Machine Learning Modeling of Aedes albopictus Habitat Suitability in the 21st Century

Abstract

The Asian tiger mosquito, Aedes albopictus, is an important vector of arboviruses that cause diseases such as dengue, chikungunya, and zika. The vector is highly invasive and adapted to survive in temperate northern territories outside its native tropical and sub-tropical range. Climate and socio-economic change are expected to facilitate its range expansion and exacerbate the global vector-borne disease burden. To project shifts in the global habitat suitability of the vector, we developed an ensemble machine learning model, incorporating a combination of a Random Forest and XGBoost binary classifiers, trained with a global collection of vector surveillance data and an extensive set of climate and environmental constraints. We demonstrate the reliable performance and wide applicability of the ensemble model in comparison to the known global presence of the vector, and project that suitable habitats will expand globally, most significantly in the northern hemisphere, putting at least an additional billion people at risk of vector-borne diseases by the middle of the 21st century. We project several highly populated areas of the world will be suitable for Ae. albopictus populations, such as the northern parts of the USA, Europe, and India by the end of the century, which highlights the need for coordinated preventive surveillance efforts of potential entry points by local authorities and stakeholders.

Keywords: habitat suitability; machine learning; vector-borne diseases.

Figure 1

Distribution of the number of examples in the training set for each month of the year (top panel) and geographical distribution of the dataset used to train and evaluate the model’s performance (bottom panel). The colour bar shows the number of examples for each grid cell.

Figure 2

The schematic, high-level overview of the procedures followed in this study to train the machine learning model and project Ae. albopictus habitat suitability.

Figure 3

The ROC curves obtained on the test dataset (left panel) and precision-recall curves (middle panel) for the Random Forest, XGBoost and, ensemble models. On the (right panel), the sensitivity of the ensemble model compared to the known presence of Ae. albopictus as a function of the number of months set as a threshold for habitat suitability. The inset in the (left panel) shows a zoomed view of the ROC curves.

Figure 4

Predicted Ae. albopictus habitat suitability in terms of months predicted as suitable by the ML model for the early part of the century (2020–2025). The normalization scenario (SSP245) is presented on the top panel, whereas the “business as usual” (SSP585) scenario is presented on the bottom panel. The colorbar shows the average number of months predicted as suitable by the machine learning model for each grid cell.

Figure 5

Comparison of the latitude profiles for the SSP245 climate scenario (left panel) and SSP585 (right panel), for the early, mid, and end of century time periods. Solid lines and the shaded areas represent the median and the 95% range, respectively.

Figure 6

Latitude profiles for the total number of months predicted as suitable by the ML model (top panels) and the transitional changes between the early to mid, end of century, and mid to end of century periods (bottom panels). The latitude profiles obtained for the SSP245 scenario are shown in red and the corresponding profiles for the SSP585 scenario are shown in blue, whereas the difference between the two is shown in gray. Solid lines and the shaded areas represent the median and the 95% range, respectively.

Figure 7

Suitable months normalised to 100 km${}^2$ (left panel), the total area over which habitat suitability is projected (middle panel) and the total number of months projected as suitable for Ae. albopictus (right panel), for the two IPCC scenarios. The blue and red lines, which correspond to the SSP245 and SSP585 scenarios, respectively, show the ensemble average for the nine climate models used for each of the two scenarios. Solid lines and the shaded areas represent the median and the 95% range, respectively.

Figure 8

Comparison between the tropical (top row of panels) and extratropical (bottom row of panels) regions of the world for habitat suitability normalised to 100 km${}^2$ (left column), total area covered (middle column) and total number of months predicted for each year (right column). Solid lines and the shaded areas represent the median and the 95% range, respectively.

Figure 9

Population at risk of Ae. albopictus-borne diseases per year for the two scenarios examined in this study (left panel). In the inset, the total population projected until the end of the century for the scenarios is presented. In addition, the increase in population at risk of Ae. albopictus-borne diseases with respect to the start of the projection window (2020) for the SSP245 scenario (blue) and SSP585 scenario (red) is shown on the right panel. The three time periods presented here correspond to the early, mid, and end of the 21st century. The median of the output from the nine climate scenarios is presented here and the shaded area (left panel) and lines (right panel) represent the 95% range.