Anopheles atroparvus density modeling using MODIS NDVI in a former malarious area in Portugal



Malaria is dependent on environmental factors and considered as potentially re-emerging in temperate regions. Remote sensing data have been used successfully for monitoring environmental conditions that influence the patterns of such arthropod vector-borne diseases. Anopheles atroparvus density data were collected from 2002 to 2005, on a bimonthly basis, at three sites in a former malarial area in Southern Portugal. The development of the Remote Vector Model (RVM) was based upon two main variables: temperature and the Normalized Differential Vegetation Index (NDVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra satellite. Temperature influences the mosquito life cycle and affects its intra-annual prevalence, and MODIS NDVI was used as a proxy for suitable habitat conditions. Mosquito data were used for calibration and validation of the model. For areas with high mosquito density, the model validation demonstrated a Pearson correlation of 0.68 (p<0.05) and a modelling efficiency/Nash-Sutcliffe of 0.44 representing the model's ability to predict intra- and inter-annual vector density trends. RVM estimates the density of the former malarial vector An. atroparvus as a function of temperature and of MODIS NDVI. RVM is a satellite data-based assimilation algorithm that uses temperature fields to predict the intra- and inter-annual densities of this mosquito species using MODIS NDVI. RVM is a relevant tool for vector density estimation, contributing to the risk assessment of transmission of mosquito-borne diseases and can be part of the early warning system and contingency plans providing support to the decision making process of relevant authorities.


Vector-borne diseases such as malaria have a global distribution and affect a growing number of people every year (Snow et al. 2005). It is estimated that in 2006 there were 247 million cases of malaria, with 881,000 deaths, of which 92% were children under the age of five, not accounting for the large indirect burden on morbidity and mortality (WHO 2008). In temperate regions, malaria is considered as a potentially re-emerging disease due to ongoing environmental and temperature changes (Lindsay and Birley 1996, Patz et al. 2005, Schroder and Schmidt 2008), land use changes (Lindblade et al. 2000, Patz et al. 2004, Munga et al. 2009) and increases in tourism fluxes (Linard et al. 2009).

Malaria epidemiology is governed by a large number of environmental factors, that come into effect at different spatial scales affecting its distribution, transmission intensity, disease outcome, small-scale variation, and seasonality, making it highly environmentally sensitive (Craig et al. 1999, Craig et al. 2004a, 2004b). Temperature is a major factor, as it affects both mosquito and parasite survival and consequently the length of the transmission cycle (Craig et al. 2004a, Depinay et al. 2004). Rainfall and standing water are other important factors, as mosquitoes require water as larvae, vegetation for sugar feeding and resting sites, and mosquito survival is favored by high relative humidity (Service and Townson 2002). However, although the relationship between mosquito density and rainfall has been repeatedly demonstrated (Craig et al. 2004a, Gomez-Elipe et al. 2007), it is not a direct one, as a certain amount of rain does not lead to a specific mosquito density. In fact, the mosquito density dependence on precipitation is only valid for areas where the availability of water is directly dependent on rainwater. In stable habitats such as lakes, ponds, swamps, marshes, or rice fields, the lack of rain may not be a factor in the survival of mosquitoes (Service and Townson 2002).

Remote sensing data have been successfully used for monitoring environmental variables such as land cover mapping, vegetation indices, landscape structure, and distribution of water bodies as conditions that influence the distribution patterns of arthropod vector-borne diseases (Beck et al. 2000). Advanced Very High Resolution Radiometer data were used to infer ecological parameters, namely vegetation, associated with Rift Valley Fever outbreaks (Linthicum et al. 1987). Satellite images from Landsat, SPOT, and IKONOS have been used to characterize and predict mosquito larval habitats (Rejmankova et al. 1998, Masuoka et al. 2003, Pope et al. 2005, Mushinzimana et al. 2006). Spatiotemporal variability of the normalized difference vegetation index (NDVI) has been positively correlated with incidence rates of vector-borne diseases such as malaria, dengue, West Nile virus, and hence epidemic outbreaks (Brown et al. 2008, Tourré et al., 2008, Ward, 2009). Studies of malaria incidence in Africa or Asia have shown their association with, or actually been modeled by, weather, rainfall, temperature, NDVI (Gomez-Elipe et al. 2007, Funk and Brown 2006, Liu and Chen 2006, Gaudart et al. 2009), and by the enhanced vegetation index (Noor et al. 2008). In order to identify areas of high malaria risk, other studies focus on topographic constraints such as elevation, wetness index, flow distance to stream, aspect of land surface, and curvature (Mushinzimana et al. 2006, Bodker et al. 2003, Githeko et al. 2006, Cohen et al. 2008).

