The study site has been the focus of intensive malariological studies and is described elsewhere 24. The area under study encompassed part of the communities of Asembo and Seme, was 3.216 × 3.216 km (10.34 km²) in area, and was located at 34°23'E, 0°11'S in Nyanza Province, western Kenya, approximately 50 km west of the city of Kisumu. Physiographically, the site falls within the Yala-Nzoia Plains, a lake lowlands region overlying granite bedrock and surrounded by Nyanza Low Plateau topography, and bordering the north shore of Winam Gulf of Lake Victoria 25. Most human inhabitants are subsistence farmers who cultivate maize, cassava, and vegetables; and husband cattle, goats, sheep, and chickens. Fishing in nearby Lake Victoria is an important local economic activity. Rainfall occurs year-round with two main peaks; the long rains falling between March and May, and the short rains between November and December 26. Total annual rainfall averages 1400 mm per year and daily temperatures range from 25.5°C to 33.0°C 24. Malaria due to infection with Plasmodium falciparum is holoendemic in the region with transmission occurring throughout the year. The principal malaria vectors in the area are two species in the Anopheles gambiae sensu latu complex (namely, A. gambiae sensu strictu and A. arabiensis), and A. funestus 2427.

Studies on the distribution, location and productivity of A. gambiae larval habitats in a hilly, highland region of western Kenya have demonstrated that productive habitats are located in valley bottoms near streams and with agricultural crops being the predominant land cover type 101314. By contrast, our study area is lowland plain with some gentle, rolling hills, and with temporary streams that become inundated during the rainy season. These streams and numerous, small, man-made water reservoirs serve as water sources in the dry season, as does the lake. Clusters of houses (i.e., compounds), streams, roads and other physical features have been mapped previously 724.

An IKONOS satellite image (Space Imaging, Atlanta, Georgia) of the study area was acquired to primarily identify land cover types associated with location of larval habitats for Anopheles mosquitoes. The IKONOS scene acquired was 100 km², of which 10.34 km² were used in this study. The image was acquired on May 17, 2005 at 8.14 GMT, and was centered on 34.435571E and 0.143949S. The image was registered and orthorectified with geographic latitude-longitude coordinates using ground reference points taken with a global positioning system (GPS). Because changes in land cover types in the study area are mainly due to seasonal cropping practices and the deciduous nature of most plant species, only one IKONOS image was used for both dry and wet seasons.

Two surveys for aquatic larval habitats were conducted in the study area from May to June for the wet season, and in October for the dry season. During each survey, all pools of standing and relatively slow moving water were quantitatively sampled using an area sampling method as described by Mutuku et al. 9. The term "potential habitat" in this study refers to any discrete body of water that is likely to remain inundated for at least 2 days. These habitats were categorized into seven types including drainage channels, burrow pits, rain pools, cluster of hoof prints, stream bed pools, wet meadows, and tire tracks 8. In the dry season survey, the small temporary pools in the form of rain pools, drainage channels, cluster of hoof prints and wet meadows were absent.

To determine the accuracy of the land cover/use classification, the study area (measured 3.216 × 3.216 km covering an area of 10.34 km²) was divided into 36 grid cells of 500 m × 500 m each. With the assumption that distribution of different types of land cover is homogeneous in the study area, 16 of these grid cells were randomly selected for ground-truthing of land cover types. One individual, familiar with the territory, traversed each cell and at each of 951 locations mapped the different land cover types. The individual identified the appropriate land cover type to category by visual inspection, and completed a GPS reading of the center location of the land cover. This procedure was repeated for each neighboring land cover. The process took two weeks and was timed to coincide with the date when the IKONOS satellite image was acquired. These points were located using differential GPS with a precision of about 1 m and were used as ground reference points for comparison with remotely sensed data and the supervised classification. The number of points mapped in each of the 16 grid cells was proportional to the land cover heterogeneity in each particular grid cell.

