Climate / Environment and Dengue

Remote-sensing environmental data for mapping entomological dengue risk levels in Martinique.

Vanessa Machault (1), André Yébakima (2), Manuel Etienne (2), Cécile Vignolles (3), Philippe Palany (4), Yves M. Tourre (5), Marine Guérécheau (1), Jean-Pierre Lacaux (1)

(1) Laboratoire d’Aérologie, Observatoire Midi-Pyrénées (OMP), Université Paul Sabatier, Toulouse, France (2) Service de Démoustication et de Lutte Anti-vectorielle, Conseil Général de la Martinique/Agence Régionale de Santé (SD-LAV), Fort-de-France, Martinique, France (3) Direction de la Stratégie et des Programmes/Terre-Environnement-Climat, Centre National d’Etudes Spatiales (CNES), Toulouse, France (4) Météo-France Direction Inter-Régionale Antilles-Guyane, Fort-de-France, Martinique, France (5) Lamont-Doherty Earth Observatory (LDEO) of Columbia University, Palisades, New York, USA

Keywords: dengue, remote-sensing, risk mapping, Aedes aegypti, entomology

Abstract

Background

Whilst in Martinique dengue is endemic, six epidemic waves did occur during the last 20 years. There is no specific treatment for dengue fever as yet and no operational vaccine is available. The only mean for controlling virus transmission is thus to effectively combat the mosquitoes. Risk maps at appropriate scales can then provide surrogate data and valuable information for a spatio-temporal assessment of entomological risk levels in order to improve dengue control efficiently.

Methodology/Principal Findings

In general, the tele-epidemiology practical concept applied here, relies on the facts that: i) a so-called experimental unit, or the smallest ‘object’ to observe must be characterized first, to properly assess the risk levels for a given disease, ii) environment and weather conditions may explain the spatio-temporal distribution and variability of diseases risk and ii) datasets from satellite images may be included into statistical models. The experimental unit for dengue vectors has been chosen here as a dwelling and its nearby surroundings. Some of its characteristics have been found to be risk or protective factors for the presence of Aedes aegypti immature stages. Those factors were environmental conditions (i.e., surface of “sparsely vegetated area”, “lawn” and “asphalt”) and weather conditions (i.e., rainfall, relative humidity). In addition, remote-sensing environmental data are used to produce dynamic high spatio-temporal resolution maps for the presence of various containers harboring the Aedes aegypti. It s found that the spatio-temporal variability of the entomological dengue risk levels may be assessed in Martinique (French Antilles).

Conclusions/Significance

The produced risk maps are the first examples of modeled entomological maps at the housing level with daily temporal resolution. This finding is an important component for preventive actions and contribution to developing operational warning systems for other vector-borne diseases such as the recently identified chikungunya in Martinique.

Introduction

Dengue is an infectious disease caused by one of the four serotypes (DEN-1 to DEN-4) of the dengue virus. It is transmitted to humans through bites from already infected female mosquitoes of the genus Aedes in urban areas. Even if mortality rate is low among human population, dengue is considered as one of the most important mosquito-borne viral disease. This is due to its extensive geographic spread, the societal cost of its burden from the 50- to 200- million annual infections and the now identified 125 endemic countries [1].

In Martinique (French Antilles), Aedes aegypti mosquito is so far the single identified vector for spreading the dengue virus. It breeds mostly in artificial domestic or peridomestic containers filled with clean water with little organic debris and low concentration of inorganic nutrients [2, 3]. Breeding sites/containers include flower pots with saucers, mounts of detritus and debris, abandoned cars and tires, badly maintained gutters, discarded old domestic appliances that may be all filled-up naturally with rainfall. In addition drum barrels may be deliberately placed under gutters or in yards to collect rain water for watering/cleaning purposes. All of the above conditions and mechanisms co-exist and favor the local environment for potential breeding sites.

Whilst in Martinique dengue is endemic, six epidemic waves did occur during the last 20 years. The penultimate epidemic, in 2010, from February to the end of the year, led to more than 41 000 clinical cases (about 10% of the island population). This outbreak, as well as the 2001 epidemic, has started in Tartane, a village located in the Caravelle Peninsula, North-East of Matinique. Entomological analyses have provided some clues for understanding the key elements for the beginning of the 2001 outbreak: higher Aedes density, with high survival rate and a gonotrophic cycle that were both favorable for virus transmission [4].

There is no specific treatment for dengue with no operational vaccine as yet [5]. The only mean for controlling virus transmission is thus to effectively combat the mosquitoes. Several techniques/strategies are now available for field testing based upon a biological approach (e.g., mosquitoes infection with Wolbachia bacteria that may interfere with dengue virus and reduce mosquitoes lifespan [6]) or a genetic approach (e.g., sterile insects techniques [7,8]. Nevertheless, source reduction of the potential larvae habitats and/or reduction of the emergence of adult specimens have long been accepted as part of dengue control strategy. Works have been done to identify the control thresholds to be attained in order to suppress transmission threats [9]. In this context, a good knowledge of the entomological conditions in a given area and a given period is thus a prerequisite. Unfortunately entomological data are seldom collected continuously. Moreover when field studies are undertaken, they often only provide a snapshot of a somewhat continuous phenomenon. Risk maps at appropriate scales can then provide surrogate data and valuable information for having a spatio-temporal evaluation of entomological risk levels in order to improve dengue control efficiently.

Informative maps of dengue vectors/cases, can be perceived as tools for: i) delivering vectors and dengue statistics, ii) driving vector control around dengue cases during outbreaks (or within risk areas identified during past epidemics) and, iii) studying space and space-time clustering of cases during outbreaks [10]. Spatial information should also allow modeling the linkages between vector presence and/or dengue cases associated with environmental and socio-economical variables [10]. Additional linkages with meteorological/weather variables should also be included to better understand the mechanisms at stake.

Risk maps have been produced for numerous diseases and for mapping a current situation or even anticipating outbreaks through Early Warning Systems (EWS) [11]. While mapping dengue at global scales is relevant for determining populations at risk, fine scale or local mapping is of major interest for setting-up local control strategies particularly when resources are limited [12]. From global/regional to local scales, the spatial and temporal distribution of vectors as well as dengue epidemiological characteristics may fluctuate along with weather/climate conditions (i.e., mainly rainfall amount, relative humidity and temperature), the environment (i.e., vegetation, soil types, among others) or human activities (i.e., human migration and local movements, transportation, urbanization or waste management procedures, among others). Modeling dengue may benefit from the use of environmental remote-sensing information. In the recent past, satellite products have been proven to provide useful information for modeling Aedes aegypti or Aedes albopictus distribution [13-21], human cases distribution [22,23] or potentialities for vectors or disease future expansion [24-26].

The present study consists essentially in mapping dengue entomological risk i.e., evaluating risk of presence of vectors, in contrast with mapping epidemiological risk. The practical and conceptual approach (CA) called tele-epidemiology (Marechal, Ribeiro et al., 2008) could then be applied to the spatio-temporal mapping of entomological dengue risk in urban settings in Martinique. This CA has been developed and patented by the French Spatial Agency (CNES) and its partners [27, 28]. It consists in monitoring and studying human and animal diseases spatio-temporal dynamics which are closely related to weather/climate and environment variability. It relies on the identification of an experimental unit (EU) that is the smallest ‘object’ that has to be identified/characterized in order to assess properly the levels of risk. This unit is based on the sound knowledge of the biological and physical processes that underline the presence/densities of immature and adult vectors. It is thus widely dependent upon the disease under study. For example this experimental unit will be a pond ( 1 ha) when studying Rift Valley Fever entomological risk [29] and a water body or aggregates of small water bodies ( 1/10 ha) when studying urban malaria entomological risk [30]. Then appropriate choices of satellite data and dynamical models must be assessed, along with extensive use of in-situ measurements.

Three hypotheses underpinned the present study. Firstly, if the potential breeding habitats could not be directly detected using satellite images even at very high spatial resolution, their specific environment could be mapped. Indeed it is thought that dwelling conditions, neighborhood and yard conditions may be associated with the presence of various containers retaining water. Indeed they provide information on local habits regarding maintenance of private yards/gardens or all kind of water storage. Secondly, estate characteristics i.e., shading and tidiness of a house and its yard, have been identified as determinants for the presence/abundance of Aedes aegypti immature stages and eggs [20, 31-33]. It is thus thought that characterizing fine scale environment could inform on the presence of the dengue vector at the immature stages in areas where containers are present. Thirdly, meteorological conditions partly drive the temporal dynamics for containers filling, as well as entomological dynamics (e.g., eggs hatching, larvae development). Consequently, the experimental unit here has been defined as being the house level with its nearby environment. The state of such units was then described with details at a given date.

The main objective here was to model in space and time houses so-called ‘positive’ for Aedes aegypti immature stages from 2009 to 2011 in Tartane (Martinique, French Antilles), by using remote-sensing environmental data and in-situ meteorological information. This was to produce high spatio-temporal resolution predictive dengue entomological risk maps.

Methods

Studied site and period

The city of Tartane (14°45’29.24“N, 60°55’10.56” W) belongs to the Caravelle Peninsula, North-East of Martinique. It is historically a fishing cove, with small and low-rise dwellings surrounded by small gardens or yards. The centre of the city is near the sea-side while other sections are located uphill. The population is of 3, 000 inhabitants. It is a tourist attraction, and includes many vacation sites. The studied site is about 8 km² (see black rectangle in Figure 1). This is exactlyl where the 2010 dengue outbreak started. Although dengue epidemics “usually” last from July to December, viral circulation in 2010 had been observed from early February, with peak in June and lasted until the end of the year. The studied period ran from June 2009 to August 2011, in order to include the 2010 epidemic and detect any heterogeneous temporal drivers.

Entomological data

In Martinique, antivectorial control and mosquito nuisance are managed by a public organism called « Service de Démoustication et de Lutte Antivectorielle » (SD-LAV). Since 1991, the SD-LAV collects information on dengue vectors with an additional effort during outbreaks. A sample of the database records has been selected for this study. Each record was associated with a given house visited at a given date. It contained information on the number and type of domestic and peridomestic containers as well as their status concerning the presence or not of Aedes aegypti immature stages (i.e., all larvae stages and pupae). Surveyed houses were consequently plotted using a Global Positioning System (GPS) device. Types of all-size containers were: drum barrels, tanks, waste-bins, flower pots and saucers, gutters, tires, discarded appliances and pools. Sampled houses are positioned in Figure 2. From their spatial distribution, six sections were identified (Figure 1).

Meteorological data

In Martinique, the summer rainy season (July to November) - with frequent and heavy rainfall and maximum temperatures of about 32°C - and the dry season (February to April) - with maximum temperatures of about 30°C - are separated by two intermediate seasons.

The year 2009 has been exceptionally hot, with low rainfall except on the Atlantic coastline. In 2010, temperatures were also high particularly during February and March. February was almost completely dry. Heavy precipitation started in early June followed by a very dry period. Again, the year 2011 experienced hot temperatures while it has been the wettest of the 2009-2011 period, with basically no dry season.

Daily temperature and humidity (i.e., minimum, maximum and mean) as well as precipitation amount, water loss and water balance were provided by Météo-France. They were recorded at the observing station located in the Caravelle Peninsula. Yearly precipitation recorded during 2009, 2010 and 2011, were respectively of 948 mm, 1, 408 mm and 1, 823 mm.

