Study areas.

Montpellier is a city of 300 000 inhabitants located in the Occitanie region in southern France (Fig 1). The city has a Mediterranean climate with an average annual temperature of 15°C and an average annual precipitation of 739 mm [37]. The city features more than 10% of vegetated areas within densely urbanized areas. It ranks among the top three greenest cities in metropolitan France [38]. The study was conducted in nine areas belonging to three different environments: two urban parks (PRK), three residential neighborhoods (RES), and four highly impervious (i.e., areas of the city center dominated by sealed surfaces such as concrete and asphalt) areas (IMP) (Fig 1). Urban parks (Botanical Garden (PRK-BOT), and Park of Aiguelongue (PRK-AGL)) and residential neighborhoods (Lemasson (RES-LEM), Aiguerelles (RES-AGR), and Soulas (RES-SOUL)) were vegetated areas. The highly impervious areas included two areas with some vegetation (Saint Charles University (IMP-SCU) and Acapulco Hotel (IMP-ACA)) and two with little to no vegetation (Buisson Bertrand Institute (IMP-BBI) and Diderot School (IMP-DID)) (Fig 1). Detailed descriptions of study areas are provided in S1 Table, and images of study areas in S1 Fig. The minimum distance between study areas was 115 m. Although Aedes mosquitoes can disperse from 75 m [39] to 290 m [40] in urban areas over 8 days, we treated the study areas as independent units due to impervious roads between them, that are known to limit mosquito dispersal [41,42].

Download:

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

Fig 1. Study areas.

Fig 1A and 1B show the location of Montpellier in France (map from Map data © OpenStreetMap contributors, licensed under ODbL) and the land cover map of Montpellier with the study areas, respectively. Sampling environments are shown in the legend of Fig 1B, delineating urban parks, residential areas, and highly impervious areas (boundaries from Montpellier Open Data). Land cover categories are shown in the legend of Fig 1B. Fig 1C, 1D, and 1E show a garden of a residential area, the Botanical Garden (PRK-BOT), and one highly impervious area respectively. Photographs and image created by the author (Colombine Bartholomée).

https://doi.org/10.1371/journal.pone.0335793.g001

Mosquito sampling.

Adult mosquitoes were captured using BG-Pro traps (BioGents, Regensburg, Germany) modified according to L’Ambert et al [43], each equipped with a BG lure—synthetic human odor mimic—and a daily replenished carbon dioxide source made from yeast, sugar, and water (S1 Fig). Two BG-Pro traps were installed in each residential area (RES-LEM, RES-AGR, RES-SOUL), and in the park of Aiguelongue (PRK-AGL). Due to their smaller size (S1 Table), only one BG-Pro trap was deployed in each of the highly impervious areas (IMP-ACA, IMP-BBI, IMP-DID and IMP-SCU). Because of its larger size (S1 Table), three BG-Pro traps were deployed in the Botanical Garden (PRK-BOT).

Traps were placed in shaded, wind-sheltered sites at a height of one meter. Therefore, 15 sites were sampled: five in the two urban parks, six in the three residential areas, and four in highly impervious areas. Every trap was separated by a minimum distance of 50 m to prevent interferences [44]. However, due to accessibility constraints, the two traps in RES-SOUL were placed 22 m apart.

One 48-hour trapping session was conducted per site per month from May to October 2023. During each session, mosquitoes were retrieved each morning, resulting in two collections per session. Mosquitoes were transported at 4°C to the laboratory (MIVEGEC laboratory, IRD regional delegation, Montpellier), and stored at −20°C. Mosquito individuals were then counted, sorted by sex, and morphologically identified using the multicriteria key MosKeyTool [45]. These data are available on GBIF [46].

Fine-scale vegetation and land cover data.

A harmonised land-cover map of Montpellier was generated at 0.5 m resolution by combining three complementary datasets (see Supplementary S2 Table): (i) the fine-scale vegetation dataset of Montpellier (Montpellier Méditerranée Métropole, 2019), which distinguishes vegetation patches using deep learning and photo-interpretation of Pléiades satellite imagery and a LiDAR-derived elevation model; (ii) the French national building database (Base de données nationale des bâtiments, BDNB); and (iii) the Copernicus Urban Atlas. Following rasterisation and integration of these sources, five land-cover categories were retained (Fig 1B): buildings, roads, low vegetation (< 3 m height), high vegetation (> 3 m height), and “others” (including railways, pathways, industrial areas, and sports facilities).

