More about ADCEM project

  1. Summary
  2. The objective of the ADCEM project is to adopt an integrated approach, based on statistical methodologies and deterministic modeling, applied to a set of relevant and validated geophysical and epidemiological data set at different spatial (local and regional) and temporal scales. The first step was to collect a set of relevant and validated epidemiological and geophysical data at different spatial and temporal scale, including quantitative information on the dust load. This study focuses on three African countries that are among the most affected by meningitis epidemics : Burkina Faso, Niger and Mali. This data set includes controlled times series of meningitis incidence at the district scale for these three countries, regional meteorological information from the European meteorological center (ECMWF), time-series of UV aerosol indexes from the TOMS and OMI instruments, from 1998 to 2008, ground based aerosol optical thickness from 2005 to 2009 in four stations of Burkina Faso and Mali, Niger and surface dust concentration measured from 2006 to 2010 in Mali and Niger.

    Compared to previous this study that investigated the link between meningitis epidemic and mineral dust (Thomson et al., 2006 and 2009), this work includes an evaluation of the capability of remotely-sensed dust measurements to reflect the surface dust concentration during the dry season (Deroubaix et al., in revision). Semi-quantitative aerosol indexes derived from remote-sensing, especially in the UV, are the only dust information available at the continental scale and for long-time periods (since the early 80’s) that can be analyzed in combination with long-term epidemiological data sets. They are generally tested against vertically integrated dust load (Aerosol Optical Thickness, AOT). Such a comparison performed for with UV Aerosol Index from the OMI instrument over the four selected Sahelian AOT stations reveals the capability of the UV Aerosol Index to represent correctly the daily AOT. A direct comparison between the OMI AI and surface dust concentrations measured by the Sahelian Dust Transect (SDT) shows (1) a good quantitative agreement at the beginning of the dry season where meningitis epidemics occurs and (2) a disagreement over longer period due to a shift of the seasonal cycle of the AI or the AOT compared to the surface concentration. This bias can be explained by the variability of the dust vertical distribution. A crude correction have been applied to the AI time-series reconstructed from 1998 to 2009 to produce a proxy of the surface concentration that better reflect its seasonal cycle. A deeper investigation of the link between the AOT and the surface concentration as a function of the regional meteorological context have been performed based on a statistical identification of typical "weather types" (Yahi et al., soumis). This study shows a good correlation between the AOT and the surface concentrations for the majority of the indentified weather types. Differences in the slope of the correlation sustains the idea that differences in the vertical distribution are linked to the regional meteorological conditions reflected by typical weather types. This work offers the perspective of retrieving surface dust concentration as a function of the weather types based on aerosol optical depth measured at the ground or derived from satellite observations such as the "deep-blue" AOTS.

    Several complementary statistical analysis of the whole set of epidemiological, dust and climatic parameters have been conducted to identify the key factors playing a role in the spatial and temporal dynamics of the epidemics at the local and regional scales. This work has been developed with the support of specialist studying the human and bacterial factors influencing the meningitis epidemics. A wavelet analyze reveals that the dust concentration was the most phased with the dynamic of the meningitis epidemic compared to other climatic factors (Agier et al., 2013a). A Principal Component Analysis allows to assess the combined effect of the climatic parameters and of the dust concentrations. The combined effect of air temperature and dust concentration explain 30% of the variance of the epidemiological data.

    The multi-variate analysis of the epidemiological data with climatic parameters and using a proxy of the dust surface concentration has lead to the establishment on an index quantifying the risk associated with this environmental factors. This index used the surface temperature and the dust concentration proxy. It is thus simple enough to be used in an early warning system, providing both parameters can be predicted or estimated with a good confidence level. Dust surface concentration simulated in a Chemistry and transport model could be used for such an application after a proper evaluation of the level of uncertainly acceptable for such an application.

  3. Main objectives and major results
  4. The main objective of the project was to investigate the link between the meningitis epidemics in the Sahel and environmental factors, with a focus on mineral dust. The first step has been to collect a set of relevant and validated epidemiological and geophysical data base at different spatial and temporal scale, including quantitative information on the dust load. A second step has been to evaluate the capability of remotely sensed aerosol index to properly represent the evolution of the surface dust concentrations, more representative of human exposure that vertically integrated aerosol products. A statistical analysis of the whole set of epidemiological, dust and climatic data has been conducted to identify the key environmental factors playing a role in the spatial and temporal dynamics of the epidemics at the local (district) and regional scales. From this analysis, an attempt to establish an index of the risks associated with climate and dust conditions in the dry season for meningitis epidemics over the Sahel was finally performed, that could be easily implemented in early warning systems.

