2.1 Study area

The study area included the West African countries located between 18°E–15°W and 4°N–16°N (Figure 1). West Africa has three major climatic zones: the Guinean zone which extends approximately between 6° and 8°N, the Sudanian zone approximately between 8° and 12°N and the Sahelian zone between 12° and 16°N (Kouassi et al., 2010).

One of the main assumptions of correlative species distribution modelling applied to invasive species is the niche conservatism of the native environmental niche into the invaded area (Pearman et al., 2008). Although usually tested using ordination techniques, here, due to lack of a sufficient number of occurrences in the native area with temporal consistency with the occurrences in West Africa, we proposed a qualitative analysis based on the similarity of the Koppen–Geiger climatic classification for the native (South East Asia, Appendix S1) and the invaded areas (South America and sub-Saharan Africa) to test this assumption (Romero et al., 2021).

2.2 Data

2.3 Predictors

A set of 13 relevant variables (Table 2) related to the environment of the tick and its host were selected. The biotic predictor selected was cattle density, as R. microplus was described as a specific and the most significant parasite for cattle (Bellgard et al., 2012). The abiotic predictors selected were temperature, rainfall, vegetation index and land cover. Vegetation (pasture for herds) was considered as dependent on climatic factors (rainfall and temperature). It plays a key role in the life of the tick host (cattle). The availability of pasture and water are determinants for the presence of cattle in an area.

Abiotic factors (temperature, humidity, vegetation index and land cover) are crucial for the survival of the tick during its off-host stages (Apanaskevich & Oliver, 2014).

The rainfall and the temperature data used in this study were climate data set for the earth land surface areas at a spatial resolution of 1 km. The rainfall data and the temperature data were annual mean patterns for the period of 1979–2013 (http://chelsa-climate.org/bioclim/; Table 2).

The cattle density dataset contains the global distribution of cattle in 2010 expressed in the total number of cattle per pixel (5 min of arc) according to the Gridded Livestock of the World database (GLW 3) (Table 2).

The normalised difference vegetation index (NDVI) used here is the calculation of the mean of the NDVI raster of the 12 months of 2015 downloaded from the MODIS NDVI (https://lpdaac.usgs.gov/products/mod13a3v006/) with a resolution of 1 km (Table 2).

The annual (2015) land cover variables (Cropland, Mosaic tree and shrub, Mosaic herbaceous cover, Shrubland, Grassland, Sparse vegetation, Urban areas, Bares areas and Water bodies) were downloaded from https://www.esa-landcover-cci.org, with a resolution of 1 km (Table 2).

2.4 Variable pre-processing

A collinearity test on the predictors was used to check for linear associations between two or more explanatory (predictors) variables. Two variables are perfectly collinear if there is an exact linear relationship between them. For the diagnosis of collinearity, we used the variance inflation factor (VIF) method to see which variables are really necessary for the computation of the different models (Akinwande et al., 2015). A VIF in the range of 5–10 shows a high correlation which can be problematic. If the VIF rises above 10, it can be argued that the regression coefficients are misestimated because of multi-collinearity, which must be treated appropriately (Akinwande et al., 2015). Variables with VIF > 5 will not be used further. The correlation of the predictors with the presence of the tick was highlighted by the multivariate regression preceding the VIF analysis and included in the collinearity analysis. This helped in the selection of relevant predictors for the modelling (Table 2). Variables with many missing values were removed from the list of predictors. The number of presence and absence data was balanced accordingly. In our data, the number of absence points was higher than the number of presence points, so we randomly sampled 316 points (total number of presence points) from all absence points. In addition, all predictors’ rasters resolutions were harmonised at the spatial resolution of 10 km.

2.5 Modelling

Six models were used in this study. The maximum entropy model (MaxEnt) was used alone with only R. microplus presence data, whereas five other models, that is generalised linear model (GLM), generalised additive model (GAM), random forest (RF) model, boosted regression tree (BRT) and support vector machine model (SVM), were used to produce an ensemble model using R. microplus presence and absence data via the R package ‘sdm’ (Naimi & Araújo, 2016).

