Content uploaded by César Capinha
Author content
All content in this area was uploaded by César Capinha
Content may be subject to copyright.
Available via license: CC BY-NC 4.0
Content may be subject to copyright.
Present habitat suitability for Anopheles atroparvus
(Diptera, Culicidae) and its coincidence with former malaria
areas in mainland Portugal
César Capinha1, Eduardo Gomes2, Eusébio Reis1, Jorge Rocha1, Carla A. Sousa3, V. E.
do Rosário2, A. Paulo Almeida3
1Centro de Estudos Geográficos, Universidade de Lisboa, Alameda da Universidade, 1600-214 Lisboa,
Portugal; 2CMDT-LA, Instituto de Higiene e Medicina Tropical, Universidade Nova de Lisboa, Rua da
Junqueira, 96, 1349-008 Lisboa, Portugal; 3Unidade de Entomologia Médica, Instituto de Higiene e
Medicina Tropical, Universidade Nova de Lisboa, Rua da Junqueira, 96, 1349-008 Lisboa, Portugal
Abstract. Malaria was a major health problem in the first half of the 20th Century in mainland Portugal. Nowadays,
although the disease is no longer endemic, there is still the risk of future endemic infections due to the continuous occur-
rence of imported cases and the possibility of transmission in the country by Anopheles atroparvus Van Thiel, 1927.
Since vector abundance constitute one of the foremost factors in malaria transmission, we have created several habitat
suitability models to describe this vector species’ current distribution. Three different correlative models; namely (i) a
multilayer perceptron artificial neural network (MLP-ANN); (ii) binary logistic regression (BLR); and (iii) Mahalanobis
distance were used to combine the species records with a set of five environmental predictors. Kappa coefficient values
from k-fold cross-validation records showed that binary logistic regression produced the best predictions, while the
other two models also produced acceptable results. Therefore, in order to reduce uncertainty, the three suitability mod-
els were combined. The resulting model identified high suitability for An.atroparvus in the majority of the country with
exception of the northern and central coastal areas. Malaria distribution during the last endemic period in the country
was also compared with the combined suitability model, and a high degree of spatial agreement was obtained (kappa
= 0.62). It was concluded that habitat suitability for malaria vectors can constitute valuable information on the assess-
ment of several spatial attributes of the disease. In addition, the results suggest that the spatial distribution of
An.atroparvus in the country remains very similar to the one known about seven decades ago.
Keywords: Anopheles atroparvus, habitat suitability, malaria, geographical information system.
Introduction
Malaria is still one of the most devastating public
health problems in the world. In 2006, there were an
estimated 247 million cases that caused nearly a mil-
lion deaths, mostly children under 5 years of age
(WHO, 2008). The disease is caused by one of four
species of the genus Plasmodium: Plasmodium falci-
parum Welch, 1897, Plasmodium vivax Grassi and
Feletti, 1890, Plasmodium ovale Stephens, 1922 and
Plasmodium malariae Laveran, 1881. Infection
occurs when the parasites are inoculated into
humans by female mosquitoes of the genus
Anopheles, designated as the vector of the disease.
Portugal, like several other European countries,
had high incidence rates of this disease until the end
of the first half of the 20th century. Although the dis-
ease has been eliminated, the former malaria vector
Anopheles atroparvus Van Theil, 1927
(Cambournac, 1942) is still abundant all through the
Corresponding author:
César Capinha
Centro de Estudos Geográficos
Universidade de Lisboa
Alameda da Universidade, 1600-214 Lisboa, Portugal
Tel. +35 1 21 7940218; Fax +35 1 21 7938690
E-mail: cesarcapinha@hotmail.com
Geospatial Health 3(2), 2009, pp. 177-187
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187
178
country (Ribeiro et al., 1988; Almeida et al., 2008.
The distribution in Portugal of An.atroparvus, and
of other members of the Anopheles maculipennis
species complex, has been the object of several stud-
ies (e.g. Pires et al., 1982; Ribeiro et al., 1992).
However, the continuous representation of this
species distribution is currently unknown. Since vec-
tor control has been and continues to be a common-
ly used strategy in malaria control (Muturi et al.,
2008), this kind of information can be very valuable
in the fight against the disease, should it occur again.
In Portugal, where malaria transmission was mos-
quito-density dependent, several anti-vector meas-
ures were adopted such as the introduction of the
mosquito fish (Gambusia sp.), intermittent rice field
irrigation, improvement of the rice field workers’
accommodations (appliance of protective nets in
beds, doors and windows) and indoor residual
spraying with DDT (Bruce-Chwatt and Zulueta,
1977; Borges, 2001). The implementation of these
measures along with patients treatment resulted in a
large reduction of the incidence of the disease,
which was declared “eradicated” from the country
by the WHO in 1973. Since then, despite an isolat-
ed case in 1975 (Antunes et al., 1987), the occur-
rence of the disease is exclusively in the form of
imported cases. The highest number of these
imported cases was reached with the end of the
Portuguese colonial war in 1974 but much later, in
the period of 1993 to 2003, Castro et al. (2004) still
reported the occurrence of 50 to 85 imported cases
per year. This situation, combined with the existence
of the local An.atroparvus vector which can be
infected with exotic strains of plasmodia such as
P. falciparum (Sousa, 2008), raises the possibility of
future endemic infections.
