To elaborate our model we removed uncertain localities such as those localities with no coordinates and all records with coordinates laying down in sites with estimated elevations below 1,600 m and above 3,900 m. Also one record in BioMap from Bolivar and three in DatAves from Amazonas, which represent corrupted information in the database were deleted.
The habitat suitability model generated in Maxent made excellent predictions, not showing any areas suitable in climatic terms for this species out of the main Andean ranges, Sierra Nevada de Santa Marta and serrania del Perija.
Assuming that the distribution of the species may have filled the complete climatic model generated, its distribution today in remnants of forest is about 47,321 km2, which corresponds to a loss of 65 % of its potential original distribution due to deforestation.
Our checklist (ABC 2016) suggests the unconfirmed possibility that subspecies quitensis is present in extreme south Colombia in the border with Ecuador. Differently, McMullan & Donegan (2014) suggest the presence of quitensis in their distribution maps as certain. Nevertheless, distribution of specimens determined as the nominate subspecies according to BioMap suggest it is unlikely that quitensis exists in Colombia.
Subspecies districta from Sierra Nevada de Santa Marta and Perija is very distinct and possibly warrant species status. This subspecies is mostly distributed in Colombia and according to our maps we estimate its extent of occurrence in Colombia in 5,257 km2, which is well below the threshold of 20,000 km2 to be catalogued as Vulnerable (VU) by the IUCN and very near the threshold of 5,000 km2 to be considered as Endangered (EN). Considering that subtropical and temperate forests have been severely transformed and fragmented, and having lost about 54 % of its potential original distribution by deforestation, we believe that possibly this subspecies must be considered as Vulnerable (VU).
Regularized training gain is 1.897, training AUC is 0.948, unregularized training gain is 1.980.
Algorithm converged after 1040 iterations (38 seconds).
The follow settings were used during the run:
552 presence records used for training.
10548 points used to determine the Maxent distribution (background points and presence points).
Environmental layers used (all continuous): bio10co bio11co bio12co bio13co bio14co bio15co bio16co bio17co bio18co bio19co bio1co bio2co bio3co bio4co bio5co bio6co bio7co bio8co bio9co
Regularization values: linear/quadratic/product: 0.050, categorical: 0.250, threshold: 1.000, hinge: 0.500
Feature types used: hinge product linear threshold quadratic
'Equal Training Sensitivity and Specificity' and 'Equate Entropy of Thresholded and Original Distributions' thresholds and omission rates:
0.109-0.150-Fractional predicted area
0.109-0.031-Training omission rate