In addition, a sensitivity analysis was carried out to assess the contribution of each predictor used in models through the variables importance analysis in each modelling technique. The variable importance was addressed in the MaxEnt model by the permutation importance values.

2.6 Cross-validation and bootstrapping

Cross-validation is a resampling procedure used to evaluate machine learning models on a limited data sample. The procedure has a single parameter called k that refers to the number of groups that a given data sample is to be split into. The model will train k models, each with k – 1 folds, and the kth (last fold) will be used to test it (Hijmans, 2012). Overall accuracy is calculated by the average of the accuracy from all k models.

Bootstrapping replicates a sampling method with replacement, each time a sample of equal size to the original data is drawn and used as training data. Observations that are not selected are used for evaluation in each round (Naimi & Araújo, 2016).

To evaluate the performance of the models, three parameters were considered: the area under the receiver operating curve (AUC), the true skill statistics (TSS) and the Boyce index. We use AUC, TSS and Boyce index for the MaxEnt model and AUC and TSS for the ensemble modelling.

A model with no predictive power would have an AUC of 0.5 (e.g. a diagonal line), whereas a perfect model would correspond to an AUC of 1 (Allouche et al., 2006; Boyce et al., 2002).

A TSS value of +1 indicates perfect agreement and values of 0 or less indicate performance no better than random (Allouche et al., 2006).

Boyce's index is an evaluation tool adapted for models predicting species distribution based on presence-only data (Di Cola et al., 2017). The Boyce index is ranged from −1 to +1 and positive values indicate consistent predictions of species presences. Values close to 0 mean that the pattern is no different from a random pattern and negative values indicate counter-predictions, that is prediction of areas of presence more frequent as it is highly relevant to the species (Boyce et al., 2002; Hirzel et al., 2006).

2.7 Software

Analyses were carried out with R software version 4.0.5 (R Core Team, 2021). The following R packages were used: dismo, ENMeval and sdm (Hijmans & Elith, 2013). The MaxEnt model was implemented using R packages rJava and dismo. The ensemble modelling with the five models (GLM, GAM, RF, BRT and SVM) was implemented using the R package ‘sdm’ (Naimi & Araújo, 2016).

The cartographic outputs were computed using the software QGIS version 3.18.1 (Quantum GIS Development Team, 2021).

3.1 Variables and models selected

The collinearity analysis revealed that only one variable (urban area) had a variance inflation factor above 10 (i.e. 56.54). This variable that was showing a high multi-collinearity furthermore did not significantly correlate with the presence of R. microplus.

Seven out of the 13 predictors were significantly (p-value < .05) correlated with the presence of R. microplus through multivariate analysis preceding the VIF. These predictors are Rainfall, Temperature, NDVI, Cattle density, Cropland, Grassland and Shrubland.

3.2 Characteristics of the predictors according to presence or absence of R. microplus in the study area

The investigation to check if there was difference between the various predictors according to areas where the tick was found or not revealed that the rainfall was significantly (p < .001) higher in areas where R. microplus was present (annual rainfall average: 1202.23 ± 252.24 mm) than areas where it was absent (annual rainfall average: 881.62 ± 231.71 mm). The tick was found in areas with higher rainfall than where it was absent (Appendix S2). Temperature was significantly (p < .001) higher in areas where R. microplus was absent (annual temperature average: 27.99 ± 0.99°C) than areas where it was present (annual temperature average: 26.44 ± 0.77°C). Areas where the tick was found are cooler than places where it was absent (Appendix S2). On the other hand, cattle density was significantly (p < .001) higher in areas where R. microplus was absent (average head/km²: 229.32 ± 196.48) than in area where it was present (average head/km²: 99.76 ± 118.02). The hosts were more concentrated in areas where the tick was absent (Appendix S2). The NDVI was significantly (p < .001) higher in areas where R. microplus is present (monthly NDVI average: 5681.17 × 10⁻⁴ ± 1024.29 ± 10⁻⁴) than in area where it is absent (monthly NDVI average: 3837.59 × 10⁻⁴). The NDVI was better in places the tick was found than where it was absent (Appendix S2). The cropland was significantly (p < .001) more important in areas where R. microplus was absent (annual cropland classes unit average: 0.73 ± 0.35) than in area where it was present (annual cropland classes unit average: 0.52 ± 0.41). The tick was found more frequently on cattle living in uncropped areas than cropped ones. The proportion of grassland was significantly higher (p < .001) in areas where R. microplus was absent (annual grassland classes unit average: 0.05 ± 0.18) than in area where it was present (annual grassland classes unit average: 0.00 ± 0.01). The tick was found more frequently on cattle living in an environment where there was less grassland (Appendix S2).