Malaria was a major factor in mainland Portugal during the first half of the 20th century but was considered to have been eliminated by the WHO in late 1973 (Bruce-Chwatt and Zulueta 1977, Cambournac 1994). Anopheles atroparvus Van Thiel 1927 was the only vector species, and it is still one of the most abundant and widespread mosquitoes, with high densities in coastal and inland regions (Almeida et al. 2008) with a spatial distribution closely resembling former malarial areas (Capinha et al. 2009). Immature stages of An. atroparvus are usually found in clean, sunlit, standing brackish and fresh water bodies (Jetten and Takken, 1994). It is a marked zoophilic species with endophagic and endophylic behaviors (Cambournac and Hill 1938, Cambournac 1942), allowing the vector to endure adverse climatic conditions. Females can overwinter, surviving for several months in animal shelters and blood feeding on the hosts that are present (Cambournac 1942).

Portugal, like southern Europe, has potential conditions for the development of vector-borne diseases, due to a large mosquito fauna, climate profile, and local characteristics such as wetlands, artificial water reservoirs, and bird sanctuaries that may offer appropriate conditions to vector population development, increasing the risk of infection to animals and humans (Almeida et al. 2008). Future climate conditions, mainly temperature, have been referred to as drivers to vector-borne disease emergence or re-emergence in temperate climate countries (Martens et al. 1997, Hay et al. 2002, IPCC 2007), such as in Portugal (Calheiros et al. 2006), as a cause for public health concern.

Autochthonous cases of malaria have been reported from malaria free countries of Italy, Germany, and Spain (Baldari et al. 1998, Kruger et al. 2001, Santa-Olalla Peralta et al. 2010). In view of this possibility, it is crucial to have a practical tool to assess, at any given moment, the areas at risk for malaria transmission. The estimation of the density of adult mosquitoes is of foremost importance, as malaria control interventions require the estimation of the intensity of transmission through calculation of the vectorial capacity and entomological inoculation rate of the local mosquito population (Macdonald 1957, Garrett-Jones 1964, Graves et al. 1990). These estimates are based on the assessment of the density of mosquitoes relative to humans, and of their survival rate, among other parameters. Modelling the malarial vector density can also highlight its relationship to environmental and climate conditions by integrating spatial and temporal patterns of occurrence and by assisting the prevention and responses in case of emergence or re-emergence of the disease.

In this paper we present the development of the Remote Vector Model (RVM) that provides estimations of the density of the former malarial vector An. atroparvus as a function of temperature, as the main climate driver and of MODIS NDVI, applied to the conditions of southern Europe. This type of tool contrasts with the laborious and time-consuming entomological surveys and can make a definite contribution toward the decision support systems of health authorities.


Study area

The study area was located in southwestern Portugal and comprised a coastal land strip 15–20 km long and 5–10 km wide, starting on the left bank of the Sado River and running south along the coast. It was a flatland area with altitudes between sea level and 60 m, presenting different land use classes (Figure 1). The north and northwest part of the study region are environmental protected areas, mainly occupied by marshlands. In the west, along the seashore and next to the sand dune system, there are 600 ha of rice fields. The south is occupied by pine forest and some semi-natural agro-forestry systems of evergreen-oaks.

Figure 1.

Geographical location of the study area.

Monitoring of adult mosquito populations was conducted at three sites: Carvalhal (38° 18′ 22.74″N, 8°45′ 3.6′'W), Comporta (38° 22′ 50.88′'N, 8° 46′ 57.18″ W), and Pego (38° 17′ 27.24′'N, 8° 45′ 32.64″ W), which are within a maximum distance of 9.61 km (Comporta and Pego) and a minimum distance of 2.42 km (Carvalhal and Pego). The sites present distinct land use characteristics according to CORINE Land Cover 2000 land use datasets (European Environment Agency 2000). Forest dominates in Comporta and Pego and rice fields in Carvalhal. Comporta also has a large area of rice fields with some salt marshes and woodland shrubs. Other agricultural areas occur mainly in Pego, which is also the site with the smallest proportion of rice fields. Urban areas represent a small fraction of all sites with the highest population concentration located in Comporta. According to the 2001 population census, the civil parishes of Comporta and Carvalhal (administrative divisions of the Grândola municipality) have a population of 2,948 human activities are concentrated in the primary and tertiary sectors of the economy with rice-growing and tourism as the leading activities.

