Background
Malaria is a major health problem in sub-Saharan Africa, where it is estimated to be responsible for over 1 million deaths every year in children younger than five and pregnant women [
1]. Out of the total human population in Africa, 15% live in highlands, where there are increasing risks for epidemics [
1]. Current strategies for malaria control involve treating infected individuals with anti-malarial drugs to clear the parasites, and reducing human-mosquito contact rates through vector control efforts. Anti-malarial drugs have little impact on the intensity of transmission at the community level because most drugs do not reduce the production of
Plasmodium gametocytes, the parasite stage responsible for initiation of infection in mosquitoes [
2]. Individuals who receive treatment can quickly become reinfected. Recent large field trials in Kenya demonstrated that insecticide-treated bed nets (ITN) can prevent 1 in 4 infant deaths in areas of intense perennial malaria transmission, if bed nets are used properly and re-treated with insecticide at appropriate intervals [
3,
4]. However, coverage and compliance are limited and emergence of insecticide-resistance genes has hindered the effectiveness of ITN programmes [
5‐
8].
In recent years, there has been renewed interest in exploring the possibility of anopheline larval control through environmental management or larvicides as additional means of reducing malaria transmission in Africa [
9‐
12]. Historically, eradication of the accidentally introduced African malaria mosquito
Anopheles gambiae from north-east Brazil in the 1930s and early 1940s succeeded, through an integrated programme that relied overwhelmingly upon larval control [
13]. A larval control programme successfully suppressed malaria for over 20 years around a Zambian copper mine [
14] and in Dar es Salaam in Tanzania [
15,
16]. Source reduction through the modification of larval habitats was an important tool for malaria eradication efforts in the United States, Israel, and Italy [
17]. However, the primary malaria vectors in sub-Saharan Africa,
An. gambiae and
Anopheles arabiensis generally utilize small temporary habitats as breeding sites [
18‐
22], which creates difficulties for environmental management. Unfortunately, identifying these mosquito larval habitats over a large geographic area based only on field survey is time-consuming and labour intensive. Therefore, better methods for rapid and accurate determination of larval habitat distribution are critical to enable larval control using bio-insecticides or environmental modification.
Remote sensing is a powerful tool for determining the landscape features and climatic factors associated with the risk of vector-borne diseases [
23‐
28]. For example, ecological parametres, particularly vegetation index, were found to be significantly associated with Rift Valley fever viral activity in Kenya through the National Oceanic and Atmospheric Administration's polar-orbiting meteorological satellites [
29]. Climatic factors associated with malaria risks in sub-Saharan Africa were identified using climate data obtained from satellites and malaria transmission distribution maps. The malaria transmission maps were developed according to biological constraints of climate on parasite and vector development [
28,
30].
Remote sensing also can be used for determining factors affecting vector abundance (e.g., [
31‐
34]) and mosquito breeding sites [
35]. For instance, Beck et al. [
31] analysed Landsat Thematic Mapper (TM) images of southern Chiapas, Mexico and found that transitional swamp and unmanaged pasture were the most important landscape elements for explaining vector abundance. Welch et al. [
33] showed that infrared aerial photos were useful in the detection of potential oviposition sites of
Psorophora columbiae, such as ditches, low-lying areas and tyre tracks in Texas. Roberts et al. [
36] used aerial photos to determine that breeding sites located at low elevations in flooded, unmanaged pastures were the most important determinants of
Anopheles albimanus adult abundance in southern Mexican villages. In general, previous studies demonstrate the utility of remote sensing technology in the risk assessment of vector-borne diseases and vector-population monitoring at a large spatial scale. Recently developed remote sensors of high spatial resolution may be particularly useful for determining mosquito larval habitat distribution and for assisting malaria vector control. For example, the spatial resolution is 1–4 metres for Ikonos [
37], 0.61–2.44 metres for QuickBird [
38], 1–4 metres for OrbView-3 [
39] and 2.5–10 metres for SPOT 5 [
40]. These resolutions compare favourably with 10–20 and 15–60 metres for Spot XS [
40] and Landsat TM 7 [
41], respectively. However, the utility of high-resolution satellite images for larval habitat identification and management have not been evaluated in African highlands where land-use pattern is highly heterogeneous and larval habitat distribution is spatially clustered [
22].
The objective of this study was to assess the potential of aerial photos, Landsat TM 7 and Ikonos images for identifying larval habitats of malaria vectors and for determining other topographic features associated with anopheline larval habitats in western Kenya highlands where frequent malaria outbreaks have been reported [
42‐
46].
