Study areas
The study took place in the cities of Basel, Lausanne, Lugano and Zurich, in Switzerland. Basel, Lausanne and Zurich are located north of the Alps, in the geographical region of the Central Plateau (Supplementary Fig. S1). This region stretches from Lake Geneva in the southwest to Lake Constance in the northeast and is the most densely populated region in Switzerland. Zurich is the largest city of Switzerland and encompasses 88 km2 with a total human resident population of 420,21741. Lausanne and Basel are smaller than Zurich, with a surface of 41 and 24 km2 and a total population of 139,408 and 173,232, respectively41. The climate in these three cities is moderately continental, with cold winters often reaching freezing temperatures in January, and warm summers. Lugano is located in Ticino, south of the Alps (Supplementary Fig. S1), where the climate is strongly affected by the Mediterranean Sea, with mild winters and summers warm and humid, sometimes hot. Lugano is the smallest of the four cities with 50,603 residents in 26 km241.
Aedes albopictus is well established in Lugano since 2009 and an integrated vector management is constantly implemented to contain the numbers of the mosquito at a manageable level. This consists of an intensive surveillance, with oviposition traps distributed according to a grid system, several control interventions, such as the removal of breeding sites and the systematic application of larvicides in public areas, mainly in catch basins, and extensive public information campaigns24,26. In Basel, two populations of Ae. albopictus are established since 2018: a first population in an area adjacent to the motorway toll on the border with France and a second population in an area near the border with Germany27. The mosquito has also been recorded repeatedly at various locations in the city of Basel and the surveillance indicates that the mosquito is spreading42. Control actions are taken exclusively within the perimeter of repeated detections of the mosquito and include regular treatment of catch basins with larvicides, distribution of flyers and door-to-door information campaigns42. In Zurich, Ae. albopictus was first detected in 2016 in a bus station for international coach services located in the centre of the city, near the main train station. Thanks to immediate surveillance and control actions (i.e., treatment of catch basins in the area with larvicides), to date there is no established population within the perimeter of the bus station despite continuous repeated introductions40. A small population was also detected in 2018 in a suburban neighbourhood in the Wollishofen district of Zurich, approximately 5 km southwest from the international bus station. Also in this case, immediate surveillance and control actions, including larval control and door-to-door information, were taken with success and no adults, eggs or aquatic stages have been found in 2020 and 202140. In Lausanne, no tiger mosquito has been reported to date (Swiss Mosquito Network, http://www.mosquitoes-switzerland.ch (accessed on 17 February 2022)).
Microclimate data
Based on a previous investigation we conducted in Ticino, Basel and Zurich20, we focused the microclimate monitoring on ordinary stormwater catch basins positioned on the side of public roads. In each city, we monitored ten catch basins located either in urban context (defined as areas with high-density development, consisting of apartment blocks, commercial or industrial units) or in residential areas consisting mainly of houses with private gardens located in peri-urban area (Supplementary Table S1, Supplementary Fig. S2). The catch basins were usually homogeneous in dimension, in the same city, although we recorded variations in depth. In Basel, we included catch basins located in the urban area near the border with France, in which Ae. albopictus is established. In Zurich, we included catch basins located in the international bus station, where Ae. albopictus was recorded in summer, and in the residential area of Wollishofen, where a small population of Ae. albopictus was detected and then likely eradicated. In Lausanne, some catch basins were selected in potential points of introduction of the mosquito (e.g., near a campsite, the main train station, etc.). In Lugano, Ae. albopictus was established in all the locations selected.
A sensor device was installed in each selected catch basin. The sensor devices were built in house. The development of the devices and the Wireless Sensor Network (WSN) has been described in detail by Strigaro et al.29. Briefly, the device consisted of a waterproof plastic box containing a LoPy Micro-Controller Unit (Pycom, Guildford, United Kingdom), a waterproof temperature probe (accuracy of ± 0.5 °C), a light sensor (measuring illuminance arriving at the sensor device, in lux), an SD card, the rechargeable batteries and other parts. The main box, with the light sensor, was hung on the inside wall of the catch basin. The temperature probe was attached to the wall at a depth ranging from 0.3 to 0.5 m, depending on the depth of the catch basin and the level of the water in the catch basin. The probe was placed in direct contact with the inside wall of the catch basin, in order to measure the microclimatic conditions where the mosquito eggs are potentially laid. The data collected was transmitted to a data warehouse based on istSOS, an open-source Python based implementation of the Sensor Observation Service standard (SOS) of the Open Geospatial Consortium (OGC)43. The data was transmitted through the Swisscom Low Power Network (LPN) LoRaWAN (Swisscom Ltd, Ittigen, Switzerland): the data sent by the sensor devices was received by a Swisscom Gateway and then sent to the data warehouse29.
In addition to the sensor devices installed in the catch basins, four devices were installed outside four catch basins in each city, except in Lugano, where three devices were installed. These external devices were placed in vegetation representing potential resting habitats for Ae. albopictus adults in the reproductive season, at 1–2 m above the ground and analyzed to confirm the close similarity between measured external temperatures and MeteoSwiss gridded temperature data. However, since the main goal of the data collection was to model the differences between MeteoSwiss gridded temperature data and catch basins’ temperatures, only a small number of external sensors were deployed. Microclimate data were collected from beginning of December 2019 to end of February 2020, a period defined as cold season, with acquisition interval set at one hour. In Lugano, data collection started on the 12th or 13th of December 2019.
