Real world description

Our decisions require adequate knowledge about the world. This knowledge, however, is far from complete. Therefore, we have to use incomplete, selected information. The model formed is a set of relationships, conditions and rules describing our understanding of what creates a specific part of our world and how it is functioning. The real world can only be described in terms of a model. There are countless possible approaches to the description and interpretation; consequently, the design of GIS systems varies according to their purpose. Every model is a simplification of the reality because it only includes selected data relevant to the model.

Spatial data

Data used to describe our world must include answers to the key questions: ‘What?’, ‘Where?’ and ‘When?’. If data are related in such a way that they can answer all three questions, they constitute spatial data.

Depending on the question, two fundamental components characterize spatial data: (1) geometric data or location data, and (2) non-geometric data or descriptive data (attributes). Geographical data are inherently in the form of spatial data. An estimated 80% of all natural sciences data has a spatial component. GIS uses geographical data in a digital format. Location data usually refer to specified spatial coordination systems of two dimensions (e.g. longitude and latitude) or three dimensions (e.g. the Cartesian orthogonal system used in satellite navigation systems). When a common mapping coordinate system is used, spatial data in GIS are termed geocoded or georeferenced data. Simple geometrical shapes are used to describe location data. A point represents the location of phenomena or objects small enough for such presentation. Sets of connected points create lines indicating the position of linear features like boundaries or communication networks. Closed lines constitute boundaries of an area such as that delineating a country, lake or forest.

Time is the third fundamental component of spatial data. It is a contextual reference. Thus, location and descriptive data relate to a given point in time.

Spatial data in maps and in GIS

Printed maps contain spatial data in a compact and easily accessible format. However, in rapidly changing regions, such maps quickly become outdated. Retrieving information from printed maps is relatively simple for small amounts of information. For larger volumes of data, as required for applications such as environmental assessments, modelling land cover changes, or climate studies, it becomes more difficult and costly. It becomes even more time consuming and cumbersome when data integration from different maps is required.

These drawbacks are overcome by the use of computing to process large volumes of data for geographical data analysis. The requirements are: (1) all data in digital format and (2) a set of relatively simple instructions (software) for data processing. Generally, there are two approaches to preparing digital spatial data for GIS.

A vector data model stores a table containing coordinates of points together with instructions on which points are alone and in which points belong to a common set. Digital data are discrete data and, hence, analogue values are eliminated when a digital data model is applied. Graphical presentation of spatial data is also based on points. To demonstrate that some points belong to one object a line connects the points. The simplest line is a vector in which both start- and end-point are defined. In a vector data model, all lines are represented by chains of vectors, and all areas by polygons. For graphical presentations, discrete values can be transformed into continuous data using mathematical functions to compute the missing values. Other types of function can be used instead of straight lines if they better describe changes of spatial data in space and/or time. Attributes are coded in separate tables using alphanumerical characters as a label for a specific class or category of properties. Association between location data tables and attribute tables is achieved by inserting the same identities in both kinds of tables for identical objects.

A raster data model uses a net of adjacent polygons (termed ‘cells’) to provide a virtual cover of a given part of a landscape. In its simplest form, cells are identical squares forming an organised grid. Each cell represents one entity of the landscape's surface. Location data for this entity is coded by the cell position in a matrix defined by row and column number. The cells are often called pixels, a term used in display technology for picture elements. Attribute values of the objects that the cells represent are assigned to corresponding pixels. Only one value can be assigned to a single cell, which requires a different raster layer for each thematic topic or period of time. Thematic layers are also used in vector models, but because of differences in attribute manipulation, raster models have more layers than vector models.