Climate change is considered a key issue for cur-
rent modifications of the distribution and abun-
dances observed for some mosquito species (e.g.
Pascual et al., 2006; Minárˇet al., 2007). These
changes may affect the transmission capability of
the vector species and are, therefore, of major
importance with regard to public health. Although
the global surface temperature has been rising in the
last century, the growth rates are not homogeneous-
ly distributed (IPCC, 2007). Recent studies for
Portugal show that the country has been experienc-
ing temperature increases in the order of 0.5°C per
decade since 1975 (Miranda et al., 2006). Due to
the many implications that this environmental fac-
tor has on the biology of An.atroparvus
(Cambournac and Hill, 1938) this modest increase
can still already have contributed to change in its
distribution.
Habitat suitability models, used to study the dis-
tribution of several species, were mainly developed
in the fields of biogeography and conservation biol-
ogy. These models are based on the quantification of
relations between the species and several environ-
mental factors considered prominent to its distribu-
tion (Guisan and Thuiller, 2005). Such quantifica-
tions enclose, in general, two possible approaches:
the mechanistic model, based on the knowledge of
the species physiology, and the correlative model
(Guisan and Zimmerman, 2000). The obtained rela-
tions are then applied to spatial data using map
algebra modules and a geographical information
system (GIS). These approaches reduce the need for
highly detailed field surveys and are one of the best
ways to achieve spatially continuous representations
of species distributions and several studies have
adopted this thinking for modelling the distribution
of disease vectors (e.g. Kuhn et al., 2002; Moffett et
al., 2007). In the present context, the study of the
propensity for vector infection through space analy-
sis processes is one of the most potentially reward-
ing approaches. Therefore, the aims of this study
were to:
(i) obtain the first spatially continuous representa-
tion of An.atroparvus habitat suitability in the
territory;
(ii) identify the degree of agreement between the
achieved distribution and former malaria areas;
(iii) verify the validity of using vectors distribution
models to assess the malaria transmission risk;
and
(iv) identify the existence of recent shifts on the
distribution of suitable areas for An.atroparvus.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187 179
In order to achieve these objectives several habitat
suitability models for An.atroparvus in mainland
Portugal were computed. Three different correlative
approaches were used relating records with the pres-
ence or absence of the species with several environ-
mental factors. The obtained results were cross-
validated and compared with the known malaria
distribution in mainland Portugal during its last
endemic period.
Materials and methods
Abundance and distribution data of Anopheles
To gather species distribution data several field
surveys were carried out. The adult mosquitoes
were collected in human and animal facilities using
electrical aspirators. The surveys took place at sev-
eral localities along mainland Portugal, between the
years 2001 and 2003. Mosquito abundance was
computed as the number of mosquitoes collected
per hour and per collector. The obtained results
were, however, inadequate to be used directly, due
to (i) the high variability in the conditions of the sur-
veyed facilities, and (ii) the lack of species-specific
abundances for the different members of the
An.maculipennis complex.
The distinct conditions of the surveyed facilities,
namely the implementation of anti-mosquito
devices, presence or absence of animal hosts and
species as well as different degrees of isolation from
the exterior environment, led to a high irregularity
of abundance values, even for fairly close sites. In
this context, it was not possible to consider the
abundance values as such, i.e. a low level of abun-
dance could not directly be taken as a representation
of a lower suitability area. On the other hand, the
opposite holds true, i.e. places where high abun-
dances were registered indicate an environmental
suitability to the species in question.
Due to the inter-species similarity of adult
Anopheles mosquitoes, the survey results com-
prised abundance records of the whole An.mac-
ulipennis complex. In order to extract abundance
values considering An.atroparvus only, it was nec-
essary to establish a prevalence value of this species
relative to the other members of the complex for
each locality. Ramos et al. (1978) identified
An.atroparvus as the only species of the complex
in Algarve, and the same was found for Alentejo
(Pires et al., 1982). However, in the mountainous
Montejunto-Estrela system and in northern areas,
the complex is also represented by An. maculipen-
nis Meigen, 1818 and some negligible records of
Anopheles melanoon Hackett, 1934 (Sousa, 2008).