The IKONOS imagery obtained from Space Imaging during the wet season was barely obscured by clouds and shadows, representing only 4% of the image area, leaving all of the study area where larval sampling occurred free of cloud cover. ERDAS IMAGINE 8.5 image processing software 28 was used for image processing and analysis. Through resolution merging, a pseudo-1- meter resolution image using the four multispectral bands and the panchromatic band was constructed using principal component analysis (PCA) in ERDAS IMAGINE 8.5 2829. This fused image (henceforth referred to as the pan-sharpened IKONOS image) was visually more interpretable than either the 4-meter multispectral image or the 1-meter panchromatic image. The IKONOS image was classified through a combination of visual, unsupervised, and supervised methods. The ISODATA (Iterative Self-Organizing Data Analysis Technique) unsupervised classification algorithm was performed to differentiate the spectral clusters corresponding to the basic land cover types. The unsupervised classification used the green, red, and near-infrared spectral bands to produce 30 spectrally distinct classes, each based on the analyst's prior knowledge of the different land cover types in the study region. With the aid of the ground-truthing reference points, the unsupervised classified image, and the pan-sharpened IKONOS image, the multispectral IKONOS image was then visually classified according to typical land cover types and vegetation. This process resulted in 27 training signatures with most of them being spectrally close. The 27 training signatures (subclasses) were amalgamated to ten major training signatures (classes) including grasslands, water, mature maize, newly cultivated fields, clouds, shadows, bare land, short shrubs, medium shrubs, and tall shrubs (including trees). The subclasses helped define each class more accurately, for example, shallow water and deep water were used for water, 3 maize fields (subclasses) were trained for mature maize and 5 patches of grassland were used to train for grassland. The supervised classification was done on the entire IKONOS scene. Supervised classification was then performed, using a maximum likelihood classifier.

To ensure that the information derived from the classification was of high quality and to deduce meaningful indications on thematic correctness 30, the classified image was assessed for accuracy using Cohen's Kappa and classification table metrics 2.C, an ArcView 3.3 extension (ESRI, Redlands, California). In the assessment, the classified IKONOS image and the ground reference data were compared. The relationship between these sets of information was summarized in two ways: (1) a confusion matrix, which describes the comparison of the remote sensing derived classification map and the ground truth reference data; and (2) the Kappa statistic, which provides a measure of agreement between the classified remotely sensed data and the ground reference data.

There were no ground reference data for the land cover classes termed "shadows", "clouds" and "water" in the classified image and hence these were assigned to a "No Data" classification. Newly cultivated fields were amalgamated with mature maize because most of the fields in the newly cultivated fields land cover class were actually maize fields. Overall accuracy and class-specific user and producer accuracies were calculated for each of the resultant six land cover classes. Producer's accuracy was obtained by taking the number of points classified correctly for a class divided by the number of ground reference points in that class, while the user's accuracy was the number of points classified correctly for a class divided by the number of points classified as that class 31. When a point was incorrectly included in a class, an error of commission has occurred. Inversely, when a point was excluded from the proper class, an error of omission has occurred.

The spatial distribution of the larval habitats was created as a layer and overlaid on the land cover layer, and the number of larval habitats in each land cover class was calculated using Zonal statistics, a tool from the Spatial Analyst extension in ArcView GIS (ESRI, Redlands, California). Accuracy in identifying the land cover class where larval habitats were located was ensured by creating a 4 meter radius buffer around habitat so that larval habitats located at the edges land cover types will associated with land cover closet to them, in which the majority land cover type was identified. A range of buffer areas (4, 8, 12, 16, and 20 meters) was examined but the 4 meter radius buffer was supported well by the majority neighborhood statistic. Chi-square analysis for categorical variables was used to examine whether there were significant differences in proportions of positive and negative larval habitats located in different land cover classes. Tukey style multiple comparison of proportions were used for post-hoc analyses 32. Seven categories of distance from streams were created using ArcGIS 9.1 (ESRI, Redlands, California), i.e., 100, 200, 300, 400, 500, 600, and >600 meter distance to the nearest stream centerline. These distance categories were overlaid on top of thematic maps of the classified IKONOS image to extract the proportions of each land cover type within the various distance categories. Regression analysis was used to determine the relationship between the proportions of the various land cover classes and the distance categories from stream channels.

The distribution of habitats in the study area is shown in Figure 1A. In the wet season survey, 86% of the 1,198 potential habitats located and sampled had Anopheles larvae; in total 19,776 larvae were collected. In the dry season, 40% of the 184 potential habitats located and sampled were positive for Anopheles larvae and in total 582 larvae were collected. Of these larvae, 2,231 were mature and could be identified morphologically to species. Overall, Anopheles gambiae s.l. (a species complex of two important malaria vector species, A. gambiae s.s. and A. arabiensis) represented 50.69% of total, A. coustani was 21.07%, A. rufipes was 10.89%, A. pharoensis (5.15%), A. squamosus (4.71%), A. maculipalpis (4.71%), A. funestus (1.97%), A. gibbensi (0.40%), and A. pretoriensis (0.40%). Analysis of 1,078 of the larvae identified morphologically as A. gambiae s.l. by the PCR method 22 revealed that 796 (73.8%) were A. arabiensis, 19.8% were A. gambiae s.s., and 69 (6.4%) did not react in the test used.

