PC : Percent contribution ; PI : Permutation importance
Response of variables to suitability
Response curves indicated the relationships between environmental variables and the predicted probability of the presence of H. monstrosus, M. torquata and E. franqueti . As stated above, the spread of E. franqueti has showed to be mostly influenced by the precipitation of the driest quarter and the temperature annual range. For this species, the occurrence probability is high if the temperature annual range is less than 8°C. The occurrence probability becomes almost zero from an temperature annual range of 20°C while E. franquetihas shown to prefer area where the precipitation of driest quarter goes from 100 mm to 300 mm. H. monstrosus has also shown to have ecological preferences that are close to those of E. franquetibut has shown, furthermore, to be influenced by Precipitation of coldest quarter. The occurrence probability has shown to be high in area where precipitation of coldest quarter is greater than 200. M. torquatahas shown to be influenced the most by the Precipitation of Driest Quarter and the Precipitation of Coldest Quarter. For this species, the increase of the Precipitation of Driest quarter and the precipitation of coldest quarter increase the probability of occurrence. The most suitable area for M. torquata are the ones where Precipitation of the driest quarter is greater than 140 mm and Precipitation of the coldest quarter is greater than 900 mm (see Figure 4).
Predicting the current distribution of E. franqueti, H. monstrosus and M. torquata in DR Congo
The result showed that the area from latitude -5° to 5° and from longitude 17° to 30° are the primary potential suitable region ofE. franqueti, H. monstrosus and M. torquata in DR Congo (see Figure 5). Current distribution models show that the most favourable areas for E. franqueti are located in the territories of Faradje, Aru, Mahagi, Djugu, Irumu, Beni, Lubero, Walikale, Rutshuru, Masisi, Kalehe, Kabare, Idjwi and Walungu. The areas favourable toH. monstrosus are the same as those of E. franqueti , in addition to which the areas located in the territories of Ango, Bambesa, Polo, Niangara, Watsa, Mambasa, and Punia must be added. Unlike these two species, M. torquata seems to find comfort in the territories located in the far north of the DR Congo, particularly in the territories of Zongo, Bosobolo, Gemena, Businga, Mobayi-Mbongo, Yakoma, Bondo, Ango, Bambesa, Niangara, Dungu, Faradje, Aru and Djugu. A small area favourable to M. torquata is also located in North Kivu province, specifically in the territories of Walikale, Rutshuru and Masisi. Overall, the ecological niche of these three species is the Eastern and Northeastern regions of DR Congo.
Predicting the future distribution of E. franqueti,H. monstrosus and M. torquata in DR Congo
2050s distribution
Under future climate scenario RCP 4.5 and RCP 8.5 (Figure 6), the suitable area of the three species will be decreasing showing that their distribution will be strongly affected. These figures show that, even in both 2050s climate scenarios the suitable of the three species will decrease, the climate of RCP 8.5 will affect the most the distribution of the three species. In this scenario, the suitable area will decrease more than in the climate scenario of RCP 4.5. Overall, highly suitable area will be concentrated in the Kivu provinces. Overall, the areas favourable to these three bat species will be located in the territories of Mahagi, Djugu, Rutshuru, Masisi, Punia, Shabunda, Kelehe, Kabare, Idjwi, Walungu, Mwenga and Uvira.
2070s distribution
As stated above for the 2050s distribution of H. monstrosus ,M. torquata and E. franqueti , in all considered future climate scenario, their suitable area will decrease severally. In 2070s, the high suitable area of H. monstrosus , M. torquata andE. franqueti will be essentially located in the Kivu provinces in DR Congo (see Figure 7). With the RCP 4.5 scenario, in general, the favourable areas for these three species will be located mainly in the territories of Mahagi, Djugu, Beni, Lubero, Rutshuru, Walikale, Masisi, Kalehe, Idjwi, Kabare, Walungu, Mwenga and Uvira. However, these areas will be greatly reduced with the RCP 8.5 scenario whereby areas favourable to these species are mainly located in the territories of Beni, Lubero, Walikale, Rutshuru, Kalehe, Idjwi, Kabare, Walungu, Mwenga and Uvira.
