A Habitat Model for Disease Vector Aedes aegypti in the Tampa Bay Area, Florida

Dengue and Zika are dangerous arboviruses endemic to tropical and subtropical regions (Kraemer et al. 2015a, Patterson et al. 2016). Approximately half the world's population is currently at risk, and 390 million people are infected by dengue annually (WHO 2021). The rapid global rise in dengue occurrence may have immense public health implications for future transmission in Florida (Wilder-Smith et al. 2019), which is particularly vulnerable. The state's proximity to arbovirus-endemic areas, the hub for tourism, and the warm and humid climate provide the ideal environment for exotic mosquitoes and mosquito-borne pathogens to be introduced and transmitted (Beeman et al. 2021, Reeves et al. 2023). Indeed, in 2022, Florida was the only state with locally transmitted cases of dengue (n = 47; CDC 2022), and local outbreaks and isolated introductions of dengue virus occasionally occur in South Florida (Rey 2014). Locally acquired dengue infections have also been documented in the Tampa Bay area in 1905, 1907, 1934, 2011, and 2019 (Rey 2014, FDH 2021). In 2016, an outbreak of Zika virus occurred in Miami-Dade and Broward counties, resulting in at least 300 human cases (Likos et al. 2016).

The primary vector that spreads dengue, Zika, and yellow fever is Aedes aegypti (F.), a highly successful invasive species (Gardner et al. 2017) with a geographic range that is very likely to expand, largely fueled by human transportation and changing climate (Kamal et al. 2018). Aedes aegypti is highly anthropophilic and strongly associated with human environments (Takken and Verhulst 2013), and typically breeds in small and often artificial outdoor containers that gather rainwater in urban and suburban areas (Kraemer et al. 2015a). Aedes aegypti has been present in the Americas since the 15th century, with seasonal distributions documented as far north as New Jersey (Donnelly 1993), and crucially is present year-round throughout Florida (Kraemer et al. 2015b, 2019; CDC 2017).

Previous studies in Florida have shown that Ae. aegypti habitats are highly dependent on surrounding land cover and are typically found in residential areas (Leisnham et al. 2014). In tropical southern Taiwan, Huang et al. (2018) found that dengue transmission was associated with urban areas and parks, as well as decreasing areas of green space, particularly in conjunction with increasing built environments. Aedes aegypti is limited by seasonal environmental conditions and is more active during specific times of the year. Monaghan et al. (2019) constructed a US-based model based on meteorological data and found the greatest Ae. aegypti abundance in September and a generally high abundance from July to October in the southeastern USA. Given the climate of Florida, there may be an even longer period associated with high Ae. aegypti abundance.

As one of the fastest growing regions in the USA, the Tampa Bay area has experienced an influx of population, resulting in the expansion of built environments after thousands of acres of agricultural, forest, and scrub lands were cleared to make way for new homes and businesses (Hartz 2022). These disturbed environments create ideal breeding habitats for Ae. aegypti and increase the potential for arbovirus transmission in counties of the Tampa Bay area (here including Hillsborough, Pasco, Pinellas, and Polk). Fortunately, each of these counties maintains year-round mosquito control agencies with high-quality historic trap data. For all of the above reasons, the Tampa Bay area is an ideal region for modeling Ae. aegypti abundance.

With no specific treatments for either dengue or Zika, prevention measures such as frequent mosquito surveillance and targeted control are paramount for public health. Our study caters to mosquito control agencies by developing an Ae. aegypti–specific abundance model using data sources that are easily accessible, to produce results that have practical, operational utility. Ultimately, the techniques and suggestions proposed in this study aim to increase the efficiency of mosquito control efforts in targeting high-abundance locations in time and space, thereby reducing vector and virus exposure to humans.

All mapping and geographic information system applications were conducted in ArcMap (10.8; Environmental System Research Institute, Redlands, CA). Mosquito abundance and date of collections provided the spatiotemporal data for the dependent variables in the habitat models. Female Ae. aegypti abundance from January 1, 2017, to December 31, 2019, was provided by the mosquito control programs from each of the 4 counties within the Tampa Bay area: Hillsborough (2,600 km²), Pasco (1,930 km²), Pinellas (710 km²), and Polk (4,660 km²). Abundance data were acquired from 497 mosquito traps of 4 common types: BG-Sentinel (n = 314), Centers for Disease Control and Prevention (CDC) light (n = 135), suction (n = 39), and Wilton (n = 9); however, only CDC light traps were used in all counties (Table 1 and Fig. 1A). Furthermore, to account for differences in productivity, traps were weighted by collection totals of female Ae. aegypti over the 3-year study period.