Classification results (Figure 1B) revealed that the landscape was highly modified by human activities typical of rural subsistence farming. Most of the study area was covered by grassland (21%), medium shrubs (16%), short shrubs (15%), mature maize (14%) and tall shrubs (13%). Shadow (mainly from tall shrubs) represented a very small proportion of the total land cover (<1%). It was combined with the tall shrubs land cover class. Bare land (10%) and newly cultivated fields (11%) were the least common land cover types. Twenty-nine percent of Anopheles larvae came from habitats sampled from the mature maize land cover type, whilst 26% were from grasslands, 17% from newly cultivated fields, and 13% from medium shrubs. Only 7% were sampled from the short shrubs land cover type, 6% from tall shrubs, and 3% from habitats located in the bare land type of land cover.

The vegetative community that dominated the Grassland land cover was primarily seasonal grasses, often grazed by cattle, and included Hyparrhania ruffa (Poaceae), Cynodon dactlylon (Poaceae), and Eragrostis tenuifolia (Poaceae). The Bare Land was a land cover in areas of exposed and typically dry soils, found along roads, near clusters of houses, and at school playgrounds. Corrugated iron roofs were captured by the supervised classification as being close to this land cover type, which is consistent with the bare soil around homesteads. Mature Maize was a dominant land cover, typified by numerous fields of maize (and occasionally sorghum) at a stage of mature growth. Maize plants in this class were generally about 2 to 3 meters tall, had tassels and were mainly found on low-lying areas along the streams. This class formed a closed canopy in most cases. In addition to crops grown in these fields, all the other vegetative communities mentioned under newly cultivated fields were also found in association with this class. The Newly Cultivated Fields land cover consisted of patches of land recently cultivated. Crops in this class were typically maize under one meter tall, sorghum, and peanuts. Land cover of this type was commonly located adjacent to or between mature maize fields. There were significantly large areas in between the crops where wet, bare soil was exposed but it was classified as newly cultivated fields given the obvious association. Non-cultivated subclasses that were spectrally similar to the newly cultivated fields included areas covered by sedges with surface water under them and fallow lands. Therefore, the vegetative communities that aside from crops were associated with this land cover included Trichodesma zeylanicum (Asteraceae), Cyperus articulatus (Cyperaceae), Commelina africana (Commelinaceae), Cyperus ajax (Cyperaceae), Pennisetum purpuneum (Napier grass), Sesbania sesban (Papillionaceae), Malva verficillata (Malvaceae) and Onagraceae spp. Vegetative communities in the Short Shrubs land cover were comprised of perennial and herbaceous plant species that were mainly less than 2 meters in height. Some of the dominant plant species included Sphaeranthus suaveolens (Asteraceae), Solanum incanum (Solanaceae), Stachytarpheta jamaicensis (Verbanaceae), Catharanthus roseus (Apocynaceae), Ocimum basilicum, Senna (Cassia) occidetalis (Ceasalpinaceae), Therettia thevetoides (Apocenaceae), and Methania angustifolia (Sterculaceae). Goats commonly browsed these plants, suggesting an agricultural use. The vegetative communities in the Tall Shrubs land cover were dominated by various species of tall shrubs and trees with an average height of over 6 meters. Justicia flava (Acanthaceae), Euphorbia tirucali (Euphorbiaceae), Acacia polyacantha (Mimosaceae) were some of the common plant species associated with this land cover type. In some cases the plants formed a closed canopy. Mango trees (Mangifera spp., Anacardiaceae) and banana plants (Musa spp., Musaceae), both cultivated for their fruit, occurred in this land cover class. Vegetative communities of the Medium Shrubs class were dominated by perennial plants with a mean height of 3–6 meters, especially Lantana camara (Verbenaceae), but also Indigofera spicata (Papillionaceae), Ipomoea spathulata (Convolvulaceae), Nerium oleander (Apocynaceae), and Senna (Cassia) floribunda (Ceasalpinaceae). One obvious use of this land cover type was in border plantings, where oleander and lantana were planted in hedgerows. The majority of streams and still water bodies (such as man-made ponds) were undetectable by the satellite imagery, either because of a closed vegetation canopy (streams) or because the spectral characteristics were similar to bare soil because of turbid water. Therefore, most true bodies of water fell within other land cover classes, and in any case Open Water per se was an uncommon land cover in the study area and was not included as a classified category. Cloud shadows were found mainly on the northwest area of the image, outside of the study grid. Areas covered by shadow within the study area were primarily as a result of shadows cast by tall trees and small buildings, so the Shadows land cover type was incorporated into the tall shrubs or bare soil land covers, respectively.