For central Portugal, Ribeiro et al. (1992), report
an abundance ratio of nine An.atroparvus for each
An.maculipennis, while Ribeiro et al. (1999a)
found the lower proportion of six to one for the
Serra da Estrela Natural Park. None of the other
studies provide exact proportion values, but they
all report a majority abundance of An.atroparvus
(e.g. Ribeiro et al., 1999b, 2002). Taking this infor-
mation into consideration, a general abundance
ratio of eight An.atroparvus for each An.mac-
ulipennis was established. The abundance records
reported from locations in the Montejunto-Estrela
system and the northern areas were then weighted
using this proportion and the ones above the 6th
decile were extracted from the resulting values.
Totalling 76 records, these areas were considered
as the ones with a higher suitability for
An.atroparvus. However, taking into account that
even these higher abundance values were also pos-
sibly biased due to the variability of the surveyed
places, they were converted to “presence only”
records, disregarding their abundance value.
Besides the presence records, absences were also
attained. These are usually surveyed locations where
the specimens were not found. In our study, and also
due to the different conditions of the surveyed
places, only those sites where other mosquitoes
species were present but An.maculipennis complex
showed null or extremely reduced values of abun-
dance (<2%) were considered. This was done to
remove false absences caused by anti-mosquito
actions or other facility-specific effects. A total of 16
absence records were attained.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187
180
Environmental factors
The gathering of the environmental data with the
highest predictive potential, supported by prior
knowledge about the prevailing factors in the stud-
ied species distribution, is foreseen as one of the
major steps to be taken in species distribution mod-
elling (Araújo and Guisan, 2006). In this sense it is
crucial to make an appropriate review of existing
knowledge concerning the ecological factors consid-
ered influential in the distribution of a species as this
allows a better and safer choice of the independent
variables eligible for integrating the predictive sta-
tistic models. Such choice is, nevertheless, many
times postponed by data unavailability and/or diffi-
culties in its spatial representation.
After compiling both the acquaintance of
An.atroparvus ecology and the available data and
accounting for the possibility that “overfitting” of
the predictive models may occur when a large num-
ber of independent variables with reduced samples
are used, five environmental factors were taken into
account:
(i) mean maximum temperature of the warmest
trimester;
(ii) mean minimum temperature of the coldest tri-
mester;
(iii) mean total annual precipitation;
(iv) wetlands density and suitability index; and
(v) agricultural density and suitability index.
In this approach, the temperature spatial variabil-
ity models are included due to temperature’s direct
influence on species physiology and behaviour
(Cambournac and Hill, 1938) or depending on an
indirect effect due to access to larval sites and vari-
ability of the depth of the water bodies. Based on
the climatological records for the five decades from
1950 to 2000, two datasets were included: the mean
maximum temperature of the warmest trimester and
the mean minimum temperature of the coldest
trimester. The use of two datasets simultaneously to
represent one unique ecological factor prevents spa-
tial and temporal homogenization resulting from
the application of a unique annual mean model. All
information concerning these variables was directly
acquired in raster structure from the WORLDCLIM
project (http://www.worldclim.org) (Hijmans et al.,
2005).
The use of precipitation as a predictor was con-
sidered due to the flushing effects on the productiv-
ity of the mosquito breeding sites, as already verified
for other Anopheles species (Paaijmans et al., 2007).
Moreover, it is strongly related to the availability
and characteristics of An.atroparvus’ habitats in the
aquatic phase of its life cycle, i.e. the depth of water
bodies and the availability of temporary ponds. This
information was also obtained from the
WORLDCLIM project for the 1950-2000 period.
In order to include information about the profu-
sion and proximity of wetlands, which are crucial for
An.artoparvus due to the direct relationship to the
aquatic life cycle (Becker et al., 2003), we employed
the Corine Land Cover 2000 (CLC2000) spatial data
sets for Portugal (scale 1:100,000) which is produced
by the European Environment Agency (EEA)
(http://www.eea.europa.eu/themes/landuse/clc-down-
load). Since this information does not cover the ter-
rain completely, a spatially continuous model based
on wetland densities was elaborated by calculating
wetland density within a circle with 10 km radius.
Different weights were also attributed to each wet-
land type considering its known suitability for
An.atroparvus. The differentiated areas were: rice
fields, inland marshes, large inland water bodies,
estuaries, coastal lagoons, intertidal flats, permanent-
ly irrigated areas, watercourses, and saltpans. Three
different levels of weighting were applied based on
expert knowledge regarding the suitability of each of
the areas to the species larval development. The final
model was obtained through the product of the den-
sity values and the factors used for weighting.