DISCUSSION
Current distribution of bat species
According to Peterson et al . (2004), african Ebola virus reservoirs would be distributed principally in evergreen broadleaf forest (rainforest) and the main focus of the geographic distribution of the reservoirs would be in the Congo Basin. this is all the more true since our results prove that the distribution areas of the three bat species in DR Congo correspond roughly to the areas covered by Tropical and subtropical moist broadleaf forests. With the results obtained in this study it is clear that the current ecological niche of the three bat species under study is the entire part from the centre to the north of DR Congo. The most favourable environments for these species are mainly the eastern and northeastern parts, border areas with southern Sudan and Uganda, two countries that have already recorded cases of Ebola hemorrhagic fever epidemics. Bats are hypothesized to be reservoirs for filoviruses among which Ebola virus. They have been identified as an important driver of outbreaks of filovirus diseases (Fiorillo et al. , 2018). Nyakarahuka et al. (2017) found that their distribution tends to correlate with that of filovirus predicted niches. Clearly, with the results of this work, the populations that appear to be most at risk are those in the areas listed above. This would be all the more true since almost all cases of Ebola in the human population in DR Congo have been recorded in these areas (from the centre to the north of DR Congo). Pigott et al. (2014) found that the vast majority of people living in suitable area of the three bat species live in rural areas and populations of DR Congo are among those the most at risk of experiencing Ebola outbreak. In addition, almost all outbreaks of Ebola in the human population have been recorded in rural areas.
Implications of the future distributions of bat species on the Ebolavirus disease risk
In this study, we found that in the future, the areas favourable to the distribution of these species will decrease considerably in terms of available surface areas. The models of future distribution of these three bat species show that the favourable environments for these species will be mainly located in the former Kivu province and the current Ituri province. These regions represent the part of the DR Congo located in the Albertine Rift. Nyakarahuka et al. (2017) suggested the Albertine Rift of East Africa to remain under heightened surveillance especially now that oil exploration will be taking place bringing an invasion of virgin lands by humans and interaction of wildlife and humans. Based on our findings, we do agree with them because in DR Congo bats will find in the future their more suitable habitat in area located in the said rift. They noted also that in this region, several other national parks as well as several forest reserves all of which harbor various species of bats and other possible reservoirs of filoviruses. The eating habits and socioeconomic status of the populations are determinant in the emergence of Ebola viruses. Increasing human encroachment and certain cultural practices sometimes linked with poverty, such as bushmeat hunting, result in increasing exposure of humans to animals which may harbour diseases including Ebola (Daszaket al. , 2000; Wolfe et al. , 2005, 2007). Indeed, bushmeat represents a natural reserve that is exploited in the absence of financial means for purchasing animal proteins. Bushmeat is considered a daily food source in rural areas. The major challenge for accessing protein sources is of economic nature. The inaccessibility of domestic animal flesh also renders the bushmeat consumption particularly important to households because it is free and they can have all the parts of the animal (Dindé et al. , 2017). Because previous outbreaks in Central Africa have been linked to reports of bushmeat consumptions and deaths of wildlife (Leroy et al. , 2004), many hypotheses have been put forward to suggest wildlife such as bats, primates, and antelopes as possible sources of infection (Osterholmet al. , 2015; Nyakarahuka et al. , 2017).
Insights on variables contributions
Variable contribution assessment showed that precipitation variables played the most important role in the distribution models. Indeed, in this study, the Precipitation of the driest quarter (43.1% - 64.3 %) played major role in the spread of H. monstrosus, M. torquata andE. franqueti (Table1 and Figure 3). In addition, the Temperature Annual Range played also a major role in the spread of E. franqueti (12.7 %) and H. monstrosus (14.6 %) while the Precipitation of Coldest Quarter has also showed to play also a major role in the spread of M. torquata (36.4%) and H. monstrosus (10.4 %). The importance of precipitation and moderate to high temperature was highlighted by (Peterson et al. , 2004) when they modeled filovirus distribution in Africa. Rainfall is important for the obvious reason that it provides water which is very important for bats survival (Adams and Hayes, 2008; Russo et al. , 2012). Rainfall also provides for the development of fruiting trees that provide roosting areas for bats as well as food for fruit bats. Because the risk of dehydration is the greatest physiological threat to life on land, drinking water is a fundamental resource for all terrestrial animals (Knut, 1997). Due to their peculiar morphology and physiology, bats often face the risk of dehydration. Much water is lost through their body surface, especially via the respiratory system and the extensive surfaces of wing membranes (Chew and White, 1960; Thomas and Cloutier, 1992). The importance of water availability has been emphasized in studies addressing the impact of climate change on bats (Adams and Hayes, 2008) as well as in those modelling bat distribution patterns (Rainho et al. , 2010). In addition, DR Congo is endowed with many water bodies and several rainforests; this could be why the areas favourable to the distribution of bats cover a very large part of the country.