For each trap location, spatial metrics were computed within circular buffer zones of 20 m, 50 m, 100 m, and 250 m radii. For each land cover category, we calculated the percentage cover (proportion of buffer area occupied by the category), total edge length (sum of the perimeters of all patches of the category), and average patch size (mean area of individual patches). For vegetation categories (low vegetation < 3 m height, high vegetation > 3 m height), we also extracted the number of patches. The landscape diversity (Shannon index) was calculated for all buffer zones. This index quantifies diversity by considering both the number of land cover categories and their relative proportions, with higher values indicating more heterogeneous landscapes (see Supplementary S2 Table for the formula). Buffer sizes were chosen according to previous works investigating the effect of urban vegetation [27,49] on the dispersal of Ae. albopictus [40]. Supplementary S3 Table provides the land cover variable values for each trap and buffer radii.

Microclimatic data.

We continuously collected hourly temperature and relative humidity near each trap using Hygro Button data loggers from May to October 2023. The minimum (lowest recorded value), maximum (highest recorded value), and mean temperatures, as well as relative humidity, were recorded for 24-hour periods starting from 48 hours prior to the beginning of each sampling session and continuing until the end of the session (S2 Table).

Meteorological and air quality data.

Hourly temperature and rainfall data were obtained from the Observatoire Départemental de l’Eau et de l’Environnement (ODEE) station located at Château d’Ô in the north-west of Montpellier (S2 Table). Relative humidity and wind speed data were obtained from the Frejorgues airport station (belonging to the Meteo France national meteorology agency network) at a 3-hour resolution (S2 Table). Growing Degree Days (GDD), a known predictor of mosquito development, were calculated from daily mean temperatures above 11°C, as this was shown to be the annual mean temperature limit for the establishment of Ae. albopictus populations [50]. Temperatures above this level are conducive to mosquito development. For each sampling event, weather data were aggregated at a weekly resolution for the sampling week (W0) and for all possible lag periods between W–6 and W0. These lag periods ranged in duration from one to seven consecutive weeks (e.g., W–2 only, W–4 to W–2, W–6 to W0). Rainfall and GDD were summed, while temperature, humidity, and wind speed were averaged over each lagged window. Precipitation data from the Château d’Ô station were also aggregated over 24-hour periods prior to and during each sampling session, following the same temporal aggregation approach used for microclimate data from the data loggers, because precipitation may influence mosquito activity and therefore sampling.

As atmospheric pressure can influence insects’ flight and activity [51–53], we evaluate the effect of daily atmospheric pressure on Ae. albopictus. For this purpose, the minimum (lowest recorded value), maximum (highest recorded value), and mean atmospheric pressure, measured every three hours at the Frejorgues airport station, were extracted per 24-hour periods before and during sampling sessions. The daily mean atmospheric pressure difference between two consecutive days was calculated per 24-hour periods before and during sampling sessions. S3 Fig depicts the temporal evolution of these meteorological variables (temperature, rainfall, wind speed, relative humidity, and atmospheric pressure).

As air pollution can influence insect behaviour and abundance [54–56], we evaluate the effect of air quality data on Ae. albopictus. For each trap location, we collected daily concentrations of nitrogen monoxide (NO), nitrogen dioxide (NO2), particulate matter (of diameter less than 10 (PM10) and 2.5 µm (PM2.5)), and ozone (O3) from the closest station of the Atmo Occitanie air quality observatory (S2 Table). As for weather data, air quality data were aggregated at a weekly resolution for the sampling week (W0) and for all possible lag periods between W–6 and W0, by averaging the daily concentrations of each pollutant over these windows.

Social and demographic data.

Socioeconomic factors such as low living standards [57], aging or degrading infrastructure [58], and population densities [59] can influence mosquito presence. Number of buildings of different ages and number of low-income households (2019) were obtained from the Filosofi database of the French Institut National de la Statistique et des Études Économiques (INSEE), provided as a 200 m resolution grid. For each buffer area, values of intersecting grid cells were summed. Human population data were obtained from the Montpellier fine-scale population dataset of 2020 (Supplementary S2 Table), in which IRIS-level population counts (~2000 inhabitants per unit) were redistributed across land plots in proportion to living area and assigned to the centroid of significant buildings. Population counts were then aggregated within the same four buffer zones (20, 50, 100, and 250 m radii).

Assessment of vector control interventions.

No municipal adulticide or larvicide treatments were implemented in the study areas during the sampling period. No individual larval habitat removal by residents was observed. However, potential actions by residents near some sampling sites could not be systematically assessed, and their effects on mosquito abundance cannot be excluded.

