In its broadest definition, the Amazon subregion is defined as the area covered by the humid tropical plain forest of South America that is shared by nine countries: Bolivia, Brazil, Colombia, Ecuador, France (French Guiana), Guyana, Peru, Suriname and Venezuela. The subregion covers some 7,200,000 sq km (Figure 1) and is populated by about 30 million people. The provision of health-care services to remote communities is often difficult and human mobility may limit malaria control 1 .

This subregion accounted for 89% of all malaria cases in the Americas that were reported by the Pan American Health Organization (PAHO) in 2008 2 . Among the Amazon countries, Brazil has the highest proportion of cases (56%). In 2011, Brazil and Colombia accounted for 68% of the cases in the Americas 3 . The three Guyanas (Guyana, Suriname and French Guiana) have the highest annual parasite index (API) of the Amazon subregion and, with Haiti, have the highest API of the Americas 2 . Four countries (Bolivia, Ecuador, French Guiana and Suriname) in this subregion have seen malaria incidence rates reduced by more than 75% between 2000 and 2011 but Guyana and Venezuela reported increased case numbers during this period 3 .

Transmission of both Plasmodium falciparum and Plasmodium vivax occurs across the Amazon (as well as some rare Plasmodium malariae infections). In 2008, P. vivax accounted for 82% of the malaria burden in this subregion but with some large disparities seen between regions. Plasmodium falciparum was responsible for about half of the cases observed in the three Guyanas but was present in smaller proportions than P. vivax in all other Amazonian countries 2 . The proportion of cases in French Guiana and Suriname due to P. falciparum are in 2012 20% lower than they were in 2000 3 .

Anopheles darlingi is the main malaria vector in Amazonian countries 4 5 6 and is the focus of most research efforts. This anopheline mosquito is widely distributed across South America and is highly anthropophilic; its biting pattern may show adaptation to human behaviour 7 8 . It is difficult to predict the occurrence of An. darlingi due to its great adaptability to different habitats and diluted presence in the environment (larva, and in some cases adults, are unlikely to be found in high densities).

Studies of An. darlingi are the most numerous but other Anopheles species (such as Anopheles marajoara) are also of interest due to their high density and entomological inoculation rates 7 9 10 . In the Amazon subregion, other anopheline species including Anopheles braziliensis, Anopheles nuneztovari and some species from the albitarsis, oswaldoi or triannulatus complexes may also be locally involved in malaria transmission 10 11 12 13 14 .

The distribution of malaria is determined by climate and other geographic factors that influence the development of mosquitoes and Plasmodium at a given time, but it is also influenced by environmental alterations over time. Ecosystem changes resulting from natural phenomena or human interventions, on a local or global scale, can alter the ecological balance and context in which vectors and their parasites develop and transmit the disease 15 . According to Patz and Olson 16 , changes in temperature patterns, due to global climate change and in variation in local land use practices, may alter malaria risk. Some authors directly relate environmental alteration to cases of malaria. Olson et al. 17 studied malaria in Mâncio Lima County, Brazil, in 2006. Adjusting for population, access to care and district size, a 4.3% increase in deforestation between 1997 and 2000 was associated with a 48% increase in malaria risk. Vittor et al. 18 19 suggested that deforestation and other human environmental alteration favour the presence of both An. darlingi larvae and adults in the Peruvian Amazon. However, Conn et al. 20 and Moreno et al. 7 suggested that human intervention could increase the presence of An. marajoara over An. darling: forest clearance and pollution may reducing the availability of larval sites for An darlingi and increase habitats preferred by An. marajoara.

Earth observation satellites permit to acquire wide ranging data concerning the continental surfaces of the Earth, with very different techniques (optical or radar imagery, radar altimetry, etc.). These data differ in their spatial, temporal, radiometric and spectral resolutions and can therefore document many environmental features at different spatial and temporal scales. The use of remote sensing (RS) to provide new insights for epidemiological studies was identified very early 21 , as many diseases have been linked to environmental features. A literature review by Herbreteau et al. 22 in 2007 found that RS was often, and increasingly, used to study parasitic diseases (59% of studies) including malaria (16% of studies). The challenge, when studying malaria, is to identify all the natural factors (such as seasonality, rainfall, temperature, humidity, surface water and vegetation) and anthropogenic elements (such as agriculture, irrigation, deforestation, urbanization and movements of populations) of the study area, and to link them with either the incidence of disease or the presence of vectors whilst also integrating temporal and spatial variations. This would then enable the identification of risk factors from the set of possible environmental parameters.

