Dengue Vector (Aedes aegypti) Larval Habitats in an Urban Environment of Costa Rica Analysed with ASTER and Quickbird Imagery

Abstract

Dengue fever is transmitted mainly by the mosquito species Aedes aegypti, which breeds in water-filled containers in tropical urban areas. In this study medium- and high-resolution satellite (ASTER and QuickBird) imagery were used to map tree cover and built-up surfaces in Puntarenas, Costa Rica, where dengue fever has been endemic since 1993. Hard and soft image classification algorithms were used to generate these variables, which were analysed with respect to vector habitat distribution as measured by the potential mosquito habitats during the wet season of 2006. A multiple linear regression model revealed that built-up surfaces derived from ASTER imagery explained 63.6% of the variance in the number of larval habitats found during the wet season when transmission and incidence of dengue were highest. The results suggest that elements of urban structure such as built-up surfaces may be used to predict the presence of an important disease vector within urban areas and that medium-resolution satellite imagery may help to inform future prevention and control strategies in Costa Rica and other countries where dengue fever is endemic.

1. Introduction and Background

Among vector-borne diseases, dengue fever is the most important arboviral disease globally with an estimated 50 to 100 million cases and 12,000 to 24,000 deaths per year (WHO 2002, Gibbons and Vaughn 2002). The principal dengue mosquito vector, Aedes aegypti, lives in close association with humans mainly in urban and suburban environments (Service 1992, Gibbons and Vaughn 2002). Female Ae. aegypti mosquitoes preferentially ingest human blood and commonly lay their eggs in water-filled artificial containers such as drums, buckets, and tires where water can collect and remain for several days (Service 1992, Gubler 1998, Focks and Chadee 1997, Calderón-Arguedas et al. 2004). Although Ae. aegypti was eliminated from Costa Rica in 1960, the vector became re-established during the 1980s and dengue virus transmission was reported in 1993 (WHO 1994). Within Costa Rica, dengue is currently the most important vector-borne disease with over 140,000 cases reported since its reintroduction, including almost 38,000 cases in 2005 (Ministerio de Salud et al. 2004, Ministerio de Salud 2006).

Remotely sensed data can provide spatial information to study the epidemiology of many vector-borne diseases (Hay et al., 1997, Beck et al. 2000, Thomson and Connor 2000, Bergquist 2001, Correia et al., 2004). Commercial imaging satellites such as IKONOS and QuickBird offer new opportunities to assess urban habitats for disease vectors by providing very high spatial resolutions (~0.05–4 m) appropriate for identification of city blocks, individual roads, tree crowns, roadways, buildings, and rooftops. Using classified QuickBird imagery, Troyo (2007) developed a series of multiple regression models that included tree cover, built area, and a variable that accounted for the number of locations sampled within an urban grid cell. These models explained between 61.5–65.0 percent of variance in containers adequate for development of Ae. aegypti mosquito larvae. Here we extend this approach to Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery to examine whether such medium-resolution data can be substituted for more costly and spatially limited commercial imagery. Small (2003) showed that the characteristic scale of urban reflectance is consistently between 10 and 20 m for a range of different cities, which suggests that soft classification of 15 m ASTER imagery may provide opportunities analyzing urban structural variables that indicate the presence of vector habitats. Specifically, we derived two urban structural variables from satellite imagery and examined how these related to the presence of larval habitats within the Puntarenas area. The same variables were obtained from 15 m ASTER and 2.4 m QuickBird imagery acquired during the same time period (2002–2003). Thus, the purpose of our study was: 1) to evaluate the suitability of medium-resolution (15 m) ASTER imagery for deriving urban structural variables that may relate to Ae. aegypti habitat distribution; 2) to examine empirical relationships between these variables derived from soft classification of ASTER imagery and larval habitats in an endemic area.

2. Study Area and Methods

Our study focused on the Greater Puntarenas area, Costa Rica, which is located on the Pacific Coast of the country (figure 1). The most densely built-up part of Puntarenas is on a narrow, low-lying, east-west trending peninsula and adjacent mainland areas that have served as a major tourist destination over the past several decades. The study area encompasses ~20 km² and census data indicate the presence of approximately 100,000 people and 20,000 houses, most of which (> 95%) are classified as urban, although there is a marked heterogeneity in sizes and quality (INEC 2002). Puntarenas is the site of dengue reintroduction to Costa Rica in 1993 (WHO 1994), and it has been greatly affected by the disease ever since. Between 2001 and 2005, more than 7,000 cases of dengue were reported by the Ministry of Health (Ministerio de Salud 2006).

Figure 1.

The study area shown at three different scales: A. Costa Rica with the approximate location of Puntarenas in the rectangle; B. Map of the localities from which dengue incidence and housing data were obtained; C. ASTER satellite image (NIR band) from 2002 showing the area analyzed in medium-resolution satellite imagery.