An important attribute of GIS is its data-handling capacity. GIS allows the input of data of different origin, content and format, and digital data can be edited and modified relatively easily. Consequently, GIS enables databases to be updated regularly. The system also offers wide flexibility in manipulating the data and displaying it in different formats. Finally, and importantly, GIS allows quantitative methods to be used in spatial data analysis. Analytical methods based on mapped analogue data are primarily qualitative and depend on visual and intuitive assessment of the data. In contrast, GIS permits qualitative methods to be expressed through mathematical relationships computed by software instructions. The real benefit of GIS is the ability to integrate data in models to give a value-added product, customised to meet user needs.

In the long term, the advantages of digital data reward the initial work and financial investment. Rapidly increasing computing power has resulted in a proliferation of GIS applications in environmental analysis during the last 20 years.

Spatial data collection

Spatial data quality

The performance of GIS depends not only on the technical characteristics of hardware and functions offered by software, but primarily on the quality of input data. Digital data are frequently considered to be of higher quality than conventional data, but this is not necessarily the case. Data quality depends on the quality of the data sources: quality of the original input data, the method of data acquisition and on the precision of all stages of data processing. Knowing the quality of data is critical to determining the appropriate use of the data. The most important measures of data quality that affects GIS applications are listed below.

Data analysis

GIS is used as an interactive analysing tool to support understanding of the environment and successive decision making as well as a database for the storage and retrieval of spatial information. Spatial analysis functions are characteristic features distinguishing a GIS from all other information systems. Ideally, a GIS should have all the functions to allow all conceivable operations with the data. In practice, a range of operations is supported by different GIS products (Table 1). These are changing constantly as the development of GIS technology provides growing numbers of increasingly sophisticated functions. The most difficult part of the analysis is defining the task or problem to be addressed and delineating the approaches to solutions. Particular regard must be given to the quality of data and its limitations.

The computing power of modern GIS allows complex analyses to be made, even with large volumes of data. This is particularly important when a model is tested and integration capabilities are needed to obtain quick and reliable analyses of alternative scenarios. Successful spatial analysis using GIS depends on the model selected, type of data relevant to the model and knowledge of how to use GIS technology appropriately. Correct use of GIS certainly requires more than pushing buttons in a GIS-menu! Results must always be evaluated for accuracy and content, and alternative approaches used or new data sought when problems arise. Ideally, analytical results should be presented as easily interpreted maps accompanied by an explanatory report.

Natural focality of diseases (NFD)

The theory of natural focality of transmissible diseases was formulated by E. N. Pavlovsky. It was based on several years' research in the taiga of the Far East (Khabarovsk region, 1936–38) studying the circulation in nature of the then newly discovered virus of tick-borne encephalitis (TBE), and the transmission pathways to humans. Pavlovsky was also influenced by observations made in expeditions to Central Asia to investigate tick-borne relapsing fever (Borrelia persica) transmitted by the argasid tick, Ornithodoros tholozani. Starting with the assumption that the disease agent and its vector form part of a biocenosis in a given habitat (biotope), Pavlovsky demonstrated that the majority of vector-borne diseases (VBD) are localized in foci located within defined geographical complexes. He called them natural foci of diseases (NFD). The term NFD differs from that of a focus in the general epidemiological sense because it involves a geographical territory that is usually precisely defined and demarcated by certain ecosystems.

Pavlovsky published on this theory of NFD over a 25-year period (Pavlovsky, 1939, 1946–48, 1964) and other researchers have contributed (Rosický, 1967; Korenberg, 1983; Kucheruk & Rosický, 1984). Today, NFD is defined as ‘A natural focus of a disease is a geographically demarcated part of the landscape which has formed during the course of natural evolution and contains an association of organisms in which there circulates a disease agent as its integral component, independently of man and domestic animals.’ Under certain conditions humans and domestic animals can become involved in the circulation of the infection. Natural focality is therefore a general biological phenomenon wherein the infective agent and its vector have, in the course of evolution, become components of a stabilized ecosystem independently of humans.

