body top image
Research

Satellite Imagery in the Study and Forecast of Malaria

(continued)


3.2 Distribution and Abundance of Mosquito Vectors

The distribution of efficiently transmitted pathogens such as those that cause malaria is generally limited by the distribution of competent vectors, which can now be predicted from satellite imagery. In Africa, the main vectors of malaria include species of the Anopheles gambiae complex (A. arabiensis, A. bwambae, A. gambiae s.s., A. melas, A. merus and A. quadriannulatus) and A. funestus, whose distributions show similarities with patterns of annual rainfall across Africa.[24] Furthermore, a relationship found in West Africa between climate (annual precipitation and annual and wet-season temperatures) and the ratio of A. gambiae s.s. to A. arabiensis was used successfully to predict the distribution and relative abundance of these two species in Tanzania (East Africa).[25] The subtlety and extensive coverage of multivariate climatic factors detectable by satellite sensors have allowed predictions of the distribution of five of the six species in this complex (Fig. 2; A. bwambae, restricted to a small site in Uganda, is not modelled). This has been achieved using maximum likelihood methods based on the relatively few studies that identified these species separately. Thus, satellites can distinguish between the habitats of species that were, until 1956, regarded as a single species.

Satellite data show strong relationships with the density of other vectors in Africa,[2][26] but mosquito abundance has only rarely been recorded, despite its importance in R0 calculations. Until such data are collected, as a matter of urgency, we speculate that abundance will be greatest in sites with conditions near to the climatic centroids defined by the satellite data, although spatially variable, densitydependent factors may confound this prediction.


Figure 2: Distributions of five mosquito species in the Anopheles gambiae complex in Africa, predicted from temporal Fourier-processed satellite data (Box 1 on page 4) and elevation (global coverage provided by the digital elevation model GTOPO30; http://eros.usgs.gov/products/elevation/gtopo30.php) at a spatial resolution of 0.05°. The colour-coded probabilities42 of presence effectively indicate the environmental suitability for each species throughout the continent. Symbols indicate sample sites for each species. Between 18° N to 30° S, each species was classed as present only within 0.15° of the sites from which it has been recorded[24] and absent only from similarly sized sites where any of the other species have been recorded. Within these sites, presence pixels (300 for A. gambiae s.s. and A. arabiensis and 100 for the other species) and absence pixels (400 for all species) were chosen at random; additional randomly selected absence pixels were chosen north of 18° N (n=200) and south of 30° S (n=50). Satellite data: middle infrared, land surface temperature and normalized difference vegetation index for 1982-1998 were derived from the National Oceanographic and Atmospheric Administration's Advanced Very High Resolution Radiometer, and cold cloud duration for 1988-1999 from Meteosat High Resolution Radiometer. Satellite and elevation data for all sample pixels were subjected to k-means clustering within the Statistics Package for the Social Sciences (SPSS, Chicago) to identify up to six natural clusters each of presence and absence pixels for each mosquito species. Within maximum likelihood discriminant analysis, stepwise selection of up to ten variables was applied to maximize predictive accuracy according to the kappa statistic,[28] sensitivity and specificity (Box 1 on page 4) and to calculate the posterior probabilities[42] with which each pixel belongs to the presence or absence classes within the training set. Sites too different from any of the training set sites are assigned to a 'no prediction' class.

issue picture

a, A. arabiensis: 86.0% correct predictions, 7.7% false positives, 6.2% false negatives, sensitivity=40.80, specificity=40.89, κ=40.679 (±50.051 95% confidence interval).

b, A. gambiae s.s.: 92.4% correct, 4.2% false positives, 3.4% false negatives, sensitivity=40.89, specificity=40.94, κ=40.826 (±50.039).

c, A. quadriannulatus: 99.1% correct, 0.9% false positives, 0% false negatives, sensitivity=41, specificity=40.99, κ=40.961 (±50.029).

d, A. melas (West Africa): 98.2% correct, 1.6% false positives, 0.1% false negatives, sensitivity=40.99, specificity=40.98, κ=40.928 (±50.039); and A. merus (East Africa): 98.4% correct, 1.6% false positives, 0% false negatives, sensitivity=41, specificity=40.98, κ=40.933 (±50.037).


3.3 Entomological Inoculation Rate

A more complete measure of malaria challenge is an estimate of the entomological inoculation rate (EIR), which is the product of the number of mosquito bites per human per unit time (ma in the R0 equation) and the proportion of mosquitoes with sporozoites (infective stages of the malaria parasites).[3] EIRs are generally derived from short- to medium-term studies, each in a relatively small area, and are therefore of limited predictive value on their own.[27] By compiling all reliable published EIRs and relating them to satellite data, we can make more extensive predictions of this measure of challenge (Fig. 3). Analysis shows that gradual changes in environmental conditions distinguish low- and high-risk areas, and do so consistently enough for an excellent overall agreement with the field data (kappa statistic (ref. 28; and see Box 1 on page 4) of 0.77). There is also good agreement between the EIR predictions (Fig. 3) and the predicted distributions of the two key malaria vectors, A. arabiensis and A. gambiae s.s. (Fig. 2), although in parts of southern Africa, where historically vector control and case-management have been effective, present-day malaria challenge is less than is shown on this map. But many areas of Africa are too different from any of the training sets to allow any predictions for them (shown in grey in Fig. 3). By identifying these regions, satellite imagery can direct new studies of this important malariometric index.


Figure 3: Satellite-derived predictions of entomological inoculation rate (EIR) in Africa. EIR data27 (map inset) were grouped into five approximately equal-sized classes of mean levels of malaria challenge. The same satellite data layers and analytical methods as used in Fig. 2 are used to define the probability with which each continental pixel belongs to one of the five challenge categories. Insufficient training data were available to define EIR in those parts of the continent marked grey.

issue picture

Descriptive accuracies: EIR range 0-4.4, 85.7% (n=21); 5.8-26.6, 81.8% (n=22); 31.0-87.0, 77.3% (n=22); 89.8-255.5, 68.2% (n=22), 259.9-703.4, 95.5% (n=22). Overall κ=40.771 (±50.064).


3.4 Resulting Malaria Incidence and Prevalence

Biologically, and for the purposes of intervention, the point of interest is to explain the relationship between the above three measures of challenge (mosquito distribution and abundance, and EIR) and the resulting disease burdens, and so be able to predict the latter. We are hindered from making even statistical predictions by the lack of good quality training sets for satellite studies, as disease risks have been determined too infrequently, and over insufficiently wide areas. With the aid of satellites, however, a small training set can produce predictions across a large area.

Numbers of monthly childhood malaria admissions in three hospitals in Kenya, expressed as a percentage of the annual totals, were found to be correlated most consistently (mean adjusted r2=40.71) with the previous month's normalized difference vegetation index (NDVI), (Box 1 on page 4)[29], which is related to plant photosynthetic activity. As minimum NDVIs of 0.35-0.40 were shown to be required before >5% of annual admissions were recorded in any one month, the duration of malaria transmission seasons across Kenya and Uganda could be predicted by counting the number of months in which these NDVI values were exceeded.[18][30] The resulting predictive maps of malaria seasons showed strong similarities to an historical map of malaria transmission periods in Kenya.[31]

page 1 | page 2 | page 3 | page 4 | top  

filler image
body bottom image