Abstract
Accurately predicting the geographic ranges of species is crucial forassisting conservation efforts. Traditionally, range maps were manually createdby experts. However, species distribution models (SDMs) and, more recently,deep learning-based variants offer a potential automated alternative. Deeplearning-based SDMs generate a continuous probability representing thepredicted presence of a species at a given location, which must be binarized bysetting per-species thresholds to obtain binary range maps. However, selectingappropriate per-species thresholds to binarize these predictions is non-trivialas different species can require distinct thresholds. In this work, we evaluatedifferent approaches for automatically identifying the best thresholds forbinarizing range maps using presence-only data. This includes approaches thatrequire the generation of additional pseudo-absence data, along with ones thatonly require presence data. We also propose an extension of an existingpresence-only technique that is more robust to outliers. We perform a detailedevaluation of different thresholding techniques on the tasks of binary rangeestimation and large-scale fine-grained visual classification, and wedemonstrate improved performance over existing pseudo-absence free approachesusing our method.