Bivariate analysis using GLMMs.

Every explanatory variable was tested using a GLMM with crossed random intercepts for trap location, with sites nested within the sampling area, and for the sampling session. We fitted binomial and zero-truncated negative binomial models to presence/absence and positive counts data, respectively. Marginal R2, representing the proportion of variance explained by the fixed effects excluding the contribution of random intercepts [63], was calculated and used for interpretation. Time-lagged effects of meteorological and air quality variables were interpreted using cross-correlation maps (CCMs) [64] of marginal R2 adjusted for the direction (+/-) of the relationship.

Multivariate analysis using random forests.

The results of the bivariate analyses were used to select the variables to be included in the multivariate analysis. First, we excluded variables poorly correlated with the response variables (p > 0.2). For each meteorological, microclimatic, air quality, land cover, and demographic variable, we retained the time lag or radii with the higher marginal R2. Multicollinearity was addressed using Pearson’s correlation coefficient with a threshold of 0.7. Then, based on empirical knowledge, an initial selection of correlated variables was made. Other collinear covariates were removed until the Variance Inflation Factor (VIF) of the remaining variables was below three [65] (S2 Fig and S1 Text). Binary classification random forests (RFs) and regression classification RFs were generated for the presence and abundance models, respectively. To improve model efficiency and reduce the number of explanatory variables, recursive feature elimination was performed for each model [66]. To account for spatial and temporal autocorrelation, sampling variables (trap location, sampling day, and sampling session) were included. For each set of hyperparameters and to assess the predictive power of each model, both a spatial and temporal cross-validation framework [67] was used, taking into account the potential spatial and temporal autocorrelation in the data. Spatial cross-validation (leave-area-out) involved training the model recursively on data from eight out of the nine sampled areas, testing it with data from the remaining area, and finally averaging the performance metric for the nine areas [61]. The temporal cross-validation (leave-sampling-session-out) involved training the model recursively on data from six out of the seven sampling sessions and testing it with data from the remaining session. The performance metrics selected were the Area Under the Curve (AUC) for the presence models [68] and the Mean Absolute Error (MAE) for the abundance models. The model results were interpreted using Variable Importance Plots (VIPs) [62] and Partial Dependence Plots (PDPs) [69]. Residual spatial autocorrelation was assessed for each model.

Softwares used.

The R version 4.2.3 programming language [70] was used. The land cover map creation and landscape metrics calculation involved the use of the ‘sf’ [71], ‘raster’ [72], ‘fasterize’ [73], and ‘landscapemetrics’ [74] packages. GLMMs were fitted using the ‘glmmTMB’ package [75]. Marginal R2 was calculated using the ‘performance’ package [76]. The ‘correlation’ package was used to evaluate multicollinearity [77]. Multivariate analyses were performed using the ‘caret’ [78] and ‘ranger’ [79] packages. The ‘CAST’ package [80] was used for the cross-validation and the interpretation plots were performed using the ‘iml’ [81], ‘pdp’ [82], ‘MLmetrics’ [83], ‘precrec’ [84] and ‘patchwork’ [85] packages. The package ‘spatialRF’ [86] was used to assess spatial autocorrelation. All scripts are available on Software Heritage [87].

Vegetation and other land cover variables.

Fig 2 shows the vegetation and other land cover variables correlated with the probability of presence or abundance of female Ae. albopictus. Two variables related to low vegetation were significantly and positively associated with the probability of presence within a 20 m radius buffer around the traps: the total length (in m) of low vegetation edges (OR =1.007, R2 = 0.042, p = 0.042, CI 95% [1.001; 1.014]) and the average patch size (in m2) of low vegetation (OR =1.01, R2 = 0.07, p = 0.006, CI 95% [1.003; 1.017]). The percentage of area occupied by buildings within a 50 m radius buffer around the traps was negatively correlated with the probability of presence (OR =0.959, R2 = 0.047, p = 0.048, CI 95% [0.920; 0.999]). Within a 20 m radius buffer around the trap, the total length edge of roads and the average size of a patch of roads were negatively correlated with the probability of presence (Fig 2).

Download:

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

Fig 2. Bivariate relationships between Ae. albopictus and land cover variables for varied buffer sizes.