Several variables have been calculated from the above data:

  • Total rainfall amount for the 2- to 30- days period before each entomological ground investigation date;
  • Number of days without rain for the 5- 10- 15- 20- 25- 30- days period before each entomological ground investigation date;
  • Average of temperature, relative humidity, water loss, water balance for the 2- to 15- days period before each entomological ground investigation date;
  • Periods when temperature/relative humidity/water loss/water balance were above some thresholds (e.g., number of days when maximum temperature exceeded 32°C), during the 15- days period before each entomological investigation date. Those thresholds were defined as the quintiles for the variables distribution.

The above data was added to the entomological database, matched by the date of the ground surveys. No spatial heterogeneity was included here since the meteorological variables were identical for all sampled houses located within the 8 km² area, with no specific reference to their location on the peninsula.

Environmental data and satellite images

A Geoeye-1 image, with clear sky and for 13/03/2011 was acquired. Data included four spectral bands at 0.41-m spatial resolution (blue, green, red and near infrared). The image was projected in WGS 84, UTM, Zone 20N and geometrically corrected using the 50-m spatial resolution elevation map (IGN BD ALTI®) from the French National Geographic Institute IGN (Institut National de l’Information Géographique et Forestière). Image processing was done using ENVI 4.8 and ENVI EX (Exelis Visual Information Solutions). Other available geographic data was: IGN topographic map (IGN BD TOPO®) and cadastral map (IGN BD ADRESSE®). Elevation, slope and objects height maps at 1-m spatial resolution were available through Litto3D® (IGN, Service Hydrographique et Océanographique de la Marine, Direction Régionale de l’Environnement, de l’Aménagement et du Logement - Martinique, Agence des Aires Marines Protégées).

Three vegetation and soil indicators were derived from the Geoeye-1 image. They are described in Table 1 [34], [35] and [36]. A three-step classification procedure allowed map production of Land Use and Land Cover (LULC) from the Geoeye-1 image. An object-oriented classification and a supervised pixel-based classification (maximum likelihood) were undertaken. As each technique proved to accurately highlight some basic elements of the LULC (by not discriminating others) a synthetic classified image was generated using the power of both classifications. For each step, validation of the classifications quality was assessed by photo-interpretation done by experts whilst some manual corrections were applied. The final classification included fourteen land-cover classes: i.e., 5 for vegetation such as “trees”, “sugar cane”, “stubbles”, “grass”, “sparsely vegetated soil”, different types of roofs (5), “sand”, “asphalt”, “swimming pools” and “ocean”.

Geographic Information System (GIS)

A GIS was built using ArcGIS 10.0 (Environmental Research Systems Institute, Redlands, CA, USA) in order to characterize the experimental units defined in the Introduction. All surveyed houses were plotted whilst environmental indicators, LULC map derived from GeoEye-1 image and elevation indicators were added as geo-referenced layers. Each single house for the Caravelle Peninsula was isolated as an object, based on the LULC map. The plot around each house was obtained from the cadastral map.

The environmental variables, i.e., the minimum/maximum/mean for the three indices, altitude, slope and object elevation, as well as the areas from each LULC class, were computed for each experimental unit. This was accomplished for each estate and for a 50-m and 100-m radius buffer zones around the individual houses. The Euclidian distance from the house to each LULC class was also calculated, as well as the inside footage of houses (assumed equal to the area of their roofs) and plots. Those data were merged with the entomological database, again for each house. When records concerned a specific house but surveyed for different periods, the same environmental data was used.

Modeling strategy

The overall database with entomological, environmental and meteorological data contained more than 300 variables: 96 variables obtained from the Geoeye-1 image, 36 altimetric variables and more than 200 meteorological variables. Each record in the database (i.e, observation) was associated with one house visited at one time.

The chosen scenario (Figure 3) included the following two steps which involved investigation of the environmental and meteorological factors that drove:

  • Step 1: the presence of one or several water-filled container(s) in a vicinity of a house at a given date, independently of the presence or not of Aedes aegypti immature stages. This was the detection of the water-filled-positive experimental units;
  • Step 2: the presence of Aedes aegypti immature stages, only in the experimental units that held one or several water-filled container(s). This was the detection of the Aedes larvae-positive experimental units. No reference to the larval density was included.

Since the number of domestic water containers was very low in the area, only the peridomestic containers were considered.

Statistical analysis and risk mapping

Statistical analyses were performed using Stata 11 (Stata Corporation, College Station, Texas, http://www.stata.com. Note that Stata 13 has just been released). Logistic regressions for explaining the outcomes from both steps above were fitted at the experimental unit level using all environmental and meteorological indicators as possible explanatory variables. For each model, the variables with p-values <0.25 from univariate analyses were candidates for multivariate analyses but only a limited number of variables was selected for multivariate analysis. Firstly, in the case of co-linearity among some explanatory variables, they were analyzed separately and the univariate model minimizing the AIC (Akaike Information Criterion) and having the best biological input was selected. Secondly, selection of variables included the exclusion of the variables for which biological input was difficult to assess, apart for the intrinsic characteristics of the different sections (e.g., distance to sugar cane field, slope). Indeed, if the section itself would have a predictive role, it would prevent models to be applied to other sections. A manual backward stepwise selection procedure was applied in the final model to keep variables with p-values <0.05. The sampling scheme implied that some autocorrelations could exist between observations since nearby observations could be more similar than distant ones as they could have more similar surroundings. In the case that the local environment would not fully taken into account by the explanatory variables, a random effect was added to the models at the level of the section. It should be noted that the small amount of observations did not allow adjusting models using a sub-set of observations only and assessing their validity with the remaining ones. Models validity was assessed using ROC (Receiver Operating Characteristic) curve [37] (i.e., representation of sensitivity against 1-specificity, or true positive rate versus false positive rate, providing the discriminative value of a test). The choice of the cut-off value was done for maximizing sensitivity and specificity. Robustness was assessed using six sub-models of each final model fitted by omitting in turns the experimental units of each section.

The linear equations derived from the regression analysis allowed predicting the outcomes at the non surveyed experimental units, i.e., other houses and other dates, for both steps. Equation of Step 2 was applied to the experimental units that were predicted as water-filled-positive in Step 1. The results were maps of the houses harboring Aedes larvae-positive container(s), for each day of year 2010. They were merged in composite monthly maps including for each house the number of days predicted as Aedes larvae-positive houses. Maps were drawn in the 6 studied sections but also in a non surveyed area (East of section 4) to test the spatial extrapolation abilities of the models.

Results

Description of data

Details for 158 experimental units were available in the final database. All of them contained geographic coordinates, plot identification, entomological, environmental and meteorological data. The observations corresponded to 10 different dates, covering different seasons of the year. A total of 117 houses were visited (i.e., 88 once, 18 twice, 10 thrice and 1 visited 4 times). The sample was about 12% of the total number of houses in the studied area: 1, 069 houses had been mapped in the Caravelle Peninsula but only 983 belonged to or were closed to the 6 studied sections.

A total of 88 experimental units were positive for the presence of peridomestic water-filled containers (from 1 to 11 containers-per-house at a given date). Thirty experimental units were Aedes larvae-positive (19% of all experimental units and 34% of all water-filled-positive experimental units): respectively 21, 5 and 4 experimental units with 1, 2 and 4 Aedes larvae-positive containers. Among all types of water-filled containers, large containers and drum barrels were the most frequently Aedes larvae-positive with respectively 57% and 51% positive observations.

Statistical modeling

Step 1. Modeling the water-filled-positive experimental units

In the univariate analysis, several environmental and meteorological variables were statistically significantly associated with the presence of one or several water-filled container(s) in an experimental unit. Those environmental parameters were: area for the class “sparsely vegetated soil” in the plot, area for the class “tiled roof” in the plot, area for the class “swimming pool” within a 50-m buffer, distance to the class “sugar cane”, maximum altitude within the 50-m buffer, slope of the plot, mean object height in the 100-m buffer (with positive sign) and area of the class “lawn” within the 50-m buffer, area of the class “sand” within the 50-m buffer and distance to the class “sparsely vegetated soil” (with negative sign).

Some rainfall, temperature, humidity and water balance variables were also associated with the presence of one or several water-filled container(s) in an experimental unit with positive or negative sign.

In the multivariate analysis, the area of the class “sparsely vegetated soil” in the plot and the total rainfall during the 4-days period before the day of the field visit were positively associated with the outcome, while the area of the class “lawn” within the 50-m buffer around the house was associated with the outcome along with a negative sign (Table 2). The section random effect was not statistically significant in the final model. In addition, the 6 sub-models fitted by omitting respectively the experimental units of each section provided very similar coefficients than the final model. The area under the ROC curve was 0.72 (95% confidence interval: 0.64 - 0.80). A total of 103 observations were correctly predicted (65%). Whilst the sensibility was 63%, the specificity was 69%. The positive predictive value was 71% and the negative predictive value was 59%.

Step 2. Modeling the Aedes larvae-positive experimental units

In the univariate analysis, several environmental and meteorological variables were associated with the presence of one or several Aedes larvae-positive container(s) in the experimental units that held water-filled container(s), with positive or negative sign. Those environmental variables were: area of the class “tree” within the 50-m buffer, distance to the class “sugar cane”, mean height of the roofs within the 50-m buffer, mean NDVI within the 50-m buffer (with positive sign) and area of the class “asphalt” within the 50-m buffer, area of the class “swimming pool” within the 50-m buffer, area of the class “tiled roof” within the 50-m buffer, area of the class “ocean” within the 100-m buffer (with negative sign). Some rainfall, temperature, humidity and water loss factors were also associated with positive or negative sign.

In the multivariate analysis, the mean of the maximum humidity recorded during the 5-days period before the day of the ground investigation was positively associated with the outcome, while the area of the class “asphalt” within the 50-m buffer around the house was associated with the outcome with negative sign (Table 3). The section random effect was not statistically significant in the final model. In addition, the 6 sub-models fitted by omitting respectively the experimental units of each section provided very similar coefficients than the final model. The area under the ROC curve was 0.74 (95% confidence interval: 0.63 - 0.86). A total of 64 observations were correctly predicted (73%). The sensibility was 70%, the specificity was 74%, the positive predictive value was 58% and the negative predictive value was 83%.

Application of the scenario (Step 1 + Step 2)

From the chosen scenario, the final model for predicting the Aedes larvae-positive experimental units (Step 2) has been applied to the experimental units that had been predicted as water-filled-positive for Step 1. Final predictions are displayed in Table 4. A total of 132 experimental units were correctly predicted (84%). The sensibility was 57% and the specificity was 90%. The percentage of correctly classified predictions was respectively 92%, 78%, 97%, 100%, 67% and 75% in sections 1, 2, 3, 4, 5 and 6.

Predictive entomological risk mapping in Tartane

Both steps of the scenario were successfully applied to each of the 983 houses of the studied area at each date of the 2010 year in order to generate the high spatio-temporal resolution entomological risk maps. The resulting composite monthly maps are displayed in Figure 4.

None of the houses were predicted with 100% of days being Aedes larvae-positive in 2010. A total of 126 houses were predicted with less than 10 Aedes larvae-positive days in the year, among which 28 were predicted as always Aedes larvae-negative (0 positive days in the year). Maximum entomological risk was predicted in section 5, mainly in the south of the area. Then, the risk was decreasing in sections 4, 2 and 1 (same risk), 6 and 3 with yearly figures ranging among sections from 77% to 15% for the number of experimental units which were Aedes-larvae positive. June and September 2010 were the months with the highest predicted entomological risk (respectively 45% and 44% of the number of experimental units were predicted as Aedes-larvae positive) while February and December were predicted with the lowest risk (respectively 17% and 18% of the number of experimental units were predicted as Aedes-larvae positive).