The structure of natural foci of diseases

The disease agent, vector and susceptible hosts are not evenly distributed within an NFD. Ticks occur in a mosaic-like pattern determined by biotic as well as abiotic factors. Locations that consistently sustain circulation of the disease agent are termed elementary foci (Pavlovsky, 1946–1948). Such places offer optimal conditions for the vector and its hosts. Elementary foci are often sites where game animals gather and conditions are favourable for small mammals, e.g. sites with mixed vegetation, glades with shrubs, as well as islets of trees in fields. Border zones (ecotones) play an important role. An elementary focus is characteristically structured. Its basis is a nucleus (synonymous with a ‘microfocus’ or ‘hot-spot’), and its centre (axis) may often be a brook (often only periodic) or a hollow with increased humidity and richer vegetation. The nucleus is usually surrounded by a transitional belt forming its boundary (Hejný & Rosický, 1965). The structure and biocenotic status of the nucleus of a long-lived focus of TBE in north-eastern Bohemia (Radvan et al. 1960) is depicted in Figs 1 and 2.

Fig. 1. Outline of the structure of a nucleus in a focus of TBE: A – axis of the nucleus; B – spruce mono-culture (without any tick occurrence); C – coat of the nucleus; D – grain field (without ticks); E – transitional phytocenoses; N – nucleus (reproduced with permission, Rosický & Daniel (1989)).

Fig. 2. Example of a nucleus in a natural focus of TBE (delineated by thick line) monitored in north-eastern Bohemia. Length of tract AC – 235 m; tract BD – 289 m; width of the nucleus proper 167 m, length 141 m (according to Radvan et al. (1960) reproduced with permission, Rosický & Daniel (1989)).

Depending on the configuration of the terrain and the local distribution of habitats, an elementary focus usually has several nuclei in which the disease agent survives inter-epizootic periods. Usually the area covered by each nucleus is not great enough to sustain circulation of the agent indefinitely; the activities of alternate nuclei are then crucial. Elementary foci are often the only type of NFD in highly exploited landscapes; identifying them in the terrain is important for epidemiological and epizootological prognoses targeted at timely preventive measures (see Figs 7 and 8). So-called bioindicators can be used to identify the presence and exact topographical localization of elementary foci in the field and in cartographical documentation. Bioindicators may be biotic or abiotic. Examples are plant associations, landscape relief, hydrological conditions, and slope gradient.

Mapping tick distribution

Methods most commonly applied in predictions

Species occurrence is primarily determined by macro-climatic factors hence macroclimate is commonly exploited in prediction models. Among the most frequently applied ‘expert’ systems is CLIMEX (Sutherst & Maywald, 1985). Such a prediction model is based on the assumption that an area of occurrence is limited by a ‘climatic envelope.’ The envelope is positively delimited by a ‘Growth Index’ which combines favourable temperature and humidity, and negatively delimited by ‘Stress Indices’ represented by extremes in temperature and humidity. In the basic layout, climatic parameters at a given site are determined by interpolation in a grid of measurements by a network of 2400 meteorological stations situated all over the world. The output of the model is an ‘Ecoclimatic Index’ of 0–100 which is interpreted as an overall measure of climatic suitability for population growth and the persistence of a given species. CLIMEX has found many applications in zoogeography including the prediction of tick occurrence (e.g. Norval et al. 1991; Perry et al. 1991; Sutherst, 2001).

Another common ‘expert’ approach is the application of geobotanical data. Certain plant associations and their structures reflect macro- and/or meso-climatic factors, geological and other conditions, but also create microclimatic, nutritional and shelter conditions for the presence of ticks and their hosts. The association of tick population density and a certain type of vegetation is thus a strong predictive criterion of occurrence that can be verified (e.g. Sonenshine, Peters & Stout, 1972; Gilot, Pautou & Moncada, 1975; Estrada-Peña, 2001a). Effective application of this principle in landscape epidemiology has been facilitated by the development of remote sensing (RS) enabling identification of tick habitats from multispectral images (Hugh-Jones, 1989; Daniel & Kolář, 1990; Hugh-Jones et al. 1992). The vegetation categories discerned in satellite-mediated imagery do not follow the classical phyto-ecological categories obtained by researchers on the ground. RS classification of tick habitats is more about the structural characteristics of the vegetation but their predictive potential is not diminished compared with ground-based characterisation (Duffy et al. 1994; Daniel et al. 1998, 1999; Beck, Lobitz & Wood, 2000).

