Species distribution modelling for Rhipicephalus microplus (Acari: Ixodidae) in Benin, West Africa: Comparing datasets and modelling algorithms

Rhipicephalus microplus is one of the most widely distributed and economically impor-tant ticks, transmitting Babesia bigemina, B. bovis and Anaplasma marginale. It was recentlyintroduced to West Africa on live animals originating from Brazil. Knowing the preciseenvironmental suitability for the tick would allow veterinary health officials to draft vectorcontrol strategies for different regions of the country. To test the performance of modellingalgorithms and different sets of environmental explanatory variables, species distributionmodels for this tick species in Benin were developed using generalized linear models, lin-ear discriminant analysis and random forests. The training data for these models were adataset containing reported absence or presence in 104 farms, randomly selected acrossBenin. These farms were sampled at the end of the rainy season, which corresponds with anannual peak in tick abundance. Two environmental datasets for the country of Benin werecompared: one based on interpolated climate data (WorldClim) and one based on remotelysensed images (MODIS). The pixel size for both environmental datasets was 1 km. Highlysuitable areas occurred mainly along the warmer and humid coast extending northwardsto central Benin. The northern hot and drier areas were found to be unsuitable. The modelsdeveloped and tested on data from the entire country were generally found to perform well,having an AUC value greater than 0.92. Although statistically significant, only small differ-ences in accuracy measures were found between the modelling algorithms, or betweenthe environmental datasets. The resulting risk maps differed nonetheless. Models based oninterpolated climate suggested gradual variations in habitat suitability, while those basedon remotely sensed data indicated a sharper contrast between suitable and unsuitableareas, and a patchy distribution of the suitable areas. Remotely sensed data yielded morespatial detail in the predictions. When computing accuracy measures on a subset of dataalong the invasion front, the modelling technique Random Forest outperformed the othermodelling approaches, and results with MODIS-derived variables were better than thoseusing WorldClim data