    1 - Geophysical and epidemiological data set

    The epidemiological data used in the project has been established by Agier et al. (2013a) based on the weekly reports of the World Health Organization for the period 1988 to 2008. It reports the meningitis incidence (number of cases/population density, also called attack rate) at the scale of the districts in the three West African countries (Burkina Faso, Niger and Mali = 141 districts) that are among the most affected by meningitis epidemics. From this data base, the weekly average meningitis incidences of each country exhibit a clear seasonal cycle, with low incidences at the very beginning of the dry season, a maximum recorded around the weeks 13 and very low incidences during the wet season, as described by Lapeysonnie (1963) (Figure 1). Observations at the district scale reveal a large variability in the three countries. The date of beginning and maximum of the epidemics as well as their intensity vary significantly from one country to the other and from year to year. Typically, the onset and maximum of the epidemics tends to occur earlier in Burkina Faso (maximum observed during week 12) than in Niger (maximum in week 16). In terms of temporal variability, in Mali the most severe epidemics (above the epidemic thresholds used by the WHO for the epidemics surveillance; WHO, 1998) have been recorded in 1994 and 1996 but the epidemic threshold has been exceeded in Burkina Faso in 1997, 2001-2002, 2006-2007 and in Niger in 1995-1996, and 2000-2003.

    Figure 1 : Weekly incidences in Niger, Mali and Burkina Faso average from 1988 to 2008 au Niger (original data from Agier et al. 2013a).The boxes depict the median value and the percentiles 25 and 75, the black lines end-up at percentile 5 and 95, the red crosses corresponds to value higher than the percentile 95, the green line corresponds to the mean value).

    The climatic and meteorological context has been documented by using meteorological fields from the European Center from Medium range Weather Forecast (ECMWF), and in particular the surface fields (wind direction and velocity, air temperature and moisture) from the ERA-interim data base, the most widely used for climatological studies (Dee et al., 2011).

    In this study, the semi-quantitative absorption Aerosol Index (AI) derived from airborne radiance measurements in the ultraviolet (Herman et al. 1997; Torres et al. 1998, 2007) has been selected as proxy of the atmospheric dust content. This choice is motivated by the fact that the TOMS (Total Ozone Mapping Spectrometer on Earth Probe 1996-2005) and OMI (Ozone Monitoring Instrument) sensors, launched in 2005) sensors allows the establishment of a long- aerosol time series overlapping the time-series of meningitis epidemics. The TOMS AI has been proved to be highly performing over continental surfaces like desert, arid or semi-arid environments because the reflectivity of these surfaces in UV is low (Eck et al., 1987; Herman and Celarier, 1997). Like most of the satellite aerosol retrievals, the TOMS and OMI AI products have been validated by comparison to the NASA AEROsol NETwork (AERONET) sun photometer Aerosol Optical Thickness (AOT) (Holben et al., 1998) at a global scale (Hsu et al., 1999; Torres et al., 2002, 2007). The AI has been widely used in the geophysical fields, for instance to characterize the dust sources over the Sahara (Prospero et al., 2002; Washington et al., 2003; Engelstaedter et al., 2006). Using the observations from different TOMS and OMI instruments, a continuous and homogeneous time series of AI has been established for the period of the epidemiological data set (1998-2008). The TOMS AI is available with a spatial resolution of 1°×1.25° (1° ~ 100km). An instrumental drift was detected for TOMS over the years 2002 to 2004 (McPeters et al., 2007), leading to a systematic bias in the retrieved AI. The OMI AI is available with a spatial resolution of 0.25°×0.25° spatial resolution since 2005, but a deterioration of the signal has been reported by the NASA since 2009. Finally, three series of AI can be defined according to the period of acquisition: 1996-2001 (TOMS), 2002-2004 (TOMS biased) and 2005-2009 (OMI). Using as a reference the annual mean and standard deviation values averaged over the last period (2005-2009), they have been standardized to produce a homogeneous time-series of AI at the district scale.

    To test the capability of the OMI AI to represent ground aerosol measurements, it can be compared to the vertically integrated aerosol content, i.e. the AOT, and to surface dust concentrations. Since the 90's, the French component of the AERONET sunphotometers networks, PHOTONS, has deployed several instruments in West Africa. Four stations have been selected based on the temporal depth of the available data sets and the common time-period with the OMI AI. Two instruments are located in Mali (Cinzana and Agoufou), one in Burkina Faso (Ouagadougou) and one in Niger (Banizoumbou). The period covered by the selected data sets is 2005-2009, except in Ouagadougou, for which the period is reduced to 2005-2007 (January). In this study, the AOT measured at 440 nm was mainly used for comparison with the UV AI. Information of the dominance of dust aerosol on the AOT was derived from the Angstrom coefficients determined between 440 and 870 nm.