A strategy different from the ‘expert’ approach is the use of covariates expected to be linked with tick occurrence in ad hoc models. Besides climatic and vegetational covariates mentioned above, there are geological, soil type, hydrological and host population covariates as well as anthropogenic and other factors co-determining the site or area of occurrence. To envisage the often subtle relationship between tick occurrence or the risk of TBD with covariates, certain statistical methods are applied (e.g. Robinson, 2000). The most common are modifications of regression analysis. Discriminant analysis, co-kriging and classification tree methods have also been applied on occasion.

Various modifications of the regression model allow investigation of the functional relationship between independent variates (or covariates) and a dependent variate. There are a number of examples of applying correlation analysis (Kitron & Kazmierczak, 1997), multidimensional linear regression (Glass et al. 1994) and logistic regression (Glass et al. 1995; Cumming, 2000a,b; Guerra et al. 2002) for constructing models of tick or TBD occurrence based on numerous covariates amassed in the GIS data base.

Bioclimatic zones and prediction scale

Predictive maps can be produced for small territories (e.g. Daniel & Kolář, 1990; De Garine-Wichatitsky, 2000) as well as at pan-continental scales across many bioclimatic zones (e.g. Estrada-Peña, 1999; Cumming, 2000a). The scale and bioclimatic zone markedly influence the choice of predictors. Variables with small spatial variance, namely macro-climatic indices, have nearly no predictive value in small territories, whereas diversified variables such as vegetation type and vegetation structure have limited value in large-scale studies. Many predictive models have been elaborated for the conditions of the tropical and subtropical climate zones of Africa, Australia and South America using mainly macroclimatic variables or their satellite surrogates and Normalized Difference Vegetation Index (NDVI). While climatic variables (monthly temperature and mean precipitation) are generally good predictors of tick occurrence in such conditions, the efficacy of NDVI and vegetation variables generally differs (Cumming, 2002). However, in the temperate zones of Europe and North America, similar studies have demonstrated the predictive potential of vegetation variables for tick and TBD occurrence (e.g. Sonenshine, Peters & Stout, 1972; Gray et al. 1998). Other effective predictors are the type and moisture of soil (e.g. Carey, McLean & Maupin, 1980; Kitron, Bouseman & Jones, 1991; Glass et al. 1994), waterways network (Kitron et al. 1991; Estrada-Peña, 2001a), terrain configuration (Merler et al. 1996), game density and even various anthropogenic factors.

Species specificity of predictive models

Predictive mapping studies have most commonly involved economically important species in tropical regions (Boophilus microplus, Amblyomma variegatum and Rhipicephalus appendiculatus) and epidemiologically-significant species of the temperate zone (I. ricinus in Europe and I. scapularis in North America). Predicability of distribution of a particular tick species is influenced by the complexity of its developmental cycle and range of hosts environmental requirements of both the tick and hosts and other factors (Randolph, 2000). For related and ecologically-close species, inter-species competition may have a confounding effect (Estrada-Peña, 2001b).