And lastly, shrublands were found more (p < .001) in areas where R. microplus is present (annual shrubland classes unit average: 0.09 ± 0.18) than areas where it was absent (annual shrubland classes unit average: 0.06 ± 0.01). The tick was found more frequently on cattle living in an environment where there is shrubland (Appendix S2).

3.3 Variable importance

The analysis of the contribution of the different predictors to the MaxEnt model highlighted that temperature and rainfall were the main variables. Temperature contributed to more than 39% and rainfall contributed more than 19% of the model accuracy. Shrubland (1.8%) and grassland (1.8%) contributed less to the model predictions (Appendix S3A). The relative variable importance based on the Pearson correlation metric in the ensemble model revealed that NDVI played an important role in the model. This variable was the most important in the five models that make up our ensemble model set. NDVI was the most important predictor in the ensemble model and the cropland and shrubland were the less important ones (Appendix S3B). In the GLM model, NDVI (88%) was the most important predictor and cattle density contributed the least to the model with 0.1% (Appendix S3C). In the GAM model, it was the rainfall that contributed the most with 39.3% and the cropland contribute the least with 2.4% (Appendix S3D). In the RF model, NDVI was the most important contributor among the predictors (20.3%) and grassland has the lowest contribution (0.4%) (Appendix S3E). The NDVI also was the largest contributor to the BRT model (62.6%) and grassland did not contribute (0%) to the model (Appendix S3F). In the fifth model (SVM), the most important predictor was NDVI (18.5%) and the least important was Grassland (1.9%) (Appendix S3G).

3.4 Habitat suitability for Rhipicephalus microplus in West Africa

The maps revealed high suitability of the coastal areas of West Africa. On the other hand, the tick R. microplus seems very poorly adapted to the environment of the Sahelian region, a desert area (Figure 2). However, the southern regions of some Sahelian countries, contiguous to the coastal areas, are also suitable for R. microplus. Some differences can be noted in these outputs according to the type of model implemented. The MaxEnt model predicts differently the presence of the tick in the Sahelian region where it is not expected at all or the probability is very low to find it (Figure 2a). On the opposite side, the other five models grouped in the ensemble model predict a significant presence of the tick in the Sahelian region (Figure 2b). Furthermore, the suitability level of the coastal areas predicted by the MaxEnt model is lower than the four other models grouped in the ensemble model.

3.5 Model evaluation

The RF model and the BRT model had the highest AUC (.95) and the GLM had the lowest AUC (.93) (Table 3). The GAM and the SVM models had the same AUC (.94). The RF model had the highest TSS (.80) and other models had the same TSS (.79) (Table 3).

The boxplot graphics in Appendix S4A highlighted the variability in the AUC among the bootstrap and cross-validation runs for each model. This figure revealed that the RF and BRT models AUC medians were above the other models for both bootstrap and cross-validation runs.

The boxplot graphics in Appendix S4B highlighted the variability in the TSS among the bootstrap and cross-validation runs for each model. Appendix S4B revealed that for the RF and GLM models, TSS medians were above the other models for bootstrap and cross-validation runs, respectively.