    In the framework of the AMMA project, a set of three stations dedicated to mineral dust monitoring has been deployed: the Sahelian Dust Transect (SDT). The stations are located along the main dust transport pathway towards the Atlantic Ocean (near 13°N), in Banizoumbou (Niger), Cinzana (Mali) and M’Bour (Senegal). The stations are equipped with a sunphotometer to determine the AOT. Surface concentrations of particulate matter smaller than 10µm (PM10) are monitored with a 5-min time step with a Tapered Element Oscillating Microbalance (TEOM) instrument and a PM10 inlet (Marticorena et al., 2010). This instrument is widely used for air quality monitoring and the PM10 inlet allows comparing the concentration measurements with air quality standards (i.e. de Longueville et al., 2012) or measurements in other places in the world. Local meteorological parameters (surface wind speed and direction, air temperature and humidity) are also monitored. The data set from the SDT is the only available multiannual time series of aerosol concentration in the Sahel. It is fully operational since January 2006 until now, with recovery rates of the order of 90%. Only the data of the stations of Cinzana (Mali) and Banizoumbou (Niger) are used in this study, since the Senegal is not markedly affected by meningitis epidemics.

    As an illustration of the dust parameter included in the data base, Figure 2 show the weekly average OMI AI, AOT and PM10 concentration at the two selected stations of the SDT from January 2006 to December 2008

    Figure 2 : Weekly averaged OMI AI (blue continuous line), AOT (blue dotted line) from AERONET/PHOTONS instruments and PM10 concentrations (Red line)from the SDT from January 2006 to December 2008 in a) Cinzana (Mali) and b) Banizoumbou (Niger). (Deroubaix et al., in revision)

    2 - Comparisons between remotely sensed dust content and surface concentration

    2.1 - A first comparison between the OMI AI, the AOT and the PM10 surface concentrations have been achieved to test the suitability of the OMI AAI to represent the surface dust concentration (Deroubaix et al, in revision). Since the AI is derived from radiance measurements, it is expected to be quite well correlated with AOT measured from the ground. On the four selected AERONET/PHOTONS Sahelian sites, the OMI-AI was found consistent with the AOT measured at the time of the satellite overpass, for the whole year as well as for the core of the dry season (i.e. January-March), (R from 0.6 to 0.7, slope from 1.42 in Banizoumbou to 1.98 in Cinzana). The correlation coefficients obtained at the daily and weekly time-scale (R= 0.73) allows to consider the OMI-AI as representative of the mean AOT at these time-scales. The correlation coefficient between the OMI AI and the PM10 surface concentration is much lower, especially at the hourly or daily scale (maximum 0.3 for the hour of satellite overpass). However, at the weekly scale, i.e. the time scale of the epidemiological data, the correlation increases to 0.49 in Cinzana (Mali) and 0.45 in Banizoumbou (Niger). This analysis shows that the OMI AI is not a very accurate proxy for the PM10 surface concentration. The time-series of the OMI AI, the AOT and the PM10 concentrations shows some discrepancies, especially during the wet season (Figure 2). For the 3 years, the PM10 concentration exhibit a seasonal cycle characterized by a maximum during the first trimester of the year (January to March) due to intense Saharan dust transport close to the surface (Marticorena et al. 2010). The seasonal cycles of the AOT and the OMI AI exhibit a maximum that occurs later (during the second trimester) than the concentration maximum. This is likely due to changes in the altitude of the dust layer since the OMI AI is more sensitive to elevated dust plumes (i.e. Yoshioka et al., 2004). Nevertheless in the core of the dry season, the OMI AI and the AOT exhibit a continuous increase consistently with the increase of the surface PM10 concentration. This simultaneous increase is observed at the period of the onset and increasing phase of the meningitis epidemic. When restricted to this period, the temporal variations of the OMI AI can be considered as reflecting the increase in the surface concentration.

    Figure 3 : Time series of initial AI (green line), homogenized AI (black line), PM10 surface concentration (blue line) and retrieved dust concentration proxy, DUST, (pink line) are plotted over the respective collection periods, for the ground-based weather stations of Cinzana (A) and Banizoumbou (B). AI scale is given on the right axis; PM and DUST scale on the left axis. The vertical thick black lines delimit the three observation periods for the AI (Agier et al. 2013b).

    As a conclusion from this comparison, the AI can be considered as a reasonable proxy for the dust concentration during the phase of increase of the epidemics. However, the seasonality of the atmospheric dust content depicted by the OMI AI appears as biased compared to the seasonal cycle of the dust surface concentration. However, dust concentrations from the SDT cannot be directly used for studying the link between environmental factors and the meningitis epidemics due to a too short overlap of the PM10 measurements and the epidemiological data set (3 years). To overcome this difficulty and produce a more representative proxy of the dust content, the AI time-series established for the period 1998-2008 has been "corrected" from this seasonal bias. The correction has consisted in deriving, from the OMI AI and PM10 concentration data set, an averaged and smoothed annual cycle of the ratio PM10/AI. This ratio reflects the week-specific proportion of the aerosols that are located at ground level as a function of time. The standardized OMI AI data were multiplied by the value of this annual ratio to the produce of a proxy of the surface dust concentrations, DUST, with a seasonal cycle phased on PM10 measurements (Figure 2). A cross-validation between the DUST data and the measured surface concentrations leads to a correlation of 0.6, i.e. a significant improvement in the representation of the measured concentrations compared to the initial AI. Obviously, the correction is crude, since it is based on two measurement stations over the whole region and 3 years of observations only. It assumes a constant seasonal cycle of the ration of surface to vertically integrated dust content, which may not be true over long periods of time. This hypothesis could have been tested using the PM10 measurements acquired after 2008, but the OMI AI suffer from a bias since 2009, preventing further investigations of the interannual variability of the ratio PM10/AI. In such conditions, this correction was the only possibility to obtain a more realistic description of the temporal variability of the surface dust concentration at the seasonal scale. This proxy is thus more appropriate for health impact studies that the initial AI data set. The relationship between the vertically integrated aerosols content (AOT) and the surface PM10 concentration must be further investigated to consider a more physical retrieval. This is the main objective of the following section.