The density model based on agricultural land use
and land cover was taken as indicator of livestock
production. For example, places such as cattle
farms have a high capability for the development
of large populations of An.atroparvus due to the
zoophilic nature of this mosquito species (Sousa,
2008). This approach is based on the broad pre-
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187 181
sumption that areas characterized by a high agri-
cultural intensity have a greater probability of con-
taining cattle herds as the main sign of economic
activity or being used for cattle breeding for
domestic/private consumption. This model includ-
ed the following classes derived from the CLC2000
classification: pastures, non-irrigated arable land,
permanently irrigated land, rice fields, vineyards,
fruit trees and berry plantations, olive groves,
annual crops associated with permanent crops,
complex cultivation patterns, land principally
occupied by agriculture, with significant areas of
natural vegetation and agro-forestry areas. The
final model regarding this factor was obtained fol-
lowing the same method used for the wetlands
model.
Predictive models calibration
When implementing predictive models, the distri-
bution records are generally split into calibration
and validation data sets. However, considering the
relatively small amount of distribution records
(n = 92) in this study, a random multiple partition-
ing of the records was used. This method avoids
excessive use of records for validation purposes,
being more suitable than a single validation set,
when only a small amount of species records is
available (Hirzel et al., 2006). In this sense three cal-
ibration-validation sets were made, each validation
fraction composed of 13 records (2 absences and 11
presences). This number of records corresponds to a
random extraction of about 15% for each record
type, which permitted the species prevalence to
remain unaffected in the calibration sets. In addi-
tion, all spatial predictor models were re-sampled
into a 1 km2grid system.
Predictive systems
Three distinct predictive models where imple-
mented: Mahalanobis distance (D2); binary logistic
regression (BLR); and a multilayer perceptron with
back propagation artificial neural network (MLP-
ANN). The use of several predictive models is sug-
gested as a method for reducing the uncertainty on
species distribution modelling (Pearson et al.,
2006).
Mahalanobis distance calibration
Mahalanobis distance (D2) corresponds to an
n-dimensional distance to the centroid of the pro-
vided values. This measure, when applied to the
habitat suitability modelling, calculates the similari-
ty between the predictors multidimensional mean
transmitted by presence records and each of the cells
composing the study area. In this sense
Mahalanobis distance requires only presence data,
thus removes the influence of possible bias and
errors present in absence sets. However, in general,
presence-only methods have lower predictive per-
formances than presence-absence models (Austin
and Meyers, 1996). Even so, in this group of mod-
els Mahalanobis distance presents some of the best
results (Tsoar et al., 2007). Besides, different meas-
urement scales and correlations amongst the predic-
tors do not affect this similarity index, so it can be
applied to non-normally distributed variables and
its outputs are easily interpreted. This measure was
implemented directly over the spatial data using the
Mahalanobis distances ArcView extension (Jenness,
2003). Due to the existence of three calibration sets
the final model was obtained trough the arithmetic
mean of their resulting outputs.
BLR calibration
Briefly, BLR is a well-known method in habitat
suitability modelling, belonging to the family of gen-
eralized linear models, which is used to estimate the
occurrence probability of a certain event, given the
values of the predictors. BLR is particularly well
suited for species distribution modelling, since it
works with a binary response like presence-absence
and does not assume normal distributions. The
resulting values are expressed in a 0 to 1 interval
and interpreted as the predicted distribution of the
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187
182
species under study. For each of the calibrated mod-
els, the Hosmer-Lemeshow tests (Hosmer and
Lemeshow, 1989) were performed, and all showed
significant values (P <0.05), indicating that the mod-
els adequately fitted the calibration data.
Artificial neural network calibration
Artificial neural networks are adaptive statistical
models based on an analogy with how the brain is
thought to function (Abdi, 2003). Simple units, the
nodes, which are connected by weighted links, con-
stitute the structure. The model calibration (which
can also be called training) takes place by applying
different weights to the signal passing these connec-
tions. In our study, an MLP-ANN composed by three
interconnected layers was used. The nodes in the
input layer receive the predictors values, while the
second layer (the hidden layer) runs the learning
back-propagation process before passing the infor-
mation on to the output layer (presence-absence data)
for prediction. The measured error between the
expected and the actual output provides the basis for
the training process which consists of running numer-
ous cycles until the desired error level is achieved.
Our model included five nodes in the hidden layer,
resulting in as many predictors. This was done to
reduce the risk of “overfitting” of the model which is
favoured by the use of an excessive number of param-
eters in non-parametric models (Liu et al., 2007). The
error measure used was the mean square error and
the networks were trained through 400,000 cycles.
The heuristic nature of artificial neural networks
implies that different modelling episodes produce
different results. In order to deal with this variabili-
ty a common approach is to calculate the mean of a
set of outputs. In this way the final model of each
calibration set was represented by the arithmetic
mean of three different calibration outputs. The
final model corresponded to the arithmetic mean of
these means.
Calibration calculations
The BLR and the MLP-ANN results come in a
range from 0 to 1, while the Mahalanobis results,
being standardized values of similarity, can vary
from 0 (maximum similarity) to any positive num-
ber of maximum dissimilarity. In this study, the final
Mahalanobis model was normalized into a 0 - 1
Fig. 1. Habitat suitability for An. atroparvus across Portugal obtained by the different correlative methods: a) Mahalanobis
distance, b) logistic regression, and c) artificial neural network.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187 183
interval based on its mean and standard deviation
values. These values were then inverted, since lower
values of this measure correspond to higher similar-
ity with the places where the species was recorded.
