pntd.0014127.t003

Temporally dynamic environmental variables

Each occurrence observation was assigned the relevant mean day- and night-time land surface temperatures; indicators of greenness of vegetation, soil moisture and soil texture; and interpolated total rainfall pertaining to the month of the observation, and the prior month. These time-varying datasets allow relative habitat suitability to be estimated as variable at a given location through time, as a function of changing resource availability, phenomena known to influence occurrence, abundance, and breeding success of ardeid hosts and mosquitoes (e.g., [86,87]). Global correlation among the covariates is presented in S4 Fig and the spatial pattern is shown for selected covariates in S5 Fig.

Strategy to minimise sampling and reporting bias

The sample or observer bias inherent in occurrence datasets used in habitat suitability modelling pose problems for their intended use [88], though some advances can mitigate these [89,90]. Since our objective was to produce continental scale models of relative host and vector habitat suitability for mainland Australia and nearby islands, the optimal datasets would result from random survey for our species of interest (mosquito trap deployment or bird observation protocol enacted), stratified across covariate space – in our case that implies repetition across months and over the years – with both presence and absence outcomes being recorded with equal probability. Such statistically ideal datasets do not exist and they would be expensive to generate. Both ardeid and mosquito datasets were spatially biased, but due to the different characteristics of the datasets, bias mitigation required different approaches.

The mosquito surveillance trap networks that generated around 98% of our vector dataset are operated or commissioned by Australian State and Territory health authorities. They were originally designed to detect pathogens of zoonoses of public health concern, primarily Murray Valley encephalitis virus (MVEV), Ross River virus and Barmah Forest virus. These trap locations are spatially biased in the sense that the majority were purposefully deployed close to likely mosquito breeding habitat (wetlands, drainage features, water holding ponds) near resident human populations of town and peri-urban areas, or other areas deemed to have elevated exposure risk. A consequence of this understandable bias is that the great majority of sampling locations were constrained to a subset of habitat types for any given latitude, though coastal and inland conditions represent considerable amplitude in their own right, e.g., [91]. Although these trap datasets included events where none of the species of interest were recorded (‘absences’), we opted to combine the trap detections with other occurrence records and treat the dataset as occurrence (presence-only) data.

There are three main options for mitigating spatial bias in occurrence (presence-only) data. 1) Include a covariate that offers a good representation of the anticipated bias in the model training, and replace that covariate in predictions with a fixed value to standardise the bias over the prediction area, e.g., [89,92]. 2) Construct a statistical model of the observation process that is presumed to generate the bias, e.g., [93]. These options complement the use of randomly selected background points, though the latter also requires at least some true presence-absence data to inform the observation model. A third option, 3) replaces random background points with target-group background points whose own generative process is presumed to have a similar bias as the occurrences, e.g., [90]. The target-group background, in reflecting the spatial (and environmental) bias present in the occurrence data, enables a prediction that is less biased towards areas over-represented in the occurrence set. We adopted the target-group background approach and fitted a model that contrasts the locations of mosquito occurrences against the locations of a target-group background dataset.

To mitigate spatial reporting bias with the target-group method we required a target group for whom observations would have been generated via a similar sampling methodology, but not sharing all similar environmental habitat requirements with the model target, otherwise the model will fail to capture those shared environmental characteristics in predictions of the species of interest. For example, if we had selected records for Culicidae, or even Insecta as the target group, it may have removed the effects of temperature and humidity from the model (since these are common environmental requirements for insect life). The resulting predictions would be biased away from these areas. Instead we obtained species occurrence records in Kingdom Animalia from GBIF [94] to use as our target group. Animalia occupy all habitats but the pattern of observation and reporting of those organisms is expected to be biased toward population centres and easily accessible areas [95,96]. By selecting records of Animalia, we therefore aimed to capture general patterns of bias in population density, accessibility and biological data recording in Australia. We arbitrarily limited the size of the Animalia background dataset to the same size as the mosquito occurrence set, and supplemented those with the mosquito trap records with zero abundance for the species of interest.

Our mosquito occurrence dataset also reflected temporal bias. Compared with the Animalia target group dataset, which suggests a steady accumulation rate through time, a quarter of the vector surveillance records we were able to compile date from after 2020 (see S6 Fig). This may be due in part to trapping having increased through time, particularly in response to new outbreaks, but primarily reflects that the data supplied to us by some state and territory authorities was limited to the period from 2021 onwards. We addressed this bias at the level of state and territory by sampling target-group background records only in years for which we had occurrence data for that state or territory.