One approach is to link malaria and the land cover (LC) and/or land use (LU) characteristics 23 . Within such a methodological framework, Ostfeld et al. 24 suggest that using more explicit landscape approaches to study eco-epidemiological systems could improve the understanding and prediction of the disease risk. Landscape composition (the number and types of patches) and configuration (the spatial relationships among patches) must be considered alongside the set of highly localized biotic and abiotic features. Within the framework of the study of landscape ecological functions (also referred to as landscape ecology), there are many ways to characterize the landscape, around point samples or LC/LU patches. This raises questions of objectivity, relevance and adequacy when carrying out environmental characterization. Some studies have therefore tried to standardize and evaluate the effectiveness of the characterization methods 25 or to objectify them 26 .

However there has been no inventory or discussion of the LC/LU classes used for malaria studies. LC concerns the physical material observed at the earth surface (such as forests, water bodies and bare rock); LU is related to the human use of the land and integrates socio-economic and cultural functions (such as agriculture and housing).

Despite their differences, LC and LU are often mapped together and often result from remotely sensed image classifications performed by RS experts and/or botanists. Such classification procedures range from totally unsupervised approach to a full visual interpretation of the images and highly depend on the availability of the remotely sensed data, the availability of experts of the application domain, the adequacy of the data for the question addressed and the competence of the technicians, engineers and/or researchers that perform the image processing. As a result, a wide variety of LC/LU typologies and methodologies can be found in the literature. Researchers interested in malaria transmission should share their approaches to ensure that landscape characterization becomes more homogeneous and standardized. This requires a full inventory of the objectives, geographical contexts, exploited data, and LC/LU classes and their impact on the malaria transmission risk. This paper reviews the articles proposing remotely sensed LC/LU mapping for the study of malaria, to identify points of consensus and divergence, and to bring out procedural limitations. It takes an interdisciplinary point of view to formalize and unify a fragmented and sometimes implicit knowledge in the field.

Referenced articles using a LC/LU characterization for the study of malaria risk in the Amazon subregion were identified by performing queries in ISI Web of Knowledge^SM databases: Web of Science®, Medline®, Journal Citation Reports® and Current Contents Connect®. The keywords and expressions chosen to construct database queries were: malaria, Anopheles darlingi, "land cover" OR "land use", "remot* sens*", satellite, environment*, natural factor*, risk factor*, deforestation, "South America", Amazon*, "Amazon basin", America*, tropical, Bolivia, Brazil, Colombia, Ecuador, French Guiana, Guyana, Peru, Surinam*, Venezuela. Double quotes were used to query expressions and asterisk was used to represent any letter(s) in the query. For example, Amazon* covered the terms Amazon, Amazonia, and Amazonian. The queries were defined by the conjunction of two or more key words and/or expressions, resulting in the identification of between zero and 108 articles. It was not possible to identify all relevant publications using a single query (see Discussion).

The publications finally selected: i) are original research articles (reviews were excluded), ii) use remotely sensed LC/LU information (the study only focused on the explicit LC/LU types and ignored numeric indexes such as the NDVI), iii) are applied to malaria (malaria vectors or malaria cases) and iv) include the Amazon subregion in the study area.

The LC/LU types used by the authors were listed (Additional file 2). For Rosa Freitas et al. 13 , Monteiro de Barros et al. 14 and Sinka et al. 6 , who exploited LC/LU maps which had high numbers of land cover types that were not initially intended to study malaria, only the types discussed in the relevant context were listed. Comparable LC/LU types were grouped, and it was indicated if these were positively or negatively associated with malaria transmission risk. Some LC/LU types appear in several lines of the table, as they can belong to different higher level LC/LU types. For example the type closed to open (>15%) broadleaved forest, regularly flooded (semi-permanently or temporarily), fresh or brackish water (Globcover channel 160) (Sinka et al. 6 ), belongs to both Forest and Water types.