The different habitat suitability models can be seen
in Figure 1.
Before using the obtained models to effectively
characterize the distribution of the suitability areas for
An. atroparvus, it was necessary to evaluate their pre-
dictive performances. Since it is only possible to eval-
uate the appropriateness of the results for the intend-
ed purpose after this procedure has been carried out,
it is considered a fundamental stage in species distri-
bution modelling (Araújo and Guisan, 2006). In the
context of the present work, the results evaluation
also allowed us to differentiate the performances
obtained by each of the modelling methods.
The kappa index (Cohen, 1960) was used as pre-
dictive evaluation measure. This index expresses the
amount of agreement between the correct classifica-
tions achieved and the ones expected merely by
chance. Its values generally vary from 0 (but can be
negative), representing results identical to the ones
expected by chance, and 1 corresponding to a per-
fect agreement between actual and obtained results.
To pro ceed with the calculation of this index it is
necessary to establish suitability threshold values. A
commonly adopted approach is to use a range of
thresholds and calculate the corresponding kappa
value for each one of them (e.g. Elith et al., 2006)
adopting the maximum value achieved as the pre-
dictive performance of the model. All three models
were partitioned in 20 intervals of equal amplitude
(0.05) and the kappa value calculated for each. The
obtained results showed BLR to be the best per-
forming model (κ= 0.77). According to Landis and
Koch (1977), this value corresponds to the existence
of an excellent agreement between the model results
and the validation records. The model obtained by
MLP-ANN presented a fairly lower result (κ= 0.51)
which even so is considered by the same authors to
be a good level of agreement. The lowest predictive
performance value was registered with the
Mahalanobis model with κ= 0.42, which is close to
the lower limit of the good-agreement range.
Considering that all three models achieved good
to excellent agreement levels, it was decided to com-
bine the output achieved, i.e. all the binary maps.
This conjugation allowed us to reduce the uncer-
tainty that would have resulted from to the output
of a single method (Pearson et al., 2006), which is
similar what has been proposed by Araújo and New
(2006) concerning uncertainty reduction for pro-
jecting species distributions in future climate scenar-
ios. Figure 2 shows the resulting model amounting
to the agreement between the three binary models.
In order to assess the spatial similarity between
the former malaria areas and habitat suitability
model, the kappa value was calculated. Taking into
account the existence of four classes in the com-
bined suitability model, different binary class com-
binations were created in order to be comparable
with the binary map of malaria distribution. These
outcomes are expressed in Table 1.
Results
The obtained kappa values (Table 1) show differ-
ent agreement levels between all of the binary suit-
ability models and the former malaria distribution
Fig. 2. Combination of binary habitat suitability models for
An. atroparvus across Portugal.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187
184
in mainland Portugal. Amongst the single-model
outputs, and in contrast to previous validation
results, the MLP-ANN achieved the highest agree-
ment (κ= 0.42), followed by BLR (κ= 0.38) and D2
(κ= 0.36). In general, all of the outputs resulting
from multi-model combination show higher predic-
tive performance than single-model binary distribu-
tions. The exception is model L1, which presents the
lowest value (κ= 0.32). The agreement value is
much higher when the disagreement areas between
the models are removed (model L4, κ= 0.62).
The resulting suitability maps for An.atroparvus
reveal a clear spatial separation in the way the coun-
try fits its habitat preferences. The northwest and
central coastal areas present lower suitability, while
the south and interior central and northern areas
comprise almost the entire distribution of higher
suitability values.
Discussion
The habitat suitability threshold where maximum
kappa is achieved can be used to produce binary
maps of suitable and unsuitable areas for the target
species. Although information is lost in this conver-
sion, a binary map based on the maximum kappa
value allows a more precise spatial discrimination
between areas of higher and lower suitability. This
transformation is useful in the present case where
the Mahalanobis model presents a clear concentra-
tion of values close to the maximum of suitability
(Fig. 1a). This is also valid for both BLR and
Mahalanobis models, where a direct interpretation
of the continuous values can be biased since the cut
values obtained by the maximum kappa (BLR =
0.65; 0.7 – D2= 0.95) are distinct from the expect-
ed theoretical value of 0.5. In this sense three bina-
ry models were built based on the suitability thresh-
old where the maximum kappa was achieved or
their mean value when more than one threshold
occurred.