Discussion and conclusion

In the present study, the practical and conceptual approach of tele-epidemiology, developed by the French Spatial Agency (CNES) and its partners allowed to draw spatio-temporal high resolution entomological predictive risk maps for the presence of Aedes aegypti immature stages in Tartane, Martinique. The experimental unit has been chosen as the single house and its close surroundings and was proven as being an appropriate scale for mapping risk. Very high spatial resolution remote-sensing environmental variables and very high temporal resolution ground meteorological variables have been used to adjust statistical models. The risk maps presented here are the first examples of modeled entomological maps at the dwelling level with daily temporal resolution. Monthly composite maps were drawn for interpretation. This kind of mapping could be seen as tools to be included in early warning and targeting vector control operational systems, based on updated satellite images and meteorological information. In the field of entomological assessment of dengue risk, ecological and meteorological data should be seen as synergetic drivers of Aedes presence and density. They were both included in the present study. During the studied period, the ecological factors were temporally static (i.e., one value for the whole period) but they brought in the spatial dynamics. Indeed, they were extracted from a very-high spatial resolution satellite image (GeoEye-1 0.41cm). On the contrary, the meteorological factors were spatially static (i.e., one value for the whole studied area) but they were extracted at fine temporal scale (daily) from a ground observation station, bringing in temporal dynamics.

The predictive composite monthly entomological risk maps drawn for the year 2010 highlighted the spatio-temporal variability among the houses that contained one or several Aedes larvae-positive container(s). Even in this small area of Tartane, heterogeneity of the entomological risk was important, spatially and temporally. Nevertheless, even if no statistical analysis has been done, some spatial clusters of entomological risk did appear: i) houses located in section 5 (see Figure 1) and the south of section 2 displayed high entomological risk along the full year; ii) houses located in the south of section 4 and the north-east of section 1 displayed higher risk than for the rest of the houses within those sections; iii) houses always Aedes aegypti-negative were mainly located in the section 2, along the seaside. Regarding temporal dynamics, June and September 2010 were the months with the highest entomological risk prediction, while February and December were the less risky months, in agreement with the rainy and dry seasons. Nevertheless, some houses remained predicted with high or very high entomological risk even during the very dry months. The latter may reflect the fact that some containers can be artificially filled with water so that the Aedes aegypti population can persist along the year.

A two-step approach

A two-step approach has been carried-out in order to closely link entomological modeling to the biological, physical and societal mechanisms that drive i) the presence of water-filled containers, and ii) larval development. The linear equations resulting from the models adjusted at both steps were applied for each of the 983 houses mapped in the studied area, for each single day of year 2010.

Modeling of the houses harboring one or several water-filled container(s) at a given date was achieved using two predictors extracted from the Geoeye-1 image and one field meteorological variable. Firstly, the surface of the class “sparsely vegetated soil” in the house plot was positively associated with the water-filled-positive experimental units. This could represent houses with low maintenance level. Association was robust since this variable within the 50-m buffer around the house was also statistically significantly associated with the dependant variable. Secondly, the surface of the class “lawn” within the 50-m buffer around the house was associated with the outcome with negative sign; houses surrounded by lawn being probably well maintained. Both variables seemed to be related to population behavior and socio-economic/socio-cultural aspects, with impacts on the presence of containers. On one hand, well maintained environment may be related to the absence of containers or waste-bins that may be filled with precipitation. On the other hand, socio-economic level may be related to the need or not to collect rainfall water in drum barrels to save money, although collecting rainfall may also be related to environmental-friendly behavior. Thirdly, the total rainfall during the 4-days period before the field visit was logically a risk factor for the presence of water-filled container(s), as some containers are intentionally or not, filled-up with rainfall.

Modeling of the houses harboring one or several Aedes larvae-positive container(s) among the sampled houses harboring water-filled container(s) was achieved using one predictor extracted from the Geoeye-1 image and one in-situ meteorological variable. Firstly, the surface of the class “asphalt” within the 50-m buffer around the house was negatively associated with the outcome. This association was robust as this variable within the 100-m buffer was also significantly linked with the independent variable. Interpretation may be helped by the fact that this explanatory variable was strongly inversely correlated with the NDVI which evaluates the density of vegetation. Plenty of studies have highlighted the association between shade - that may be brought by vegetation - and the presence of Aedes immature stages [31, 38, 39]. Shade may lower any very high water temperature that has been specifically recognized as a negative factor for the presence of Aedes aegypti larvae [3]. Vegetation may also provide nutriments for larvae by the leaves falling in the water [40] whilst nectar could be food for adult mosquitoes [41]. In Martinique, Aedes aegypti is mainly endophilic so vegetation probably does not serve as resting site (Yebakima A., personal communication). While scales of the associations between shade and Aedes aegypti have been found at 2-3m on the ground [39], the results of the present work corroborate previous studies having included remote-sensing images. Indeed, the presence of medium height trees within a 30-m radius buffer zones was associated with adult Aedes aegypti abundance in Arizona [20] and niche modeling using Landsat 7 images (with 30-m spatial resolution) allowed predicting the areas suitable for Aedes aegypti breeding sites in Colombia [13]. Nonetheless, the class “asphalt” probably brought more information than the NDVI variable since it has been retained in the statistical final model. Indeed, asphalt surroundings may not be favorable for the breeding sites to persist long enough to be suitable for full cycle larval development.

Secondly, the mean of maximum humidity for the 5-days period before the field visit was positively associated with the presence of one or several Aedes larvae-positive container(s) in the experimental units. The association was robust since the variables for the 2- to-14 days period before the field visit were also significantly associated with the outcome. Very similar results were found in Brazil [42] and Australia [43]. Humidity is correlated to precipitation and evaporation that are related to longer presence of water in the containers, raising the possibilities for larval development. Studies have already highlighted the linkages between rainfall and Aedes aegypti [44-47] whilst dengue risk models have included this rainfall data as a predictive variable [26].

Accuracy and validity of models

It is recognized that, since number of observations was somewhat small, the validation was not based upon datasets different to those used to fit the models. Nevertheless, the final number of observations that were correctly predicted was 85%. A total of 89% of the Aedes larvae-negative houses and 67% of the Aedes larvae-positive houses were correctly predicted. From an operational point of view, this scenario is very powerful to limit the amount of time spent on the ground but it should be improved in terms of detecting positive houses.

Bias could have been introduced in the models since ground data collection was not accomplished at the same dates for each section and some section were preferentially investigated during rainy/dry seasons. Nevertheless analysis of the risk maps showed that the sections that had been followed mainly during respectively rainy/dry seasons had no particular pattern in terms of prediction of Aedes aegypti larvae presence. Indeed, while an entirely environmental model could not have been fitted due to this bias, the presence of meteorological variables in the models served as a kind of seasonal adjustment. In addition, the six sub-models which were fitted at both steps by omitting successively the experimental units of each section provided similar estimates. The latter means that none of the section could have deeply modified the modeling results.

From entomological risk maps to ground control actions

The type of risk maps drawn in the present study could be included in an operational warning system and targeting vector control preventive actions. It has been actually argued that the use of models should participate in spatially and temporally prioritizing prevention where the risk is the greatest [12]. On one hand, entomological data are rarely collected on a routine in a given area, and when field studies are undertaken, they often provide only a snapshot of a continuous phenomenon. Risk maps are expected to bring this continuity for better evaluation of risks. On the other hand, maps that are available at the household level could facilitate detection of “key estates” [48, 49] for efficient control. Indeed, targeting could be done in the houses that were predicted with high number of Aedes larvae-positive days in a month, which could be seen as a surrogate for productivity.

Conclusion

As it has been often advocated, focusing interventions in the places and periods where the risk is maximal is paramount for better allocate limited resources and improve dengue control. Entomological risk maps may be seen as one of the tools available and tele-epidemiology may be applied in this context. The present study showed that environmental information at fine spatial scale - obtained using satellite images - coupled with field meteorological data also at fine temporal scale were successfully included as explanatory variables in the statistical models. It allowed drawing daily entomological predictive risk maps for the presence or not of one or several Aedes aegypti larvae-positive containers, at the level of individual dwellings. Finally the approach presented in this paper can be applied immediately to the emerging chikungunya entomological risk levels in Martinique since the Aedes aegypti is also the vector for this disease [50].

Acknowledgments

The authors would like to thank David Flamanc from the Direction Regionale de l’Environnement, de l’Aménagement et du Logement – Martinique, for providing Litto3D® data, along with their full use. They acknowledge the SERTIT (SErvice Régional de Traitement d’Image et de Télédétection; Hervé Yésou, Carlos Uribe, Claire Huber) for providing inputs for the definition of the very high resolution indices. VM, CV and JPL thank Sanofi-Pasteur for supporting this study. VM and JPL also acknowledge CNES (Centre National d’Etudes Spatiales) for additional support. This is an LDEO contribution # XXXX.