Data were collected and analyzed at two levels of spatial aggregation: administrative districts or ‘localities’, and individual cells, where field sampling was conducted. A larval field survey was performed during the wet season in last week of July, which involved a sampling method detailed in Troyo et al. (2008a) to select the locations for larval habitat analysis. This entailed the establishment of variably sized grids that resulted in a sample cell size of 1 ha, an area sufficient to contain 13±6 houses, which is a sample large enough for field team to survey within a half day (Troyo et al. 2008a). A stratified random sample of 36 cells (10% from each locality) was selected where all the locations included would be searched for mosquito larval habitats. A ‘location’ was any legally limited section of land that may have included a house or building such as parks, streets and sidewalks, households, lots, construction sites, etc. All locations surveyed in a sample cell were searched and a total of 829 larval habitats were sampled for mosquito larvae during the wet season (Troyo et al. 2008b). These habitats included cans, small plastic food containers, buckets, tires, drums, wash tubs, roof gutters, and any other objects that might contain mosquito larvae. The presence or absence of mosquito immature stages was noted for each habitat and when present, samples were collected and processed for species identification.

Satellite imagery and data

Four ASTER scenes (15 m) were acquired during the dry and wet seasons respectively of 2002 and 2003, and two QuickBird scenes available for March 2002 and October 2003 were mosaicked to produce single high-resolution images, which covered the districts of Puntarenas and most of Chacarita (10 localities in total). The QuickBird scenes were individually georeferenced to increase their geospatial accuracy (RMS = 2.9 m and 3.1 m for the 2002 and 2003 scenes, respectively) by using 38 ground control points obtained with a hand held global positioning system.

Urban land cover maps were produced from the QuickBird scenes by applying supervised image classifiers to each QuickBird image. Classification algorithms included the maximum likelihood (MLC) and back-propagation artificial neural network (BPNN). These classifiers assigned each pixel to specific, predetermined land cover classes including tree, asphalt, building, bare soil, etc. and the resultant classified images were mosaicked. Classification accuracy was assessed by using points selected at random from the original panchromatic QuickBird scene and mean percent cover for each class was calculated for the QuickBird and 15 m ASTER pixels covering the cells sampled for vector habitats as well as localities in which locations were examined for containers. The panchromatic image provided a set of mutually exclusive training and validation points for the automated image classifiers. The proportion of built area and tree cover was extracted for the individual localities of Puntarenas as these had acceptable class accuracies (reported below) and were assumed to relate to mosquito habitat distribution based on descriptions in the epidemiological literature (Gubler 1998, Focks and Chadee 1997, Bisset-Lazcano et al. 2006, Vezzani et al. 2005). Geographic information system (GIS) operations were applied to the final classified products to extract data at the locality and sample cell levels.

Owing to limited spatial coverage and high cost of the QuickBird imagery, we applied a soft classification approach to the ASTER scenes to estimate fractional tree and built-up cover for 36 sample cells in Puntarenas. We used a multi-layer perceptron (MLP) back-propagation algorithm to decompose pixel end-members into proportional tree cover, built-up surfaces (paved and buildings), and open/green spaces. Although neural networks function largely as a ‘black box’, research suggests they may produce superior results relative to linear unmixing and regression trees (Liu and Wu 2005). The back-propagation MLP parameters such as learning rate, layer and node numbers, and iterations were adjusted experimentally to produce a set of soft classification outputs for visual and quantitative assessment. To incorporate seasonality in the classification, the first three MS ASTER bands (15 m) from three relatively cloud-free ASTER scenes (< 10 percent cloud) from both wet and dry seasons were included as inputs to the MLP and training data were derived from the panchromatic bands of the QuickBird imagery described above. Changes in urban structure from 2002–2006 were assumed to be minimal; therefore, the imagery was considered suitable for analysis of larval habitat data collected in 2006. Both classified and panchromatic imagery from QuickBird were used along with housing data aggregated to the locality level to assess the soft classification outputs from the ASTER imagery and percent built surfaces and tree cover was averaged for each of the 1 ha grid cells where larvae were sampled.

3. Results

The most accurate thematic map resulting from classification of the 2002/2003 QuickBird images correctly classified 80% of the urban land cover and possessed a Kappa index of 0.74. In this map the “built” class had 24% errors of omission and 20% errors of commission, while “tree” class had 7% errors of omission and 10% errors of commission. Values for each locality extracted from the ASTER imagery revealed a significant correlation (r = 0.682, p < 0.05) between proportional built-up derived from ASTER and the same variable extracted from the QuickBird (figure 2a). Further, the ASTER-based product produced a higher estimate of proportional built-up area than the QuickBird product. On the other hand, agreement between the two products was higher for proportional tree cover (r = 0.970, p < 0.01), although the QuickBird product produced a slightly higher estimate of tree cover per locality than the ASTER-based product (figure 2b). Using house density data from the Ministry of Health, we found a positive relationship (r = 0.742, p < 0.01) between one element of the built class (houses) and the proportion built-up derived from ASTER.

Figure 2.