Because we interpreted the curated datasets from eBird and BirdLife Australia’s as equivalent to presence-absence survey data, the treatment of spatial and temporal bias was simplified. In effect, we assumed that observers were equally as likely to record a ‘complete checklist’ in sites where our ardeid cohort occurred as in other locations, over space and through time. Consequently, whatever spatial reporting bias that applied to the presences would have been shared by the absences. Therefore, other than thinning of records described earlier, we modelled the distributions of the ardeid cohort without further accounting for spatial bias.

Modelling approach

A combined model of potential JEV vectors and vertebrate hosts

BRT and maxnet algorithms have both been shown to yield similarly accurate predictions for common datasets (e.g., [92,99]), but our host and vector habitat suitability models return metrics of suitability that are not directly comparable. The presence-background modelling of Culex species estimates relative abundance of species in the cohort (after accounting for spatial bias in reporting), whilst the presence/absence modelling of Ardeids provides an estimate of the probability of species occurrence. Both vector and host suitability metrics would be expected to be non-linearly, positively related to suitability for the pathogen, which depends on the abundance of both vectors and hosts. Therefore, using predictions for February 2022– the month before the declaration of a Communicable Disease Incident of National Significance when detections were simultaneously emerging across eastern Australian jurisdictions [108], we produced a combined niche model after rescaling each output by association with publicly available data on confirmed JEV detection locations [3,109]. We first defined an empirical cumulative distribution function for hosts and vectors, respectively, based on the extracted point suitability values (for the prediction to February 2022) coinciding with the location of each confirmed JEV detection in humans or commercial pig premises. The resulting values have a common interpretation related to the proportion of detected JEV infections associated with that value or lower (e.g., a rescaled suitability value of 0.75 indicates that 75% of JEV detections were associated with the corresponding level of habitat suitability or lower, and conversely, 25% of the detections were associated with that suitability value or higher). These cumulative distribution functions for vectors and for hosts impose a common conceptual mapping between the arbitrary vector/host suitability indices and the vector/host suitability for JEV transmission (specifically the intensity of detections). We combined them into a composite suitability layer based on the minimum value of the rescaled host and vector predictors for each pixel, with the result being an estimate of the joint distribution of suitable habitat for both vectors and vertebrate hosts.

Relative habitat suitability for vectors

Mosquito trap and occurrence data for the target Culex mosquito species were obtained from Australia’s northern extreme in the Torres Strait Islands down to Flinders Island and Tasmania south of the mainland (Table 1, Figs 1A and S1 Fig). The surveillance trap locations that comprised 98% of our occurrence dataset were biased toward populated areas and major cities (see also S8 Fig), and in water-intensive agricultural areas, with relatively little penetration into the semi-arid and arid interior. The Animalia target-group occurrence (Fig 1B) also reflect population centres, but also tracked transportation corridors and seasonally popular tourism areas. The distribution of travel-time accessibility values for our Animalia target group background and the Culex surveillance records were comparable, and distinctly less biased toward major population centres than the Culex records obtained from public databases (see S8 Fig). Compared with the target background, there were relatively higher densities of mosquito occurrence data near remote populations in the north and north-west of the mainland, reflecting targeted mosquito surveillance priorities. None of the Culex cohort have ever been recorded south of the mainland, including from traps recently established in Tasmania and Flinders Island in 2023 (Fig 1A). The model of habitat suitability for selected Culex species was built on 15539 vector occurrences and 16239 background points, including trap absences (Fig 1A and 1B). Cx annulirostris was the most commonly detected species, present in 90% of all cohort occurrences in the model training data. The next most common was Cx quinquefasciatus, which was detected in 47% of all cohort occurrences in the training data, including 85% of the occurrences where Cx annulirostris was not detected.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Overview of input occurrence data and penalised logistic regression model of relative habitat suitability for selected Culex species in Australia.