Uncertainties related to the models
The models we obtained have shown to be accurate. The Area Under the Curve (AUC) proportion was high, about 0.96. Therefore, the models are considered to perform better than random. Accurate predictive models are of particular importance for effective and adaptive management and conservation, ecological research and prediction. Thus, predictive accuracy is an important feature that is sought in species distribution modeling (Jarnevich et al. , 2015). Although these models have shown to be accurate, certain reasons may prevent complete reliance on their predictions. Variation in the output of SDMs may arise due to errors and uncertainties related to the SDMs themselves, characteristics of species life histories and future climate models (Beaumont et al. , 2008).
There is a debate criticizing the prediction results of these models. What is really modeled, the fundamental niche or the realized niche? There is ongoing debate in the ecological literature regarding exactly what these distributions represent (Warren and Seifert, 2011; McInerny and Etienne, 2013; Warren et al. , 2013). The lack of consensus on how distribution modelling relates to niche concepts is probably caused not only by inconsistency of niche definitions, but also the variability in data, methods and scale across studies (Araújo and Guisan, 2006; Soberón, 2007). Some studies have suggested that, if the realized niche is the subset of abiotic environmental space to which a species is restricted by biotic interactions (Hutchinson, 1957), then, by definition, known occurrence points used to generate distribution models represent the realized niche (Phillips et al. , 2006). Fundamental or environmental niches are only considered to be approximated by distribution models when occurrence data are drawn from a broad geographical extent (relative to the total range of the species in question) (Phillips et al. , 2006). Other studies caution against such generalizations (Elith and Leathwick, 2009), arguing that the different niches quantified using observed occurrences of species reflect an unknown conjunction of the environmental niches of the species, the biotic interactions they experience and the habitats available to species and colonized by them (Soberón, 2007).
There are uncertainties that would result from biases in the data used to develop the models. Indeed, the uncertainty associated with ecological data is great challenge in species distribution modeling. It must be accounted for if results are to be appropriately interpreted or if they are the basis of a decision-making process (Elith et al. , 2002; Barry and Elith, 2006). Several factors can affect the precision of SDMs among which we can find factors such as spatial autocorrelation, data sampling bias, varying detection probabilities (between species and observers), non-representative data prevalence, mismatched scales data misregistration or the failure to incorporate critical habitat variables in the models that damage severally the quality of data to be used in modeling (Pearce et al. , 2001; Boyce et al. , 2002; Kadmonet al. , 2003; Gu and Swihart, 2004; Barry and Elith, 2006; Johnson and Gillingham, 2008; Osborne and Leitão, 2009). In addition, the ecological characteristics of the species to be modeled can also have an effect on its SDMs’ precision (McPherson and Jetz, 2007; Franklin et al. , 2009). Imperfect detection can, for instance, mislead inference about drivers and extent of species distributions, quantifications of diversity and conclusions about environmental changes (Guillera-Arroita, 2017). This is a major weakness of these models because most surveys of natural populations, including opportunistic surveys that produce presence only observations, are prone to imperfect detection (Yoccoz et al. , 2001; Chen et al. , 2013). Furthermore, Dorazio (2012) and Lahoz-Monfort et al. (2014) have shown that failure to account for imperfect detectability in models of Presence-only data induces bias in estimates of SDMs. Observed counts of abundance are biased when individuals are imperfectly detected and this lead to biased occurrence probabilities and stated occupancy (Royleet al. , 2005). In spite of this, some researchers have recommended ignoring this challenge (Johnson and Gillingham, 2008; Banks-Leite et al. , 2014; Stephens et al. , 2015). Hence, there is no consensus about the importance of accounting for this sort of measurement error in SDMs (Guélat and Kéry, 2018).