The number of studies that assume or conclude a positive, negative or unknown relationship between malaria and each LC/LU type are given in Figure 2. The LC/LU types correspond to those presented in Additional file 2. The positive associations between the different types of non-anthropized forests and malaria in Rosa-Freitas 13 were counted only once. Considering each forest type separately would bias the results as this study considered a much greater number of forest types than the other studies. There were no significant differences between these forest types regarding the presence of the primary malaria vectors (An. darlingi and Anopheles albitarsis).

Defining simple database queries to provide all the relevant papers according to the objective was challenging; this reflects the diversity of terms used in the field. The variety of disciplinary domains interested in this topic (including epidemiology, entomology, ecology, RS and modelling) can explain such variation.

There was also potential paper selection bias due to the dominant use of the English language in the databases considered. Relevant publications in Spanish or Portuguese, underrepresented in ISI Web of Knowledge^SM system, could have been missed out.

The geographical criteria chosen for paper retrieval excluded some pioneering studies such as those conducted in Central America (more precisely, in Belize 37 38 39 40 and in Chiapas, Mexico 41 42 ). These studies investigated the relationships between Anopheles abundance or larval habitats and landscape elements characterized by RS in order to predict areas at risk for malaria transmission. Studies in Central America have focused on different anopheline species from those in the Amazon subregion (Anopheles albimanus, Anopheles vestitipennis, Anopheles punctimacula and Anopheles pseudopunctipennis) with one exception 40 , making it difficult to directly compare their findings with the LC/U types associated with malaria risk in this review. However, the proposed methodologies may be useful for future studies based on entomological data in the Amazon.

Some interesting and normative works concerning LC/LU mapping in the region were also ignored as this review focused solely on studies dealing with malaria. For example, the digital land cover map of South America (1 km spatial resolution) produced by Eva et al. in 2004 43 could not be considered.

The availability, cost and national distribution policies of RS data (for research purposes and public health) influence the number of peer-reviewed, published papers produced. It is therefore not straightforward to interpret the distribution of published papers over time.

The great variety of data and approaches proposed by authors justifies this review but makes it difficult to identify the underlying common aspects of the studies. Even if studies focused on a relatively homogeneous biome, local differences could result in different observations between authors. The main differences seem to originate from the disparities in objectives, study scale, data, resolution, environmental characterization and data processing approaches. Despite such variety, it was attempted to compare these papers and to identify their common points.

It should be noted that the incidence of malaria is not equivalent to the level of transmission but also depends on the level of immunity, prevention measures and treatment. The link between LC/LU characteristics and malaria data (such as incidence and prevalence) is therefore not a direct one. Concerted and effective malaria control action may initially bring down the incidence of malaria. If the action is sustainable, the transmission could be reduced due to the human population being less infective for vectors. Anthropization is often presented as a factor favouring malaria. However, when this phenomenon is present over a length of time, medical services and associated projects to prevent and fight malaria risks are often established. Human impact can therefore become a factor which reducing malaria risk without necessarily landscape modification.

Environmental features such as land cover or land use can be identified by methods other than RS. For example, some field studies have identified environmental characteristics associated with malaria risk 44 45 . However, satellite imagery can have advantages for environmental health studies as it allows a spatially complete and almost continuous characterization of the earth’s surface at increasingly high spatial and temporal resolutions. Ideally, the results of image classification should be corroborated by field validation. A landscape epidemiology approach would greatly benefit from botanist expertise in providing better characterization of landscape patches.

Malaria cases are usually geo-localized to the localities of residence of the patients or, at more local scale, to patients' home. When identifying the environmental determinants of malaria at such a very local scale, such locations may differ from those of the point of transmission, making identified relationships between environmental features and epidemiological data inaccurate. It is necessary to consider the vector or to make an assumption of suspected transmission sites in such cases. The two reviewed studies at such a very local scale 26 34 considered the links between environmental features and epidemiological data by assuming a domiciliary or peri-domiciliary transmission. This highlights difficulties in obtaining entomological data of sufficient quantity and quality, possibly due to the cost and logistic efforts required for their collection.

The choice of satellite images must not only be driven by logistical constraints; images need to be specifically selected to suit the scale of the biological phenomenon being investigated 46 . Pope et al. 42 proposed a hierarchical approach to determine the appropriate scale at which RS predictions of mosquito production are made.