A) the distribution of the combined occurrence data for Culex spp; B) the distribution of background occurrence data for Kingdom Animalia; C) the relative ‘importance’ of top seven covariates as calculated by permutation on ensemble fit; D) prediction of relative habitat suitability to the outbreak period (February 2022); E–G) the shape of the marginal relationship between relative habitat suitability and the three most influential predictor variables; H–J) prediction of relative habitat suitability from December 2021 and June 2022, and standard deviation of predictions across the temporal sequence from July 2021 to June 2023. Base maps for A, B and H–J obtained from Australian Bureau of Statistics (CC BY 4.0) https://www.abs.gov.au/statistics/standards/australian-statistical-geography-standard-asgs-edition-3/jul2021-jun2026/access-and-downloads/digital-boundary-files.

https://doi.org/10.1371/journal.pntd.0014127.g001

The variable most strongly associated with the pattern of habitat suitability for vectors was night-time land surface temperatures (Fig 1C). The interval of 5–15 C marked a steep increase in predicted suitability (Fig 1E). Higher grassland cover and lower-to-moderate forest and woodland cover and were associated with higher habitat suitability, though were influential over a much smaller range in the response scale (Fig 1F and 1G respectively).

The resulting spatial expression of predictions from the habitat suitability model for the selected Culex species was high suitability over large swathes of mainland Australia during the austral Summer of 2021/22 (December–February, Fig 1D and 1H, and see the full sequence of monthly predictions (S1 Animation in S1 File). In December 2021 high habitat suitability was predicted to extend over much of Eastern Australia except areas of high altitude and the southern island of Tasmania. Very high suitability was predicted inland of the Great Dividing Range through sub-tropical and temperate latitudes, with much more localised areas of high suitability along eastern and south western coastal margins. The pattern for February of 2022 at the peak of the outbreak was more granular in comparison. Areas of very high suitability still mapped to the area inland of the Great Dividing Range but they were less expansive than the preceding December, tracing major river systems and wetland complexes. Higher suitability was predicted for the expansive monsoon-fed savannas in the Gulf Country, and across toward the Kimberley region in tropical northwestern Australia in a broad band. Equally high habitat suitability, though less expansive in area, was predicted for long stretches of the coast across Australia’s north. Extensive woodland and forest areas of tropical Queensland and north-eastern areas of the Northern Territory, which are at least seasonally wet, were predicted to have relatively low suitability. With the notable exceptions of the inland river systems in Australia’s southeast, and Greater Western Sydney, the zones of high predicted habitat suitability have relatively low human population density. While some areas of highest suitability corresponded with well-sampled arbovirus surveillance locations, for others, such as the “Channel Country” around the intersection of SA, NT, QLD and NSW, there were very few Culex occurrence records in our dataset. Areas of higher elevation or extreme aridity stand out as unfavourable even during the outbreak period. The results of our Culex cohort model and those for Cx. annulirostris alone showed high congruence (compare Figs 1 and S9).

By winter (June 2022) suitability was restricted to areas at the tropical and equatorial latitudes (Fig 1H and 1I) due to the strong influence of night-time temperatures. Variability in predictions across the full temporal sequence (S1 Animation in S1 File) was most pronounced at sub-tropical and temperate latitudes (Fig 1J).

Post-hoc evaluation of temporal pattern in habitat suitability predictions to mosquito occurrence locations (S10 Fig) shows that the seasonality in predictions corresponds well with gross pattern in Culex available species abundance data. The correspondence was better in temperate and sub-tropical latitudes than in the tropical north. The vector habitat suitability model had an average AUROC on validation folds of 0.84 (range 0.77–0.89), indicating excellent predictive performance, and mean continuous Boyce index of 0.99 (range 0.98–1). The omission rate for the minimum training presence was 0 in all but one training fold, and for the 10th percentile of training presences the omission rate of the model was 0.14 on average (sd 0.08, range 0.08–0.27).

Relative habitat suitability for hosts

Terrestrial occurrence records for the selected ardeid host species were concentrated in accessible and highly populated areas in South and Eastern Australia (Fig 2A and 2B) though at least some presence and absence records were available over the majority of the country. Presence records were abundant in the regions around Australian population centres and along the accessible coastline, and sporadically following established transport networks throughout all but the arid zone (Fig 2A). The distribution of absence records was largely similar to the presence records, though absence records fill some gaps in the presence data that are associated with large areas of intact woodland and forest in Australia’s alpine zone and semi-arid parks and reserves. Absence records also effectively trace the few transport corridors that traverse the arid zone (Fig 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Overview of input occurrence data and BRT model of relative habitat suitability for selected ardeid host species in Australia.