Discussion
The aim of using remote sensing is to identify geographic features associated with mosquito breeding habitats, and ultimately to predict the spatial distribution of aquatic habitats, especially anopheline-positive habitats, for vector control programmes. Remote sensing images with different spatial and spectral resolutions are available. Among the satellite sensors used in this study, Landsat TM 7 has a spatial resolution of 30 × 30 metres and seven spectral bands (15 metres when using the panchromatic band), while Ikonos has a 1 × 1 metre sensor and four spectral bands. It was demonstrated that supervised classification based on Ikonos images had improved accuracy in determining land-use and land-cover types – a geographic feature significantly associated with the occurrence of anopheline larval habitats [
47,
57] – than aerial photos and Landsat TM 7 images. Eighty-nine percent of the area in this study site was correctly classified using Ikonos, while only 77% and 61% of the area was correctly identified by aerial photos and a Landsat TM 7 image, respectively. When aquatic habitats were visually identified using these three types of images, the Ikonos image had a high spatial and spectral resolution. About 41% of the aquatic habitats were identifiable based on the Ikonos image while only 11.6% were correctly identified using aerial photos and none were correctly identified using the Landsat TM 7 image. Landsat TM image under the 15 m panchromatic band could not be used for identification of aquatic habitats because most of habitats (81.8%) in this study site were less than 100 m
2 (Table
3), much less than the one-pixel size of Landsat image (225 m
2). In principle, the detectable object has to be 1.5 the size of one pixel. Although more stable aquatic habitats are more productive to
An. gambiae mosquitoes,
An. gambiae larval habitats are generally smaller than the habitats identifiable by Landsat TM image (58, 59). These results suggest that the Ikonos sensor is superior to Landsat TM 7 image and aerial photos for determining potential anopheline larval habitats.
Several studies have used remote sensing to identify potential mosquito breeding habitats. Anyamba et al. [60] found that normalized vegetation difference index anomalies were associated with the availability of
Aedes mosquito breeding habitats and Rift Valley Fever risks in Kenya. Sithiprasasna et al. [61] delineated stream networks from Ikonos satellite images and found that the risk of malaria infections was negatively correlated with distance-to-streams in Thailand. In Egypt, Hassan and Onsi [62] combined limited ground surveys with remote sensing techniques to identify mosquito breeding habitats in the Natroun Lakes area. In several past studies, vegetation cover, landscape structure, and distribution of water bodies were found to be associated with malaria risks [
32]. In this study, it was shown that most aquatic habitats were close to streams and rivers (< 100 m). For example, 78.1% out of 314 and 68.3% out of 779 of anopheline-positive habitats were located within 50 metres of streams in the dry and rainy seasons respectively. The Ikonos image was able to identify 86.4% of the rivers and streams. Although Ikonos could only visually identify 40.7% of the aquatic habitats, the image and topographic derivatives could be used to improve the identification rates of aquatic habitats significantly. The computer model derived from topographic maps and remotely sensed parametres showed improved accuracy in determining the spatial distribution of aquatic habitats. The spatial distribution of more than 75% of aquatic habitats was predicted correctly. Whether these results are site-specific and whether they could be extrapolated to wide areas in the highlands is of interest.
Each remote sensor type has advantages in availability and utility for mosquito vector habitat determination. Aerial photos can be obtained for any area of interest, and they are generally not limited by cloud coverage because aerial photos can be taken when the sky is clear. Flying closer to the ground or using appropriate lens can increase the ground resolution of aerial photos. However, each photo covers a small area, image assembly based geo-referenced ground points is required to produce a mosaic of images for a large area, and thus may introduce geo-reference errors. In addition, aerial photos may suffer edge distortions caused by inappropriate camera position. Landsat TM 7 images are easily available at low cost in digital formats and a single scene covers a large area (swath width = 185 km), but the resolution is coarse. Ikonos images have the best spatial resolution for identification of major breeding habitats and human settlements, but they are more costly and often the availability of images is limited in areas where cloud coverage is significant and frequent. Each remote sensor type also had its limitations. The utility of panchromatic aerial photos is limited by the lack of infrared band capability, which is particularly useful for the identification of water bodies, especially when they are partially or completely covered by vegetation. While the Landsat TM 7 images have seven spectral bands, their 30 × 30 m ground resolution undermines their utility for identifying the patchy and discontinuous land-use characteristic of the western Kenya highlands. Although the Ikonos images have the advantage of high spatial and spectral resolution, this advantage is eroded when identifying land-cover types with similar spectral reflectance (e.g., fallow land vs. pasture, young forest vs. shrubs). Recognizing the advantages and disadvantages of these remote sensor types can help the selection of appropriate remote sensing images so that assist mosquito vector control efforts.
Successful prediction of the spatial distribution of anopheline mosquito habitats would allow vector control efforts to target the most productive larval habitats, resulting in a reduction of operational costs [
25]. The statistical model used showed an 18–25% misclassification rate for anopheline-positive habitats. This classification error for anopheline-positive habitats could be due to inadequate understanding of factors regulating habitat productivity. Further knowledge of the underlying mechanisms of habitat productivity will help predict the spatial distribution of anopheline larvae in Africa.
Acknowledgements
This paper is published with the permission of the Director, Kenya Medical Research Institute and the Director, International Center for Insect Physiology and Ecology. We thank M. Okonji, B. Omboko, M. Abuom, and L. Atieli for technical assistance, and anonymous reviewers for critical comments. This research was supported by NIH grants R01 (AI) 50243 and D43 TW01505.