Climate in the study area is typically Mediterranean with an Atlantic influence. Temperatures range from a monthly average of 10° C in January to 21° C in August, with a total annual precipitation between 500 and 700 mm concentrated from October to May. In the summer season, the total precipitation is about 20 mm per month and the temperatures are relatively high. Climate data, daily temperature, and precipitation were gathered from the Portuguese Meteorological Institute monitoring station located in Comporta village.

The anomalies for monthly temperature and precipitation during the sampling period, compared with the 1960–2000 climatological series, are shown in Figure 2. The total precipitation registered for the hydrological year of 2003–2004 and 2004–2005 was 50% below the normal series, with 2005 presenting values lower than 200 mm.

Figure 2.

Monthly temperature and precipitation anomalies for 2002–2005, with reference to the Climatological Normal 1960–1990. (Source: Portuguese Meteorological Institute).

Collection of data and data sets used

Adult mosquitoes were monitored from indoor resting collections using hand mechanical aspirators, twice a month between May, 2001 and October, 2005. Mosquitoes were identified with keys for the Portuguese fauna (Ribeiro and Ramos 1999). The places monitored were all low height animal shelters, where hosts were present the entire year. Two neighboring animal shelters were chosen per locality (Comporta, Carvalhal, and Pego), and selected based on mosquito data collected during a whole year before the beginning of the study, in which this method also proved to be the best to monitor the population of this species when compared with CDC traps and human landing collections. Using these preliminary data, shelters that gathered the best characteristics to act as An. atroparvus resting places were selected, as these would provide representative estimates of their population densities, represented as the averaged number of female mosquitoes captured per collector per min (fem.col–1.min-1). Collections were carried out for a standard period of 10 min in each shelter and each collection.

The NDVI, as a measure of the photosynthetic activity of plants, was used as a proxy for suitable conditions of mosquito development, as it refers to spatial and temporal dynamics of different vegetation types that are present naturally around the areas where immature stages of the vector are found (Cambournac and Hill 1938).

MODIS/Terra Vegetation 16-Day L3 Global 250m (MOD13Q1 NDVI was selected due to: i) compatibility between the 16 days compositing time with mosquito development; ii) appropriateness of 250 m spatial resolution to study changes in vegetation dynamics at a local scale that was assumed in the model as a proxy of the vector habitat, and iii) suitability of data series ranging from 2000 onwards. Climate data in the form of daily temperatures and precipitation were gathered from the Portuguese Meteorological Institute monitoring station located in Comporta village. As no other meteorological station is present in the study area and the sample sites are relatively close, the climate data weres used for all sites without any spatial interpolation.

Data processing and integration

The Remote Vector Model (RVM) attempts to estimate An. atroparvus density using two drivers: i) temperature and ii) NDVI. These variables affect the vector density in different ways: temperature serves mainly as a limiting factor of mosquito development; NDVI is used as a proxy for the quantification of mosquito development as it characterizes vegetation development; and indirectly, water presence, is a sign of physical conditions favorable for mosquito development. The model development was supported by mosquito data collection on the three sampling sites mentioned above.

Precipitation was also a target of a preliminary analysis regarding its possible integration in the model. Pearson correlations were conducted between An. atroparvus density values and different time spans of precipitation values. Time lags between precipitation and mosquito densities were also used to assess a potential influence of rainfall and subsequent mosquito development. The results were inconclusive for all tests and no relation was found (data not shown). This led to the exclusion of precipitation as a variable for the RVM.

As part of the preliminary analysis, Pearson's correlations between vector samples and averaged daily temperatures for several time spans, between two and 30 days prior to mosquito collection, were analyzed. The average of the mean daily temperature for the 15 days showed the best Pearson correlation coefficient (r = 0.60, p<0.05) with An. atroparvus densities and was used as the temperature variable in the model, designated T15. Temperature is considered a limiting factor for mosquito density, as under low temperatures there is no development even though all other habitat conditions may be present. However, high temperatures are no guarantee that high mosquito densities may occur, as these depend also on other environmental factors.