From tick maps to TBD prediction

Prediction of TBD risk is dependent on the prediction of occurrence of competent tick species; however, other factors contributing to the existence of natural foci may have a substantial effect. When the influence of such factors is negligible, the potential distribution of TBDs can be estimated on the basis of a simple combination of maps of major vectors with the tools of GIS. In other cases, GIS models should combine additional data sources. Special modelling techniques are required when variations through time of abiotic and biotic factors have an influential effect. Such dynamic factors include tick phenology, cyclic climatic phenomena and population cycles of the hosts. Interactions of these factors can sometimes be complicated (Jones et al. 1998). An elegant technique is the application of Fourier transformation to represent the periodicity of the temporal series of covariates. This approach reduces the amount of data and renders the model relatively easy to present (e.g. Rogers, 2000). Seasonal dynamics of climatic factors plays a key role in one model of TBE distribution (Randolph et al. 1999). The model assumes that virus persistence in a natural focus depends on ‘co-feeding’ of I. ricinus larvae and nymphs which, owing to the phenology of these stages, occurs only where there is a significant annual temperature gradient. This model uses the Land Surface Temperature (LST) index derived from data of the National Oceans and Atmospheric Administration (NOAA) satellites as a surrogate value for near-ground temperature (Green, Rogers & Randolph, 2000; Randolph & Rogers, 2000).

Validation of predictive models

Rapid developments in computer technologies and the increasing supply of RS data have led to wider applications of predictive mapping. Sophisticated GIS software packages offer the user a number of powerful modelling tools. However, their application has the shortcoming of being a ‘black box’ and can generate maps from input data that do not respect the elementary limitations of the mathematical model applied. Paradoxically, most such outputs look meaningful at first sight. Therefore, it is critical to verify the reliability of such predictions, for which a number of statistical procedures are available (Robinson, 2000).

Population denominator

Any geographical pattern of disease incidence combines two spatial variables: the intensity of disease risk and the density of the population at risk. To rate the risk from incidence data, some information about the spatial distribution of the population at risk is required. There are two main approaches to this task in spatial epidemiology. The first relies on census tract counts, and the second employs some control events assumed to be uniformly distributed in the population, such as incidence of a common disease. With TBD, however, the risk is associated with a visit to rural locations, so the population at risk does not necessarily reflect residential population density and is not distributed uniformly. Instead, the interplay of behaviour or activities that may lead to exposure to ticks (demographic factors, road networks, and landscape features) and can shape the population at risk. Moreover, these factors can markedly vary from place to place (Rand et al. 1996). For example, tourism and outdoor recreational activities, which have been recognised as important risk factors in Lyme borreliosis (LB) (O'Connell et al. 1998), can exaggerate disease incidence in recreationally-exploited areas resulting in overestimation of risk. Assessment of the population at risk depends on the scale. Studies of large areas are less likely to be influenced by variations in population denominator, whereas small-area studies require special modelling techniques (e.g. Thomsen, 1991; Whittie, Drane & Aldrich, 1996). For example, paired data can be used on locations of residence and disease acquisition to model patterns of peridomestic exposure (Fig. 3).

Fig. 3. The frequency distribution of distances between residential addresses and places of infection in a set of patients of Lyme borreliosis fitted with power regression illustrates a peri-residential exposure of the population at risk (reproduced with permission, Zeman (1997)).

Analysis of spatial clustering

One important question in spatial epidemiology is whether an observed pattern of cases indicates clusters or whether it might have arisen by chance alone. As stated previously, TBD cases tend to be aggregated so, rather than presence of clusters, it is important to gain an insight into the ‘character of dispersion’, which may contribute to ecological studies of a particular disease. Several statistical methods for spatial clustering have been adopted to analyse TBD, namely methods for quadrat counts, second-order methods, Moran statistic, and Kulldorff-Nagarwalla statistic.

Second-order methods

These are suited for analysing point-pattern disease maps and are methods based on a mathematical theory of distribution of between-point distances (Diggle, 1983). The crucial statistic is the K function, which characterizes how the number of neighbours varies with distance in an analysed point pattern. A plot of K(r)−πr² versus r (where r is a radius) illustrates the mean excess of cases over a homogeneous Poisson model of the same point density within r about any case. For a ‘random’ pattern, it coincides with the x axis (no excess) while positive and negative shifts indicates clustering and regularity, respectively (Fig. 4). Another useful statistic is

, sometimes referred to as the L function. Testing significance in these methods relies on Monte Carlo simulations (Fig. 5). An example of applications of these methods in the study of TBD includes cattle anaplasmosis (Hungerford, 1991), I. scapularis infested deer (Kitron et al. 1992), and human LB and TBE (Zeman et al. 1990; Zeman, 1997).