4.1 Model evaluation

According to the classification of Swets (1988), the MaxEnt model is ‘Useful’ and all the ensemble models are ‘Highly accurate’.

For the ensemble models, TSS values indicate that the selected models were accurate.

The MaxEnt model was built with presence data only. These two accuracy parameters (AUC and TSS) are known to use sensitivity and specificity in their computation (Shabani et al., 2018).

The MaxEnt model has a lower AUC value than other models built with presence and absence data, so it was less accurate than these models according to this parameter. However, on the other hand, the MaxEnt model had a high Boyce index that is an adapted parameter for presence-only models evaluation. The Boyce index is used to measure how well a model can predict the presence of species. The MaxEnt model was confirmed by its Boyce index value to be a model that does not require absence data to model species distribution.

The RF was the most accurate model among the ensemble models regarding its AUC and TSS means level and variance between both cross-validation and bootstrap runs. Cross-validation tends to be less biased but has a relatively large variance. On the other side, bootstrapping tends to reduce the variance considerably but gives more biased results. Combining the two methods, as is the case here in a model set, could lead to fairly accurate prediction results without too much bias and very little variance.

4.2 Characterisation of predictors and their importance in the models

The process of selecting variables for the construction of the models allowed the selection of seven variables including the annual average rainfall, the annual average temperature, the annual average vegetation index and the density of the host (cattle) and three land cover variables (cropland, grassland and shrubland).

Ticks are found in areas with significantly higher rainfall and moderate temperature. Water and temperature are important factors in tick biology, especially in one-host ticks because the non-parasitic (and therefore environment-dependent) stage is the egg and larva stage. Ticks are very sensitive to desiccation while searching for a host, as they constantly lose water through the integument of their body surface in dry conditions through transpiration. They also lose water through their spiracles in connection with respiration during locomotor activity (Herrmann & Gern, 2015; Randolph & Storey, 1999). The development and mortality of ticks are strongly influenced by these two factors (Estrada-Peña et al., 2015). Our models show quite clearly that areas that are wetter and cooler are more suitable for R. microplus. Previous studies have shown in Benin and Eastern Burkina Faso that the R. microplus tick prefers areas of moderate temperature and high rainfall (De Clercq et al., 2015; Zannou et al., 2021). The results seem to show that high temperatures are a limiting factor for the presence of the tick in an environment of West Africa. However, it is important to note that other factors such as humidity (represented here by rainfall and the vegetation index) also contribute to the presence of the tick. Therefore, in an environment with good rainfall and adequate vegetation, the temperature could not be the only limiting parameter. It is not uncommon to find in areas where the tick is present temperatures exceeding 28°C at certain times of the year. But generally, in West Africa the regions that often experience high temperatures are the countries of the Sahel which are generally very dry (low rainfall) and therefore unsuitable for the tick.

The NDVI of areas where the tick was found was higher than where it was not. The vegetation index, itself dependent on temperature and humidity (Hao et al., 2012), is an important variable in that it provides information on the presence, vitality and photosynthesis activity of vegetation. Engorged R. microplus female after dropping off the host need shelter under vegetation with convenient temperature and humidity to lay eggs (Latif & Walker, 2004). Eggs are sensitive to heat and risk desiccation when exposed to high temperatures and dry areas. Incubation and hatching of eggs are only possible under some particular climatic conditions (Pfäffle et al., 2013). Our results indicate that this variable is positively correlated with the presence of the tick and the models developed show that it took a large part in predicting the presence of the tick. De Clercq et al. (2015) in Benin and Zannou et al. (2021) in Benin and Burkina Faso have also evidenced the importance of the NDVI on a smaller scale.