The overall accuracy of the land cover classification from the IKONOS multispectral image was 89% (Kappa statistic, 0.86). The confusion matrix showed that the IKONOS classification was best at distinguishing mature maize and grassland land cover types (Table 1). Generally, the frequency of confusion of classification was low but did occur; for instance, medium shrubs were sometimes confused with both maize and tall shrubs, and turbid water was confused with bare land. Bare land had the lowest classification accuracy because it was largely confused with mature maize, grassland and short shrubs (Table 1). Overall, instances of confusion were minimal and did not affect user's classification accuracy (Table 2). The user's accuracy ranged between 77 and 95%, with relatively low errors of commission (excesses), varying between 5 and 23% (Table 2). The producer's accuracy ranged between 85 and 93%, with similarly relatively low errors of omission (deficits) depending on the class (7–15%).

Most Anopheles larval habitats were located within the mature maize land cover (28%), followed by grasslands (25%), newly cultivated fields (15%), medium shrubs (13%), tall shrubs (9%), short shrubs (7%), and bare land (3%) (Figure 2). There were far more habitats in the wet season compared to the dry season, and more potential habitats were occupied by larvae in the wet season compared to the dry season as well (Figure 2). Assuming a random association, the number of the potential habitats located on each land cover type should be proportional to the total area covered by each land cover type. The expected number of potential habitats by land cover type was calculated within the study area using the method of proportions and using total area of each type as the denominator. It was compared with the observed number of each habitat type with a χ² test. Overall, there were significant differences between the numbers of the observed and expected potential habitats (χ² = 324.2, df = 6, P < 0.0001). Approximately 68% of all potential habitats were associated with three of the land cover types, namely mature maize, newly cultivated fields, and grassland. The three land covers had a higher number of the habitats than the expected by chance alone (Table 3). If the habitats had been randomly distributed, 46% of them would be expected to be found in these land cover types. For the remaining four land covers, i.e., short shrubs, medium shrubs, tall shrubs and bare land, there were fewer larval habitats than expected by chance (Table 3).

Larval habitat location varied with distance from stream channels in both wet and dry seasons (Figure 3). In the wet season, 82% of all potential habitats and 72% of Anopheles-positive larval habitats were located within 200 meters of stream channels. In the dry season, 94% of all potential habitats and 93% of Anopheles-positive larval habitats were located within 200 meters of stream channels. Nonlinear regression showed that the number of habitats y per 100 m distance from streams x fit well the exponential equation: y = 1580e^-0.0084x (R² = 0.97). For the wet season, χ² analysis for categorical variables showed that there were significant differences in the proportions of Anopheles-positive larval habitats among the distance categories (χ² = 25, df = 6, P = 0.0003). χ² analysis for categorical variables was not done for the dry season data because most of the distance categories had fewer than 5 potential habitats.

The relationship between the proportion of the various land cover classes and the distance categories from stream channels was examined to determine whether the association of particular land cover types with habitat locations might also be explained by distribution of land cover types relative to streams. Regression analysis showed that area of land covered by grassland, newly cultivated fields, and mature maize were all higher nearer streams and decreased away from streams (Figure 4). By contrast, the area of land covered by bare land, short shrubs, and medium shrubs increased with increasing distance from streams (Figure 4). There was no relationship between area of land covered by tall shrubs and distance from stream channels (r = 0.45, df = 5, P > 0.05).

Within the Asembo study area (area = 10.34 km²), supervised classification of the multispectral 4-meter IKONOS image identified water on only 128 pixels. These pixels were extracted from the classified image and converted into polygons using ArcGIS 9.1 (ESRI, Redlands, California), a process that resulted in 27 water bodies (Table 4). To confirm that the IKONOS image truly identified water bodies that were potential larval habitats, a verification procedure was carried out in the field. Maps of the water bodies were prepared and loaded in a personal digital assistant (PDA), which was then used as a navigation guide to each of them. The verification process revealed that 44% (12/27) of the water bodies were rain pools, 30% (8/27) were burrow pits, 22% (6/27) were stream courses, and 1/27 (4%) was a puddle in an automobile tire track.