Fig. 4. Principal types of spatial distribution: (a) random, (b) regular, and (c) aggregated, and corresponding shapes of the K(r)−πr2 diagram clearly distinguishing between them (reproduced with permission, Zeman, Vitkova & Markvart (1990)).

Fig. 5. Plot of K(r)−πr2 statistic versus radius in km for tick-borne encephalitis cases (solid) along with confidence envelopes of a Monte Carlo model of complete spatial randomness (dotted) documents that spatial clustering is well pronounced in this disease (reproduced with permission, Zeman (1997)).

Computing risk maps

Once it is established that disease clustering occurs, it is worth proceeding with a statistical mapping procedure in order to highlight any underlying risk pattern. Most methods are based on a model of continuous space change that can be approximated by a smooth, mathematically-defined surface. These techniques have been reviewed elsewhere (e.g. Lam, 1983; Van Beurden & Hilferink, 1993; Lawson, 2001). The literature includes several examples of maps of TBD risk assessed by means of kernel estimator, kriging and splines. Sometimes it is also desirable to compare two spatial patterns of disease cases. By means of the Kelsall-Diggle estimator it is feasible to compute a relative-risk map.

Kernel interpolation at an inserted point is achieved by averaging surrounding values weighted with a distance-decaying function, e.g. Gaussian. This method is applicable for both continuous measures and point pattern data (Bithell, 1990; Lawson & Williams, 1993). It requires judgement concerning a parameter that controls the smoothness of the estimated surface (so called ‘window width’). Even if ‘objective’ methods for its selection are available (e.g. cross-validation), an empirical ‘trial-and-error’ approach sometimes leads to maps that are more acceptable from a practical point of view. Kernel smoothing of points of infection was applied to compute the risk maps of LB and TBE (Zeman, 1997).

Kriging is an interpolation method based on a generalised linear least-squares algorithm, using variograms as weighting functions. Variograms depict the average squared difference between data values as a function of distance and orientation between data locations (Isaaks & Srivastava, 1989). Kriging is a popular and widely implemented method of surface mapping in GIS packages; however, it lacks a unique solution and needs a large amount of interactive input and subjective assessment during the processing. This method was used to create a model that predicts LB transmission risk in Rhode Island, USA based on a survey of nymphal I. scapularis population density (Nicholson & Mather, 1996).

Splines are techniques of piecewise polynomial interpolation (Journel & Huibregts, 1978). An interpolated surface is composed of a number of elementary ‘tiles’ characterized by polynomials of a low degree fitting the enclosed data-points and extended to adjacent polynomials so that the whole surface is continuous and smooth. Splines usually tend to intersect the observed values, and require rather subjective tuning up of the control lattice (an analogy with the ‘window width’ parameter of the kernel estimator). A modification of this method was tested with epidemiological data for TBE (Krejcir, 2000).

Kelsall-Diggle relative-risk estimation method is a double kernel smoothing procedure for two point patterns (Kelsall & Diggle, 1995). It is suitable for analysing spatial variation of two types of health events if the question of interest is whether they are distributed equally. If not, an identification of areas of unusually high or low incidence of one of the two types is possible. In a surveillance of LB, for example, patients were invited to specify likely infection origin as either tick-bite, or insect-bite or unknown. Spatial pattern of their residential addresses was then analysed by the Kelsall-Diggle method assuming that all cases originated from tick-bite transmissions even if they were unrecognised (Fig. 6). The analysis, however, showed significant differences between the distribution of ‘tick’ cases and that of ‘insect’ and ‘unknown’ cases (Zeman, 1998, 1999).