Bibliography

1. Murray NE, Quam MB, Wilder-Smith A (2013) Epidemiology of dengue: past, present and future prospects. Clin Epidemiol 5: 299-309. 2. Christophers S (1960) Aedes Aegypti (L.) The Yellow Fever Mosquito: Its Life History, Bionomics and Structure; Press CU, editor. 739 pp. p. 3. Hemme RR, Tank JL, Chadee DD, Severson DW (2009) Environmental conditions in water storage drums and influences on Aedes aegypti in Trinidad, West Indies. Acta Trop 112: 59-66. 4. Etienne M (2006) Etude de la bioécologie d’Aedes aegypti à la Martinique en relation avec l’épidémiologie de la dengue. Montpellier. 5. Wan SW, Lin CF, Wang S, Chen YH, Yeh TM, et al. (2013) Current progress in dengue vaccines. J Biomed Sci 20: 37. 6. Iturbe-Ormaetxe I, Walker T, SL ON (2011) Wolbachia and the biological control of mosquito-borne disease. EMBO Rep 12: 508-518. 7. Dame DA, Curtis CF, Benedict MQ, Robinson AS, Knols BG (2009) Historical applications of induced sterilisation in field populations of mosquitoes. Malar J 8 Suppl 2: S2. 8. Bellini R, Medici A, Puggioli A, Balestrino F, Carrieri M (2013) Pilot field trials with Aedes albopictus irradiated sterile males in Italian urban areas. J Med Entomol 50: 317-325. 9. Focks DA, Brenner RJ, Hayes J, Daniels E (2000) Transmission thresholds for dengue in terms of Aedes aegypti pupae per person with discussion of their utility in source reduction efforts. Am J Trop Med Hyg 62: 11-18. 10. Eisen L, Lozano-Fuentes S (2009) Use of mapping and spatial and space-time modeling approaches in operational control of Aedes aegypti and dengue. PLoS Negl Trop Dis 3: e411. 11. Bergquist NR (2001) Vector-borne parasitic diseases: new trends in data collection and risk assessment. Acta Trop 79: 13-20. 12. Scott TW, Morrison AC (2010) Vector dynamics and transmission of dengue virus: implications for dengue surveillance and prevention strategies: vector dynamics and dengue prevention. Curr Top Microbiol Immunol 338: 115-128. 13. Arboleda S, Jaramillo ON, Peterson AT (2012) Spatial and temporal dynamics of Aedes aegypti larval sites in Bello, Colombia. J Vector Ecol 37: 37-48. 14. Neteler M, Roiz D, Rocchini D, Castellani C, Rizzoli A (2011) Terra and Aqua satellites track tiger mosquito invasion: modelling the potential distribution of Aedes albopictus in north-eastern Italy. Int J Health Geogr 10: 49. 15. Roiz D, Neteler M, Castellani C, Arnoldi D, Rizzoli A (2011) Climatic factors driving invasion of the tiger mosquito (Aedes albopictus) into new areas of Trentino, northern Italy. PLoS One 6: e14800. 16. Estallo EL, Lamfri MA, Scavuzzo CM, Almeida FF, Introini MV, et al. (2008) Models for predicting Aedes aegypti larval indices based on satellite images and climatic variables. J Am Mosq Control Assoc 24: 368-376. 17. Rotela CH, Espinosa MO, Albornoz C, Lafaye M, Lacaux J-P, et al. Desarrollo de mapas predictivos de densidad focal de Aedes aegypti en la ciudad de Puerto Iguazú (Argentina), basados en información ambiental derivada de imágenes SPOT 5 HRG1; 2008; La Habanna, Cuba. 18. Fuller DO, Troyo A, Calderón-Arguedas O, Beier JC (2009) Dengue vector (Aedes aegypti) larval habitats in urban environment of Costa Rica analysed with ASTER and QuickBird imagery. Int J Rem Sens 31: 3-11. 19. Vanwambeke SO, Bennett SN, Kapan DD (2011) Spatially disaggregated disease transmission risk: land cover, land use and risk of dengue transmission on the island of Oahu. Trop Med Int Health 16: 174-185. 20. Landau KI, van Leeuwen WJ (2012) Fine scale spatial urban land cover factors associated with adult mosquito abundance and risk in Tucson, Arizona. J Vector Ecol 37: 407-418. 21. Sarfraz MS, Tripathi NK, Tipdecho T, Thongbu T, Kerdthong P, et al. (2012) Analyzing the spatio-temporal relationship between dengue vector larval density and land-use using factor analysis and spatial ring mapping. BMC Public Health 12: 853. 22. Van Benthem BH, Vanwambeke SO, Khantikul N, Burghoorn-Maas C, Panart K, et al. (2005) Spatial patterns of and risk factors for seropositivity for dengue infection. Am J Trop Med Hyg 72: 201-208. 23. Rotela C, Fouque F, Lamfri M, Sabatier P, Introini V, et al. (2007) Space-time analysis of the dengue spreading dynamics in the 2004 Tartagal outbreak, Northern Argentina. Acta Trop 103: 1-13. 24. Neteler M, Metz M, Rocchini D, Rizzoli A, Flacio E, et al. (2013) Is Switzerland Suitable for the Invasion ofAedes albopictus? PLoS One 8: e82090. 25. ECDC (2009) Development of Aedes albopictus risk maps ECDC technical report. Stockholm: European Centre For Disease Prevention and Control. 26. Rogers DJ, Suk JE, Semenza JC (2014) Using global maps to predict the risk of dengue in Europe. Acta Trop 129: 1-14. 27. Vignolles C, Lacaux JP, Tourre YM, Bigeard G, Ndione JA, et al. (2009) Rift Valley fever in a zone potentially occupied by Aedes vexans in Senegal: dynamics and risk mapping. Geospat Health 3: 211-220. 28. Vignolles C, Tourre YM, Mora O, Imanache L, Lafaye M (2010) TerraSAR-X high-resolution radar remote sensing: an operational warning system for Rift Valley fever risk. Geospat Health 5: 23-31. 29. Lacaux J-P, Tourre Y-M, Vignolles C, Ndione J-A, Lafaye M (2006) Classification of ponds from high-spatial resolution remote sensing: Application to Rift Valley Fever epidemics in Senegal. Rem Sens Environ 106: 66-74. 30. Machault V, Vignolles C, Pagès F, Gadiaga L, Tourre YM, et al. (2012) Risk Mapping of Anopheles gambiae s.l. Densities Using Remotely-Sensed Environmental and Meteorological Data in an Urban Area: Dakar, Senegal. PLoS ONE 7. 31. Tun-Lin W, Kay BH, Barnes A (1995) The Premise Condition Index: a tool for streamlining surveys of Aedes aegypti. Am J Trop Med Hyg 53: 591-594. 32. Nogueira LA, Gushi LT, Miranda JE, Madeira NG, Ribolla PE (2005) Application of an alternative Aedes species (Diptera: culicidae) surveillance method in Botucatu City, Sao Paulo, Brazil. Am J Trop Med Hyg 73: 309-311. 33. Peres RC, Rego R, Maciel-de-Freitas R (2013) The use of the Premise Condition Index (PCI) to provide guidelines for Aedes aegypti surveys. J Vector Ecol 38: 190-192. 34. Rouse JW, Hass RH, Schell JA, Deering DW (1973) Monitoring vegetation systems in the Great Plains with ERTS. Third ERTS Symposium, NASA SP 351. 309-317. 35. Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens Environ 8: 127-150. 36. McFeeters SK (1996) The use of the normalised difference water index (NDWI) in the delineation of open water features. Int J Remote Sens 17: 1425–1432. 37. Mason SJ, Graham NE (2002) Areas beneath the relative operating characteristics (ROC) and levels (ROL) curves: statistical significance and interpretation. . Q J Roy Meteor Soc 128: 2145-2166. 38. Vezzani D, Albicocco AP (2009) The effect of shade on the container index and pupal productivity of the mosquitoes Aedes aegypti and Culex pipiens breeding in artificial containers. Med Vet Entomol 23: 78-84. 39. Vezzani D, Rubio A, Velazquez SM, Schweigmann N, Wiegand T (2005) Detailed assessment of microhabitat suitability for Aedes aegypti (Diptera: Culicidae) in Buenos Aires, Argentina. Acta Trop 95: 123-131. 40. Reiskind MH, Greene KL, Lounibos LP (2009) Leaf species identity and combination affect performance and oviposition choice of two container mosquito species. Ecol Entomol 34: 447-456. 41. Martinez-Ibarra JA, Rodriguez MH, Arredondo-Jimenez JI, Yuval B (1997) Influence of plant abundance on nectar feeding by Aedes aegypti (Diptera: Culicidae) in southern Mexico. J Med Entomol 34: 589-593. 42. Favier C, Degallier N, Vilarinhos Pde T, de Carvalho Mdo S, Yoshizawa MA, et al. (2006) Effects of climate and different management strategies on Aedes aegypti breeding sites: a longitudinal survey in Brasilia (DF, Brazil). Trop Med Int Health 11: 1104-1118. 43. Azil AH, Long SA, Ritchie SA, Williams CR (2010) The development of predictive tools for pre-emptive dengue vector control: a study of Aedes aegypti abundance and meteorological variables in North Queensland, Australia. Trop Med Int Health 15: 1190-1197. 44. Wee LK, Weng SN, Raduan N, Wah SK, Ming WH, et al. (2013) Relationship between rainfall and Aedes larval population at two insular sites in Pulau Ketam, Selangor, Malaysia. Southeast Asian J Trop Med Public Health 44: 157-166. 45. Baruah S, Dutta P (2013) Seasonal prevalence of Aedes aegypti in urban and industrial areas of Dibrugarh district, Assam. Trop Biomed 30: 434-443. 46. Duncombe J, Clements A, Davis J, Hu W, Weinstein P, et al. (2013) Spatiotemporal patterns of Aedes aegypti populations in Cairns, Australia: assessing drivers of dengue transmission. Trop Med Int Health 18: 839-849. 47. Stewart Ibarra AM, Ryan SJ, Beltran E, Mejia R, Silva M, et al. (2013) Dengue Vector Dynamics (Aedes aegypti) Influenced by Climate and Social Factors in Ecuador: Implications for Targeted Control. PLoS One 8: e78263. 48. Tun-Lin W, Kay BH, Barnes A (1995) Understanding productivity, a key to Aedes aegypti surveillance. Am J Trop Med Hyg 53: 595-601. 49. Chadee DD (2004) Key premises, a guide to Aedes aegypti (Diptera: Culicidae) surveillance and control. Bull Entomol Res 94: 201-207. 50. Simon F, Savini H, Parola P (2008) Chikungunya: a paradigm of emergence and globalization of vector-borne diseases. Med Clin North Am 92: 1323-1343, ix.

Figure legends

Figure 1. The Martinique Island, the studied area and the six studied sections on the Tartane Peninsula. Figure 2. Map of the sampled houses. Figure 3. Scenario retained for dengue entomological risk mapping. Figure 4. Monthly entomological risk maps from January until December 2010, from the modeling experiment. The number of days when the 983 houses within the studied area were predicted as Aedes larvae-positive, are given (see colored code at bottom left).

Supporting information legends

Supporting Information 1. Monthly entomological risk maps resulting from modeling and including the number of days when the 983 houses of the studied area were predicted as Aedes larvae-positive. Power point presentation.

Table 1. Environmental indicators calculated from the Geoeye-1 image at 0.41-m spatial resolution. Environmental indicator Spectral bands combination* Description NDVI (Normalized Difference Vegetation Index) [34,35]

A value superior to 0.2 usually corresponds to a vegetated area, which gets denser when this value rises. Negative values indicate non-vegetated features such as barren surfaces (rocks and soils), water, built-up areas or asphalt. NDWI Mac Feeters (Normalized Difference Water Index) [36]

It delineates open water features while eliminating the presence of soil and terrestrial vegetation features. Its value increases with the presence of water and decreases with the presence of vegetation. It is also suggested that it may provide turbidity estimations of water bodies. ANDWI (Adapted NDWI Mac Feeters Index) Using the blue band, this adapted NDWI Mac Feeters maximizes the detection of water. * NIR : Near infrared, SWIR : Short wave infrared

Table 2. Remote-sensing environmental and ground meteorological variables significantly associated with the presence of one or several water-filled container(s) in an experimental unit. Multivariate logistic regression with section random effect is given (Step 1). Explanatory variables Coefficient 95% confidence interval p-value 158 experimental units, 6 groups Surface of the class “sparsely vegetated soil” in the house plot (per 10 m²) 0.10 [0.02 ; 0.19] 0.017 Surface of the class “lawn” within the 50-m buffer around the house (per 100 m²) -0.10 [-0.17 ; -0.03] 0.007 Total rainfall during the 4 days preceding the field visit (per 10 mm) 0.26 [0.07 ; 0.46] 0.007 Section random effect 0.498

Table 3. Remote-sensing environmental and ground meteorological variables significantly associated with the presence of one or several Aedes larvae-positive container(s) in an experimental unit that holds water-filled container(s). Multivariate logistic regression with section random effect is given (Step 2). Explanatory variables Coefficient 95% confidence interval p-value 88 experimental units, 6 groups Surface of the class “asphalt” within the 50-m buffer around the house (per 100 m²) -0.10 [-0.17 ; -0.04] 0.003 Mean of maximum humidity during the 5 days preceding the field visit (per 1%) 0.27 [0.07 ; 0.48] 0.008 Section random effect 0.209

Table 4. Final predictions of the scenario versus field data: number of Aedes larvae-positive experimental units versus number of Aedes larvae-negative experimental units (i.e., without water-filled container(s) or with Aedes larvae-negative container(s)). Prediction from scenario Number of Aedes larvae-negative experimental units Number of Aedes larvae-positive experimental units Total Field data Number of Aedes larvae-negative experimental units 115 13 128 Number of Aedes larvae-positive experimental units 13 17 30 Total 128 30 158

Remote-sensing environmental data for mapping entomological dengue risk levels in Martinique.