NDVI datasets were re-projected using the MODIS reprojection tool. The dataset was then evaluated for geo-location problems, but comparison with Eurosion Study GIS coastline (European Commission 2004) and CORINE Land Cover 2000 land use datasets (European Environment Agency 2000) showed no significant shifts in the dataset for the study area.

For each sampling site, the pixels presenting a VI Usefulness Index quality flag above “0010 – Good quality” were selected using the “VI Quality Assessment Science Data Sets” on a pixel basis. The 16-days compositing NDVI values were linearly interpolated through time to match the actual date of the vector sampling.

NDVI values registered in the area surrounding each sampling site were analyzed through different window sizes, ranging from 250 to 2500 m (Figure 3), with the purpose of assessing the spatial variability of NDVI with An. atroparvus density values. For each NDVI window, a set of statistics was calculated: i) Percentiles 50, 75, 80, 90, and 95; ii) average, and iii) mode. Pearson correlation coefficients between these values and the An. atroparvus density were calculated excluding samples with low temperature values. The five pixel window (1250 m) presented the most consistent results, with the 75th percentile (3rd quartile) selected as the best compromise with high correlation value (r= 0.65; p<0.005) and spatial representativeness (number of pixels = 96), as shown in Figure 4. Percentile 75 represents the 3rd quartile of the NDVI distribution within each window. There is an assumption that five pixel windows that have a high percentile 75 value will have appropriate conditions for having high vector density. Furthermore, field observations with mark-release-recapture revealed the maximum dispersion distance of An. atroparvus in the study area to be within the chosen window of 1,250 m (CAS unpublished observations).

Figure 3.

Example of two NDVI analysis windows for the Comporta and Carvalhal sample sites.

Figure 4.

Pearson correlation between the P75 NDVI values and An. atroparvus density values, for different window sizes.

Assessment of independence between the two main variables was conducted to exclude possible colinearity issues, because in certain environmental conditions temperature and NDVI can present a strong correlation. Covariance values between NDVI and T15 were calculated for the period where T15 > 15° C, presenting a low correlation (Pearson correlation coefficient of 0.20, p<0.05), thus being sufficiently independent and suitable for integration in the model.

Remote Vector Model

The influence of temperature was integrated in the model through its effect on mosquito density. The assumption made was that low temperatures prevent vector development and high temperatures are not directly responsible for high density values. The effect of temperature on mosquito density varies from zero, when temperature inhibits all vector growth even if all the appropriate habitat conditions are suitable, to one, where temperatures present no restriction to vector development. The point where there is the inflection between high temperature restriction and low or no temperature restriction is given by the Tmax that will be subject to calibration. The temperature effect (TE) affects vector density through a simplified logistic curve defined by:


where T15 is the average of the mean daily temperature (° C) for the 15 days prior to mosquito collection; and Tmax is the temperature where TE increases exponentially. At low temperature values, mosquito density values are low. As temperature increases and crosses the Tmax value, high vector density values are possible, since its limitation effect decreases.

Regarding possible limitations to vector development caused by high temperatures these were unable to be assessed due to the meteorological conditions in the study area. The highest temperatures registered at the three sites were not enough to assess any impact in vector density. This prevented the development of a second set of equations that would model its limiting effect.

The RVM model considers the NDVI component through i) the NDVI value, a specific moment given by the percentile 75 of the 1,250 m analysis window, ii) a NDVImax, where it is assumed that the vector density reaches its maximum growth, iii) the density–NDVI sensitivity parameter (B) that determines how vector density behaves with a change in the NDVI, and iv) the estimated maximum vector density (C) that a specific site can sustain. NDVImax and B were subject to calibration in the model development stages. Parameters were integrated using a generalized logistics curve (Richards 1959), which is a widely used function for growth modeling, according to:


The C parameter is a ceiling value for vector density based on the assumption that any area, due to its habitat conditions, can support vector density up to a certain carrying capacity. Although this value is unknown and its estimation constituted a challenge, the approach to its calculation involved the use of historical data from Portugal to establish a relationship between high NDVI and high An. atroparvus density values. Seven sample points from the Alentejo area (a southern Portuguese region) where campaigns were conducted during the spring-summer of 2004 and 2005, as well as the three Comporta region sample sites were used in the determination of the C value (data not shown).

