WhatFrog: Development of app and web interface for automated anuran recognition and mapping

Background/Question/Methods

The modern decline in anuran (frog and toad) populations is a well-documented phenomenon. To help identify signs of further decline as well as foresee potential remediation directions, we propose an automated system for monitoring anuran populations. In our solution, volunteers gather anuran vocalizations using their own mobile phones (and electronic tablets), which are equipped with software that automatically recognizes the frog species present in the recordings. In addition to on-device recognition capabilities, our system is supported by Web-based services that volunteers can use for uploading the recordings, tagging, location, and other relevant information. Our anuran-recognition algorithms use machine-learning and speech-recognition techniques. Given an anuran call as input, our algorithm obtains a distinct acoustic signature (acoustic fingerprint) that is then compared to a database of pre-learned signatures from various species. These acoustic fingerprints are similar to the ones used by music-recognition software. However, our method implements some extra flexibility and robustness to noise needed for working in the field. For this, we use a combination of powerful statistical-classification techniques such as support-vector machines. In addition to using the likelihood matching from acoustic signatures, our algorithm uses prior models of geographical location to help disambiguate classification. 

Results/Conclusions

A prototype of our software is ready. On experiments performed on a dataset containing 100 calls of fifteen frog species, our automatic method achieved a promising nearly 90% correct classification, which is similar on average to that achieved by trained volunteers. The dataset of calls used for testing the algorithm includes background noise, groups of frogs of various species, as well as insect noise.