    2.2 - A deeper analysis of the link between the AOT and the surface concentration has been achieved by accounting for the influence of the regional meteorological context on this link (Yahi et al., soumis). To account for the regional meteorological context, the tridimensional meteorological fields from ECMWF has been analyzed to identify typical "weather types" during the dry season (October to May 2006 to 2010). This identification is based on a non-supervised neuronal classification method (Yahi et al., 2010). The method is applied to the vertical profile of wind velocity and direction and temperature extracted over the pixels containing the station, the 8 surrounding pixels and 2 additional pixels aligned on the N-NE, N-NW, S-SW and S-SE axes, to reach a regional extend.

    Figure 4 : Mean surface temperature and wind fields at the station of Banizoumbou (Niger) (black dot) for 3 of the identified weather types, on corresponding to a southwestern flux (WT 1) and the two other to typical Harmattan conditions (WT4 and 5). For each weather type, the wind direction and temperature in Banizoumbou are given in red.

    Five weather types are finally identified in Cinzana and Banizoumbou, but appear very similar at the two stations (example on figure 4). Three of these weather types are clearly associated to the Harmattan regime while the two others correspond to incursions of the monsoon flow during the phases of installation or retreat of the Monsoon. Despite a typical succession of these weather types during the dry season, their relative occurrence the different weather types also exhibit a strong interannual variability (Figure 5).

    Figure 5 : Time series for the five years of study (2006-2010) of the monthly occurrence of the different weather types at Banizoumbou (Niger). (Yahi et al., soumis).

    The relationship between the AOT and the PM10 concentrations have been examined separately for each weather type (Figure 5). For most of the weather types, the correlation between the PM10 concentrations and the AOT are higher than that obtained for the whole data set. In Banizoumbou, the correlation coefficients are high (ranging from 0, 58 to 0,79) for the Weather Types related to the Harmattan regime and lower for those associated with monsoon flow (R2 0.25 and 0.51). The winds associated with the weather types related to monsoon flow (WTB1 and WTB2) are lower and the concentration in particular matter lesser. For these two weather types, the PM10 distribution is biased toward low values (less than the median value)while the distribution of the AOT is slightly biased toward high values (higher than the median). This suggests a loss of correlation between the PM10 concentrations at the ground level and the columnar atmospheric dust load. In this case the relationship between AOD and PM10 is not expected to be linear and the correlation does not present any statistical significance (R²<0.4). The same general comments apply to the results obtained at Cinzana, except that one of the Harmattan weather type, which represents situations with an important percentage of small PM10 values shows a very small linear correlation coefficient. The correlations for the weather types related to the monsoon flow are also higher than in Banizoumbou. These results suggest that, in most of the cases, a specific and significant relationship can be found between the surface PM10 concentrations and the AOTs when accounting for differences in the regional meteorological context described by the weather type classification. The relationships obtained for the different weather types differ not only in terms of correlation but also in terms of the slope of the linear regression (see Figure 6). These differences imply that for a given measured AOD, the corresponding surface PM10 concentration can vary by up to a factor two depending on the meteorological conditions.

    Figure 6 : Hourly mean PM10 concentrations as a function of AOT for the whole data set and for the each weather type identified in Banizoumbou (Niger). (Yahi et al., soumis).

    These changes in the relationship between the PM10 surface concentration and the AOT can reflect differences in the depth of the dust layer or the dust distribution inside the layer. During the dry season, aerosol vertical profiles derived from lidar measurements indicate the persistence of low level transport for desert dust (Léon et al., 2009; Cavalieri et al., 2010). The temporal succession of the different weather types during the dry suggests a progressive increase of the dust concentration at very low altitude from the beginning to the core of the dry season. This can be due either to a decrease in the thickness of the dust layer or to an increase in the dust load. Both phenomena can be related to an intensification of the eastern Low Level Jet (LLJ), consistently with the vertical profiles of wind velocity. Both at the end and the beginning of the dry season, the surface aerosol layer appears thicker, as sustained by the wind profiles associated with these weather types, which explain the lower slope of the regression for these weather types.