Relationships are analysed based on Ae. albopictus presence or abundance and land cover variables for different buffer sizes around traps. The adjusted marginal R2 reflects the variance explained by the explanatory variable, adjusted for correlation direction. Boxes are colored if the p-value was < 0.2 (No asterisk: p-value ∈ [0.05; 0.2], *: p-value ∈ [0.01, 0.05], **: p-value ∈ [0.001; 0.01]; ***: p-value ∈ [0; 0.001]). Box color depends on the direction of the relationship (blue: negative, red: positive). Color intensity varies according to the marginal R2 value.

https://doi.org/10.1371/journal.pone.0335793.g002

Regarding abundance, positive associations were observed with all vegetation-related variables except for the average patch size of low vegetation within a 50 m radius buffer that showed a negative and significant association (RR = 0.996, R2 = 0.098, p = 0.0001, CI 95% [0.994; 0.998]). The most remarkable positive association was for the total edge length of low vegetation within a 100 m buffer (RR = 1.0008, R2 = 0.190, p = 0.007, CI 95% [1.0002; 1.0014]). Variables representing the building land cover category consistently showed negative relationships with female Ae. albopictus abundance, especially for the percentage area variable (across all buffer radii) and for all variables within the 100m buffer radius. RR for the percentage of building areas in the 100m buffer was 0.945 (R2 = 0.240, p = 0.0007, CI 95% [0.914; 0.976]) (Fig 2).

Microclimatic, meteorological, and air quality variables.

Fig 3 shows the relationship between presence or abundance and microclimatic or meteorological variables recorded from 48h before the start of sampling and during sampling. Local relative humidity (RH) variable showed positive associations with presence when recorded before sampling (with the strongest association for maximum hourly relative humidity 24-48h before sampling; OR = 1.194, R2 = 0.201, p = 0.002, CI 95% [1.070; 1.332]) and negative association when recorded during sampling (with the strongest association for minimum RH; OR =0.924, R2 = 0.134, p = 0.001, CI 95% [0.880;0.969]). Significant associations of RH data with abundance were negative whatever the period considered.

Download:

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

Fig 3. Bivariate relationships between Ae. albopictus and microclimatic and meteorological variables 48h before and during sampling.

Relationships are analysed based on Ae. albopictus presence or abundance and variables recorded 48h before the start and during sampling sessions. Microclimatic variables include minimum hourly relative humidity (rhmin), maximum hourly relative humidity (rhmax), mean hourly relative humidity (rhmean), minimum hourly temperature (tmin), maximum hourly temperature (tmax) and mean hourly temperature (tmean). Meteorological variables include daily cumulated rainfall, minimum atmospheric pressure (Patmin), maximum atmospheric pressure (Patmax), mean atmospheric pressure (Patmean), and the difference of mean atmospheric pressure between two consecutive days (Patdiff). The adjusted marginal R2 reflects the variance explained by the explanatory variable, adjusted for correlation direction. Boxes are colored if the p-value is < 0.2 (No asterisk: p-value ∈ [0.05; 0.2], *: p-value ∈ [0.01, 0.05], **: p-value ∈ [0.001; 0.01]; ***: p-value ∈ [0; 0.001]). Box color depends on the direction of the relationship (blue: negative, red: positive). Color intensity varies according to the marginal R2 value. bef=before.

https://doi.org/10.1371/journal.pone.0335793.g003

Local temperatures during and before sampling consitently showed positive correlations with presence, particularly for mean and maximum hourly temperatures recorded during sampling (tmean: OR =1.592, R2 = 0.408, p = 0.0002, CI 95% [1.251; 2.026]; tmax: OR =1.497, R2 = 0.402, p = 0.0001, CI 95% [1.226; 1.828]). Abundance was also positively associated with tmean (RR = 1.143, R2 = 0.197, p = 0.0002, CI 95% [1.064; 1.228]) and tmax (RR = 1.057, R2 = 0.061, p = 0.07, CI 95% [0.996; 1.122]) during sampling, but with lower R2.

Daily total precipitation was negatively associated with presence during sampling (OR =0.863, R2 = 0.137, p = 0.003, CI 95% [0.783; 0.951]). Aedes albopictus presence was positively associated with all variables describing atmospheric pressure recorded during the sampling hours, but negatively with lagged data. The strongest effect size was for the difference in mean daily atmospheric pressure (Patdiff) between the sampling day and the previous day (OR = 1.008, R2 = 0.194, p = 0.008, CI 95% [1.002; 1.014]). Average (patmean) and maximum (patmax) atmospheric pressure during the 24-48h period preceding sampling showed negative association with abundance (RR = 0.997, R2 = 0.243, p = 0.018, CI 95% [0.994; 0.999] and RR = 0.998, R2 = 0.177, p = 0.0001, CI 95% [0.996; 0.999], respectively).

