Aedes mosquito data collection
We adopted the Aedes mosquito collection database established in our previous study [33]. We updated the database by reviewing some recently published work [28, 29, 51–53]. The new Aedes surveillance system covers 23 provinces, plus surveillances conducted by provincial CDCs, resulting in a total of > 1000 sampling sites [29, 32, 54]. Most of the surveillances are done in areas where Aedes data is already available, especially in southern China. New study sites are few, especially in areas where Aedes has not been observed. For example, in Anhui Province, Aedes surveillance was conducted in 16 cities, all of them within existing areas where Aedes has been found [53]. In Sichuan province, Aedes samplings since 2016 have covered 21 cities, three-quarters of which were in areas where Aedes had been found [55]. However, there are several newly Aedes-invaded places, especially at high altitude (Fig. 1). In this study, we used data for Ae. albopictus only; other Aedes mosquitoes were not included [54, 56, 57].
Environmental And Climatic Suitability Modeling
We have also updated the machine learning classification and regression tree (CART) model based on the updated Aedes database and WorldClim 2.0 data [33, 58]. Details of the climatic and environmental data have been described in our previous study [33]. Briefly, the 1970–2000 average monthly climate data for minimum, mean, and maximum temperature and precipitation were downloaded from WorldClim version 2.1 (https://www.worldclim.org/data/worldclim21.html). The environmental regions were divided into four categories: humid, sub-humid, semiarid, and arid [33]. Climatic zones comprised nine categories: south subtropical, mid-subtropical, north subtropical, warm temperate, mild temperate, cool temperate, plateau subtropical, plateau temperate, and plateau subfrigid [33].
We conducted univariate analyses to examine the relationship between Ae. albopictus prevalence and climatic variables using Chi-square automatic interaction detection (CHAID) [59, 60]. The aim was to examine how does change in climatic variables affect the presence/absence of Ae. albopictus. CHAID is similar to logistic analysis, but CHAID produces the critical cutoff of predictors and allows for nonlinear combination of predictors [61]. In addition to overall prediction accuracy, we also measured the sensitivity and specificity to examine the prediction skewness (bias to presence/absence) and used Yale’s coefficient to measure the association between the observed and predicted prevalence of Ae. albopictus by each climatic variable [62, 63]. We did not examine the impact of summer temperature (June–September) on Ae. albopictus prevalence and we only examined the climatic effect for the same month and the following four months.
For the multivariate analysis, the detailed multi-step modeling process has been described in our previous study [33]. Briefly, after data pre-processing, CART models were developed using a 10-fold cross-validation method to predict the potential seasonal (or monthly) distribution ranges of Ae. albopictus in China at high resolution based on environmental-climatic conditions (refer to Supplement A for modeling detail). Since Ae. albopictus was only found in northern China from June to September, June to September were aggregated as one season for risk analyses. Environmental-climatic suitability for Ae. albopictus was predicted as the average predicted suitability probability of the 10 models developed during the 10-fold cross-validation modeling process, and spatial resolution was 30 arcsec or approximately 1 km.
Model performance was measured using prediction accuracy, sensitivity (present predicted as present), specificity (absent predicted as absent), and Cohen’s Kappa coefficient [64]. Kappa measures the reliability of agreement between observed and predicted qualitative data and considers the possibility of the agreement occurring by chance.
Trends In Climate Change In China, 1970–2021
To examine the heterogeneity of climate trends in China, we collected daily meteorological records for the period of 1970–2021 from 90 meteorological stations (Supplement Figure S1). Since Aedes mosquitoes exist nearly everywhere in southern China except the Qinghai-Tibetan Plateau, we selected only a representative subset of stations for climate trend analysis. We selected as many stations as possible from northern China, especially near the current margin of Aedes distribution [33], but excluded some stations to avoid oversampling; i.e., if two stations were located very close to each other (< 200 km) we selected only one of the two stations.
Daily records were summarized as monthly mean maximum/minimum/mean temperature and monthly cumulative precipitation. Trends of monthly data at each station were analyzed using linear regression analysis. Climate trends were measured as the rates of annual change in monthly maximum/minimum/mean temperature and annual cumulative precipitation. Due to the large variation in monthly precipitation in different years, trends in monthly precipitation were not analyzed. Climate change trends in China were analyzed by month (temperature) or annually (precipitation) and aggregated based on latitude.
Climate change and its impact on Aedes distribution
To predict future climate distribution, we needed to create a climate trend map of China. Based on our climate trend analyses, we produced the trend distribution map using the geostatistical spatial interpolation method of Universal Kriging (refer to Supplement B for modeling detail) [65], which assumes a three-order polynomial trend model, i.e., trends in climate change may be linearly or nonlinearly correlated with latitude/longitude. Using this climate trend map and 1970–2000 mean climate as the baseline, we predicted the temperature and precipitation distributions in China for 2020, 2050 and 2080, a typical risk projection framework [40, 41]. We compared the projected temperature increase in 2050 and 2080 between this study and the GCMs using 2000 as baseline [48].
We used the suitability models established earlier to predict the Aedes distribution in each month based on the 2020, 2050 and 2080 climatic projections. Ae. albopictus risk was measured as the probability of presence of Ae. albopictus.
All data analyses were conducted using R 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) except Universal Kriging, which was done using ArcGIS Pro 3.0.0 (ESRI, Redlands, CA, USA). The following R packages were used in this study: for raster image reading and risk mapping, we used the raster and crop methods within the rasterImage and sp packages; and for regression tree modeling, we used the ctree and rpart methods within the rpart, party, and caret packages.