Similarity-based large-scale distribution mapping of orchids

Abstract

The expected presence/absence of a target species outside the area of actual observations is commonly estimated using statistical models or decision criteria. This investigation demonstrates a similarity-based solution for predictive mapping as an alternative to generalizing models. The maps of the expected distribution of 12 orchids were created using find sites from field observations and absence sites generated onto the observation track. The expected presence/absence of a species in a location was calculated according to the similarity between the location and selected examples of presence and absence sites. A machine learning system selected the best predictive sets for each species out of 161 cartographic and remote sensing features. The usefulness of the predictive distribution maps was expressed as the ratio of the density of find sites per track in the predicted presence area relative to the density per track in the predicted absence area. The predictive mapping was more efficient for Dactylorhiza incarnata, D. russowii, Gymnadenia conopsea, and Goodyera repens. Soil properties and the proportion of find sites for the other species in the vicinity were the most indicative site characteristics. The rarer species were found to be better indicators of the occurrence of the other species than were the more common orchids. The proposed approach—to direct subsequent field observations to sites where the occurrence of the target species was predicted but has not yet been recorded—helped discover new populations of orchids and enhance the representativeness of absence sites.