Model evaluation
We found variation in performance of the four algorithms used, with some
models having TSS < 0.7 for several species. RF was utilised
for all 110 species, while MARS, MAXENT, and ANN were limited to 93, 90,
and 89 species respectively. For 17 species (Cynopterus sphinx ,Hipposideros fulvus , Lyroderma lyra , Pipistrellus
ceylonicus , P. coromandra , P. tenuis , Pteropus
medius , Rhinolophus lepidus , Rhinopoma hardwickii ,R. microphyllum , Rousettus leschenaultii ,Scotophilus heathii , S. kuhlii , Scotozous dormeri ,Sphaerias blanfordi , Taphozous longimanus , and T.
nudiventris ) RF was the only algorithm represented in the final
ensemble. There was a significant negative correlation between the
number of occurrence points and the number of models used in the final
ensemble (Pearson’s r = -0.791; p < 0.001). RF was the
only model represented in the final ensemble for most species with more
than 120 occurrences; Taphozous nudiventris and Sphaerias
blanfordi were the two exceptions to this, with 97 and 17 occurrences
respectively. RF models also had the highest mean AUC and TSS scores
across all species (AUC: 0.998 ± 0.001 [0.991 - 1.000]; TSS: 0.980 ±
0.014 [0.940 - 1.000]), followed by MARS (AUC: 0.970 ± 0.016
[0.920 - 0.997]; TSS: 0.897 ± 0.055 [0.761 - 0.995]), MAXENT
(AUC: 0.958 ± 0.019 [0.904 - 0.994]; TSS: 0.859 ± 0.052 [0.754 -
0.978]), and ANN (AUC: 0.955 ± 0.024 [0.887 - 1.000]; TSS: 0.869 ±
0.059 [0.750 - 0.999]).