Local climate data
We used two types of local climate data. The first type is the momentary hourly free-air temperatures recorded at 2 m above ground level by permanent weather stations. The weather stations belong to SwissMetNet, the automatic monitoring network of MeteoSwiss. For each city, we selected the weather station closest to the study area (Supplementary Table S1, Supplementary Fig. S2) and temperature data were retrieved from https://gate.meteoswiss.ch/idaweb (source: MeteoSwiss, Zurich-Airport, Switzerland; accessed on 12 August 2021).
The second type of local climate data is the MeteoSwiss spatial climate daily datasets (source: MeteoSwiss). These temperature datasets are constructed through interpolation of daily minimum, maximum, and mean temperatures from a network of approximately 90 SwissMetNet permanent weather stations to a 1 km resolution grid in the Swiss coordinate system CH190344,45. This results in three temperature datasets describing the km-scale distribution of day-to-day temperature variations in Switzerland. We referred to them as gridded temperature data. Each monitored catch basin and external device was assigned, based on its geographical position, to the corresponding 1 km × 1 km cell of the climate grid. Each cell was identified with its MeteoSwiss (MS) number (Supplementary Table S1).
Data analysis
The hourly temperatures were used to compute daily mean, maximum and minimum temperatures and daily temperature ranges, which were calculated as the difference between the maximum and minimum daily temperature. Temperatures of catch basins and external habitats were compared to temperatures of permanent weather stations and to the gridded temperatures both graphically and using the nonparametric Mann–Whitney U-test, for which a P value of < 0.05 was determined as significant. For these calculations and comparisons, we used the softwares Microsoft Excel (Microsoft Corporation, Redmond, USA) and IBM SPSS Statistics for Windows, Version 26.0 (IBM Corporation, Armonk, USA).
The relation between catch basins and gridded temperatures was further investigated, using scatterplots, for each data series (i.e., daily average, minimum and maximum temperatures). The association between each temperature series obtained from the sensors of the WSN and the corresponding gridded temperature series was analyzed independently for each catch basin using a simple linear regression model:
$$ T_{i} = m_{i} T_{MS} + q_{i} , $$
where ({T}_{i}) is the temperature of the i-th catch basin, ({T}_{MS}) the gridded temperature of the corresponding cell and ({m}_{i}) and ({q}_{i}), respectively, the slope and intercept of the linear model. For comparison, similar models were built also for the relation between gridded temperatures and weather stations and external sensors temperature measurements. The coefficients (slope and intercept) of the individual catch basin microclimate models were then plotted on a scatterplot to compare their distribution across cities. The differences between these coefficients in the different cities were also jointly analyzed using a MANOVA hypothesis test.
After the independent analysis of each individual catch basin, we built, for each temperature series, a global model of the relation between gridded temperatures (predictor) and catch basin temperatures (response). To account for the difference between catch basins and between cities, we considered a hierarchical model grouping data based on the city (first level) and the catch basin (second level). We assume random effects intercept and slope both for the city and the catch basin predictors. Let yc,m be the daily temperature (either minimum, maximum or average) observed in catch basin m of city c, and x the temperature from the corresponding cell of the gridded temperature model. The resulting mixed effect model can be written as follows
$$ y_{c,m} = a_{0} + b_{0} x + a_{c} + b_{c} x + a_{m} + b_{m} x , $$
where ({a}_{0}) and ({b}_{0}) are the fixed effects, ({a}_{c}) and ({b}_{c}) the random effects associated to the city (c) and ({a}_{m}) and ({b}_{m}) the random effects associated to the individual catch basing (m). To further confirm the presence of a significant effect of cities on catch basins temperatures, we computed the normalized relative likelihood30 of the mixed effects model including both the city and catch basin random effects versus the mixed effects model including only catch basin random effects.
To describe the average relation between microclimatic and gridded temperatures in each city, we dropped the terms associated to the catch basin random effects, i.e., ({a}_{m}) and ({b}_{m}x), and produced from the available gridded temperatures, estimates of the average catch basin temperatures in each cell of city (c) as follows:
$$ y_{c} = a_{0} + b_{0} x + a_{c} + b_{c} x. $$
(1)
The corrected temperatures were incorporated into the previously developed ecological niche model, which has been described in detail in Ravasi et al.28. Briefly, the ecological niche model consists in an ensemble of Lasso-regularized logistic regression models that use the long-term dataset of Ae. albopictus presence–absence records in Ticino from 2005 to 2012 as response variable in the training process. The model evaluated 79 explanatory variables, which included indicators of terrain morphology, land use coverage, total human population, meteorological temperatures and precipitations and travel distance (by car) from the nearest 200 m × 200 m cell with Ae. albopictus establishment. The performance of the model in predicting the probability of establishment of Ae. albopictus in Ticino was assessed both by cross validation over the training years 2007 to 2012 and on the test years 2013, 2014, and 2015. On the year 2007 to 2012, which represent the period of diffusion of Ae. albopictus in Ticino, the model obtained an AUC above 0.85, whereas on years 2013 to 2015, which refers to a period were mosquitoes were well established (less interesting, therefore, for the purpose of predicting establishment), the AUC was almost 0.75. The ensemble model was then used to extrapolate the probability of establishment in the cities of Basel and Zurich starting from the initial observations of mosquitoes’ establishment collected in 2019, and using meteorological data from the years 2015, 2016, 2017 and 2018 as scenarios for the meteorological conditions in the coming years28. For each of the four scenarios and each of the eight models of the ensemble, a risk prediction was produced for each 200 × 200 m cell of the areas of interest. Finally, a single risk prediction for each cell and a measure of its uncertainty were obtained, respectively, as the average and the standard deviation of these 32 risk estimates and projected onto a suitability map showing the average risk estimates and an uncertainty map showing the standard deviations.