    These results show that the surface concentrations cannot be simply retrieved from the measured AOTs without accounting for the local and regional meteorological context. A statistical inversion of the PM10 surface concentration based on the AOT and the local meteorology as a function of the Weather types allows a very good retrieval of the measured PM10 concentrations. This work open perspectives for the retrieval of long-time series of PM10 concentrations over the Sahel, based on AOT measured from the ground measurements or derived from satellite observations, such as the recent "deep-blue" aerosol products from the MODIS-Aqua instruments (2003-2012) extended to the measurements from the Seawifs instrument (1997-2012).

    3 - Analysis of the link between meningitis epidemics, climate and dust conditions

    3.1 - Wavelet analysis: The AI time-series produced from 1998 to 2008 and converted into a proxy of surface dust concentrations, referred as DUST, the meteorological parameters from the ERAi data base and the epidemiological data set have been used for a wavelet analysis. This analysis allowed to quantify the coherence of the temporal variations of the selected parameters and to detect time-lags between the parameters that were found consistent (Agier et al., 2013b). Such an analysis was conducted for the first time at the district scale in Niger with a weekly temporal resolution over one decade. Results highlighted the special case of dust in comparison to wind, humidity or temperature: a strong similarity between districts is noticed in the evolution of the time-lags between the seasonal component of dust and meningitis incidence. Despite this consistency across districts, the time-lag between the seasonal component of dust and meningitis evolves across the study period: this is likely due to another time-varying factor, which interaction with dust would trigger an increase in meningitis cases. A specific analysis of the results obtained using either the initial standardized AI data set or the DUST proxy reinforce the confidence in this proxy. The pattern of similarity among districts in the evolution of the phase difference observed for the AI versus meningitis is reproduced for DUST versus meningitis; the average phase difference becomes positive, but its variability is maintained. This indicates that the applied correction was not too strong and did not constrain all district-level annual curves to peak simultaneously. The average phase difference between dust concentration proxy and meningitis is 11 days (1.55 week), which is consistent with the time for dust to damage the respiratory tract along with the incubation period of the disease (1–14 days) (Stephens et al., 2007). On the opposite meningitis epidemics not phased with other climatic parameters (Figure 7). This study sustain the fact that dust acts as a trigger of the meningitis epidemics: when the levels of dust in the low layers of the atmosphere start increasing (decreasing), meningitis incidence will start increasing (decreasing) approximately 10 days later. This hypothesis is also sustained by Martiny & Chiapello (2013).

    These results, together with the assumption of dust damaging the pharyngeal mucosa and easing bacterial invasion, reinforce the hypothesis that mineral dust act as a forcing on meningitis seasonality and prompts considering dust as a major predictor of the timing of meningitis epidemics. A limitation of this approach is that it does not allows to account for the effect of dust and climate variables on the amplitude of the epidemics, and in particular the possible effects of the local intensity and persistence of dust events. In addition, this analysis cannot account for combined effects of the different parameters. Finally, the use of continuous time-series mainly highlights the strong seasonality of the local climatic conditions. The following step was thus to develop a multi-variate analysis focused on the dry season.

    Figure 7: Phase difference (in weeks) between the meningitis incidence and : (a) the standardized AI, (b) the DUST concentration proxy derived from the standardized AI, (c) the surface wind direction, (d) air humidity, (e) air temperature, (f) surface wind speed; Values on the figures correspond to the average phase difference over the whole period (Results from Agier et al., 2013b).

    3.1 Multiparameter statistical analysis

    3.1 - A first step in this analysis has been to classify independently the spatial distribution of the epidemiological and the geophysical parameters. The objective of this preliminary analysis is to prepare and to facilitate the interpretation of the statistical analysis combining all parameters. The classification, or spatial clustering, was based on an automatic ascending classification method. For meningitis incidences (Figure 8), the identified classes are not spatially continuous, since the incidence is impacted by a contrasted distribution of population density. The northern class corresponds to a vast area with low incidences and low population density. On the opposite the southernmost class corresponds to a region with high meningitis incidence and epidemics occurring earlier than in the other classes). The spatial structure of the meningitis incidence differs from those indentified for the climatic parameters. The climatic parameters exhibit a meridian pattern while the distribution of the dust concentration proxy classed has both a zonal and a meridian structure. The annual cycle of the climatic parameters are very similar in all the districts and differs only by their amplitude.

    Figure 8: Spatial clustering of the weekly meningitis incidences in the 141 districts of Burkina-Faso, Mali and Niger from 1998 to 2008. (Deroubaix, Thesis manuscript, in prep.).

    3.2 The statistical method selected for this multi-parameter analysis is the Principal Component Analysis, also called Empirical Orthogonal Function analysis (EOF), which enables to highlight the similarities or the differences between the characteristics of the considered individuals. This technique has been applied to different combinations of data (including both the AI and the DUST proxy, or the DUST proxy only) and period (whole year compared to dry season only) at three different scales (country, regions, districts). We summarize here the final results obtained using the DUST proxy during the dry season when meningitis epidemics take place.