A) the distribution of filtered checklists for the period 2000–2023 where the selected species were confirmed sighted or heard; B) the corresponding distribution of comprehensive checklists where those species were neither seen nor heard; C) a summary of the relative influence of the top seven predictor variables submitted to the model; D) prediction of relative habitat suitability to the outbreak period (February 2022); E–G) the shape of the marginal relationship between relative habitat suitability and the three most influential predictor variables; H–J) prediction of relative habitat suitability to December 2021 and June 2022, and standard deviation of predictions across the temporal sequence from July 2021 to June 2023. Base maps for A, B, and H–J obtained from Australian Bureau of Statistics (CC BY 4.0) https://www.abs.gov.au/statistics/standards/australian-statistical-geography-standard-asgs-edition-3/jul2021-jun2026/access-and-downloads/digital-boundary-files.

https://doi.org/10.1371/journal.pntd.0014127.g002

The bias evident in sampling intensity around population centres, being common to both presences and absences, did not strongly influence the predicted habitat suitability over the entire country. The predicted habitat suitability for ardeid hosts in February 2022 based on the model was broadly characterised by a strong coastal and hinterland emphasis on the one hand, most evident around bays and estuaries around much of Australia’s coast, and the major inland drainage basins and wetland complexes of the Murray and Darling Rivers (Fig 2D). The distribution of suitable habitats comparing three sample months from 2021–2022 differs more in intensity than spatial pattern, with the inland areas predicted to be less suitable in winter (June 2022) compared with broad, enhanced suitability across northern Australia. The greatest variation over the sequence of monthly predictions was areas immediately inland from the Great Dividing Range and across parts of northern Australia (Fig 2J, and see the full sequence of monthly predictions at S2 Animation in S2 File.

The strongest covariate influences were fractional cover of herbaceous wetlands, areas that over the long term have tended to support surface water (as distinct from those with high proportional cover of “permanent water”), and mean night time surface temperatures, with a clear threshold around 10–15 C over which habitat suitability improves (Fig 2C).

The BRT model of habitat suitability for ardeid hosts had an AUROC of 0.84 when predicting to the 10% of observations (2406 presences, 15839 absences) held out from model building, indicating very good predictive performance. The model also appears to be a credible comparison against the accumulated aerial survey records of occurrence and abundance for ardeid species in the Eastern Australian Waterbird Survey (EAWS) [105], which were not used to inform our model (S11 Fig). Locations where high counts of ardeid species have been recorded in the EAWS survey corresponded to areas of higher habitat suitability predicted by our model.

Combined model: relative habitat suitability for JEV vectors and vertebrate hosts

The model of the combined suitability for vectors and vertebrate hosts for February 2022 (Fig 3) takes the minimum value of ardeid host and vector habitat suitability at each pixel for the respective predictions to that month. At the continental scale it reflects a combination of the breadth of the habitat suitability across land cover types from vectors and concentration of the highest intensity areas in the vicinity of permanent and periodically wet drainage features more attributable to the ardeid habitat suitability model. These combined influences highlight small areas of high suitability in coastal areas across Northern Australia, and broader areas of moderately high suitability in an arc across the tropical north and down on the inland side of the Great Dividing Range, particularly near the major river courses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Prediction of combined habitat suitability for Culex vectors and ardeid hosts of JEV (joint minimum rescaled suitability) for Australia to February 2022.

Point locations indicating likely infection location for confirmed detections of JEV in human cases (squares) and commercial pig premises (circles) published in the public domain to August 31, 2022 [109] used to rescale the original model outputs are overplotted and detailed in panels. Base maps obtained from Australian Bureau of Statistics (CC BY 4.0) https://www.abs.gov.au/statistics/standards/australian-statistical-geography-standard-asgs-edition-3/jul2021-jun2026/access-and-downloads/digital-boundary-files.