Ten explanatory features were identified as most informative for the prediction of establishment of Ae. albopictus in Ticino28, by at least half of the eight ensemble models; among them, two were related to cold-season temperatures. It appeared therefore important to account for the results obtained from the microclimate analysis when predicting the risk of the establishment. To incorporate these results, gridded temperatures relative to the cold season should have been transformed based on the model in Eq. (1) before deriving the cold-season temperature predictors used in Ravasi et al.28. However, since these predictors are linear functions of the original temperature series, the linear transformation in Eq. (1) can be applied directly to them. The natural approach to account for cold-season catch basin temperatures would then be the following: (i) transform the predictors related to the cold-season temperatures in the training dataset from Ticino using Eq. (1) with the random effect coefficients ({a}_{c}) and ({b}_{c}) estimated for c = Lugano; (ii) re-train the individual models of the environmental niche ensemble model using the transformed training dataset; (iii) transform the cold-season temperature related predictors for Basel, Lausanne and Zurich of the years 2015, 2016, 2017 and 2018 (i.e., the reference years used to develop the risk maps) using Eq. (1) with the random effect coefficients ({a}_{c}) and ({b}_{c}) estimated for c = Basel, Lausanne, Zurich; (iv) apply the model trained in (ii) to the dataset including the transformed cold-season temperature predictors for Basel, Lausanne, Zurich to obtain corrected risk predictions of Ae. albopictus establishment in those cities. However, since in Ravasi et al.28 the original features were rescaled in the range [0, 1] (minmax scaling) before learning the individual models of the ensemble, steps (i) and (ii) would not affect their parameters nor change the set of predictors selected but would only result in different values of the minmax scaler parameters related to the cold-season temperature predictors. Therefore, we simplified the above procedure by skipping steps (i) and (ii) and modifying step (iii) as follows. First, the cold-season temperature related predictors for the cities of Basel, Lausanne or Zurich are converted into average catch basin temperatures using Eq. (1) with the random effect coefficients ({a}_{c}) and ({b}_{c}) of each city; then, the average catch basin temperatures estimated for Basel, Lausanne and Zurich are inversely transformed according to Eq. (1) using, this time, the random effect coefficients ({a}_{c}) and ({b}_{c}) found for Lugano. Such transformed predictors should be interpreted as the gridded cold-season temperatures that, in Lugano, would produce (estimated) catch basins condition equal to those obtained for Basel, Lausanne or Zurich in correspondence of the gridded cold-season temperatures considered for these cities (i.e., those observed in the years 2015, 2016, 2017 and 2018).
For Basel and Zurich, the feature “car distance to establishment” was computed using the records of positive oviposition traps for 2019 (37 points of first establishment for Basel and 16 for Zurich; data kindly provided by the Inspection body for chemical and biosafety (KCB) of the Cantonal Laboratory of Basel-Stadt, the Swiss Tropical and Public Health Institute, the Urban Pest Advisory Service of the City of Zurich, and the Office for Waste, Water, Energy and Air (AWEL) of the Canton of Zurich). The feature was built by computing the smallest distance between each cell of Basel or Zurich and the identified position of 2019 establishments. In Lausanne, Ae. albopictus has not been observed yet. Therefore, it was not possible to use positive oviposition traps to compute the car distance to establishment feature. Coherently with what was done in Ravasi et al.28, we kept a fixed car distance to establishment of 0.5 min (30 s) for all cells, implying that establishment has already occurred in the neighborhood. This provided an overview of the areas in Lausanne which are mostly at risk in case a first colony of mosquitoes arrives.
Four types of maps were produced for each city, by projecting the results onto Swiss national maps (Federal Office of Topography swisstopo): (1) a suitability map estimated from the original temperatures (not shown here); (2) a suitability map based on the transformed temperatures; (3) a map showing the difference between risk estimates in (1) and (2); (4) an uncertainty map for the predictions obtained using the transformed temperatures, providing an indication about the reliability of the risk estimates displayed in the suitability map.
Scatterplots, risk maps, MANOVA tests, and regression models were created in Python (version 3.8.12, Python Software Foundation, Wilmington, United States) using packages matplotlib, statsmodels.
Source: Ecology - nature.com
