Malaria is one of the leading public health problems in most of sub-Saharan Africa, particularly in Ethiopia. Almost 75% of the landmass of Ethiopia is estimated to be affected by malaria, and about 68% of the population of Ethiopia (approximately 52 million people in 2007) is at risk of malaria corresponding to the human population living in areas below 2000 m elevation[1–3]. Malaria transmission is seasonal in Ethiopia and varies across the country depending on climatic and ecological factors favourable to disease-transmitting vector and parasite development, including elevation, rainfall, and temperature[1, 3, 4]. The major malaria transmission season occurs primarily during September to December following the main rainy season, which occurs from June to September with peak precipitation in July and August[1–4]. Large-scale malaria epidemic outbreaks occur particularly in the Ethiopian highlands where transmission is unstable, immunity of the population is low, and almost all age groups of the population are at risk of severe morbidity and mortality from malaria[1, 3]. In Ethiopia epidemics occur in 5–8 year cycles; several epidemics have been reported in recent years, for example, in 2003 and 2005[1, 2]. Accordingly, there is a pressing need to develop malaria early warning to enhance public health decision making for control and prevention of malaria epidemics.

A malaria early-warning system that provides predictions of the temporal and spatial pattern of epidemics could help to control and prevent malaria epidemics. If an effective malaria early-warning system were operational, it could assist public health decision makers in prioritizing scarce resources to areas and periods most at risk[5–7]. One approach to developing a malaria early-warning system is to use statistical forecasting models based on historical malaria cases and environmental risk indicators. Thus, quantifying the links between meteorological and environmental variables and malaria is a key step towards effective malaria early warning. However, a major limiting factor is the lack of adequate environmental data covering the spatial and temporal variability of key environmental risk factors. The availability of spatially extensive and temporally consistent data from earth-observing sensors offers a source of environmental information for the development of epidemiological forecasting models. Measurements of the spatial and temporal patterns of environmental conditions that influence parasite development and the mosquito life cycle can be obtained by using remotely sensed data[8].

The influences of climatic variables on malaria risk have been well documented in previous research. Temperature, rainfall, and humidity are key determinants for malaria risk. Rainfall is associated with the presence and persistence of pools for the female anopheline mosquitoes to lay their eggs and for the completion of larval development[5, 9]. Although rainfall is necessary to produce breeding sites for the mosquito to complete its life cycle, excessive rainfall can have an adverse effect by washing away mosquito larvae[10]. Temperature is associated with the durations of larval development, mosquito survival, and parasite development[5, 9]. Humidity is associated with the persistence of the breeding sites and mosquito survival. Moreover, vegetation is not only a surrogate variable for the presence of moisture, but also provides a resting site for adult mosquitoes. Previous efforts to develop malaria early warning in sub-Saharan Africa have tried to link temporal patterns of morbidity with climatic variability[11–13]. Studies in the East African highlands have demonstrated significant lagged associations of temperature and rainfall with malaria cases[14, 15]. The current research will expand on these studies to examine the temporal variability of malaria risk in response to satellite-derived meteorological and environmental variables in the Ethiopian highlands.

In Ethiopia, only a few prior studies quantified the effects of environmental variables on malaria risk[5, 13, 16, 17]. A common limitation of these studies is their reliance on meteorological station data alone. Meteorological stations tend to be sparsely distributed in developing countries[18]; thus, it is difficult to obtain meteorological and environmental variables that span larger geographic areas. Earth observation data from space-borne sensors with high temporal resolution (daily to weekly) can provide alternative sources of information on meteorological and environmental variability[9, 18]. Despite the advances in terrestrial remote sensing over the past three decades, only recently have these data been widely applied to tackle problems in public health[9, 19–22].

Remote sensing-derived environmental variables, including indices like the normalized difference vegetation index (NDVI), the enhanced vegetation index (EVI), and land surface temperature (LST), have been used to quantify vector-borne disease risk and for malaria early warning[8, 9, 19, 23, 24]. Satellite-based precipitation measurements along with vegetation indices and LST have been shown to be useful for malaria prediction[9, 24]. However, there are other types of remotely sensed environmental variables, such as actual evapotranspiration (ETa), that have not yet been explored in the context of malaria early warning. ETa measures the flux of moisture at and below the soil surface that transpires through plants and into the atmosphere as water vapour and the moisture at the surface that directly evaporates into the atmosphere[25]. Because ETa is limited by water availability, a high ETa is often indicative of high levels of surface water and soil moisture. Thus, remotely sensed estimates of ETa may indicate suitable environmental conditions for mosquito breeding and larval growth and development.

