Overview and computational strategy
To build and evaluate the recognisers for eight target frog species, we
used a large database of annotated calls from the study area in
Koondrook-Perricoota (KP) forest. The Forestry Corporation of NSW– the
body commissioning the study – had previously annotated 831 files from
20 sites and found varying levels of presence for the different species.
From these files, we extracted between 100 to 200 reference calls per
species, from all sites where the species was detected. Following manual
inspections for call clarity and variation, we used these reference
calls to build approximately 20 candidate recognisers per species (i.e.,
one template equals one recogniser per species). Some recognisers were
based on templates of the same reference call, but were created using
different amplification settings, which is modifiable in monitoR. We
then ran the recognisers on the pre-annotated files to calibrate the
score cut-off (similarity threshold between reference call templates and
the sound files) and estimate omission and commission errors. We then
chose the final recognisers based on the best ROC (Receiver Operator
Criterion, Zou et al., 2007), thus minimising both false positive and
false negative detections.