Host (cattle) density was higher in areas where the tick is absent than in areas where it was present. In West Africa, the cattle population is more concentrated in the Sahelian countries, which are less watered and warmer. As the tick is more inclined to areas with more water and less heat, a negative correlation between these two variables appears. Nevertheless, the frequent and periodic movements of herds between these two ecosystems enabled by transhumance should be considered. Indeed, in dry season herds in search of pasture and water are forced to reach areas that are still wet at that time and therefore heavily infested by ticks (Zannou et al., 2021). These regular movements and stays expose the animals to R. microplus and the main pathogen it transmits: that is Babesia bovis and Babesia bigemina. These pathogens are responsible of extensive production losses (Jonsson, 2006; Waldron & Jorgensen, 1999).

Croplands were significantly higher in areas where R. microplus was absent than in areas it was present. Cropland is known to be the site of much human activity. Before the crops are planted, the land is well ploughed to help destroy ticks’ nests. During the vegetative cycle of the different crops, farmers carry out several weeding operations or use herbicides. Pesticides are also regularly used to protect these plants and guarantee a good harvest at the end of the season. All these anthropic actions do not allow the establishment of the non-host phase of the ticks in such environments. It must also be said that R. microplus is a specific tick and should be found in the areas most frequented by its bovine host. Cattle are kept away from the farmlands to avoid the destruction of crops by them. They are only tolerated in cropping areas in the dry season to take advantage of crop residues and fertilise the fields.

Rainfall was by far the most important predictor in the MaxEnt model followed, respectively, by NDVI, cropland, temperature, cattle density and shrub. The grassland did not play a role in the MaxEnt model prediction. The ensemble model revealed that NDVI was by far the most important predictor followed, respectively, by rainfall, temperature, cattle density, cropland, shrubland and grassland.

Although most existing studies (De Clercq et al., 2015; Sungirai et al., 2018) on the projection of the tick distribution only take into account climatic variables, our results showed the importance of other predictors such as cropland and host density in the modelling. The type of vegetation does not seem to be very important in predicting the presence of R. microplus. However, it should be noted that shrub vegetation seems to have a better predictive value than grassland. Shrubland provides a better canopy for the ticks and protects them from desiccation during the off-host phase. Grass is only really useful during the quest phase of the host. However, it is important to put the importance of these two variables into perspective because the collections were made on the animals and not on the ground by flagging. Similar results on vegetation preference were observed in the study of Omodior et al. (2021) in the United States.

4.3 Habitat suitability for R. microplus in West Africa

This work qualitatively provides a broader view compared to the previous study by De Clercq et al. (2015) and according to the model outputs the coastal areas seem globally more adapted to the tick. This previous R. microplus distribution modelling used occurrences data from 2011. Recent work (Ouedraogo et al., 2021) had revealed the presence of the tick in areas where it was previously absent (2015). Therefore, the present work has taken into account presence data from several countries, compared to the work from De Clercq et al. (2015), which only sampled and modelled data from Benin. When looking at habitat suitability maps generated in the current study in Benin in comparison to the study of De Clercq et al. (2015), it can be noticed that our models predicted suitable habitats for the species also in the northern part of Benin, whereas De Clercq et al.’s (2015) predictions emphasise mostly the southern one. This may be due to the larger dataset of observations used, as well as different predictors and the tuning of the MaxEnt model. The MaxEnt model as well as the ensemble model predicts a higher risk of tick presence in northern Benin than De Clercq et al.’s (2015) study. This work provides a larger view of the situation of R. microplus in West Africa. It also uses different methods—the ensemble models which are based on the consensus of several models—to predict the habitat suitability for R. microplus. The interest of the ensemble modelling is that it relies on the principle of the ‘wisdom of the crowd’ to make a better prediction than those models used individually (Hao et al., 2019).