To accommodate the estimated maximum flight range of Ae. aegypti, we created a 500-m buffer around each trap, increasing the likelihood of capturing the overall ecological and physical features of the surrounding environment (Hausermann et al. 1971, Honório et al. 2003, Harrington et al. 2005, Maciel-De-Freitas et al. 2007, Moore and Brown 2022). This buffered region provided locations by which the values of all subsequent model predictors were extracted and averaged for interpretation into the model assessments.

Full information on the habitat modeling predictor datasets is provided in the Supplemental Material Datasets. All processed raster data, aggregated as means within each 500-m-buffered trap location, were exported and merged into a collective dataset. Predictors were evaluated as detailed in Supplemental Table S1. Previous Ae. aegypti modeling literature consistently reveals 4 types of predictors as most strongly associated with mosquito habitat: human population density, percent impervious, land use/land cover (LULC), and surface vegetation indices (in particular, normalized difference vegetation index [NDVI] and enhanced vegetation index [EVI]; Fig. 1) (Leisnham et al. 2014, Sallam et al. 2017, Ducheyne et al. 2018, Hamlet et al. 2018, Hopperstad et al. 2021).

All statistical analyses were conducted in SAS (version 9.3; SAS Institute Inc., Cary, NC) and JMP (version 16.0.0; SAS Institute Inc., Cary, NC) software. A total of 36 potential model predictors were assessed in modeling female Ae. aegypti habitat suitability. Prior to model evaluation, all predictors were tested for multicollinearity (Fig. 2). The EVI and NDVI were further assessed separately by average monthly value. Epidemiologic week trap totals were averaged by county and assessed for unequal variances using Welch's test. Trap group mean totals were analyzed using Tukey's honest significance test (for categorical variables) and generalized linear regression (for continuous variables). Inclusion of parameters within global models was first screened by backward stepwise regression, using minimum Bayesian Information Criterion (BIC) as the stopping rule. Candidate covariates were then analyzed by standard least squares regression. The final model with the lowest overall BIC was selected as the best-fit global model.

Final models were developed for 4 scenarios: 1) Best-fit global model (all submissions from 2017 to 2019 in the Tampa Bay area, FL); 2) Best-fit global model, using only mosquito submissions collected via CDC light traps (the only type used across all 4 counties); 3) Pasco County best-fit model (only using submissions provided by Pasco County from 2020 collections [PC20]); and 4) PC20 submissions and Tampa Bay area citizen science observations best-fit model.

The 2 best-fit global models (all trap types and only CDC light trap types) were validated using PC20 and citizen science submissions. We used Ae. aegypti adults collected from CDC light traps that were placed in new locations within PC20 (n = 856 collected mosquitoes; Supplemental Fig. S1A). Research Grade Ae. aegypti citizen science observations (n = 36; 16 unique locations from 2019 to 2022) from the 4-county region, submitted through iNaturalist.org and acquired from the Global Mosquito Observation Dashboard (http://www.mosquitodashboard.org; Carney et al. 2022), were also applied to model validations (Supplemental Fig. S1B). Overall model strength was assessed by least squares regression of predicted versus actual Ae. aegypti abundance.

A total of 44,212 female Ae. aegypti were collected from 497 traps by the mosquito control agencies of the 4 Tampa Bay counties between 2017 and 2019. Overall, BG-Sentinel traps produced the greatest number of collected mosquitoes (n = 41,060). Aedes aegypti collections significantly differed by trap type (P = 0.0342), with BG-Sentinel producing the greatest average number of collections per trap ( = 130.8), significantly different from both CDC light ( = 14.8, P = 0.0033) and suction ( = 27.4, P = 0.0162) traps. Average mosquito collections peaked during epidemiologic week 26 and remained high through week 39 (Fig. 3A). These epidemiologic weeks coincide with the summer months of June, July, and August, producing the highest average mosquito collections for all counties and trap types (Table 1). By county, Hillsborough collected the greatest number of mosquitoes both in total and per trap ( = 34.8), statistically different from the other 3 counties (Pasco, Pinellas, and Polk counties; = 1.6, 14.2, and 6.9, respectively; P < 0.0001 for each county).