Fig. 6. Log relative risk of acquiring Lyme borreliosis through the ‘tick’ or ‘insect’ bite according to self-reported data in cases from Central Bohemia computed by the Kelsall-Diggle's method; test of non-constant ratio gave p-value of 0·023, and solid and dashed lines indicate where the 95% tolerance limits were surpassed for the ‘insect’ and ‘tick’ reports, respectively (reproduced with permission, Zeman (1997)).

This is not an exhaustive review of all work in the field of spatial analysis of TBD data. Some other approaches such as Bayesian and Markoff Chain Monte Carlo (MCMC) methods have been used. The purpose here is to provide an overview of the main analytical approaches currently available which can be employed in TBD research. Despite wide interest in the methodology of GIS and spatial epidemiology, users of the methodologies usually have some statistical knowledge.

Description of remote sensing data

Remote sensing is an information technology providing information on landscape (i.e. biophysical properties of the real world) through measured values of electromagnetic radiation from the landscape surface.

Electromagnetic radiation is characterized by two parameters: intensity of electromagnetic energy carried by the radiation and wavelength of the radiation. The energy interacts with an object and is reflected or emitted from the object. The energy is modified in a unique way by a given object that distinguishes it from the radiation energy measured on a particular wavelength. It is this energy that is measured and converted into data for subsequent processing. Information about the surface of the earth can be derived from spatial and spectral distributions of energy emanating from the surface, and also from temporal variations in the distributions.

The optical region of the spectra (0·3 μm–1000 μm) is commonly used for remote sensing. For practical reasons the optical region is divided into several shorter wavelength intervals of the visible (0·3–0·7 μm), near-infrared (0·7–1·1 μm), middle-infrared (1·1–5 μm), thermal infrared (5–100 μm) and submillimetre infrared (100–1000 μm) region of the spectrum. Due to atmospheric absorption, only wavelengths shorter than about 15 μm are used from the optical spectrum. Other portions of the electromagnetic spectrum are also useful in remote sensing, e.g. waves from the microwave region (0·01 m–1·0 m) penetrate clouds. Remote sensing measurements are particularly discriminating when using several very narrow spectral bands and producing a multispectral data set. In contrast, monospectral data are obtained if only one spectral band is used for measurement (e.g. panchromatic band covering the entire or a substantial part of the visible radiation). The number of channels or bands defines a dimension of the acquired data set. Increased utility of narrow spectral bands has resulted in the progressive narrowing of bands and increased the number of bands used 10- to 100-fold. This hyperspectral remotely sensed data allows more accurate feature identification. Radiation energy is measured using a radiometer. The size of the surface area (pixel) from which a radiometric value is determined defines the spatial resolution, while the number and width of the spectral bands measured define the spectral resolution of the radiometer. Individual pixels captured in continuous lines and columns make up an image. A line of detectors or a mirror enables pixels to be measured in a row, whilst the movement of a platform within the radiometer generates a second dimension of the image. Imaging radiometers are called scanners or multispectral scanners when producing a multispectral data set.

Scanner platforms use various types of carrier, but operational remote sensing systems are based on satellites. At present, Earth-observation satellites circle in specific orgbits. Near-polar, sun-synchronous orbits carry satellites over both poles. As the Earth rotates below the satellite it brings all sites on the globe within the satellite's field of view. The plane of the orbit is fixed with respect to the sun. As a result, measurements are taken by the radiometer at approximately the same solar illumination conditions for each daytime overpass. Orbiting around the Earth brings a satellite repeatedly over the same site on the surface. The time interval between two successive measurements of the same site determines the time resolution of the scanner.

Scanners are passive remote sensing systems measuring reflected or emitted radiation originating from a natural source (e.g. Sun, Earth), whereas an active remote sensing system generates radiation from artificial sources within the system (e.g. radar, lidar).