In the obtained combined model, a clear distribu-
tion pattern emerges. Unsuitability agreement cov-
ers approximately all the northwest of the country,
while the south and the interior north are almost
entirely occupied by the least uncertain suitable
areas. This spatial pattern is fairly coincident with
the distribution of precipitation and temperature
variations along the country. Unsuitable areas
match coarsely to the wettest places with mild tem-
peratures in the summer, while dryer areas reaching
higher temperatures are almost entirely suitable
areas. Unsuitable areas are seen here not as places
with complete absence of the species, but instead as
least suitable places where lower abundances occur.
This is in agreement with the distribution pattern
achieved by several studies based on direct interpre-
tation of field data. For example, Almeida et al.
(2008), found that despite being present along the
entire country An. maculipennis s.l. presented lower
abundances in the districts of Braga, Leiria, Lisbon,
Oporto and Viana do Castelo, which correspond
closely with the unsuitable areas obtained, and high-
er abundances in the south and north-east of the
country, also coinciding with the least uncertain
suitable areas. Higher abundances of
An.atroparvus have previously been recorded in
Alentejo and Algarve (the two southernmost regions
of the country) by Ramos et al. (1978) and Pires et
al. (1982). However, this spatial pattern is some-
what divergent with the known distribution of the
species outside the Iberian Peninsula. In fact,
An.atroparvus is found mainly in northern coun-
tries with colder and wetter climate conditions such
Threshold L1 L2 L3 L4 D2BLR MLP-ANN
Kappa 0.32 0.43 0.45 0.62 0.36 0.38 0.42
Table 1. Kappa values of former malaria distribution and present habitat suitability models for An. artoparvus in Portugal.
L1 = unsuitable in all models vs remaining area; L2 = unsuitable for 2 models vs remaining area; L3 = suitable in all models vs remaining
area; L4 = unsuitable in all models vs suitable in all models; D2= Mahalanobis distance using best kappa threshold; BLR = binary logistic
regression using mean best kappa threshold; MLP-ANN = multilayer perceptron artificial neural network using mean best kappa threshold.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187 185
as the Netherlands, United Kingdom, Germany or
Poland, being absent from some of the warmer
Mediterranean regions, such as south Italy, Greece
and Turkey. Considering the higher suitability of
this species in the southern areas of Portugal, its
absence from many of the other Mediterranean
areas can be due to competitive exclusion involving
other species. In this sense, it is possible to infer that
this species has reached, not only in Portugal but
also in places along the Iberian Peninsula, some of
the most highly suitable conditions for its existence.
This would justify its former dominant role in
malaria transmission in these areas.
The combined suitability model was also used for
comparison with the known distribution of malaria
during its last endemic period. The spatial distribu-
tion of the disease was documented by Cambournac
(1942) who presented a map based on known val-
ues of spleen rates (i.e. the proportion of people of
a given population with enlarged spleens) by munic-
ipality, totalling six distinct regions (Fig. 3).
Although this author considers the existence of dif-
ferent spleen rates between the areas, he also refers
that they were all related with the existence of high
abundance levels for An.atroparvus.
With the exception of the L1 model, all the results
from the multi-model combination show higher pre-
dictive performance than single-model binary distri-
butions. However, the agreement value is far higher
when the disagreement areas between models are
removed. The higher kappa value, achieved by the
L4 binary model, points to two different assump-
tions, i.e. the combination of different suitability
models can be a good possibility for achieving bet-
ter predictive results and the An.atroparvus distri-
bution pattern in mainland Portugal is still substan-
tially similar to the one existing when the last
endemic malaria episodes occurred. This suggests
that despite the recent temperature increase in
Portugal, the distribution of this species remains
unaltered from the one registered nearly seven
decades ago. This invariance is important when one
considers that the country has registered as much as
a mean temperature increase of 0.5ºC/decade since
1975 (Miranda et al., 2006). Since An.atroparvus
reaches some of the highest temperatures in its dis-
tribution range in the southernmost part of the
Iberian Peninsula and survives there, it can be
expected that northern populations will remain
resilient to possible temperature increments in the
next decades. It is also clear that vector habitat suit-
ability models for vector species can act as one of
the components of information in health risk analy-
sis studies.
The option to apply more than one predictive
model and their subsequent combination for achiev-
ing the interpreted habitat suitability model allowed
a more precise spatial discrimination of its distribu-
tion. This achieved distribution presented higher
agreement with the former malaria areas than indi-
vidual models, which suggests higher appropriate-
ness of this technique to achieve species habitat suit-
ability models.
Conclusions
The current An.atroparvus distribution is fairly
similar to former malaria distribution in Portugal,
suggesting that habitat suitability models of vectors
can be a good surrogate in the spatial assessment of
Fig. 3. Former malaria distribution in continental Portugal
(Cambournac, 1942).