The relationship between the 99th percentile of vector density data and 95th percentile of NDVI, excluding winter time values that are temperature limited, thus presents high values of NDVI associated with low mosquito densities. A linear relationship was found for the maximum vector densities, based on high NDVI values was established with statistical significance, C = 42.3*P95NDVI-14.2, R2= 0.89 (r = 0.94, p ≤ 0.005), suggesting the best conditions for vector growth. NDVImax represents the point where the mosquito density increase is greater, influencing how the NDVI relates with the vector density.

Different NDVImax values will change the range of NDVI values that affect vector density. For low values of NDVImax, density increase occurs at low levels of NDVI, reaching the C value (maximum density possible for the specific location) at medium NDVI values. High values of NDVImax indicate that density response only starts at high values of NDVI. This parameter expresses a key model concept as it establishes the range of NDVI values that induce changes in vector population.

The B parameter (vector sensitivity to NDVI changes) relates the way vector density reacts to a NDVI change around the NDVImax. An increase of B will increase the sensitivity of the vector density to NDVI changes around the NDVImax Point. High B values establish a sharper density growth curve as they reduce the scope of values where there is change in vector density. Vector density, expressed as fem.col-1.min-1, can be estimated through:


Model calibration and validation

Model calibration relied on data from three sample sites for the years 2002 to 2005. An exhaustive simulation exercise was performed varying all parameters at increments, namely: i) NDVImax, from 0.5 to 0.7, by 0.01 intervals; ii) Tmax from 14 to 20, by 0.1° C intervals; and iii) B, from 5 to 20, by unit increments

Assessment of accuracy was achieved through the computation of various statistical coefficients (Janssen and Heuberger 1995, Beven 2000), namely: i) Pearson correlation coefficient (r), to assess the global model accuracy; ii) relative mean bias (RMB), to assess the bias between averaged observed data to averaged modelled data; iii) variance ratio (VR), by comparing the variance of the observed data with the variance of modeled data to evaluate how the model expresses variance compared with the field observations; and iv) Nash-Sutcliffe modeling efficiency (NS) and normalized root mean square error (NRMSE), to allow a pair-wise comparison integrating absolute errors in the analysis.

Model development workflow

The model development and implementation, including the calibrating and validation phase can be simplified in four main stages:

  • 1) Preliminary analysis where the input variables were selected.
  • 2) Function selection where the set of equations that would model the effect of temperature and NDVI on the vector density were selected.
  • 3) Model parameterization and calibration when all the internal model parameters were subject to calibration in order to determine the optimal values for each of the parameters.
  • 4) Model validation using the parameterization calculated in the calibration phase, the model was initialized and the results compared to observed data. The data used for calibration purposes were removed from this process. All these processes are presented in Figure 5.
Figure 5.

Simplified workflow of the model development.



The calibration procedure combined all three sample sites in six simulation sets. The simulation sets included data from the Carvalhal (sample size = 80), Comporta (sample size = 77), and Pego (sample size = 62) sites. Furthermore, combinations of the sites were also created (Comporta + Carvalhal, Comporta + Pego, Carvalhar + Pego). The model calibration exercise entailed 40,000 model runs per calibration dataset and consisted of changing all the input parameters with the purpose of identifying convergence in the model runs and determining an optimized calibration parameterization. For visualization purposes, calibration results in the assessment parameters in Figure 6a–f (r, NRMSE, and NS) were all averaged.

Figure 6.

Assessment coefficients: a) and b) Assessment coefficients for the NDVImax parameter; c) and d) Assessment coefficients for the Tmax parameter; e) and f) Assessment coefficients for the B parameter.

Calibration exercise 1

The Carvalhal sample site was used as the calibration dataset for the first exercise, being the assessment of convergence for NDVImax, Tmax, and B shown in Figure 6a–f. This calibration dataset presents a good model adjustment for all the assessment coefficients. For NDVImax=0.6 there is a maximization of the NS values, minimization of NRMSE, and a relative mean bias around zero. Tmax convergence occurs around 16.5° C and the optimum value for parameter B results from a trade-off of assessment coefficients, namely the maximum NS and minimum NRMSE, with around 13 as the selected one.

These convergence parameter values resulted in a good adjustment of predicted vs observed density with an r of 0.89 (p<0.05), NS of 0.8, NRMSE of 0.6, and a RMB of 0.02. Optimum values were NDVImax=0.59, B=12, and Tmax= 16.5° C (Table 2). Comparisons between modeled data and observed data are shown in Figure 7.

Table 2.  Validation dataset results using the calibration parameters.
Validation datasetSample sizer (p<=0.05)NSNRMSEVRRMB
Comporta + Pego1390.640.371.190.660.02
Figure 7.