To date, there has not been a study examining the direct application of remote sensing data to model malaria risk in Ethiopia. Thus, the research presented here will help to fill this knowledge gap by (1) quantifying the relationships between satellite-derived climatic and environmental variability and temporal patterns of malaria risk and (2) developing a model of malaria risk that incorporates both historical surveillance and environmental data. In this study the following research questions were addressed. First, are remotely sensed variables significantly associated with temporal patterns of malaria risk? Second, what are the temporal lags at which each environmental variable is associated with malaria risk? Third, does the addition of remote sensing covariates improve time series model fits compared to models based only on historical case data? A time series modelling approach, which has been widely used in previous research to quantify associations between environmental variability and disease risk, was used to answer these questions[5, 9, 14, 16, 26].

Rainfall was associated with malaria cases at a lag of one to three months in five of the study sites. This result is consistent with previous studies. A study in neighbouring Eritrea showed a correlation of satellite-based rainfall estimates with two to three months lead time and malaria cases[24]. The current findings are also consistent with a study in Ethiopia, which demonstrated that rainfall measurements from meteorological stations were significantly associated with malaria cases in lags of 10–12 weeks in highland districts and six to 10 weeks in lowland districts, respectively[5]. Another study in Ethiopia showed that there was a strong association of rainfall based on meteorological station data with malaria at a lag of two to three months[13]. A study in Rwanda similarly demonstrated that malaria incidence was significantly associated with rainfall[28]. The current research extends these previous studies by demonstrating that remotely sensed precipitation estimates can be used to model temporal patterns of malaria cases. It should be noted that the accuracy of satellite precipitation estimates are affected by sampling and spatial and temporal heterogeneity such as topography, location, and rainfall type. However, a previous study found that TRMM based precipitation estimates over the complex terrain of the Ethiopian highlands outperformed other satellite rainfall products[18].

The relationship between air temperature measured at meteorological stations and malaria risk has been highlighted in several previous studies[14, 29], including studies of climatic variability and malaria risk in Ethiopia[5, 16]. For example, one of these studies found positive significant associations of minimum temperature with malaria cases in the highland regions of Ethiopia at lags ranging from seven to 10 weeks[5]. In this study, LST was positively associated with malaria incidence at a lag of one month in six districts. LST is estimated from thermal radiance that is emitted from the land surface[30]. Satellite-based measurements of LST are sensitive to thermal characteristics of the ground, atmospheric effects of spectral radiation and bulk emissivity of the mixture of materials within the scene. As a result, temporal patterns of LST may not be tightly correlated with the near surface air temperature[31]. Despite these caveats, results indicate that LST is a potential indicator of malaria risk in the highlands of Ethiopia. In the current application, LST was suitable for prediction because despite its limitations, it captured relative values of temperature through time and among sites.

Vegetation greenness is expected to respond positively to rainfall in moisture-limited regions of East Africa; thus, it could serve as an indirect indicator of surface water and near-surface humidity. Several other studies have reported positive associations of vegetation indices and malaria risk[24, 32, 33]. For example, a study in Eritrea showed association of NDVI with malaria at a lag of four months[24]. In the present study, there was significant association of EVI and malaria risk in three districts at a lag of three months. In contrast, NDVI exhibited no significant association with malaria cases in the final models. Although the NDVI is a widely used vegetation index that can provide insights into vegetation dynamics, it loses sensitivity over denser vegetation[34–37]. The EVI offers an alternative index that retains more sensitivity than the NDVI over denser vegetation, thereby enabling it to capture more variation and change in a mature canopy[34, 35, 37]. Land cover across the Amhara region is mostly cropland and herbaceous vegetation, and a comparison of NDVI and EVI time series exhibits the expected loss of sensitivity of NDVI at higher EVI values (data not shown).