Scatter diagrams showing the relationships between tree cover and built-up area derived from ASTER and QuickBird imagery, averaged by locality for 10 localities covered by both image sources: a. ASTERBUILT versus percent built-up from QuickBird; b. ASTERTREE versus percent tree cover obtained from QuickBird imagery.

To explain variation in positive containers (#HABITATS), we employed a multiple regression model based on cell-level data that included a correction factor to account for the spatial heterogeneity of locations evaluated (LOCE) in each cell, and the independent variables proportional tree cover (ASTERTREE) and built-up (ASTERBUILT) derived from ASTER. The LOCE corresponds to the total number of accessible locations (households, parks, streets, lots, fields, buildings, etc.) surveyed completely for mosquito larval habitats. The following equation gives the result for the wet season, when vector abundance is greatest:

$$\mathrm{\#}\mathrm{HABITATS}=-39.17*\left(\mathrm{ASTERBUILT}\right)+1.71*\left(\mathrm{LOCE}\right)+17.05$$

$$R^{\mathit{2}}=0.636,p=0.000,\text{d}.\text{f}.=31$$

The inclusion of ASTERTREE to the multiple regression models did not improve results and was not found to be significant (Table 1). Furthermore, we tested for normality of the residuals using the Kolmogorov-Smirnov d statistic (K-S d = 0.087, p > 0.2) and Lillefors p (p > 0.2), and both indicated a normal data structure. We also assessed the model residuals with the Durbin-Watson statistic (D-W = 1.71), which suggested no serial autocorrelation.

Table 1.

Detailed results of the multiple regression analysis involving the independent variable, #HABITATS, and the dependent variables ASTERTREE, ASTERBUILT and LOCE.

	Beta	Std.Err.	B	Std.Err.	t(30)	p-level
Intercept			21.550	11.117	1.938	0.062
ASTERTREE	−0.087	0.178	−16.857	34.42709	−0.489	0.627
ASTERBUILT	−0.407	0.167915	−46.124	18.98943	−2.428	0.021
LOCE	0.749	0.120	1.659	0.267	6.213	0.000

4. Discussion and Conclusions

Although the spatial resolution of imagery used in this study was inappropriate for resolving Ae. aegypti larval habitats, the results are comparable to those that Troyo (2007) achieved using QuickBird imagery. Therefore, medium-resolution data from ASTER appeared to provide a proxy variable (ASTERBUILT) for statistical modelling of vector habitats in an endemic urban area. Previous studies that have used satellite imagery have shown that dengue is related to urban or built up areas as opposed to forested areas (e.g. Nakhapakorn and Tripathy 2005). However, our results are inconsistent with this finding, perhaps because the field survey found few positive containers reported in the most densely built-up areas of Puntarenas (e.g. El Centro), where there was low tree cover, fewer houses, with small yards and more commercial and industrial buildings or relatively good sanitation. Thus, although such areas may be densely built-up, industrial and commercial areas may have low human population density and few mosquitoes. Troyo’s (2007) analysis using QuickBird imagery indicated that moderately built-up residential areas with moderate tree cover were likely to contain relatively high numbers of habitats positive for Ae. aegypti larvae, which was evidenced through multiple regression models developed for the total number of larval habitats as well as for the proportion of habitats positive for Ae. aegypti. This is also supported by studies that suggest that tree crowns are likely to provide some benefit to Ae. aegypti larvae by reducing evaporation from containers (Vezzani et al. 2005, Bisset-Lazcano et al. 2006, Barrera et al. 2006). Other studies have shown that high amounts of tree cover were inversely related to dengue incidence, which suggests that such areas lack sufficient containers for breeding sites and may be isolated from human hosts, as the Ae. aegypti mosquito typically does not travel far (less than 200 m) in search of blood meals (Muir and Kay 1998, Harrington et al. 2005, Russell et al. 2005). Further analysis of site microclimatic factors may reveal more complex relationships between urban mosquito habitats and tree cover. In addition, analyses of container profiles in Greater Puntarenas have shown that some locations are more likely to harbor larval habitats (Troyo et al. 2008b). Including different types of location in this model may reveal further effects on habitat abundance, although insufficient sample size of 36 cells in our study precluded the use of more independent variables in the analysis.

Several limitations of our study are worth noting. Even though sample cells were spatially bounded and independence was assumed, their boundaries cannot be considered barriers to mosquito or human dispersal, as they have no discernable influence on movement of individuals. Classification errors may also have introduced unexplained variation in the regression models and therefore improvements in classification accuracy are warranted, possibly through the use of image-segmentation algorithms. Nonetheless, these results should be of interest to public health officials seeking to improve dengue prevention and control in sites where tree cover, impervious surfaces and open areas are mixed within the urban environment. In addition, medium-resolution imagery, which typically covers 60×60 km or greater, may include large urban areas or multiple municipalities that are at risk of dengue. Urban land-cover maps derived from such imagery may provide a foundation for the development of dengue risk maps that support prediction and identification of priority zones for vector control activities, such as insecticide spraying, education, and source reduction campaigns, especially in areas where prompt action is required and limited or no detailed epidemiological and entomological data is available.