This project successfully developed, tested and implemented call recognisers for eight species of frog in the Murray-Darling Basin. Recognisers for all but one species performed well and substantially better than many species recognisers reported in the literature .We achieved this through a comprehensive development phase, which carefully considered and refined the representativeness of training data, as well as the construction (amplitude cut-off) and the similarity thresholds (score cut-offs) of each call template used.
We demonstrated the utility of automated frog call detection to monitor ecosystem response at all watered sites. At most sites the response was almost instant – calls were close to a peak around one week after watering. The sharp response curves by the recognisers demonstrate the utility of multi-species call recognisers. This is demonstrated in the diagram below – as soon as water reaches a site, frog calls increase significantly. We tested this using a Linear Mixed Model, in which we contrasted watered and control sites – showing that only watered sites increase in activity.
While slightly more involved than building recognisers using commercial packages, the workflows ensure that a high quality recogniser can be built for which the performance can be fine-tuned using multiple parameters. Using the same framework, recognisers can be fine-tuned in future iterations. We believe that multi-species recognisers are a highly effective and precise way to detect the effects of ecosystem restoration – in this case environmental watering - delivering a much sharper response signal than previous index based analyses.