Vanessa Machault (1), André Yébakima (2), Manuel Etienne (2), Cécile Vignolles (3), Philippe Palany (4), Yves M. Tourre (5), Marine Guérécheau (1), Jean-Pierre Lacaux (1)

(1) Laboratoire d’Aérologie, Observatoire Midi-Pyrénées (OMP), Université Paul Sabatier, Toulouse, France (2) Service de Démoustication et de Lutte Anti-vectorielle, Conseil Général de la Martinique/Agence Régionale de Santé (SD-LAV), Fort-de-France, Martinique, France (3) Direction de la Stratégie et des Programmes/Terre-Environnement-Climat, Centre National d’Etudes Spatiales (CNES), Toulouse, France (4) Météo-France Direction Inter-Régionale Antilles-Guyane, Fort-de-France, Martinique, France (5) Lamont-Doherty Earth Observatory (LDEO) of Columbia University, Palisades, New York, USA

Keywords: dengue, remote-sensing, risk mapping, Aedes aegypti, entomology

Abstract

Background

Whilst in Martinique dengue is endemic, six epidemic waves did occur during the last 20 years. There is no specific treatment for dengue fever as yet and no operational vaccine is available. The only mean for controlling virus transmission is thus to effectively combat the mosquitoes. Risk maps at appropriate scales can then provide surrogate data and valuable information for a spatio-temporal assessment of entomological risk levels in order to improve dengue control efficiently.

Methodology/Principal Findings

In general, the tele-epidemiology practical concept applied here, relies on the facts that: i) a so-called experimental unit, or the smallest ‘object’ to observe must be characterized first, to properly assess the risk levels for a given disease, ii) environment and weather conditions may explain the spatio-temporal distribution and variability of diseases risk and ii) datasets from satellite images may be included into statistical models. The experimental unit for dengue vectors has been chosen here as a dwelling and its nearby surroundings. Some of its characteristics have been found to be risk or protective factors for the presence of Aedes aegypti immature stages. Those factors were environmental conditions (i.e., surface of “sparsely vegetated area”, “lawn” and “asphalt”) and weather conditions (i.e., rainfall, relative humidity). In addition, remote-sensing environmental data are used to produce dynamic high spatio-temporal resolution maps for the presence of various containers harboring the Aedes aegypti. It s found that the spatio-temporal variability of the entomological dengue risk levels may be assessed in Martinique (French Antilles).

Conclusions/Significance

The produced risk maps are the first examples of modeled entomological maps at the housing level with daily temporal resolution. This finding is an important component for preventive actions and contribution to developing operational warning systems for other vector-borne diseases such as the recently identified chikungunya in Martinique.

Introduction

Dengue is an infectious disease caused by one of the four serotypes (DEN-1 to DEN-4) of the dengue virus. It is transmitted to humans through bites from already infected female mosquitoes of the genus Aedes in urban areas. Even if mortality rate is low among human population, dengue is considered as one of the most important mosquito-borne viral disease. This is due to its extensive geographic spread, the societal cost of its burden from the 50- to 200- million annual infections and the now identified 125 endemic countries [1].

In Martinique (French Antilles), Aedes aegypti mosquito is so far the single identified vector for spreading the dengue virus. It breeds mostly in artificial domestic or peridomestic containers filled with clean water with little organic debris and low concentration of inorganic nutrients [2, 3]. Breeding sites/containers include flower pots with saucers, mounts of detritus and debris, abandoned cars and tires, badly maintained gutters, discarded old domestic appliances that may be all filled-up naturally with rainfall. In addition drum barrels may be deliberately placed under gutters or in yards to collect rain water for watering/cleaning purposes. All of the above conditions and mechanisms co-exist and favor the local environment for potential breeding sites.

Whilst in Martinique dengue is endemic, six epidemic waves did occur during the last 20 years. The penultimate epidemic, in 2010, from February to the end of the year, led to more than 41 000 clinical cases (about 10% of the island population). This outbreak, as well as the 2001 epidemic, has started in Tartane, a village located in the Caravelle Peninsula, North-East of Matinique. Entomological analyses have provided some clues for understanding the key elements for the beginning of the 2001 outbreak: higher Aedes density, with high survival rate and a gonotrophic cycle that were both favorable for virus transmission [4].

There is no specific treatment for dengue with no operational vaccine as yet [5]. The only mean for controlling virus transmission is thus to effectively combat the mosquitoes. Several techniques/strategies are now available for field testing based upon a biological approach (e.g., mosquitoes infection with Wolbachia bacteria that may interfere with dengue virus and reduce mosquitoes lifespan [6]) or a genetic approach (e.g., sterile insects techniques [7,8]. Nevertheless, source reduction of the potential larvae habitats and/or reduction of the emergence of adult specimens have long been accepted as part of dengue control strategy. Works have been done to identify the control thresholds to be attained in order to suppress transmission threats [9]. In this context, a good knowledge of the entomological conditions in a given area and a given period is thus a prerequisite. Unfortunately entomological data are seldom collected continuously. Moreover when field studies are undertaken, they often only provide a snapshot of a somewhat continuous phenomenon. Risk maps at appropriate scales can then provide surrogate data and valuable information for having a spatio-temporal evaluation of entomological risk levels in order to improve dengue control efficiently.

Informative maps of dengue vectors/cases, can be perceived as tools for: i) delivering vectors and dengue statistics, ii) driving vector control around dengue cases during outbreaks (or within risk areas identified during past epidemics) and, iii) studying space and space-time clustering of cases during outbreaks [10]. Spatial information should also allow modeling the linkages between vector presence and/or dengue cases associated with environmental and socio-economical variables [10]. Additional linkages with meteorological/weather variables should also be included to better understand the mechanisms at stake.

Risk maps have been produced for numerous diseases and for mapping a current situation or even anticipating outbreaks through Early Warning Systems (EWS) [11]. While mapping dengue at global scales is relevant for determining populations at risk, fine scale or local mapping is of major interest for setting-up local control strategies particularly when resources are limited [12]. From global/regional to local scales, the spatial and temporal distribution of vectors as well as dengue epidemiological characteristics may fluctuate along with weather/climate conditions (i.e., mainly rainfall amount, relative humidity and temperature), the environment (i.e., vegetation, soil types, among others) or human activities (i.e., human migration and local movements, transportation, urbanization or waste management procedures, among others). Modeling dengue may benefit from the use of environmental remote-sensing information. In the recent past, satellite products have been proven to provide useful information for modeling Aedes aegypti or Aedes albopictus distribution [13-21], human cases distribution [22,23] or potentialities for vectors or disease future expansion [24-26].

The present study consists essentially in mapping dengue entomological risk i.e., evaluating risk of presence of vectors, in contrast with mapping epidemiological risk. The practical and conceptual approach (CA) called tele-epidemiology (Marechal, Ribeiro et al., 2008) could then be applied to the spatio-temporal mapping of entomological dengue risk in urban settings in Martinique. This CA has been developed and patented by the French Spatial Agency (CNES) and its partners [27, 28]. It consists in monitoring and studying human and animal diseases spatio-temporal dynamics which are closely related to weather/climate and environment variability. It relies on the identification of an experimental unit (EU) that is the smallest ‘object’ that has to be identified/characterized in order to assess properly the levels of risk. This unit is based on the sound knowledge of the biological and physical processes that underline the presence/densities of immature and adult vectors. It is thus widely dependent upon the disease under study. For example this experimental unit will be a pond ( 1 ha) when studying Rift Valley Fever entomological risk [29] and a water body or aggregates of small water bodies ( 1/10 ha) when studying urban malaria entomological risk [30]. Then appropriate choices of satellite data and dynamical models must be assessed, along with extensive use of in-situ measurements.

Three hypotheses underpinned the present study. Firstly, if the potential breeding habitats could not be directly detected using satellite images even at very high spatial resolution, their specific environment could be mapped. Indeed it is thought that dwelling conditions, neighborhood and yard conditions may be associated with the presence of various containers retaining water. Indeed they provide information on local habits regarding maintenance of private yards/gardens or all kind of water storage. Secondly, estate characteristics i.e., shading and tidiness of a house and its yard, have been identified as determinants for the presence/abundance of Aedes aegypti immature stages and eggs [20, 31-33]. It is thus thought that characterizing fine scale environment could inform on the presence of the dengue vector at the immature stages in areas where containers are present. Thirdly, meteorological conditions partly drive the temporal dynamics for containers filling, as well as entomological dynamics (e.g., eggs hatching, larvae development). Consequently, the experimental unit here has been defined as being the house level with its nearby environment. The state of such units was then described with details at a given date.

The main objective here was to model in space and time houses so-called ‘positive’ for Aedes aegypti immature stages from 2009 to 2011 in Tartane (Martinique, French Antilles), by using remote-sensing environmental data and in-situ meteorological information. This was to produce high spatio-temporal resolution predictive dengue entomological risk maps.

Methods

Studied site and period

The city of Tartane (14°45’29.24“N, 60°55’10.56” W) belongs to the Caravelle Peninsula, North-East of Martinique. It is historically a fishing cove, with small and low-rise dwellings surrounded by small gardens or yards. The centre of the city is near the sea-side while other sections are located uphill. The population is of 3, 000 inhabitants. It is a tourist attraction, and includes many vacation sites. The studied site is about 8 km² (see black rectangle in Figure 1). This is exactlyl where the 2010 dengue outbreak started. Although dengue epidemics “usually” last from July to December, viral circulation in 2010 had been observed from early February, with peak in June and lasted until the end of the year. The studied period ran from June 2009 to August 2011, in order to include the 2010 epidemic and detect any heterogeneous temporal drivers.

Entomological data

In Martinique, antivectorial control and mosquito nuisance are managed by a public organism called « Service de Démoustication et de Lutte Antivectorielle » (SD-LAV). Since 1991, the SD-LAV collects information on dengue vectors with an additional effort during outbreaks. A sample of the database records has been selected for this study. Each record was associated with a given house visited at a given date. It contained information on the number and type of domestic and peridomestic containers as well as their status concerning the presence or not of Aedes aegypti immature stages (i.e., all larvae stages and pupae). Surveyed houses were consequently plotted using a Global Positioning System (GPS) device. Types of all-size containers were: drum barrels, tanks, waste-bins, flower pots and saucers, gutters, tires, discarded appliances and pools. Sampled houses are positioned in Figure 2. From their spatial distribution, six sections were identified (Figure 1).

Meteorological data

In Martinique, the summer rainy season (July to November) - with frequent and heavy rainfall and maximum temperatures of about 32°C - and the dry season (February to April) - with maximum temperatures of about 30°C - are separated by two intermediate seasons.

The year 2009 has been exceptionally hot, with low rainfall except on the Atlantic coastline. In 2010, temperatures were also high particularly during February and March. February was almost completely dry. Heavy precipitation started in early June followed by a very dry period. Again, the year 2011 experienced hot temperatures while it has been the wettest of the 2009-2011 period, with basically no dry season.

Daily temperature and humidity (i.e., minimum, maximum and mean) as well as precipitation amount, water loss and water balance were provided by Météo-France. They were recorded at the observing station located in the Caravelle Peninsula. Yearly precipitation recorded during 2009, 2010 and 2011, were respectively of 948 mm, 1, 408 mm and 1, 823 mm.

