Suitable seasons: Global monthly habitat suitability for the arbovirus vectors Aedes aegypti and Aedes albopictus in 1975-2024

This paper introduces the Climademic Suitability Model, a machine learning framework that maps global monthly habitat suitability for *Aedes aegypti* and *Aedes albopictus* from 1975 to 2024, revealing that climate-driven expansion has already exposed over 5 billion people to these vectors, surpassing previous 2050 projections.

The Gist

Imagine the world as a giant, shifting puzzle. For decades, scientists have been trying to figure out where two specific types of mosquitoes—Aedes aegypti and Aedes albopictus—can survive. These aren't just any mosquitoes; they are the "delivery trucks" for dangerous viruses like Dengue, Zika, and Chikungunya.

This paper introduces a new, super-smart tool called the Climademic Suitability Model. Think of it as a high-tech weather forecast, but instead of predicting rain, it predicts where mosquitoes can build a home.

Here is the breakdown of how it works and what it found, explained simply:

1. The Problem: The Mosquitoes Are Moving In

These mosquitoes used to stay mostly in the hot, tropical parts of the world. But thanks to climate change, global travel, and trade, they are packing their bags and moving into new neighborhoods (like Europe and North America) where they weren't seen before.

The problem is that we don't have a perfect map of where they are right now, month by month. We know they are there, but we don't always know when they are active or how far they have spread.

2. The Solution: A "Mosquito Crystal Ball"

The researchers built a machine learning model (a type of AI) that acts like a crystal ball.

• How it learns: Instead of needing to know where the mosquitoes aren't (which is hard to prove), the AI only looked at millions of records of where mosquitoes were seen between 1975 and 2024. It learned the "recipe" for a perfect mosquito home.
• The Ingredients: The AI mixed together data on temperature, humidity (dew point), rain, wind, land use (forests vs. cities), and human population.
• The Output: It produces a monthly map for every corner of the globe, showing the probability of a mosquito being able to survive there.

3. The Big Discovery: The "5 Billion" Threshold

The most shocking finding is about population.

• The Old Prediction: Experts used to think that by the year 2050, about 5 billion people would live in areas suitable for these mosquitoes.
• The New Reality: This model says we have already passed that number.
• The Analogy: Imagine a rising tide. Scientists thought the water would reach the 5-billion mark in 2050. This study says, "Look down! The water is already at the 5-billion mark, and the tide is rising faster than we thought."
• Why? Because the areas where these mosquitoes can live are expanding, and those same areas are also where the human population is growing the fastest (especially in Asia and Africa). It's a "double whammy" of more people and more mosquitoes in the same places.

4. What Drives the Mosquitoes? (The "Secret Sauce")

The researchers used a special tool called SHAP to peek inside the AI's brain and see what mattered most.

• The Main Drivers: It turns out Temperature and Dew Point (how humid the air feels) are the VIPs. If it's too cold or too dry, the mosquitoes can't survive.
• The Surprising Guest: You might think "Cities" (urban areas) are the most important factor because these mosquitoes love human blood. However, the model found that at a global scale, the weather is a much bigger boss than whether you live in a city or a farm. The weather sets the stage; the city just provides the actors.

5. Seasonal Shifts: The "Dance" of the Seasons

The model doesn't just show a static picture; it shows a movie.

• Summer in the North: As the Northern Hemisphere warms up (March to June), the "safe zone" for mosquitoes moves north, covering more of Europe and the US.
• Winter in the South: As the Southern Hemisphere cools down, the mosquitoes retreat, only to come back when their summer arrives.
• The Trend: Over the last 50 years, the "safe zones" have expanded. Places that used to be too cold for 6 months of the year are now suitable for 8 or 9 months.

6. Why This Matters

This isn't just about mosquitoes; it's about preparation.

• Early Warning: Public health officials can use these maps to know before an outbreak happens. If the map says a region in Germany or Canada is becoming suitable in June, health workers can start spraying and educating people before the mosquitoes arrive.
• The Future: As the climate continues to change, this model acts as a compass, helping us navigate a world where the risk of mosquito-borne diseases is moving into our backyards faster than expected.