    The analysis for the dry seasons for all years and districts (Table 1) reveals a first component characterized by the opposition of the wind speed with the wind direction, relative humidity and the temperature. The second component is linked to the dust concentration (DUST) and the third component is mainly related to the meningitis incidence. The first component (CP1) explains 41% of the total variance. It reflects the opposition between situations of strong Harmattan flux (high winds from the North) and situations with low winds and high temperature that may result from the influence of the Saharan Heat Low. The second component (CP2) explains 22% of the total variance of the database and corresponds to the presence of dust which is correlated with temperature and anti-correlated with relative humidity. The third (CP3) explains 15% of the dataset variability.

    A second analysis was performed by selecting only epidemic years, i.e. years and district for which the WHO thresholds were exceeded (Table 1). In this case, the meningitis incidence does not participate anymore to the third component (CP3) but is correlated with CP1 and CP2. The correlation of the climatic parameters and of the DUST parameter with the three components is almost unchanged. The projection of the different parameters on the principal plane defined by these two components shows the co-variation of the incidence and climate variables on one hand (CP1) and the dust on the other hand (CP2). This analysis was then refined by separating the data for which different epidemic thresholds have been reached and by introducing cumulated dust concentration with variable time-lags. For the epidemic years, the links between the incidence and the dust proxy appears as reinforced when averaging the dust concentrations over 3 to 5 weeks with a 1-week time-lag.

     


    CP1

    CP2

    CP3


    Correlation

    Correlation

    Correlation

    No criteria

    Incidence

    0,34

    0,43

    -0,83

    Temperature

    0,68

    0,53

    0,23

    Humidity

    0,73

    -0,43

    0

    Wind speed

    -0,83

    -0,1

    -0,08

    Wind dir.

    0,83

    -0,1

    0,16

    DUST

    -0,25

    0,76

    0,32

    Epidemic threshold

    Incidence

    0,59

    0,56

    -0,03

    Temperature

    0,78

    0,35

    0,37

    Humidity

    0,66

    -0,62

    -0,34

    Wind speed

    -0,84

    0,12

    -0,34

    Wind dir.

    0,83

    -0,36

    -0,25

    DUST

    0,27

    0,75

    0.49

    Table 1 : Correlation of the selected parameters at the weekly scale for the whole districts for the 3 first identified principal component for the whole period ("no criteria") or when selecting epidemic years ("epidemic threshold) .

    The results of the EOF suggest that the variability of the incidence during the epidemic phase can be explained by the atmospheric conditions such as the increase of the temperature associated with the Saharan Heat low (Lavaysse et al., 2009) as Saharan dust transport events. The independence of CP1 and CP2 suggest a different influence of the temperature and of the dust. Similar level of correlation between the meningitis incidence and the temperature (R= 0.62) and a dust index derived from horizontal visibility (R=0.48) were evidenced on a 3 years' time-series obtained in the town of Zaria, in Nigeria (100km south of our area of study) (Greenwood, 1984). At first sight, these results can appear inconsistent with Martiny and Chiapello (2013) who recently show that low humidity is a necessary condition in the beginning and the development phases of the epidemics. In our analysis, the relative humidity is highly correlated to the temperature and associated to CP1, the influence of humidity may is thus accounted for in CP1.

    4 - Determination of an index of the risk due to dust and climatic conditions.

    Based on the results from part 3.2, it appears that meningitis epidemics can be related to atmospheric conditions. A stepwise statistical regression technique has been used to detect the most relevant explanatory variables and to try and produce a statistical model to predict the meningitis incidence during the epidemics. All parameters are tested step by step and their level of correlation with the incidence or their redundancy with other parameters is evaluated to retain the most relevant parameters. This method is applied to each district considered as epidemics (incidence >10 or 20, i.e. 10 or 20 cases per 100 inhabitant). The highest correlations are obtained with the temperature and the dust concentration proxy, whatever the spatial scale (i.e. district or region) and the epidemic threshold (Table 2). A statistical model has thus been established to model the incidence as a function of these two parameters.

     

     

    Region-year (Incidence > 10)

    Region-year (Incidence > 20)


    N

    % in the model

    R

    N

    % in the model

    R

    Temperature

    13

    0,76

    0,47

    8

    0,73

    0,45

    Humidity

    3

    0,18

    -0,05

    1

    0,09

    -0,07

    Wind speed

    5

    0,29

    -0,02

    4

    0,36

    -0,02

    Wind dir.

    5

    0,29

    0,13

    3

    0,27

    0,17

    Dust

    10

    0,59

    0,25

    7

    0,64

    0,32

     

    Districts-year (Incidence > 10)

    Districts-year (Incidence > 20)


    N

    % in the model

    R

    N

    % in the model

    R

    Temperature

    53

    0,72

    0,42

    33

    0,69

    0,40

    Humidité

    9

    0,12

    -0,05

    5

    0,10

    -0,05

    Wind speed

    10

    0,14

    -0,04

    3

    0,06

    -0,07

    Wind dir.