A total of 13 LULC types were present in the Tampa Bay area (Fig. 3B). Mosquito traps placed in developed, low-intensity built environments produced the greatest number of female Ae. aegypti. Conversely, mosquito traps placed in developed, high-intensity built environments caught the fewest. All other trap types did not differ statistically in the total number of mosquitoes collected. Overall, locations with traps in areas of increasing built environment (e.g., imperviousness and human population density) resulted in significantly greater mosquitoes (P < 0.0001; Fig. 3C).

Of the 36 covariates evaluated, temperature-specific variables (e.g., current collection day temperature and 1-, 3-, and 4-wk lags) were 4 of the 5 highest-ranked predictors of Ae. aegypti abundance (Table 2). Other notable predictors included total housing units, average human population density, and median household income, all ranked within the top nine in strength of association. Comparatively, EVI had an approximately 3-fold stronger association with weighted Ae. aegypti trap collections versus NDVI. However, a breakdown of each vegetation index by month showed similar outcomes in Ae. aegypti abundance: the warmest months (August, June, July, and September) produced the greatest trap collections (Table 3).

Best-fit global models evaluating the weighted average of all female Ae. aegypti collections included 5 covariates: NDVI, year, current day temperature, White population, and median household income (P < 0.0001, R² = 0.094; Table 4). Stratifying the best-fit model to the weighted average of female Ae. aegypti collected by CDC light traps maintained 5 distinct covariates (P < 0.0001, R² = 0.081; Table 5): 28-day temperature lag, total housing units, housing built between 1970–79, and housing built between 1960–69. By month, model outputs displayed clustered areas of relatively high- and low-weighted mosquito abundance (indicated by red and blue colors, respectively) that shifted in response to seasonal changes to temperature (Fig. 4).

Validation results of best-fit global models 1 and 2 were as follows: all mosquito trap collections in PC20, P = 0.11; CDC light trap–only collections in Pasco County in 2020, P = 0.19, R² = 0.03. Combining the PC20 mosquito collections with the Tampa Bay area citizen science observations resulted in the following: all mosquito trap collections in PC20, P = 0.11, R² = 0.054; CDC light trap-only collections in PC20, P = 0.16, R² = 0.031 (Table 5 and Supplemental Fig. S2).

The results of this study demonstrate both the plausibility and practical value of information provided by Ae. aegypti abundance models. All predictors assessed in this study are free and publicly available. With the exception of NDVI and EVI, all data are analysis-ready, requiring no additional processing or preparation for use. Our methodology and results are consistent with those of previous literature on assessing vector mosquito risk, in highlighting the importance of utilizing climatic data, particularly mean temperature and its associated weekly lags (Uelmen et al. 2021). Also important are socioeconomic, race, and household age predictors in Ae. aegypti risk assessments (Little et al. 2017, Goodman et al. 2018). Previous literature has established one or more of these predictors as valuable in mosquito ecology, but rarely do these nonenvironmental forces occupy >50% of a final model's total covariate composition (Karki et al. 2020).

Another key finding is the importance of quality and spatiotemporal resolution of data. The 4-county global models assessed in this study were significant, but contained greater root mean square errors, higher model assessment criterion (AICc; Akaike information criterion corrected) and BIC (Bayesian information criterion), and poorer mean fit (as indicated by R²) compared with the smaller models in Table 5. Despite far lower sample sizes than the 4 county assessments, both best-fit models were strongly significant (2020 Pasco County collections alone as well as paired with citizen science observations; P = 0.0001 and P = 0.0009, respectively). Out of all best-fit and validation models analyzed, the best-fit 2020 Pasco County collections had the strongest correlation (R² = 0.124) and lowest model assessment criterion (AICc = 1,058.21, BIC = 1,073.12).

While all predictors herein are freely available to the public, these models cater to mosquito control agencies and require mosquito abundance data as the primary independent variable. Additionally, NDVI and EVI do require a modest computational skillset. To alleviate that burden to novice users, we provide the publicly available websites where these layers can be acquired, as well as a copy of all raw and processed imagery (https://drive.google.com/file/d/1_2QHlHqYVDYTKwmXtsHN87Y3ftkNZWq7). Fortunately, at least for the Tampa Bay area, high-quality abundance models can be generated without the use of any vegetation indices, although this is likely highly variable and specific to geography, time, and species of interest. Results from this analysis demonstrate that simplified models—that is, data layers consisting of at least one temperature lag (7, 21, and/or 28 days), one socioeconomic (number of housing units), and one demographic (population density) variable—produce surprisingly robust estimates of mosquito abundance that can be used as stratified approximations for exposure risk.