Characteristic properties of remote sensing data

Both optical and microwave systems have their limitations for Earth surface applications. Two major categories of data distortion can be defined: (1) radiometric distortions caused by, e.g. atmospheric effects and changes in detector sensitivity; and (2) spatial or geometric distortions due to e.g. oblique view, altitude changes along the orbit, instability of the spacecraft and Earth rotation. The consequence is that repeated measurements taken by different or even the same sensor will have a component of variation owing to the pixel size, variable atmospheric path length, variable atmospheric conditions and the time-space variability of the environmental conditions being measured. These distortions can be identified and corrected by using an array of in situ measurement sites to provide ‘ground truth’ data.

Satellite remote sensing systems

Data from remote sensing satellites have been used successfully to monitor environmental conditions suited to the emergence and outbreak of infectious diseases. In several projects, the relationships between environmental parameters sensed by satellites (e.g. water, temperature and vegetation cover), the occurrence of disease vectors (e.g. mosquitoes, ticks and flies), disease reservoirs (e.g. deer and rodents) and patterns of human settlements, migration and land use, have been established. Based on such relationships, predictive models can be developed to aid public health efforts to control specific diseases.

The earliest family of polar-orbiting meteorological satellites came on-line in the 1960s supplying valuable data for national meteorological services around the world. The best known satellites of this type are the NOAA series of satellites, managed by the US National Oceanic and Atmospheric Administration. Besides meteorology, data from the main imaging sensor, the Advanced Very High Resolution Radiometer (AVHRR), of NOAA satellites have found widespread applications in oceanography, agriculture, hydrology, forestry and other fields. A broader range of satellites, with low spatial resolution for environmental monitoring, has been introduced over the years (Table 2).

Typical features of this group are scanners with a large swath width implying a high revisit rate and large pixel size. These properties enable investigations of dynamic phenomena on a global and continental scale. As low-resolution data are the cheapest satellite data available, they are the primary data sources for global change studies and are useful for real-time information where timing is crucial. However, these features have drawbacks for other types of analysis. The geometry of a large swath width results in substantial enlargement of the pixel size at the picture edges compared to pixels in the middle of the line. A pixel size of 1 km for vertical pixels can reach several kilometres for oblique measurements. In addition to the need of geometrical corrections, a pixel of this size often covers too much heterogeneous landscape. When analysing pixels with much inherent variability of classes, it is impossible to assign an average radiometric value for the whole pixel area to just one class. The same applies to other values derived from original measurements (e.g. NDVI, LST). Analysis of remote sensing data requires either defined classes of homogeneous areas in the pixel or a more complex knowledge of radiometric values corresponding to all possible area ratios of class combinations within the pixel. With pixels of one kilometre, NOAA-type data cannot be used for local or even regional scale environmental studies.

The opportunity to improve mapping from space of biotic and abiotic components followed the launch of the first Landsat satellite in 1972. The MSS scanner on this US satellite measured 60 m pixels in four spectral bands. For 14 years, Landsat satellites were the only source of remote sensing data with such resolution. French Spot 1 added a new scanner (HRV) providing multispectral data of 20 m pixel and 10 m pixel in panchromatic mode. At present, there are many potentially useful sensors for mapping vegetation type (Table 3). The latest satellite of the Landsat series, Landsat-7, has an advanced ETM+ scanner providing 30 m multispectral data, a 15 m panchromatic band, and an improved 60 m thermal infrared band. The latest Spot 5 satellite has improved spatial resolution in the panchromatic band down to 2·5 m and multispectral data for 10 m pixels. Ikonos, a commercial satellite, has started a new generation of very high-resolution satellites (1 m pixels in the panchromatic band and 4 m in visible and near infrared spectral bands) providing data for detailed studies on large scales.

In addition to the passive systems shown in Tables 2 and 3, there are also active remote sensing systems in orbit (Table 4). Synthetic aperture radars (SAR) are particularly important as they can penetrate cloud and vegetation canopy cover and can detect soil moisture and surface roughness.