The ensemble model seems to predict a stronger fit of the Sahelian zone to the tick than the MaxEnt model. When we consider the contribution of the variables to the prediction of the models, we notice that the pattern is not the same. For the MaxEnt model, it is the temperature variable that holds the greatest importance, whereas for the ensemble model it is the NDVI. This may have contributed to this difference in prediction, because the Sahel countries are significantly warmer than the coastal countries, but there is still vegetation in some parts of the Sahel around the oases. It should also be underlined that the ensemble model is a consensus model among several models, so its predictions must be given special attention. Precautions must be taken in some Sahelian countries that are at risk due to the extensive livestock farming methods used. These farming methods could favour the importation of the tick, which could adapt and continue its invasion.

According to the updated Koppen–Geiger climatic map provided by Beck et al. (2018), the coastal areas of West Africa show climatic similarities with southern Asia and South America, native area and previously invaded area of the tick R. microplus, respectively. These climatic similarities could be part of the factors that enabled the rapid adaptation and spread of the tick in the environment of some West African countries.

All six models clearly show that countries in the southern part of West Africa (coastal countries) are more suitable for R. microplus than those in the Sahel. As expected, temperature, humidity and vegetation influenced most the biology of the tick. Indeed, the Sahelian countries are drier and hotter than the coastal countries. Moreover, the vegetation is less developed there. In spite of the higher density of cattle in the Sahel, the off-host life will be difficult for R. microplus larvae because of the unfavourable climatic conditions. However, as the habitat suitability maps show, the risk of having the tick in some parts of the Sahel is not null. In some Sahelian countries, there are areas such as permanent water points that could provide conditions for the tick to establish itself, especially when the species is known to be highly invasive and adaptive. Recent unpublished work has revealed its establishment in the central plateau of Burkina Faso, a fairly hot area compared to the usual places where it is found (Abel Biguezoton, personal communication). In recent years, Burkina Faso has developed water reservoirs, which have the dual function of being used for the cultivation of vegetables and for watering cattle herds. These environments can present risks of the introduction and dissemination of ticks from herds returning from transhumance in coastal countries. The valleys of perennial rivers also represent biotopes that can favour the establishment of the invasive tick as they are very frequented during the dry season by herds in search of water and pasture.

Although native to Asia, R. microplus has invaded sub-Saharan Africa and is probably replacing progressively the indigenous Rhipicephalus decoloratus and Rhipicephalus annulatus (Adakal et al., 2013; Gomes & Neves, 2018; Madder et al., 2011). This rapid expansion was also favoured by the extensive livestock farming practised by more than 70% of cattle farmers in sub-Saharan Africa (Boka et al., 2017).

4.4 Management options faced to R. microplus in West Africa

This study provides a broader view of habitat suitability for R. microplus. It allows decision makers in West African countries to have a better idea of the level of suitability of their area for the invasive tick. It is therefore a tool to identify areas where the emphasis should be placed on raising awareness among livestock stakeholders.

In areas favourable to the tick and not yet infested, measures must be taken to prevent or detect quickly its introduction. For instance, management options should be awareness campaign, avoiding the purchase of cattle originating from infested areas, avoiding herd movement to already infested areas and capacity building for detection of R. microplus. This capacity building should concern farmers, para-veterinarians, private veterinarians and veterinary officers. Other actions could consist to strengthen surveillance of the veterinary acaricides drug market.

Furthermore, in areas where the tick is already established, it will be necessary to raise awareness of the measures to be implemented to limit its spread and, ultimately, its control (e.g. campaign of information, adequate protocol of acaricide treatment, follow-up of acaricide resistance, training veterinary agents in tick identification, strengthen the surveillance of the veterinary drug market).

In areas identified here as not favourable to the tick, veterinary services should put in place measures to avoid the introduction and the establishment of the invasive tick (e.g. avoid the introduction of animals from infested areas without appropriate measures, control the movement of animals from infested areas and strengthen the capacity of veterinary services agents in tick detection).

This study is a tool to help in the development of sub-regional and local control plans for R. microplus tick and consequently the pathogens it transmits. The direct consequence is the improvement of cattle production and productivity and in turn food security. Good knowledge of the risk for ticks with veterinary and/or medical importance establishment in an area allows for careful management of financial resources dedicated to their control.