Another limitation of SDMs is that they do not correct for sampling bias. Unrepresentative presence only locations of the region of interest induce biased SDMs’ estimates (Phillips et al. , 2009; Yackulicet al. , 2013). Accounting for the effects of geographical sampling bias in the acquisition of data can be critical to the accuracy of SDMs made using presence only data (Phillips et al. , 2009). Most of surveys of natural populations, including opportunistic surveys that produce presence only data are prone to sampling bias. For instance, if survey locations are selected based on their accessibility or convenience induce bias in datasets. Most of time, presence only data are collected based on accessibility of sampled locations. Thus, samples are located near urban settlements, rivers and roads instead of being collected systematically or randomly. Hence, their sample localities are not representative of the real range of environmental conditions in which the species of interest occurs (Reddy and Dávalos, 2003; Kadmonet al. , 2004). Such geographical sampling bias is a characteristic of most specimen locality data available from open access data portals (Hortal et al. , 2008). Failure to correct for geographical sampling bias can result in a SDMs that reflects sampling effort rather than the true distribution of a species (Phillips et al. , 2009).
Species occurrence data suffer from the disadvantage of containing the problem of spatial auto-correlation. We cannot afford to ignore this bias in the modeling of species distribution. Species distribution models implicitly assume that the geographical data points for species records are independent and the environmental layers used as hypothetical predictive variables and associated to the geographical records of species also show problems of spatial auto-correlation (Segurado et al. , 2006; Cruz-Cárdenas et al. , 2014). The spatial auto-correlation caused by colonized sites tending to cluster around initial invasion foci inflate not only model accuracy, but also the estimated explanatory power of environmental predictors (particularly when these are distally related to the requirements of the focal species) and underestimate uncertainty in model parameters (Dormann, 2007). Disregard and not avoid spatial auto-correlation has consequences such as incurring in biased model and hence inflate type I errors (Dormann, 2007; Cruz-Cárdenas et al. , 2014). Therefore, where biological or population processes induce substantial auto-correlation in the species distribution, and this is not modeled, model predictions will be inaccurate.
Another source of uncertainty in the future model is related to the climate model used. Climate models are currently the best tools we have for simulating future climate scenarios. However, variation within and among alternate climate models poses problems for end users trying to identify optimal models from which to obtain simulations (Martinez-Meyer, 2005; Perkins et al. , 2007). To date there is no clear guidance on how to select the most appropriate simulations for a given application (Beaumont et al. , 2008). Species distribution models have been very criticized for their weaknesses in predicting climate change impact on the geographic species dispersion. Among these weaknesses, we can cite: uncertainties related to the models used, difficulties in ecological interactions setting, individual idiosyncratic responses of the species to climate change, limitations of species-specific dissemination, plasticity of physiological limits and disseminating agents responses (Fandohan et al. , 2013).
CONCLUSION
Although the models used and their predictions are highly criticized, they nevertheless shed light on the management of current and future risks of Ebola epidemics by mapping the areas where most attention and prevention measures should be focused. Species distribution models are useful tools for, among other things, informing the conservation management of wildlife and their habitats under a rapidly changing climate. They can provide decision makers with information about the likely degree of change in a species climatic domain and geographic distribution. The MaxEnt model is potentially useful for forecasting the future adaptive distribution of the three bat species under climate change, and it provides important guidance for comprehensive management of the Ebolavirus risk. The results obtained in this study showed that climate change will significantly reduce the areas favourable to the distribution of these three species, reducing the risk they represent in the emergence of the Ebola epidemic. However, since not all reservoirs of the Ebolavirus are yet well identified and the other components of the emergence of the Ebola epidemic have not been taken into account in this study, it will not be possible to say that the effects of climate change will reduce the risk associated with Ebola. But it is still possible that this may be the case. It is therefore necessary for future work to take into account these components to assess the effects of climate change on the risk of Ebola occurrence in DR Congo.