Several variables have been calculated from the above data:

  • Total rainfall amount for the 2- to 30- days period before each entomological ground investigation date;
  • Number of days without rain for the 5- 10- 15- 20- 25- 30- days period before each entomological ground investigation date;
  • Average of temperature, relative humidity, water loss, water balance for the 2- to 15- days period before each entomological ground investigation date;
  • Periods when temperature/relative humidity/water loss/water balance were above some thresholds (e.g., number of days when maximum temperature exceeded 32°C), during the 15- days period before each entomological investigation date. Those thresholds were defined as the quintiles for the variables distribution.

The above data was added to the entomological database, matched by the date of the ground surveys. No spatial heterogeneity was included here since the meteorological variables were identical for all sampled houses located within the 8 km² area, with no specific reference to their location on the peninsula.

Environmental data and satellite images

A Geoeye-1 image, with clear sky and for 13/03/2011 was acquired. Data included four spectral bands at 0.41-m spatial resolution (blue, green, red and near infrared). The image was projected in WGS 84, UTM, Zone 20N and geometrically corrected using the 50-m spatial resolution elevation map (IGN BD ALTI®) from the French National Geographic Institute IGN (Institut National de l’Information Géographique et Forestière). Image processing was done using ENVI 4.8 and ENVI EX (Exelis Visual Information Solutions). Other available geographic data was: IGN topographic map (IGN BD TOPO®) and cadastral map (IGN BD ADRESSE®). Elevation, slope and objects height maps at 1-m spatial resolution were available through Litto3D® (IGN, Service Hydrographique et Océanographique de la Marine, Direction Régionale de l’Environnement, de l’Aménagement et du Logement - Martinique, Agence des Aires Marines Protégées).

Three vegetation and soil indicators were derived from the Geoeye-1 image. They are described in Table 1 [34], [35] and [36]. A three-step classification procedure allowed map production of Land Use and Land Cover (LULC) from the Geoeye-1 image. An object-oriented classification and a supervised pixel-based classification (maximum likelihood) were undertaken. As each technique proved to accurately highlight some basic elements of the LULC (by not discriminating others) a synthetic classified image was generated using the power of both classifications. For each step, validation of the classifications quality was assessed by photo-interpretation done by experts whilst some manual corrections were applied. The final classification included fourteen land-cover classes: i.e., 5 for vegetation such as “trees”, “sugar cane”, “stubbles”, “grass”, “sparsely vegetated soil”, different types of roofs (5), “sand”, “asphalt”, “swimming pools” and “ocean”.

Geographic Information System (GIS)

A GIS was built using ArcGIS 10.0 (Environmental Research Systems Institute, Redlands, CA, USA) in order to characterize the experimental units defined in the Introduction. All surveyed houses were plotted whilst environmental indicators, LULC map derived from GeoEye-1 image and elevation indicators were added as geo-referenced layers. Each single house for the Caravelle Peninsula was isolated as an object, based on the LULC map. The plot around each house was obtained from the cadastral map.

The environmental variables, i.e., the minimum/maximum/mean for the three indices, altitude, slope and object elevation, as well as the areas from each LULC class, were computed for each experimental unit. This was accomplished for each estate and for a 50-m and 100-m radius buffer zones around the individual houses. The Euclidian distance from the house to each LULC class was also calculated, as well as the inside footage of houses (assumed equal to the area of their roofs) and plots. Those data were merged with the entomological database, again for each house. When records concerned a specific house but surveyed for different periods, the same environmental data was used.

Modeling strategy

The overall database with entomological, environmental and meteorological data contained more than 300 variables: 96 variables obtained from the Geoeye-1 image, 36 altimetric variables and more than 200 meteorological variables. Each record in the database (i.e, observation) was associated with one house visited at one time.

The chosen scenario (Figure 3) included the following two steps which involved investigation of the environmental and meteorological factors that drove:

  • Step 1: the presence of one or several water-filled container(s) in a vicinity of a house at a given date, independently of the presence or not of Aedes aegypti immature stages. This was the detection of the water-filled-positive experimental units;
  • Step 2: the presence of Aedes aegypti immature stages, only in the experimental units that held one or several water-filled container(s). This was the detection of the Aedes larvae-positive experimental units. No reference to the larval density was included.

Since the number of domestic water containers was very low in the area, only the peridomestic containers were considered.

Statistical analysis and risk mapping

Statistical analyses were performed using Stata 11 (Stata Corporation, College Station, Texas, http://www.stata.com. Note that Stata 13 has just been released). Logistic regressions for explaining the outcomes from both steps above were fitted at the experimental unit level using all environmental and meteorological indicators as possible explanatory variables. For each model, the variables with p-values <0.25 from univariate analyses were candidates for multivariate analyses but only a limited number of variables was selected for multivariate analysis. Firstly, in the case of co-linearity among some explanatory variables, they were analyzed separately and the univariate model minimizing the AIC (Akaike Information Criterion) and having the best biological input was selected. Secondly, selection of variables included the exclusion of the variables for which biological input was difficult to assess, apart for the intrinsic characteristics of the different sections (e.g., distance to sugar cane field, slope). Indeed, if the section itself would have a predictive role, it would prevent models to be applied to other sections. A manual backward stepwise selection procedure was applied in the final model to keep variables with p-values <0.05. The sampling scheme implied that some autocorrelations could exist between observations since nearby observations could be more similar than distant ones as they could have more similar surroundings. In the case that the local environment would not fully taken into account by the explanatory variables, a random effect was added to the models at the level of the section. It should be noted that the small amount of observations did not allow adjusting models using a sub-set of observations only and assessing their validity with the remaining ones. Models validity was assessed using ROC (Receiver Operating Characteristic) curve [37] (i.e., representation of sensitivity against 1-specificity, or true positive rate versus false positive rate, providing the discriminative value of a test). The choice of the cut-off value was done for maximizing sensitivity and specificity. Robustness was assessed using six sub-models of each final model fitted by omitting in turns the experimental units of each section.

The linear equations derived from the regression analysis allowed predicting the outcomes at the non surveyed experimental units, i.e., other houses and other dates, for both steps. Equation of Step 2 was applied to the experimental units that were predicted as water-filled-positive in Step 1. The results were maps of the houses harboring Aedes larvae-positive container(s), for each day of year 2010. They were merged in composite monthly maps including for each house the number of days predicted as Aedes larvae-positive houses. Maps were drawn in the 6 studied sections but also in a non surveyed area (East of section 4) to test the spatial extrapolation abilities of the models.

Results

Description of data

Details for 158 experimental units were available in the final database. All of them contained geographic coordinates, plot identification, entomological, environmental and meteorological data. The observations corresponded to 10 different dates, covering different seasons of the year. A total of 117 houses were visited (i.e., 88 once, 18 twice, 10 thrice and 1 visited 4 times). The sample was about 12% of the total number of houses in the studied area: 1, 069 houses had been mapped in the Caravelle Peninsula but only 983 belonged to or were closed to the 6 studied sections.

A total of 88 experimental units were positive for the presence of peridomestic water-filled containers (from 1 to 11 containers-per-house at a given date). Thirty experimental units were Aedes larvae-positive (19% of all experimental units and 34% of all water-filled-positive experimental units): respectively 21, 5 and 4 experimental units with 1, 2 and 4 Aedes larvae-positive containers. Among all types of water-filled containers, large containers and drum barrels were the most frequently Aedes larvae-positive with respectively 57% and 51% positive observations.

Statistical modeling

Step 1. Modeling the water-filled-positive experimental units

In the univariate analysis, several environmental and meteorological variables were statistically significantly associated with the presence of one or several water-filled container(s) in an experimental unit. Those environmental parameters were: area for the class “sparsely vegetated soil” in the plot, area for the class “tiled roof” in the plot, area for the class “swimming pool” within a 50-m buffer, distance to the class “sugar cane”, maximum altitude within the 50-m buffer, slope of the plot, mean object height in the 100-m buffer (with positive sign) and area of the class “lawn” within the 50-m buffer, area of the class “sand” within the 50-m buffer and distance to the class “sparsely vegetated soil” (with negative sign).

Some rainfall, temperature, humidity and water balance variables were also associated with the presence of one or several water-filled container(s) in an experimental unit with positive or negative sign.

In the multivariate analysis, the area of the class “sparsely vegetated soil” in the plot and the total rainfall during the 4-days period before the day of the field visit were positively associated with the outcome, while the area of the class “lawn” within the 50-m buffer around the house was associated with the outcome along with a negative sign (Table 2). The section random effect was not statistically significant in the final model. In addition, the 6 sub-models fitted by omitting respectively the experimental units of each section provided very similar coefficients than the final model. The area under the ROC curve was 0.72 (95% confidence interval: 0.64 - 0.80). A total of 103 observations were correctly predicted (65%). Whilst the sensibility was 63%, the specificity was 69%. The positive predictive value was 71% and the negative predictive value was 59%.

Step 2. Modeling the Aedes larvae-positive experimental units

In the univariate analysis, several environmental and meteorological variables were associated with the presence of one or several Aedes larvae-positive container(s) in the experimental units that held water-filled container(s), with positive or negative sign. Those environmental variables were: area of the class “tree” within the 50-m buffer, distance to the class “sugar cane”, mean height of the roofs within the 50-m buffer, mean NDVI within the 50-m buffer (with positive sign) and area of the class “asphalt” within the 50-m buffer, area of the class “swimming pool” within the 50-m buffer, area of the class “tiled roof” within the 50-m buffer, area of the class “ocean” within the 100-m buffer (with negative sign). Some rainfall, temperature, humidity and water loss factors were also associated with positive or negative sign.

In the multivariate analysis, the mean of the maximum humidity recorded during the 5-days period before the day of the ground investigation was positively associated with the outcome, while the area of the class “asphalt” within the 50-m buffer around the house was associated with the outcome with negative sign (Table 3). The section random effect was not statistically significant in the final model. In addition, the 6 sub-models fitted by omitting respectively the experimental units of each section provided very similar coefficients than the final model. The area under the ROC curve was 0.74 (95% confidence interval: 0.63 - 0.86). A total of 64 observations were correctly predicted (73%). The sensibility was 70%, the specificity was 74%, the positive predictive value was 58% and the negative predictive value was 83%.

Application of the scenario (Step 1 + Step 2)

From the chosen scenario, the final model for predicting the Aedes larvae-positive experimental units (Step 2) has been applied to the experimental units that had been predicted as water-filled-positive for Step 1. Final predictions are displayed in Table 4. A total of 132 experimental units were correctly predicted (84%). The sensibility was 57% and the specificity was 90%. The percentage of correctly classified predictions was respectively 92%, 78%, 97%, 100%, 67% and 75% in sections 1, 2, 3, 4, 5 and 6.

Predictive entomological risk mapping in Tartane

Both steps of the scenario were successfully applied to each of the 983 houses of the studied area at each date of the 2010 year in order to generate the high spatio-temporal resolution entomological risk maps. The resulting composite monthly maps are displayed in Figure 4.

None of the houses were predicted with 100% of days being Aedes larvae-positive in 2010. A total of 126 houses were predicted with less than 10 Aedes larvae-positive days in the year, among which 28 were predicted as always Aedes larvae-negative (0 positive days in the year). Maximum entomological risk was predicted in section 5, mainly in the south of the area. Then, the risk was decreasing in sections 4, 2 and 1 (same risk), 6 and 3 with yearly figures ranging among sections from 77% to 15% for the number of experimental units which were Aedes-larvae positive. June and September 2010 were the months with the highest predicted entomological risk (respectively 45% and 44% of the number of experimental units were predicted as Aedes-larvae positive) while February and December were predicted with the lowest risk (respectively 17% and 18% of the number of experimental units were predicted as Aedes-larvae positive).