Local temperatures during and before sampling consitently showed positive correlations with female presence, particularly for mean and maximum hourly temperatures recorded during sampling (tmean: OR =1.662, R2 = 0.439, p = 0.0001, CI 95% [1.298; 2.129]; tmax: OR =1.553, R2 = 0.424, p = 0.0001, CI 95% [1.262; 1.991]). Abundance was also positively associated with tmean (RR = 1.134, R2 = 0.197, p = 0.0001, CI 95% [1.055; 1.218]) and tmax (RR = 1.052, R2 = 0.050, p = 0.098, CI 95% [0.991; 1.117]) during sampling, but with lower R2.

Daily total precipitation was negatively associated with presence during sampling (OR =0.859, R2 = 0.133, p = 0.003, CI 95% [0.777; 0.949]). Aedes albopictus presence was positively associated with all variables describing atmospheric pressure recorded during the sampling hours, but negatively with lagged data. The strongest effect size was for the difference in mean daily atmospheric pressure (Patdiff) between the sampling day and the previous day (OR = 1.008, R2 = 0.217, p = 0.003, CI 95% [1.003; 1.015]). Average (patmean) and maximum (patmax) atmospheric pressure during the 24-48h period preceding sampling showed negative association with abundance (RR = 0.997, R2 = 0.217, p = 0.049, CI 95% [0.994; 0.999] and RR = 0.998, R2 = 0.183, p = 0.0003, CI 95% [0.996; 0.999], respectively).

Fig 4 shows the effect size and significance (p < 0.2) of weeks-lagged meteorological variables on the presence and abundance of female Ae. albopictus. The relationship between weekly cumulated rainfall (CUMRF) and both the probability of presence and abundance was positive, with a statistically significant association for abundance and the highest effect size observed the 6th week before sampling (for presence: OR = 1.430, R2 = 0.334, p = 0.168, CI 95% [0.860; 2.375]; for abundance: RR = 1.075, R2 = 0.127, p = 0.04, CI 95% [1.003; 1.152]). The probability of presence was positively and significantly associated with minimum (TMIN), average (TMEAN), and maximum (TMAX) temperatures. The highest effect size was observed for TMIN one week before sampling (OR =1.620, R2 = 0.372, p = 0.001, CI 95% [1.209; 2.172]. For abundance, relationships with TMIN, TMEAN, and TMAX were also positive and significant for time interval 0–2 weeks before sampling, with the highest effect size observed for TMAX (RR = 1.161, R2 = 0.192, p = 0.0006, CI 95% [1.067; 1.267]). Both responses were positively and significantly correlated with the weekly growing degree days (GDD). The highest effect on presence and abundance was observed for GDD during the week preceding sampling (OR = 1.085, R2 = 0.357, p = 0.002, CI 95% [1.031; 1.141]) and for time interval 0–2 weeks before sampling (RR = 1.011, R2 = 0.189, p = 0.0009, CI 95% [1.004; 1.017]), respectively. Both responses were negatively and significantly correlated with relative humidity (RH), with the highest effect observed for RH recorded 6 weeks before sampling (OR = 0.792, R2 = 0.472, p = 0.006, CI 95% [0.670; 0.937] for presence; RR = 0.948, R2 = 0.183, p = 0.007, CI 95% [0.912; 0.986] for abundance). The relationship between wind speed (WS) and abundance was negative, with the highest effect observed for time interval 0–5 weeks before collection (RR = 0.343, R2 = 0.086, p = 0.157, CI 95% [0.078; 1.510]).

Download:

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

Fig 4. Cross correlation maps showing the lagged effect of meteorological variables on Ae. albopictus presence or abundance.