Comparison between observed and predicted data for Carvalhal site using the calibration optimal values.

Calibration exercises 2–6

A similar methodology was used for the remainder calibration exercises with results shown in Table 1. Analysis of results showed that calibration exercise 1, using the Carvalhal dataset, presented the best convergence between predicted and observed density. The resulting parameter values (NDVImax=0.59, Tmax=16.5, and B=12) were then used as input parameters for the validation process.

Table 1.  Assessment coefficients for the convergence parameter value.
Calibration ExerciseCalibration DatasetNDVImaxTmaxBr (p<=0.05)NSNRMSEVRRMB
4Carvalhal + Comporta0.5916.5140.840.700.790.80–0.02
5Carvalhal + Pego0.6016.5130.850.730.760.750.03
6Comporta + Pego0.6016.5120.640.381.180.660.05

Validation results

Validation was conducted using the sample sites of Comporta and Pego. The model was initialized with the optimum parameterization calculated during the calibration exercise. As the Carvalhal site was selected as the calibration site, it was removed from the validation to avoid any contamination of the results (over-calibration of the model to fit the Carvalhal trends and observations). The validation dataset resulted from the combination of the Pego and Comporta sites. Validation dataset results are shown in Table 2.

Validation for the Comporta site presented a Pearson correlation coefficient of 0.68 (p<0.05) for a modeling efficiency/Nash Sutcliffe of 0.44. The high value of the Pearson correlation coefficient shows a good overall global fit of the predicted data against the observations. Furthermore, the NS value demonstrates that the pair-to-pair comparison between predicted and observed data is also robust. In Figure 8a, the predicted results for a three-year period are plotted against the observations. The model predicts the trend within each year and also provides good fit when comparing different years (peak values of different years are well predicted), demonstrating its capacity to cope with inter-annual condition changes. Intra-annual results are also encouraging as the model is able to predict with accuracy the time when the peak occurs and also the decline of the vector density. The results of this validation show a good intra- and inter-annual adjustment of the model by correctly predicting the intra-annual tendencies and overall peak values, with some displacement between the predicted and observed peak densities. There is a slight underestimation of the predicted data, but near zero values are predicted accurately and overall there is a good adjustment between predicted and observed values.

Figure 8.

a) Comparison between observed and predicted for the Comporta site. b) Comparison between observed and predicted data for the Pego site.

The Pego site, where overall mosquito density values were lower, presented sub-optimal validation results when compared to the Comporta site. Pearson correlation coefficient (r=0.59, p<0.05) showed a good overall adjustment between predicted and observed. Regarding the pair-wise comparison assessed by the NS results, the results, while positive, are less satisfactory (NS = 0.11). These can be explained by the difficulties in the model to predict low vector densities. Inter-annual density (Figure 8b) shows that the model is able to predict the density trend, although with reduced accuracy when compared to the Comporta site. For higher values the same underestimation tendency reported in the Comporta site is seen with the coinciding temporal shift in the estimation of intra-annual density peak.


The Remote Vector Model (RVM) aims to predict the density of An. atroparvus at a local scale in a former malarial area in Portugal, contributing to the calculation of vectorial capacity, crucial to malaria epidemiology models. The study area has a heterogeneous landscape including natural and human-managed land uses integrating different intra- and inter-annual NDVI trends and behaviors, and four years of continuous bimonthly sampling of An. atroparvus density values in three sites supports the model development, constituting a robust data series.

In our study, we observed a strong correlation between An. atroparvus densities and NDVI values. Thus, the underlying concept behind the RVM model is that NDVI can be used as a proxy of favorable vector habitat conditions as it indicates vegetation and, indirectly, water presence, which are conditions necessary for high vector density values as they constitute appropriate breeding conditions.