Discussion and conclusion

In the present study, the practical and conceptual approach of tele-epidemiology, developed by the French Spatial Agency (CNES) and its partners allowed to draw spatio-temporal high resolution entomological predictive risk maps for the presence of Aedes aegypti immature stages in Tartane, Martinique. The experimental unit has been chosen as the single house and its close surroundings and was proven as being an appropriate scale for mapping risk. Very high spatial resolution remote-sensing environmental variables and very high temporal resolution ground meteorological variables have been used to adjust statistical models. The risk maps presented here are the first examples of modeled entomological maps at the dwelling level with daily temporal resolution. Monthly composite maps were drawn for interpretation. This kind of mapping could be seen as tools to be included in early warning and targeting vector control operational systems, based on updated satellite images and meteorological information. In the field of entomological assessment of dengue risk, ecological and meteorological data should be seen as synergetic drivers of Aedes presence and density. They were both included in the present study. During the studied period, the ecological factors were temporally static (i.e., one value for the whole period) but they brought in the spatial dynamics. Indeed, they were extracted from a very-high spatial resolution satellite image (GeoEye-1 0.41cm). On the contrary, the meteorological factors were spatially static (i.e., one value for the whole studied area) but they were extracted at fine temporal scale (daily) from a ground observation station, bringing in temporal dynamics.

The predictive composite monthly entomological risk maps drawn for the year 2010 highlighted the spatio-temporal variability among the houses that contained one or several Aedes larvae-positive container(s). Even in this small area of Tartane, heterogeneity of the entomological risk was important, spatially and temporally. Nevertheless, even if no statistical analysis has been done, some spatial clusters of entomological risk did appear: i) houses located in section 5 (see Figure 1) and the south of section 2 displayed high entomological risk along the full year; ii) houses located in the south of section 4 and the north-east of section 1 displayed higher risk than for the rest of the houses within those sections; iii) houses always Aedes aegypti-negative were mainly located in the section 2, along the seaside. Regarding temporal dynamics, June and September 2010 were the months with the highest entomological risk prediction, while February and December were the less risky months, in agreement with the rainy and dry seasons. Nevertheless, some houses remained predicted with high or very high entomological risk even during the very dry months. The latter may reflect the fact that some containers can be artificially filled with water so that the Aedes aegypti population can persist along the year.

A two-step approach

A two-step approach has been carried-out in order to closely link entomological modeling to the biological, physical and societal mechanisms that drive i) the presence of water-filled containers, and ii) larval development. The linear equations resulting from the models adjusted at both steps were applied for each of the 983 houses mapped in the studied area, for each single day of year 2010.

Modeling of the houses harboring one or several water-filled container(s) at a given date was achieved using two predictors extracted from the Geoeye-1 image and one field meteorological variable. Firstly, the surface of the class “sparsely vegetated soil” in the house plot was positively associated with the water-filled-positive experimental units. This could represent houses with low maintenance level. Association was robust since this variable within the 50-m buffer around the house was also statistically significantly associated with the dependant variable. Secondly, the surface of the class “lawn” within the 50-m buffer around the house was associated with the outcome with negative sign; houses surrounded by lawn being probably well maintained. Both variables seemed to be related to population behavior and socio-economic/socio-cultural aspects, with impacts on the presence of containers. On one hand, well maintained environment may be related to the absence of containers or waste-bins that may be filled with precipitation. On the other hand, socio-economic level may be related to the need or not to collect rainfall water in drum barrels to save money, although collecting rainfall may also be related to environmental-friendly behavior. Thirdly, the total rainfall during the 4-days period before the field visit was logically a risk factor for the presence of water-filled container(s), as some containers are intentionally or not, filled-up with rainfall.

Modeling of the houses harboring one or several Aedes larvae-positive container(s) among the sampled houses harboring water-filled container(s) was achieved using one predictor extracted from the Geoeye-1 image and one in-situ meteorological variable. Firstly, the surface of the class “asphalt” within the 50-m buffer around the house was negatively associated with the outcome. This association was robust as this variable within the 100-m buffer was also significantly linked with the independent variable. Interpretation may be helped by the fact that this explanatory variable was strongly inversely correlated with the NDVI which evaluates the density of vegetation. Plenty of studies have highlighted the association between shade - that may be brought by vegetation - and the presence of Aedes immature stages [31, 38, 39]. Shade may lower any very high water temperature that has been specifically recognized as a negative factor for the presence of Aedes aegypti larvae [3]. Vegetation may also provide nutriments for larvae by the leaves falling in the water [40] whilst nectar could be food for adult mosquitoes [41]. In Martinique, Aedes aegypti is mainly endophilic so vegetation probably does not serve as resting site (Yebakima A., personal communication). While scales of the associations between shade and Aedes aegypti have been found at 2-3m on the ground [39], the results of the present work corroborate previous studies having included remote-sensing images. Indeed, the presence of medium height trees within a 30-m radius buffer zones was associated with adult Aedes aegypti abundance in Arizona [20] and niche modeling using Landsat 7 images (with 30-m spatial resolution) allowed predicting the areas suitable for Aedes aegypti breeding sites in Colombia [13]. Nonetheless, the class “asphalt” probably brought more information than the NDVI variable since it has been retained in the statistical final model. Indeed, asphalt surroundings may not be favorable for the breeding sites to persist long enough to be suitable for full cycle larval development.

Secondly, the mean of maximum humidity for the 5-days period before the field visit was positively associated with the presence of one or several Aedes larvae-positive container(s) in the experimental units. The association was robust since the variables for the 2- to-14 days period before the field visit were also significantly associated with the outcome. Very similar results were found in Brazil [42] and Australia [43]. Humidity is correlated to precipitation and evaporation that are related to longer presence of water in the containers, raising the possibilities for larval development. Studies have already highlighted the linkages between rainfall and Aedes aegypti [44-47] whilst dengue risk models have included this rainfall data as a predictive variable [26].

Accuracy and validity of models

It is recognized that, since number of observations was somewhat small, the validation was not based upon datasets different to those used to fit the models. Nevertheless, the final number of observations that were correctly predicted was 85%. A total of 89% of the Aedes larvae-negative houses and 67% of the Aedes larvae-positive houses were correctly predicted. From an operational point of view, this scenario is very powerful to limit the amount of time spent on the ground but it should be improved in terms of detecting positive houses.

Bias could have been introduced in the models since ground data collection was not accomplished at the same dates for each section and some section were preferentially investigated during rainy/dry seasons. Nevertheless analysis of the risk maps showed that the sections that had been followed mainly during respectively rainy/dry seasons had no particular pattern in terms of prediction of Aedes aegypti larvae presence. Indeed, while an entirely environmental model could not have been fitted due to this bias, the presence of meteorological variables in the models served as a kind of seasonal adjustment. In addition, the six sub-models which were fitted at both steps by omitting successively the experimental units of each section provided similar estimates. The latter means that none of the section could have deeply modified the modeling results.

From entomological risk maps to ground control actions

The type of risk maps drawn in the present study could be included in an operational warning system and targeting vector control preventive actions. It has been actually argued that the use of models should participate in spatially and temporally prioritizing prevention where the risk is the greatest [12]. On one hand, entomological data are rarely collected on a routine in a given area, and when field studies are undertaken, they often provide only a snapshot of a continuous phenomenon. Risk maps are expected to bring this continuity for better evaluation of risks. On the other hand, maps that are available at the household level could facilitate detection of “key estates” [48, 49] for efficient control. Indeed, targeting could be done in the houses that were predicted with high number of Aedes larvae-positive days in a month, which could be seen as a surrogate for productivity.

Conclusion

As it has been often advocated, focusing interventions in the places and periods where the risk is maximal is paramount for better allocate limited resources and improve dengue control. Entomological risk maps may be seen as one of the tools available and tele-epidemiology may be applied in this context. The present study showed that environmental information at fine spatial scale - obtained using satellite images - coupled with field meteorological data also at fine temporal scale were successfully included as explanatory variables in the statistical models. It allowed drawing daily entomological predictive risk maps for the presence or not of one or several Aedes aegypti larvae-positive containers, at the level of individual dwellings. Finally the approach presented in this paper can be applied immediately to the emerging chikungunya entomological risk levels in Martinique since the Aedes aegypti is also the vector for this disease [50].

Acknowledgments

The authors would like to thank David Flamanc from the Direction Regionale de l’Environnement, de l’Aménagement et du Logement – Martinique, for providing Litto3D® data, along with their full use. They acknowledge the SERTIT (SErvice Régional de Traitement d’Image et de Télédétection; Hervé Yésou, Carlos Uribe, Claire Huber) for providing inputs for the definition of the very high resolution indices. VM, CV and JPL thank Sanofi-Pasteur for supporting this study. VM and JPL also acknowledge CNES (Centre National d’Etudes Spatiales) for additional support. This is an LDEO contribution # XXXX.