In a nutshell: The world is getting warmer and more humid in places where millions of people live. The mosquitoes are following the heat. We used to think we had until 2050 to worry about 5 billion people being at risk, but this new "Mosquito Crystal Ball" tells us that risk is already here, and we need to act now.

Technical Summary

1. Problem Statement

The global spread of Aedes aegypti and Aedes albopictus, the primary vectors for dengue, Zika, and chikungunya, poses an escalating public health threat. While these mosquitoes were historically endemic to tropical and subtropical regions, climate change and environmental shifts have accelerated their expansion into previously non-endemic areas.

Key Challenges Identified:

• Data Limitations: Existing predictive models are often constrained to annual temporal resolutions due to a lack of temporally resolved mosquito observation data. This obscures critical seasonal activity shifts and local seasonal cycles.
• Methodological Gaps: Traditional binary classification models struggle with class imbalance (presence-only data vs. absence data). Many studies rely on artificially generated pseudo-absence data, which can introduce bias, especially for invasive species with underreported surveillance.
• Lack of Explainability: Previous models often function as "black boxes," failing to provide quantifiable insights into the specific environmental drivers (e.g., temperature, land use) governing habitat dynamics.

2. Methodology

The authors introduce the Climademic Suitability Model, a data-driven, spatio-temporally dynamic framework designed to predict monthly global habitat suitability at a 0.25° spatial resolution (approx. 31 km) for the period 1975–2024.

A. Data Integration

The model integrates four distinct data streams:

1. Mosquito Observations:
   • Historical: Kraemer et al. (2015) compendium (1960–2014) containing ~42,000 sightings.
   • Recent: Global Mosquito Observations Dashboard (GMOD) (2015–2023) aggregating citizen science data (GLOBE Observer, Mosquito Alert, iNaturalist).
   • Validation: 2024 data from ECDC, GBIF, and local monitoring programs (e.g., Taiwan, Brazil).
2. Climate Data: ERA5 reanalysis data (ECMWF) providing monthly means of surface air temperature, dew-point temperature, precipitation, and wind speed (1940–present).
3. Land Use/Land Cover (LULC): HILDA+ dataset (1960–2019) classifying land into six categories (urban, cropland, pasture, forest, etc.). Note: 2020–2024 LULC data were approximated using 2019 values.
4. Population Data: GHS-POP dataset (1975–2030) providing population density.

B. Machine Learning Architecture

The core algorithm is a One-Class Support Vector Machine (OCSVM), chosen because it learns the characteristics of a single class (presence) without requiring labeled absence data.

• Yearly Model Development:
  • Trained on data from 1975–2014.
  • Features: 48 climate variables (12 months × 4 variables), 8 LULC variables, and 1 population variable.
  • Hyperparameters: Radial Basis Function (RBF) kernel with $\nu=0.01$ (contamination) and $\gamma=0.01$ (kernel coefficient), determined via median heuristic.
• Monthly Model Development:
  • To enable monthly predictions, the model uses a heuristic approach: for each sighting, it extracts the 75th percentile of monthly climate values (Jan–Dec) to approximate the optimal conditions during which the sighting likely occurred.
  • Features: 4 climate variables (single monthly mean), 8 LULC, 1 population (Total 13 features).
  • Hyperparameters: $\nu=0.03$ and $\gamma=0.03$.
• Incremental Learning (The "Climademic" Model):
  • The initial OCSVM monthly model was refined using an incremental learning approach. New observations from GMOD (2015–2023) were sequentially fed into the model.
  • Only misclassified data points were retained to update the decision boundary iteratively, creating the final "Climademic Suitability Model."
• Probability Calibration:
  • OCSVM outputs decision scores (inlier vs. outlier). These were converted to probability values (0–1) using quantile binning (10 bins) based on the method by Que and Lin (2024), which mimics the distribution of training decision scores.