Lagged meteorological variables include weekly cumulated daily rainfall (CUMRF), mean daily temperature (TMEAN), daily minimal temperature (TMIN), daily maximal temperature (TMAX), the weekly Growing Degree Days (GDD), mean daily relative humidity (RH), and mean daily wind speed (WS). Time lags are expressed in weeks prior to sampling. The adjusted marginal R2 reflects the variance explained by the explanatory variable, adjusted for correlation direction. Red-bordered squares highlight the time lag with the highest marginal R2, with the interval indicated in the top left corner. Black-bordered squares denote marginal R2 values close to the highest (within 10%). Gray squares represent correlations with p-values > 0.2.

https://doi.org/10.1371/journal.pone.0335793.g004

Daily nitrogen dioxide (NO2) concentrations were positively associated with the probability of presence of female Ae. albopictus, with the highest effect size observed for the sampling week (OR = 1.109, R2 = 0.126, p = 0.163, CI 95% [0.959; 1.282]; S5 Fig). Particulate matter concentrations (PM2.5 and PM10) were also positively correlated with the probability of presence and with the abundance. However, PM2.5 and PM10 concentrations were strongly correlated with all time-lagged temperature metrics (Pearson coefficient > 0.8). Tropospheric ozone (O3) concentration (three weeks before sampling) was negatively correlated with the probability of presence and positively (one week before sampling) with abundance. Daily O3 concentrations were also highly correlated with weekly cumulated rainfall, wind speed, and daily relative humidity (Pearson coefficient > 0.8).

Socio-demographic variables.

Human population density in a 250 m radius buffer around traps was positively correlated with the Ae. albopictus probability of presence (OR=1.0013, R2 = 0.056, p = 0.029, CI 95% [1.0001; 1.0026]; S6 Fig).

Presence model.

The most contributing variables to presence probability were a combination of real-time and time-lagged microclimatic and meteorological variables (Fig 5A), with only a minimal contribution of vegetation variables. The first contributing variable was the local minimum hourly temperature recorded 24–48 hours before sampling. The probability of presence showed a slight increase between 11.9°C and 16°C, rose sharply from 16°C to 25°C, and plateaued beyond 25°C (Fig 5B). Minimum atmospheric pressure during sampling was the second most influential variable, with a unimodal relationship: presence probability increased up to 1012.91 HPa before declining. Cumulated rainfall 6 weeks before sampling emerged as the third key variable, displaying a linear positive association with presence probability, from 0 mm to 15 mm. The fourth major factor was the mean daily atmospheric pressure difference between the sampling and previous day. Presence probabilities decreased for descreasing pressures and were stable for stable or increasing pressures. Secondary contributors included mean local hourly relative humidity 24–48 hours before sampling, daily NO₂ concentrations, and vegetation metrics within a 20 m buffer around traps (average patch size of low vegetation, number of high vegetation patches, and total length of low vegetation edges). These variables slightly increased presence probability, whereas the total length of edges of roads in a 20 m radius buffer weakly reduced it. The model demonstrated strong predictive accuracy (AUC = 0.89; S7A Fig). Cross-validation with the leave-area-out method produced consistent predictions, except in the Diderot School area (IMP-DID). In contrast, leave-sampling-session-out cross-validation exhibited greater variability between sampling areas (S7B Fig).

Abundance model.

The most contributing variables to Ae. albopictus abundance comprised a combination of microclimatic, weeks-lagged meteorological, vegetation and other land cover variables. The first contributing variable was the average size of low-vegetation patches within a 50 m radius buffer around traps (Fig 6A). Abundance increased markedly for areas between 0 and 120 m2, plateaued up to 237 m2, and subsequently decreased slightly (Fig 6B). The second most important variable was the local maximum hourly temperature during sampling. A strong positive correlation with abundance was observed between 24°C and 38.7°C, followed by a negative correlation beyond 40.6°C. The third key variable was the total length of road edges within a 250 m buffer around traps, with a strong negative correlation between 8230 m and 11 000 m. The fourth contributor was the percentage of low vegetation within a 250 m buffer, which showed a positive effect between 10% and 23%, beyond which it had no observable effect. The fifth and sixth contributors were the cumulated rainfall in the sixth week before sampling and the weekly GDD between the sampling day and two weeks prior, both of which were positively associated with mosquito abundance. The average wind speed between the day of sampling and five weeks before had no effect on abundance between 3.7 m/s and 4.3 m/s, but showed a negative correlation beyond this range. Finally, the last two contributors were the average patch size of roads within a 250 m buffer and the maximum hourly relative humidity during the 24 hours preceding sampling, both of which were slightly negatively correlated with mosquito abundance.

The abundance model demonstrated good overall accuracy, except for high abundance cases (>11 mosquitoes, Mean Absolute Error (MAE)=8.00, S8A Fig). Spatial cross-validation retained overall abundance trends, except in RES-LEM and PRK-AIG, and showed an underestimation of high abundance and an overestimation of low abundance. Temporal cross-validation, using leave-sampling-session-out, did not preserve abundance trends but captured relative differences in Ae. albopictus abundance between areas (S8B Fig).