Calibration for the Carvalhal site achieved a good adjustment of the modeled against the observed density, with an r of 0.89 (p<0.05) demonstrating the overall accuracy of the adjustment and a NS of 0.8 showing that the pair-wise comparison between observed and modeled data was also extremely positive. Using these parameters for the validation sites, we could demonstrate RVM's ability to predict seasonal density behavior with a Pearson correlation coefficient of 0.68 (p<0.05) and a NS of 0.44, demonstrating RVM's ability to predict seasonal vector behavior. It also accounted for inter-annual trends resulting from different climate and habitat conditions, although density values below five female mosquitoes per collector per min were inaccurately predicted. This was observed in 2005, a very dry year with a significant decrease in favorable habitat conditions and a consistent decrease in mosquito density that is corroborated by the model results. The year 2005 demonstrates the model limitation in predicting low peak density values. However, this does not seem to be a serious drawback to the model as density values for this species are usually above ten female mosquitoes per collector per min, in all districts of Portugal, as recent surveys have shown (Almeida et al. 2008). The model presents an increased accuracy for higher density values as its main application to areas considered “hot spots” for mosquito development, such as Carvalhal, which is the site with the highest percentage of rice fields.

The Pego site, where the vector density registered is consistently lower than in the other sites, possibly because this is the area with the smallest proportion of rice field and fewer larval habitats, presented suboptimal results and highlighted the necessity of future development regarding the increase of model accuracy for “non-hotspot” areas. The development of a second set of equations that adjust the model sensitivity when overall environmental site conditions are less favorable can lead to an improvement in results, although this hypothesis would require further field data and validation. Another model limitation concerns a slight temporal shift of the predicted peak density against observed values, that can be explained when drastic changes in population occur, expressing a model sensitivity issue to abrupt climate and habitat changes. The temporal resolution of the input data used (16 days for NDVI, 15 days for temperature) tends to smooth environmental changes limiting model response. This constitutes a challenge to further work regarding the exploratory study of other temperature and NDVI datasets. Sudden drops in observed density values were sometimes observed in the three sampling sites. These can be explained by the use of pesticides (organophosphates) in the rice crop fields that occupy the largest proportion of the study area. As pesticides cause larval destruction while not producing a measurable environmental response, the change is not expressed in the predicted density. These types of oscillations are outside the model scope and study objectives.

Overall, the main assumptions of the model, that the mosquito population depends directly on climate and habitat conditions and that the NDVI can serve as a surrogate for those conditions, seem to be well supported by the results. Further developments include the identification of spatial constraints combined with new sample sites presenting different suitable areas, e.g., wetlands, in order to develop a two-dimensional model component. A study comparing the use of NDVI and enhanced vegetation index as the Vegetation index variable of the RVM model can also be relevant due to possible advantages this product has regarding reduced soil background effect problems. Studies on other remote sensing variables, as evapotranspiration or middle infrared reflectance, can also lead to improvement in modelled results. Finally, the use of land surface temperature derived from Earth observation satellites as the model input temperature is a short-term goal, as it will increase spatial variability of temperature data and improve operational applicability.

The results of this study constitute a reasonable basis for local scale modeling of vector-borne diseases using remote sensing data, NDVI, which can provide a significant contribution to malaria epidemiologic modeling. There is an increased relevance for Mediterranean areas where malaria and other vector-borne diseases pose a risk of re-emergence, such as the outbreaks of Chikungunya and West Nile viruses in Italy, respectively in 2007 and 2008–2009 (Angelini et al. 2007, Barzon et al. 2009), or West Nile in Greece and Spain in 2010 (Papa et al. 2010, Promed 2010), or the most recent cases of autochthonous dengue and Chikungunya in France (Gould et al. 2010, La Ruche et al. 2010). Thus, the RVM presented here can serve as a practical tool to assess what areas are at risk, and constitute an early-warning system that might avoid conventional monitoring campaigns that are costly and time and labor intensive. Additionally, the fact that RVM uses input data that is easily acquirable (temperature data and NVDI) and has overall low computational demands makes it a good option to be integrated in any decision support system. Such a tool can thus be used to help policy makers in defining priorities in the control of related public health emergencies.


This study was supported by the RARIMOSQ project funded by Fundação Calouste Gulbenkian, under contract nº Proc. 35–60624. Project EDEN, (EU grant GOCE-2003–010284 EDEN) partially supported the funding of monitoring campaigns during 2005. (The paper is cataloged by the EDEN Steering Committee as EDEN0062 ( The contents of this publication are the sole responsibility of the authors and can in no way be taken to reflect the views of the European Union. We acknowledge the work of João Rodrigues and Ricardo Alves regarding mosquito collection and processing during 2004/2005. The authors acknowledge the contribution of Dr. João Pedro Nunes and Dr. Nuno Carvalhais on the revision of the modelling calibration methodologies and general comments regarding paper organization and structure. We would like to thank the anonymous reviewers for their comments and contributions that significantly improved this manuscript.