All statistical models are inherently error-prone and do not capture all details of reality. That being said, a useful model is a good model. Limitations of the data and methodology used in our assessment include the fact that our independent variable, weighted Ae. aegypti abundance by trap location, is a presence-only event, and therefore the inclusion of absence data may strengthen the model output. Our modeling approach incorporated annual, monthly, and weekly data at varying resolutions. As a general guide, higher resolutions in both space and time are preferable (Little et al. 2017). Overall, our spatial resolution ranged from 4 km (PRISM climate data) to 30 m (impervious and LULC). Lastly, we produced monthly model abundance maps as a visual guide for our readers. Following suit with the prior example, higher temporal resolutions would be more useful for mosquito control personnel, and the maps herein could be modified to produce weekly results, for example.

A key advantage of the 4 counties assessed was the longstanding operation of dedicated mosquito control agencies. However, each agency deploys their own methods for surveillance and control, including unique protocols that prompt action. These differences result in inconsistencies in final data collections (frequency, trap type and locations) and likely yield weaker model fits for a multicounty conglomerate model compared to individual county models

Our results mirror similar findings (Leisnham et al. 2014) by demonstrating an overall peak abundance trend during the hottest months in the Tampa Bay area (June–September), when temperatures consistently reach a high in the 90s (°F) (>32°C) and occasionally reach 100s (°F) (>37°C). These temperatures exceed reported thermal maximums for both dengue and Zika viruses within Ae. aegypti (Shocket et al. 2020). Additional studies should investigate the thermal limitations of arboviruses within Ae. aegypti to assess seasonal disease activity (Monaghan et al. 2019), including the plausibility that overall transmission may decrease, while overall mosquito abundance may increase.

As expected, the predictors included in our final models are those strongly associated with temperature and human-built environment (Leisnham and Juliano 2009, Leandro-Reguillo et al. 2017). In general, locations that were more urban (and thus have higher population densities) with older homes tend to have the highest predicted mosquito abundances. Of the 4 counties, Pinellas had the highest abundance of Ae. aegypti per trap as a proportion of the total county area. Pinellas also has the greatest mean population density among the 4 counties. Conversely, Polk County, the most rural and least densely populated county out of the four, had the overall lowest mosquito abundance. However, the location of highest abundance shifts to downtown Tampa and its immediate suburbs in the summer months, which has public health implications for the residents there.

Citizen science has enormous potential for vector monitoring going forward (Palmer et al. 2017, Carney et al. 2022). If current monitoring trends are mostly trap-based, then there is the potential for spatial gaps in our understanding of new extents of invasive or expanding arbovirus vectors. For example, a citizen science monitoring project recently yielded the first iNaturalist observations of Ae. vittatus (Bigot) globally and Ae. scapularis (Rondani) in the USA (Carney et al. 2022). These 2 species in particular, along with Ae. aegypti in the Tampa Bay area for the present study, were the 3 targets of the campaign (mosquitoAI.org). Our analyses herein took advantage of 36 Research Grade observations of Ae. aegypti from 16 unique locations within the study area, providing a litmus test for our models' overall performance and robustness. Ideally, increasing the number of citizen science observations could improve the data needed to conduct even more generalizable models.

Due to the strong association of Ae. aegypti with urban and suburban landscapes, combined with a relatively short flight range (500 m) of the mosquito, Ae. aegypti–borne viruses are primarily an urban threat. Warming climates and range expansion (Kamal et al. 2018) necessitate updating urban risk models to include additional latitudes within theoretical suitability of Ae. aegypti. Lastly, ports of entry are vital in facilitating Ae. aegypti introductions. Given that the Miami Zika outbreak originated a few miles from the Miami International Airport (Gardner et al. 2017, Marini et al. 2017), future efforts should examine the distance of high-risk areas to possible Ae. aegypti introduction sources, including various ports of entry. For example, Leandro-Reguillo et al. (2017) examined the potential for Zika virus transmission to proliferate in Barcelona, and of the 3 highest risk areas determined, one (Old Town) is near the port of Barcelona. Looking forward, technological advances will continue to greatly assist mosquito surveillance and control activities. From high-resolution imagery remotely sensed from space to mobile device-based mosquito observations on the ground, such technologies can be leveraged for the prediction, detection, and monitoring of invasive vectors, and ultimately help to combat mosquito-borne diseases well into the future.

Supplemental material is available online with the article.

This research was funded by the National Science Foundation under Grant No. IIS-2014547 to R.M.C.