    12

    0,16

    0,05

    10

    0,21

    0,06

    Dust

    42

    0,57

    0,25

    34

    0,71

    0,34

    Table 2 : Number (N) of district, percentage (%) of district for which a variable (temperature, wind speed and direction, dust) is included in the model, and average correlation coefficient (R) of the multiple linear regression between the incidence and the environmental for the regions and years or the districts and years that reached the selected incidence thresholds (10 or 20).

    The capability of the model to reproduce the observed incidence has been tested by separating the data set into a learning set and a validation set. The modeled incidences are significantly correlated with the measured ones, with correlation coefficient ranging from 0.48 to 0.61. One third of the weekly incidence is thus explained by the model, with a dominant influence of the temperature and the dust concentration. This correlation is quite insensitive to the spatial scale (district to region) and to the epidemic threshold. It increases when the dust proxy is averaged over a period of 3 to 5 weeks.

    The capability of the model to reproduce the timing of the epidemics (onset date, date of maximum) has been tested at the scale of the districts, the regions and the country. The level of agreement with the observed incidence is higher at the regional or national scales than at the district scale. This shows the relevance of the model but also reflect the influence of other factors at the district scale, as already highlighted by the wavelet analysis.

    The model also allows reproducing a spatial pattern of the incidence that is very similar to the observed one and in particular to locate the regions with high epidemics incidence. Differences in the timing of the epidemics, and in particular the earlier occurrence of the epidemics in Burkina-Faso compared to Mali and Niger, are also reproduced. This highlights the importance of the climatic conditions on the temporal and spatial pattern of the meningitis epidemics.

  5. Conclusion and prospects
  6. Several conclusions can be drawn from this project both on methodological aspects and on the link between meningitis epidemics, mineral dust and climatic conditions in the Sahel.

    - From a methodological point of view, important efforts have been dedicated in the project to the investigation on the capability of satellite derived product to represent the aerosols surface concentration over the Sahel. This was possible because of the unique surface concentration time series acquired by the stations of the Sahelian Dust Transect deployed since 2006 in the Framework of the AMMA program. Our results point out the fact that semi-quantitative aerosols index (AI) and the aerosols optical thickness (AOT) can be used only with some precaution or restriction for health impact studies. The relationship between the surface concentration and the vertically integrated dust content was found dependant on the regional meteorological context that influences the vertical distribution of the dust layers. As a result, the surface dust concentration cannot be retrieved unambiguously from the AOT or the OMI AI with a good precision, except for some selected periods. The surface dust concentrations and the OMI AI exhibit consistent increases at the beginning of the dry season. The AI can thus be used as an indicator of the increase of the surface concentration during the increasing phase of the epidemics. Differences in the seasonal cycle of the AI and AOT have been evidenced that also prevent the direct use of these parameters to investigate the link between the dust surface content and the meningitis epidemics. A crude but pragmatic approach have been develop to correct this seasonal bias based on the average seasonal ratio between the surface concentration measured at two stations of the Sahelian Dust Transect (Mali, Niger) and the OMI AI for the period 2006-2008. A dust concentration proxy time-series has thus been established from an homogenized AI time series derived from the TOMS and OMI instruments. This time series allows to investigate the possible link between the dust and climatic conditions over a ten year period (1998-2008) including two major meningitis epidemics.

    - The link between climatic and dust conditions and the meningitis epidemics have been investigated using several different but complementary approaches. Compared to previous studies on the subject, in this work these links have been investigated at various spatial scales, from the country to the district scale, using a specific proxy of the dust surface concentration. A first approach, based on wavelet analyses, showed that this dust proxy was the parameter that was the most coherent with the meningitis incidences at the seasonal scale in all the documented districts. The average difference of phase between the dust and meningitis signal is 11 days. This is consistent with the time for dust to damage the respiratory tract along with the incubation period of the disease (1–14 days) (Stephens et al., 2007). A multi-variable statistical approach has been developed to investigate the combined effect of the dust and the climatic parameters on the dynamic and intensity of the epidemics. The two parameters that have been found the most correlated to the incidence are, by order of importance, the surface air temperature and the dust concentration. For dust concentration, the correlation is increased when it is averaged over 3-5 weeks and including a 1 week time lag. This reinforces the idea of a triggering effect of the dust concentration on the epidemics, and suggest an impact of dust persistency on the epidemic intensity which is in agreement with the recent study by Martiny & Chiapelli (2013).