Bibliography

1. Murray NE, Quam MB, Wilder-Smith A (2013) Epidemiology of dengue: past, present and future prospects. Clin Epidemiol 5: 299-309. 2. Christophers S (1960) Aedes Aegypti (L.) The Yellow Fever Mosquito: Its Life History, Bionomics and Structure; Press CU, editor. 739 pp. p. 3. Hemme RR, Tank JL, Chadee DD, Severson DW (2009) Environmental conditions in water storage drums and influences on Aedes aegypti in Trinidad, West Indies. Acta Trop 112: 59-66. 4. Etienne M (2006) Etude de la bioécologie d’Aedes aegypti à la Martinique en relation avec l’épidémiologie de la dengue. Montpellier. 5. Wan SW, Lin CF, Wang S, Chen YH, Yeh TM, et al. (2013) Current progress in dengue vaccines. J Biomed Sci 20: 37. 6. Iturbe-Ormaetxe I, Walker T, SL ON (2011) Wolbachia and the biological control of mosquito-borne disease. EMBO Rep 12: 508-518. 7. Dame DA, Curtis CF, Benedict MQ, Robinson AS, Knols BG (2009) Historical applications of induced sterilisation in field populations of mosquitoes. Malar J 8 Suppl 2: S2. 8. Bellini R, Medici A, Puggioli A, Balestrino F, Carrieri M (2013) Pilot field trials with Aedes albopictus irradiated sterile males in Italian urban areas. J Med Entomol 50: 317-325. 9. Focks DA, Brenner RJ, Hayes J, Daniels E (2000) Transmission thresholds for dengue in terms of Aedes aegypti pupae per person with discussion of their utility in source reduction efforts. Am J Trop Med Hyg 62: 11-18. 10. Eisen L, Lozano-Fuentes S (2009) Use of mapping and spatial and space-time modeling approaches in operational control of Aedes aegypti and dengue. PLoS Negl Trop Dis 3: e411. 11. Bergquist NR (2001) Vector-borne parasitic diseases: new trends in data collection and risk assessment. Acta Trop 79: 13-20. 12. Scott TW, Morrison AC (2010) Vector dynamics and transmission of dengue virus: implications for dengue surveillance and prevention strategies: vector dynamics and dengue prevention. Curr Top Microbiol Immunol 338: 115-128. 13. Arboleda S, Jaramillo ON, Peterson AT (2012) Spatial and temporal dynamics of Aedes aegypti larval sites in Bello, Colombia. J Vector Ecol 37: 37-48. 14. Neteler M, Roiz D, Rocchini D, Castellani C, Rizzoli A (2011) Terra and Aqua satellites track tiger mosquito invasion: modelling the potential distribution of Aedes albopictus in north-eastern Italy. Int J Health Geogr 10: 49. 15. Roiz D, Neteler M, Castellani C, Arnoldi D, Rizzoli A (2011) Climatic factors driving invasion of the tiger mosquito (Aedes albopictus) into new areas of Trentino, northern Italy. PLoS One 6: e14800. 16. Estallo EL, Lamfri MA, Scavuzzo CM, Almeida FF, Introini MV, et al. (2008) Models for predicting Aedes aegypti larval indices based on satellite images and climatic variables. J Am Mosq Control Assoc 24: 368-376. 17. Rotela CH, Espinosa MO, Albornoz C, Lafaye M, Lacaux J-P, et al. Desarrollo de mapas predictivos de densidad focal de Aedes aegypti en la ciudad de Puerto Iguazú (Argentina), basados en información ambiental derivada de imágenes SPOT 5 HRG1; 2008; La Habanna, Cuba. 18. Fuller DO, Troyo A, Calderón-Arguedas O, Beier JC (2009) Dengue vector (Aedes aegypti) larval habitats in urban environment of Costa Rica analysed with ASTER and QuickBird imagery. Int J Rem Sens 31: 3-11. 19. Vanwambeke SO, Bennett SN, Kapan DD (2011) Spatially disaggregated disease transmission risk: land cover, land use and risk of dengue transmission on the island of Oahu. Trop Med Int Health 16: 174-185. 20. Landau KI, van Leeuwen WJ (2012) Fine scale spatial urban land cover factors associated with adult mosquito abundance and risk in Tucson, Arizona. J Vector Ecol 37: 407-418. 21. Sarfraz MS, Tripathi NK, Tipdecho T, Thongbu T, Kerdthong P, et al. (2012) Analyzing the spatio-temporal relationship between dengue vector larval density and land-use using factor analysis and spatial ring mapping. BMC Public Health 12: 853. 22. Van Benthem BH, Vanwambeke SO, Khantikul N, Burghoorn-Maas C, Panart K, et al. (2005) Spatial patterns of and risk factors for seropositivity for dengue infection. Am J Trop Med Hyg 72: 201-208. 23. Rotela C, Fouque F, Lamfri M, Sabatier P, Introini V, et al. (2007) Space-time analysis of the dengue spreading dynamics in the 2004 Tartagal outbreak, Northern Argentina. Acta Trop 103: 1-13. 24. Neteler M, Metz M, Rocchini D, Rizzoli A, Flacio E, et al. (2013) Is Switzerland Suitable for the Invasion ofAedes albopictus? PLoS One 8: e82090. 25. ECDC (2009) Development of Aedes albopictus risk maps ECDC technical report. Stockholm: European Centre For Disease Prevention and Control. 26. Rogers DJ, Suk JE, Semenza JC (2014) Using global maps to predict the risk of dengue in Europe. Acta Trop 129: 1-14. 27. Vignolles C, Lacaux JP, Tourre YM, Bigeard G, Ndione JA, et al. (2009) Rift Valley fever in a zone potentially occupied by Aedes vexans in Senegal: dynamics and risk mapping. Geospat Health 3: 211-220. 28. Vignolles C, Tourre YM, Mora O, Imanache L, Lafaye M (2010) TerraSAR-X high-resolution radar remote sensing: an operational warning system for Rift Valley fever risk. Geospat Health 5: 23-31. 29. Lacaux J-P, Tourre Y-M, Vignolles C, Ndione J-A, Lafaye M (2006) Classification of ponds from high-spatial resolution remote sensing: Application to Rift Valley Fever epidemics in Senegal. Rem Sens Environ 106: 66-74. 30. Machault V, Vignolles C, Pagès F, Gadiaga L, Tourre YM, et al. (2012) Risk Mapping of Anopheles gambiae s.l. Densities Using Remotely-Sensed Environmental and Meteorological Data in an Urban Area: Dakar, Senegal. PLoS ONE 7. 31. Tun-Lin W, Kay BH, Barnes A (1995) The Premise Condition Index: a tool for streamlining surveys of Aedes aegypti. Am J Trop Med Hyg 53: 591-594. 32. Nogueira LA, Gushi LT, Miranda JE, Madeira NG, Ribolla PE (2005) Application of an alternative Aedes species (Diptera: culicidae) surveillance method in Botucatu City, Sao Paulo, Brazil. Am J Trop Med Hyg 73: 309-311. 33. Peres RC, Rego R, Maciel-de-Freitas R (2013) The use of the Premise Condition Index (PCI) to provide guidelines for Aedes aegypti surveys. J Vector Ecol 38: 190-192. 34. Rouse JW, Hass RH, Schell JA, Deering DW (1973) Monitoring vegetation systems in the Great Plains with ERTS. Third ERTS Symposium, NASA SP 351. 309-317. 35. Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens Environ 8: 127-150. 36. McFeeters SK (1996) The use of the normalised difference water index (NDWI) in the delineation of open water features. Int J Remote Sens 17: 1425–1432. 37. Mason SJ, Graham NE (2002) Areas beneath the relative operating characteristics (ROC) and levels (ROL) curves: statistical significance and interpretation. . Q J Roy Meteor Soc 128: 2145-2166. 38. Vezzani D, Albicocco AP (2009) The effect of shade on the container index and pupal productivity of the mosquitoes Aedes aegypti and Culex pipiens breeding in artificial containers. Med Vet Entomol 23: 78-84. 39. Vezzani D, Rubio A, Velazquez SM, Schweigmann N, Wiegand T (2005) Detailed assessment of microhabitat suitability for Aedes aegypti (Diptera: Culicidae) in Buenos Aires, Argentina. Acta Trop 95: 123-131. 40. Reiskind MH, Greene KL, Lounibos LP (2009) Leaf species identity and combination affect performance and oviposition choice of two container mosquito species. Ecol Entomol 34: 447-456. 41. Martinez-Ibarra JA, Rodriguez MH, Arredondo-Jimenez JI, Yuval B (1997) Influence of plant abundance on nectar feeding by Aedes aegypti (Diptera: Culicidae) in southern Mexico. J Med Entomol 34: 589-593. 42. Favier C, Degallier N, Vilarinhos Pde T, de Carvalho Mdo S, Yoshizawa MA, et al. (2006) Effects of climate and different management strategies on Aedes aegypti breeding sites: a longitudinal survey in Brasilia (DF, Brazil). Trop Med Int Health 11: 1104-1118. 43. Azil AH, Long SA, Ritchie SA, Williams CR (2010) The development of predictive tools for pre-emptive dengue vector control: a study of Aedes aegypti abundance and meteorological variables in North Queensland, Australia. Trop Med Int Health 15: 1190-1197. 44. Wee LK, Weng SN, Raduan N, Wah SK, Ming WH, et al. (2013) Relationship between rainfall and Aedes larval population at two insular sites in Pulau Ketam, Selangor, Malaysia. Southeast Asian J Trop Med Public Health 44: 157-166. 45. Baruah S, Dutta P (2013) Seasonal prevalence of Aedes aegypti in urban and industrial areas of Dibrugarh district, Assam. Trop Biomed 30: 434-443. 46. Duncombe J, Clements A, Davis J, Hu W, Weinstein P, et al. (2013) Spatiotemporal patterns of Aedes aegypti populations in Cairns, Australia: assessing drivers of dengue transmission. Trop Med Int Health 18: 839-849. 47. Stewart Ibarra AM, Ryan SJ, Beltran E, Mejia R, Silva M, et al. (2013) Dengue Vector Dynamics (Aedes aegypti) Influenced by Climate and Social Factors in Ecuador: Implications for Targeted Control. PLoS One 8: e78263. 48. Tun-Lin W, Kay BH, Barnes A (1995) Understanding productivity, a key to Aedes aegypti surveillance. Am J Trop Med Hyg 53: 595-601. 49. Chadee DD (2004) Key premises, a guide to Aedes aegypti (Diptera: Culicidae) surveillance and control. Bull Entomol Res 94: 201-207. 50. Simon F, Savini H, Parola P (2008) Chikungunya: a paradigm of emergence and globalization of vector-borne diseases. Med Clin North Am 92: 1323-1343, ix.

Figure legends

Figure 1. The Martinique Island, the studied area and the six studied sections on the Tartane Peninsula. Figure 2. Map of the sampled houses. Figure 3. Scenario retained for dengue entomological risk mapping. Figure 4. Monthly entomological risk maps from January until December 2010, from the modeling experiment. The number of days when the 983 houses within the studied area were predicted as Aedes larvae-positive, are given (see colored code at bottom left).

Supporting information legends

Supporting Information 1. Monthly entomological risk maps resulting from modeling and including the number of days when the 983 houses of the studied area were predicted as Aedes larvae-positive. Power point presentation.

Table 1. Environmental indicators calculated from the Geoeye-1 image at 0.41-m spatial resolution. Environmental indicator Spectral bands combination* Description NDVI (Normalized Difference Vegetation Index) [34,35]

A value superior to 0.2 usually corresponds to a vegetated area, which gets denser when this value rises. Negative values indicate non-vegetated features such as barren surfaces (rocks and soils), water, built-up areas or asphalt. NDWI Mac Feeters (Normalized Difference Water Index) [36]

It delineates open water features while eliminating the presence of soil and terrestrial vegetation features. Its value increases with the presence of water and decreases with the presence of vegetation. It is also suggested that it may provide turbidity estimations of water bodies. ANDWI (Adapted NDWI Mac Feeters Index) Using the blue band, this adapted NDWI Mac Feeters maximizes the detection of water. * NIR : Near infrared, SWIR : Short wave infrared

Table 2. Remote-sensing environmental and ground meteorological variables significantly associated with the presence of one or several water-filled container(s) in an experimental unit. Multivariate logistic regression with section random effect is given (Step 1). Explanatory variables Coefficient 95% confidence interval p-value 158 experimental units, 6 groups Surface of the class “sparsely vegetated soil” in the house plot (per 10 m²) 0.10 [0.02 ; 0.19] 0.017 Surface of the class “lawn” within the 50-m buffer around the house (per 100 m²) -0.10 [-0.17 ; -0.03] 0.007 Total rainfall during the 4 days preceding the field visit (per 10 mm) 0.26 [0.07 ; 0.46] 0.007 Section random effect 0.498

Table 3. Remote-sensing environmental and ground meteorological variables significantly associated with the presence of one or several Aedes larvae-positive container(s) in an experimental unit that holds water-filled container(s). Multivariate logistic regression with section random effect is given (Step 2). Explanatory variables Coefficient 95% confidence interval p-value 88 experimental units, 6 groups Surface of the class “asphalt” within the 50-m buffer around the house (per 100 m²) -0.10 [-0.17 ; -0.04] 0.003 Mean of maximum humidity during the 5 days preceding the field visit (per 1%) 0.27 [0.07 ; 0.48] 0.008 Section random effect 0.209

Table 4. Final predictions of the scenario versus field data: number of Aedes larvae-positive experimental units versus number of Aedes larvae-negative experimental units (i.e., without water-filled container(s) or with Aedes larvae-negative container(s)). Prediction from scenario Number of Aedes larvae-negative experimental units Number of Aedes larvae-positive experimental units Total Field data Number of Aedes larvae-negative experimental units 115 13 128 Number of Aedes larvae-positive experimental units 13 17 30 Total 128 30 158

In the same section…

Keywords

Related articles