Listening for Frogs

August 17th, 2016

In recent times the use of biological and ecological indicators has become a standard water quality assessment tool, and opportunely, frogs have been used widely as bioindicators/indicators of ecosystem health.

As such, CSIRO partnered with Seqwater to develop a biological monitoring project in the Somerset catchment with three important and overlapping aims:

  1. establish a water quality monitoring program in the upper Somerset catchment based around the presence/absence of Frog species and changes in the composition of Frog communities;
  2. locate extant Frog populations and identify threatening processes in order to implement measures to ameliorate their impact;
  3. undertake an assessment of areas in the upper catchments of the Stanley River and upper Kilcoy creek areas to provide information on the distribution of endangered, threatened and presumed extinct Frogs.

The Challenge

The project involves the recording of sounds in Somerset and Nerang areas near Brisbane with 6 spatially distributed audio nodes. The audio is then analysed post facto to label the recorded sounds and develop methods to classify frog species. The main challenge is to classify frogs among the presence of multiple other species that also make sounds, which creates a noisy environment. The Figure below shows the spectrogram of recorded audio highlighting at least 4 species that make sounds during overlapping periods. audio_spectrogram

Our Solution

A method for detecting vocalization of giant barred frogs (Mixophyes iteratus) in noisy audio is proposed. Audio recordings from remote wireless sensor nodes were segmented into individual sounds and from each sound a small set of features was extracted. Feature vectors were compared to those of example calls using a Euclidean distance formula as a detection system. The system achieved a sensitivity of 0.85 with specificity of 0.92 when distinguishing M. iteratus calls from other species’ calls and sensitivity of 0.88 with specificity 0.82 against background noise.

More details on this work are available in this paper:

Benjamin Croker & Navinda Kottege (2012), Using feature vectors to detect frog calls in wireless sensor networks. Journal of the Acoustical Society of America. 131(5), pp. EL400-EL405. ISSN: 0001-4966