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187
186
malaria risk. Opposite to what usually happens in
other malaria-endemic areas, the incidence of the
disease in the country was dependent on vector den-
sity. Thus, to spatially model the current abundance
of the vector is, indirectly, to model malaria trans-
mission rates and therefore also the receptivity (and
risk) for malaria re-emergence. This similarity sug-
gests that An. atroparvus still presents the same
preferential distribution as nearly seven decades ago
and that the species is still not affected by the
increasing temperatures recorded in the country for
the last decades. This resilience of the species in one
of the warmer and dryer area of its distribution indi-
cates that northern populations may indeed not
present, at least not in the short-term, large distrib-
utional variations due to a direct influence of poten-
tial temperature increases.
Acknowledgements
This publication was partially funded by: (i) EU grant
GOCE-2003-010284 EDEN and is catalogued by the
EDEN Steering Committee as EDEN0144 (www.eden-
fp6project.net), being the contents of this publication of the
sole responsibility of the authors and do not necessarily
reflect the views of the European Commission, and;
(ii) “Arbovirus dos mosquitos de Portugal”
(POCTI/35775/ESP/2000).
References
Abdi H, 2003. Neural networks. In: Encyclopedia for
research methods for the social sciences. Sage, Thousand
Oaks, pp. 725-728.
Almeida A, Galão R, Sousa C, Novo M, Parreira R, Pinto J,
Piedade J, Esteves A, 2008. Potential mosquito vectors of
arboviruses in Portugal: species, distribution, abundance
and West Nile infection. Trans R Soc Trop Med Hyg 102,
823-832.
Antunes F, Forte M, Tavares L, Botas J, Carvalho C,
Carmona H, Araújo F, 1987. Malaria in Portugal 1977-
1986. Trans R Soc Trop Med Hyg 81, 561-562.
Araújo M, Guisan A, 2006. Five (or so) challenges for species
distribution modelling. J Biogeogr 33, 1677-1688.
Araújo M, New M, 2006. Ensemble forecasting of species
distributions. Trends Ecol Evol 22, 42-47.
Austin M, Meyers J, 1996. Current approaches to modeling
the environmental niche of Eucalyptus: implications for
management of forest biodiversity. Forest Ecol Manag 85,
95-106.
Becker N, Petric D, Zgomba M, Boase C, Dahl C, Lane J,
Kaiser A, 2003. Mosquitoes and their control. Kluwer
Academic/Plenum Publishers, 518 pp.
Borges F, 2001. A malária no vale do rio Sado. In: Memórias
do Instituto de Malariologia de Águas de Moura. Da luta
anti-palúdica ao museu, Câmara Municipal de
Palmela/Centro de estudos de vectores e doenças
infecciosas, Palmela, pp. 59-64.
Bruce-Chwatt LJ, Zulueta J, 1977. Malaria eradication in
Portugal. Trans R Soc Trop Med Hyg 71, 232-240.
Cambournac F, 1942. Sobre a epidemiologia do sezonismo
em Portugal, Lisboa: Sociedade Industrial de Tipografia,
235 pp.
Cambournac F, Hill R, 1938. The biology of Anopheles
atroparvus in Portugal. Transactions of the Third
International Congress of Tropical Medicine and Malaria 2,
178-184.
Castro L, Cardoso A, Queirós L, Gonçalves G, 2004. Malária
na Região Norte de Portugal (1993-2002). Caracterização
epidemiológica. Acta Med Port 17, 291-298.
Cohen J, 1960. A coefficient of agreement for nominal Scales.
Educ Psychol Meas 20, 37-46.
Elith J, Graham C, Anderson R, Dudik M, Ferrier S, Guisan
A, Hijmans R, Huettmann F, Leathwick J, Lehmann A, Li J,
Lohmann L, Loiselle B, Manion G, Moritz C, Nakamura
M, Nakazawa Y, Overton J, Peterson T, Phillips S,
Richardson K, Scachetti-Pereira R, Schapire R, Soberon J,
Williams S, Wisz M, Zimmermann N, 2006. Novel meth-
ods improve prediction of species distributions from occur-
rence data. Ecography 29, 129-151.
Guisan A, Thuiller W, 2005. Predicting species distribution:
offering more than simple habitat models. Ecol Lett 10,
993-1009.
Guisan A, Zimmermann N, 2000. Predictive habitat
distribution models in ecology. Ecol Model 135, 147-186.
Hijmans R, Cameron S, Parra J, Jones P, Jarvi A, 2005. Very
high resolution interpolated climate surfaces for global land
areas. Int J Climatol 25, 1965-1978.
C. Capinha et al. - Geospatial Health 3(2), 2009, pp. 177-187 187
Hirzel A, Lay G, Helfer V, Randin C, Guisan A, 2006.
Evaluating the ability of habitat suitability models to pre-
dict species presences. Ecol Model 199, 142-152.
Hosmer DW, Lemeshow S, 1989. Applied logistic regression,
John Wiley & Sons Inc, New York, USA 397 pp.
IPCC, 2007. Climate change 2007: the physical science basis.