C. Explainability

The authors utilized SHAP (SHapley Additive exPlanations) with the KernelSHAP algorithm to interpret the model. This allowed for:

• Global Interpretation: Ranking feature importance across the entire dataset.
• Local Interpretation: Visualizing how specific features (e.g., temperature in Berlin vs. Sao Paulo) drive predictions for individual instances via waterfall plots.

3. Key Contributions

1. High-Resolution Spatio-Temporal Modeling: The first global model to provide monthly habitat suitability predictions for both Ae. aegypti and Ae. albopictus spanning 50 years (1975–2024).
2. One-Class Learning Framework: Successfully applied OCSVM to overcome the "absence data" problem in mosquito surveillance, utilizing presence-only data to learn the ecological niche without artificial pseudo-absences.
3. Incremental Learning Pipeline: Demonstrated a method to continuously update the model with new surveillance data (GMOD), allowing for near-real-time adaptation.
4. Explainable AI: Provided a transparent view of the drivers of habitat suitability, moving beyond black-box predictions to quantify the impact of temperature, dew point, and land cover.

4. Key Results

A. Seasonal and Spatial Dynamics

• Seasonality: The model accurately captures seasonal shifts. In the Northern Hemisphere, suitability expands northward in summer (June) and contracts in winter (December). The reverse pattern is observed in the Southern Hemisphere.
• Regional Differences:
  • Europe/North America: Ae. albopictus shows greater suitability than Ae. aegypti, particularly in northern latitudes, reflecting its higher cold tolerance.
  • Tropics: Ae. aegypti dominates in Brazil, Madagascar, and Australia.
• Year-Round Suitability: In 2024, year-round suitability is predicted for large parts of Brazil, Central America, Sub-Saharan Africa, India, Southeast Asia, and Eastern China.

B. Long-Term Trends (1975–2024)

• Expansion: Significant expansion of suitable habitats occurred in Central America, Eastern US, Southern Brazil, parts of India, Eastern China, SE Asia, Japan, and Eastern Australia.
• Contraction: Notable decreases in suitability were observed in Western US, Northern Brazil, large parts of Africa, Central China, and Western Australia.
• Drivers:
  • Temperature & Dew Point: Identified as the dominant drivers. Areas with declining dew point temperatures correlated strongly with declining habitat suitability.
  • Land Use: Shifts from pasture to forest in Southern Africa and parts of Europe/Asia correlated with increased suitability.

C. Population Impact

• 5 Billion Threshold Surpassed: The study estimates that over 5 billion people currently live in areas that have become more suitable for these vectors since the mid-1970s.
• Demographic Shift: This threshold, previously projected to be reached by 2050, has been reached earlier due to the convergence of habitat expansion and rapid population growth in Asia and Africa.
• Year-Round Risk: Approximately 34% of the global population (2.7 billion people) now lives in areas suitable for these mosquitoes year-round.

5. Significance and Implications

• Public Health Preparedness: The model provides a framework for near-real-time vector surveillance and proactive planning. It allows health authorities to identify emerging risk zones before colonization occurs.
• Policy Relevance: The finding that 5 billion people are now at risk (surpassing 2050 projections) underscores the urgent need for coordinated, large-scale vector control strategies.
• Epidemiological Modeling: The high-resolution monthly data can be integrated into transmission models for dengue, Zika, and chikungunya to improve epidemic forecasting.
• Limitations & Future Work:
  • LULC Lag: The model assumes static land use for 2020–2024.
  • Temporal Dependencies: The model treats months independently, potentially missing carry-over effects (e.g., cold winters affecting overwintering eggs).
  • Presence vs. Establishment: The model predicts suitability (viability upon introduction), not necessarily established populations, potentially leading to over-prediction in areas with no current transmission.

In conclusion, the Climademic Suitability Model represents a significant advancement in vector ecology, offering a robust, explainable, and high-resolution tool to track the shifting landscape of arboviral risk in an era of accelerating environmental change.