    - A statistical model of the meningitis epidemics have been established based on the air temperature and the reconstructed dust concentration. These parameters are the input parameters of a statistical model that allows evaluating the risk of meningitis epidemics due to climatic factors and to high dust loads. This model explains 30% of the variance of the meningitis incidence, which is quite high compared to other pathologies considered as sensitive to climate. This means that climate and dust are predominant factors for meningitis , but highlights, once again, that other factors that the climatic conditions strongly influence the meningitis epidemics in the Sahel. This can be compared to the climatic or synoptic meteorological indicators established in previous studies. Climate indexes have been correlated with the intensities of the epidemics explaining 25% of the year to year variability in Niger (Yaka et al., 2008). Our model provide a slightly higher level of understanding and does not concerns Niger only. In particular, it reproduce differences in the intensity and timing of the epidemics from on country to the other. The Harmattan winds have been proved to play on the timing of the meningitis epidemics seen at the national large scale (Sultan et al., 2005). In our statistical model, the first component is clearly linked to the Harmattan regime, and characterized by an opposition of the wind intensity with the wind direction, relative humidity and the temperature. Similarly, the highest dust load are associated with the Harmattan regime (Yahi et al. soumis). As a result, in the developed model, the temperature and dust concentration bring the same level of information than a regional indicator of the Harmattan regime. The combined effect of temperature and dust have been found to correlate with meningitis incidences recorded over 3 years in the town of Zaria, in Nigeria (100km south of our area of study) (Greenwood, 1984). In this study, a dust index was derived based on synoptic records of horizontal visibility and a lag-time between dust and meningitis occurrence was introduced. Our results confirm the conclusion from this local study but extends them to a 10 years period and 141 districts of Burkina Faso, Mali and Niger.

    In terms of prospect, this work allows proposing some recommendations for a deeper investigation of the link between dust, climate and meningitis epidemics in the Sahel, but it also open some perspective of applications in the short term.

    The simplicity of the proposed statistical model makes it easily implemented in any environment or institution as a help for early warning systems. Its application requires an estimation of the dust surface concentrations in almost real-time. Such information could be derived either from a regional dust forecast model or from a retrieval based on satellite AOT and meteorological conditions. Additional work would be required first to evaluate the respective capability of these two approaches to correctly reproduce the dust surface concentration with a relevant precision and second to estimate the technical aspect and feasibility of such approaches.

    The results from this study clearly highlight the need to "correct" satellite aerosols products for health applications. During the project, a corrected AI time series have been produced however, a more geophysical approach can be developed. Indeed, the analysis of the link between the AOT and the surface concentration as a function of the weather type open the perspective of a surface concentration retrieval allowing to account for the interannual variability of the seasonal cycle of the atmospheric dust content and vertical distribution. This analysis was based on ground-based AOT measurements but could be extended to satellite derived Aerosol index or AOTs. In particular the "Deep-blue" algorithm used to derived the AOT over land in arid regions with the MODIS measurements has recently been applied to the data from the SeaWifs instrument, available since 1998. The use of this data set in a retrieval based on a statistical analysis of weather types should lead to the most realistic re-construction of the dust surface concentration in the Sahel. As a result, if the link between dust and meningitis epidemic is stronger than evidenced in this project, it could be properly quantified by the use of such a dust time-series. This approach can also be applied to present satellite data set to produce regional estimation of the dust surface concentration.

    Finally, the results from this project and related work (De Longueville et al, 2012) point out the importance of the high dust concentrations measured in the Sahel and reinforce the interest of a warning system for this public health problem in this region.

    Publications of the project or related to the project

    - Agier, L., Broutin, H., Bertherat, E., Djingarey, M.H., Lingani, C., Perea, W., and Hugonnet, S. , Timely detection of bacterial meningitis epidemics at district level: a study in three countries of the African Meningitis Belt. In Transaction Royal Society Tropical Medicine Hygiene (pp. 30-36): Oxford University Press,.DOI: 107(1): 30-36 doi:10.1093/trstmh/trs010, 2013a

    - Agier L, Deroubaix A, Martiny N, Yaka P, Djibo A, Broutin H., Seasonality of meningitis in Africa and climate forcing: aerosols stand out, J. R. Soc. Interface, 10: 20120814, http://dx.doi.org/10.1098/rsif.2012.0814, 2013b.

    - A. Deroubaix, N. Martiny, I. Chiapello , B. Marticorena, Suitability of OMI aerosol index to reflect surface conditions for studying the impact of mineral dust on health in the Sahel: Preliminary application to meningitis epidemics, Int. J. Remote. Sens., in revision.

    - de Longueville, F., Hountondji, Y-C., Ozer P., Marticorena, B., Chatenet B., and S. Henry, Saharan Dust Impacts on Air Quality: What Are the Potential Health Risks In West Africa? Human and Ecological Risk Assessment : an International Journal, DOI:10.1080/10807039.2012.716684, 2012.

    - Martiny, N. and I. Chiapello, Assessments for the impact of mineral dust on the meningitis incidence in West Africa, Atmospheric Environment, sous presse.

    - H. Yahi, B. Marticorena, S. Thiria, B. Chatenet, C. Schmechtig, J.L. Rajot and M. Crepon, Influence of the weather types identified in the dry season in the Sahel on the mineral dust atmospheric content, J. Geophys. Res., soumis.