Summary for policy makers working group I to the
Intergovernmental Panel on Climate Change. Fourth
Assessment Report (AR4).
Jenness J, 2003. Mahalanobis distances (mahalanobis.avx)
extension for ArcView 3.x. Jenness Enterprises. Available:
http://www.jennessent.com/arcview/mahalanobis.htm.
Kuhn K, Campbell-Lendrum D, Davies C, 2002. A continen-
tal risk map for malaria mosquito (Diptera: Culicidae) vec-
tors in Europe. J Med Entomol 39, 621-630.
Landis J, Koch G, 1977. The measurement of observer
agreement for categorical data. Biometrics 33, 159-174.
Liu Y, Starzyk JA, Zhu Z, 2007. Optimizing number of hid-
den neurons in neural networks. In: Proceedings of the 25th
conference on Proceedings of the 25th IASTED
International Multi-Conference: artificial intelligence and
applications. IASTED/ACTA Press, Innsbruck, Austria, pp.
138-143.
Minárˇ J, Halgosˇ J, Bartalová A, Jalili N, 2007. Current
climatic change and its impact on mosquito fauna in
conditions of Slovakia and Czech Republic. Acta Zool Univ
Comen 47, 177-182.
Miranda P, Valente A, Tomé AR, Trigo R, Coelho F, Aguiar
A, Azevedo F, 2006. O clima em Portugal nos séculos XX e
XXI. In: Alterações climáticas em Portugal. Cenários,
Impactos e Medidas de Adaptação, Projecto SIAM II,
Gradiva, Lisboa, Portugal, pp. 45-113.
Moffett A, Shackelford N, Sarkar S, 2007. Malaria in Africa:
vector species’ niche models and relative risk maps. PLOS
One 2, e824.
Muturi E, Burgess P, Novak R, 2008. Malaria vector man-
agement: where have we come from and where are we
headed? Am J Trop Med Hyg 78, 536-537.
Paaijmans K, Wandago M, Githeko A, Takken W, 2007.
Unexpected high losses of Anopheles gambiae larvae due to
rainfall. Plos One 2, e1146.
Pascual M, Ahumada J, Chaves L, Rodo X, Bouma M, 2006.
Malaria resurgence in the East African highlands: tempera-
ture trends revisited. P Natl Acad Sci USA 103, 5829-5834.
Pearson R, Thuiller W, Araújo M, Martinez-Meyer E,
Brotons L, Mcclean C, Miles L, Segurado P, Dawson T, Lees
D, 2006. Model-based uncertainty in species range predic-
tion. J Biogeogr 33, 1704-1711
Pires C, Ribeiro H, Capela R, Ramos H, 1982. Research on
the mosquitoes of Portugal (Diptera, Culicidae) VI - The
mosquitoes of Alentejo. An Inst Hig Med Trop 8, 79-101.
Ramos H, Ribeiro H, Pires C, Capela R, 1978. Research on
the mosquitoes of Portugal (Diptera, Culicidae) II - The
mosquitoes of Algarve. An Inst Hig Med Trop 1, 79-101.
Ribeiro H, Alves-Pires C, Ramos H, Capela R, 2002.
Research on the mosquitoes of Portugal (Diptera,
Culicidae) XII - The mosquitoes of Minho and Douro
Litoral. Garcia de Horta - Série Zoológica 24, 31-50.
Ribeiro H, Pires C, Capela R, 1988. An annotated checklist
of the mosquitoes of continental Portugal (Diptera,
Culicidae). Actas do III Congresso Ibérico de Entomologia,
233-253.
Ribeiro H, Ramos H, Capela R, Pires C, 1992. Research on
the mosquitoes of Portugal (Diptera, Culicidae) II - The
mosquitoes of the Beiras. Garcia de Horta - Série Zoológica
16, 137-161.
Ribeiro H, Ramos H, Pires C, 1999a. Os mosquitos do
Parque Natural da Serra da Estrela (Insecta, Diptera,
Culicidae). Garcia de Horta - Série Zoológica 23, 119-168.
Ribeiro H, Ramos H, Pires C, 1999b. Os mosquitos do
Parque Natural de Montesinho (Insecta, Diptera,
Culicidae). Garcia de Horta - Série Zoológica 23, 23-67.
Sousa CA, 2008. Malaria vectorial capacity and competence
of Anopheles atroparvus Van Thiel, 1927 (Diptera:
Culicidae): implications for the potential re-emergence of
malaria in Portugal. PhD Thesis, New University of Lisbon,
Portugal, 187 pp.
Tsoar A, Allouche O, Steinitz O, Rotem D, Kadmon R, 2007.
A comparative evaluation of presence-only methods for
modelling species distribution. Divers Distrib 13, 397-405.
WHO 2008. World Malaria Report 2008. World Health
Organization, Geneva, Switzerland, 215 pp.
