Abstract

Acoustic techniques are becoming a popular alternative to monitor bat species. However, monitoring programs depend upon the effective recognition of species or any other ‘recognizable unit’ set as the monitoring target. In spite the evolution of patterns in call design that make some bat species recognizable using acoustic features, acoustic identification also imposes several challenges. Here, we study the problem of identifying bat species from echolocation calls in order to build automated acoustic monitoring algorithms. The main purpose of this study was to assess the accuracy and utility of two Machine Learning classification techniques faced with a number of bat call discrimination tasks that reflect phylogenetic and ecological constraints on bat calls design. We constructed a reference search-phase echolocation bat calls library for Mexican bat species and applied Random Forest and Dynamic Time Warping algorithms to classify 59 species. We selected 4538 full-spectrum reference calls from a set of 21154 collected by the authors, donated by colleagues and from Echobank. We trained both algorithms with a number of hierarchical tasks. First we train them to classify calls to Family, Guild or Genus then to species and compared general species classification accuracy to a non-hierarchical classification done directly to species. Results show good general classification accuracy, being this study with the highest number of species classified so far. This suggests there is potential to develop an automated classification system for bat calls in megadiverse countries even if data is limited. These methods can provide key information on distribution and diversity changes through time and will set the foundations for a future national acoustic monitoring program.