Findings from this study show novel results on the lagged positive association between ETa and malaria cases. No other study has examined the temporal relationship between ETa and malaria occurrence. Results showed a significant positive association of ETa and malaria cases in three of the 12 study sites at lags of one to three months. ETa provides information on the flux of moisture at and below soil, plants, and water bodies, and is therefore an indicator of water availability at the soil surface. Thus, ETa may offer a more proximal measurement of environmental conditions suitable for mosquito breeding and can provide a new avenue for monitoring environmental risk factors for malaria and, perhaps, other vector borne diseases.

Although this study has practical utility towards developing a malaria early-warning system, several limitations should be noted. First, the current models include only remotely sensed environmental indicators. However, there are other confounding factors that affect malaria risk such as land use/land cover, population mobility, local hydrology, socio-economic factors, and public health interventions; these processes are not captured in the current model. Additional research will be necessary to quantify these influences and incorporate them into future modelling efforts. Second, it should be noted that this type of time-series model is strongly data-driven, requiring a sufficient time series of historical data for model parameterization. Furthermore, consistent collection and timely reporting of malaria surveillance data, preferably at weekly rather than monthly intervals, would be needed to enable operational forecasting using this type of autoregressive model, which includes surveillance case data as predictor variables.

More specifically, the univariate SARIMA modelling approach (Model I) can be viewed as a type of malaria early-detection strategy, which uses temporal patterns of historical cases as a baseline to identify anomalies that may indicate the early stages of an emerging epidemic[38, 39]. As a type of early detection, Model I can use historical case data to make short-term forecasts of whether malaria cases are likely to increase or decrease in upcoming months. In contrast, disease-forecasting efforts based on environmental data alone are typically referred to as early warning[5–7]. The potential for combining early detection and early warning to leverage the advantages of both approaches has not been widely recognized or explored. The improved fit of Model II highlights the feasibility of this blended approach by demonstrating that predictions from early detection models can be improved by incorporating remotely sensed environmental variables.

Overall, temporal patterns of malaria were associated with satellite-derived meteorological and environmental variables. However, there is spatial variation in the impact of environmental drivers across the study sites. This spatial variation may reflect other factors such as differences in land use/land cover, local hydrology, topography, immunity of the population, population movement, vector control, and health access. Moreover, there was a geographical pattern of the statistically significant predictor variables in the models. For instance, malaria was associated with vegetation (EVI) in three neighbouring sites (Ankasha Guagusa, Dembecha, and Fagita Lekoma) on the southwest of the Amhara region. On the other hand, LST was found to be a statistically significant predictor of malaria risk across broader geographic areas. Thus, there is a potential for identifying geographical areas where similar environmental drivers predict malaria risk. The resulting information could be used to spatially stratify the region and develop models for malaria early warning that are tailored to specific areas.

Prediction skill varied among the districts. The results and their potential application for operational early warning system should therefore be interpreted cautiously. It should also be noted that there are limitations to the clinically diagnosed malaria data due to inconsistent registration of cases, incomplete reporting, and lack of centralized data sharing platforms. In some cases, patients visiting a health centre may come from neighbouring districts and may not be exposed to the same climate and environment. These uncertainties likely contributed to the variable levels of error in the current model predictions. To set up a robust malaria early warning system, there is a need for routine surveillance data at weekly temporal resolution and district level that is disseminated in a digital format. The current study contributes to quantifying the association of climatic and environmental variability with malaria risk by using remotely sensed variables and historical morbidity data in an effort towards the development of a malaria early-warning system in the Amhara region.

Development of operational Malaria Early Warning Systems (MEWS) has been proposed by the World Health Organization (WHO) to combat malaria epidemics particularly in climate sensitive regions[7]. In particular, an integrated malaria early-warning system can be developed using environmental monitoring and epidemic surveillance. One of the key steps to develop effective malaria early warning is to quantify the relationship between malaria cases and meteorological and environmental determinants. In this study a time series regression SARIMA modelling approach was used to quantify the lagged association of environmental variables with malaria cases and predict malaria cases among 12 districts in the Amhara Region of Ethiopia. The major finding in the current study was that there was a strong lagged association between malaria cases and satellite-derived meteorological and environmental variables across the 12 sites in the Amhara region of Ethiopia. This lead time of one to three months can be used to develop a malaria early-warning system for the region. Equally important, this study highlighted the potential for integrating modelling approaches based on historical case data (early detection) and environmental data (early warning) to enhance the effectiveness of malaria risk forecasting efforts.