https://doi.org/10.1371/journal.pntd.0014127.g003

Combining host and vector habitat suitability models

Our use of an empirical cumulative distribution function to combine the response scale of the target-group background model for vectors with the presence absence model for ardeid hosts is a notable contribution to the habitat suitability literature. Integrated distribution models are increasingly applied to combine heterogeneous datasets for a single target species [93,135,136], but our approach extends this logic to combine functionally distinct ecological groups into a representation of the fundamental reservoir required for pathogen transmission. The resulting surface, based on the minimum rescaled suitability of vectors and hosts, provides a principled way to identify areas where both components of the transmission cycle coincide. Walsh and colleagues [19] produced an alternative geospatial model of JEV suitability in Australia using the same detection dataset but different modelling approaches and covariates. While both models indicate relatively high suitability inland of the Great Dividing Range—coinciding with the epicentre of the 2022 outbreak—the model of Walsh et al. also suggested very high suitability along southern and eastern coastal margins. The relative absence of detected human infections in those densely populated areas, despite presumably greater opportunity for detection, weighs against that emphasis.

Data limitations

The JEV locations used to rescale our combined surface were obtained from public domain sources and represent only a subset of confirmed infections and locations. We lacked information on spatial precision or expert confidence in attributed locations and extracted single point values from corresponding raster cells in creating the empirical cumulative distribution functions. Bootstrapping or areal summaries might have been a more appropriately conservative approach to represent location uncertainty. Critically, because we used these points to rescale the component layers, they cannot be used for independent validation of the combined surface. A more comprehensive dataset incorporating detections across multiple host species, with appropriate characterisation of surveillance effort, would enable more robust estimation and evaluation. Such work is underway in collaboration with relevant agencies.

The mosquito occurrence dataset we used was compiled under time pressure during an active outbreak. We were only able to model occurrence rather than abundance, though more dynamic models would be attainable if reliable abundance data were available at national scale (e.g., [137]). Critically, abundance data support estimation of important parameters influencing transmission such as the force of infection from mosquitoes to hosts, or vectorial capacity [138]. Our post-hoc comparison suggested good correspondence between predicted suitability and observed abundance per calendar month in temperate and sub-tropical latitudes, but poorer correspondence in the tropical north where predictions missed nuance in seasonal abundance patterns. In the subset of data with abundance attributes, interpretation was complicated by varying trapping methods, counting protocols, and jurisdictional differences in data management. A coherent approach and protocols for collating and harmonising mosquito abundance data would be an important next step for future outbreak preparedness.

Model limitations

Our target-group background of Animalia records reflected a similar pattern of accessibility bias to the mosquito surveillance network, but accessibility (travel time to a large city) is only part of the bias picture. Surveillance networks do distribute traps according to human population centres but strategies may also reflect allocation according to past detections, transmission risk factors, vulnerable populations, and likely larval sources. These nuances are not captured by general accessibility metrics. Future modelling could invest more effort in background sampling that more closely reflects the specific bias signature of mosquito surveillance.

Our temporally dynamic models, while capturing month-to-month variation in environmental suitability, include no mechanistic representation of population dynamics, dispersal, breeding phenology, or immunity. Temporal variability was implicitly modelled as covariate combinations with no consideration of residence time or movement scales in geographic space. As mentioned earlier such correlative approaches cannot explain why the 2022–2023 summer, with similarly favourable environmental conditions, produced only relatively isolated detections. Mechanistic transmission models incorporating vector, host, and immunological dynamics would allow exploration of hypotheses regarding the stochastic nature of JEV risk in Australia; several have been published for JEV (reviewed in [124]) though to date they have not included realistic spatial variation. Integrating mechanistic and distributional approaches should be a priority for future work.

Host species scope

We excluded feral pigs from our host models due to the absence of reliable distribution data at the time of analysis. Given the numerous detections in feral pigs during 2022, their capacity for high viraemia [55,118], and reported abundance and extent [36], there were conceptual grounds for their inclusion. Where feral pigs and ardeid birds share habitat preferences their exclusion would not be consequential, but it may have contributed to us underestimating host suitability in wooded habitats in Australia’s tropical north where feral pigs are reportedly abundant but ardeid density may be lower (though modelled suitability for the Culex was also low in those habitats, which would limit changes to the combined model). On the other hand, since pigs are not thought likely to contribute to the transmission of MVEV (reviewed in [139]), our models with only waterbird hosts may offer greater flavivirus generality by virtue of that absence. A recently published national feral pig distribution model [140] could be incorporated into future JEV assessments.